Predictive soil mapping a reviewP Sculla J Franklina OA Chadwickb andD McArthuraaDepartment of Geography San Diego State University San Diego CA 92182-4493USAbDepartment of Geography University of California Santa Barbara Santa BarbaraCA 93106 USA

Abstract Predictive soil mapping (PSM) can be defined as the development of a numerical orstatistical model of the relationship among environmental variables and soil properties which isthen applied to a geographic data base to create a predictive map PSM is made possible by geo-computational technologies developed over the past few decades For example advances ingeographic information science digital terrain modeling remote sensing fuzzy logic has createda tremendous potential for improvement in the way that soil maps are produced The StateFactor soil-forming model which was introduced to the western world by one of the earlyPresidents of the American Association of Geographers (CF Marbut) forms the theoretical basisof PSM PSM research is being driven by a need to understand the role soil plays in thebiophysical and biogeochemical functioning of the planet Much research has been published onthe subject in the last 20 years (mostly outside of geographic journals) and methods have variedwidely from statistical approaches (including geostatistics) to more complex methods such asdecision tree analysis and expert systems A geographic perspective is needed because of theinherently geographic nature of PSM

Key words GIS soil geography soil survey

I Introduction

Soil is a fundamental natural resource it is the basis of human agriculture Civilizationsrise in regions blessed with rich soil they fall when humans fail to treat soil withrespect Further soil plays an essential role in the biophysical and biogeochemicalfunctioning of the planet On the continents soil forms a porous boundary where thebiosphere hydrosphere lithosphere and atmosphere interact Understanding the

Progress in Physical Geography 272 (2003) pp 171ndash197

copy Arnold 2003 1011910309133303pp366ra

Author for correspondence Tel 315-228-7864 fax 315-228-7726 e-mail pscullmailcolgateedu

172 Predictive soil mapping a review

spatial distribution and management of soil is critical to maintain a productive societyand to understand the complex balance of chemical and physical processes that makelife possible on Earth

While primarily referred to as pedology and studied in soil science departments suchthemes are the focus of many different academic disciplines For example soilgeography is the study of the location distribution and pattern of soils on the landscape(Buol et al 1997) Geography and pedology share a common history in this centuryCF Marbut a geographer by training was responsible for communicating pedologicalideas developed by the Russian school of pedology within the USA (Cruickshank1972) Marbut was a student of WM Davis the first president of the AmericanAssociation of Geographers (AAG) and later became president himself in 1924 As aresult soil was an active area of research within geography during the early years of theAAG While soil geography is still occasionally published in geography journals (seeBarrett 1999 for a recent perspective) a great deal of geographic soil research hasappeared over the last 20 years outside of geographic journals

From an applied perspective pedology and soil geography form the basis of soilsurvey which remains the primary means by which information on the spatialproperties of soil is collected presented and archived in the USA and throughout theworld The United States Department of Agriculture (USDA) in cooperation with otherfederal and state agencies began soil survey work in 1896 The combined state andfederal government effort became known as the National Cooperative Soil Survey(Indorante et al 1996) Its mission was to furnish private landowners land managersand consultants with soil maps to aid in land-use decision making More recentlyhowever soil maps have been used to provide chemical and physical data input withinecological and hydrological process models (Burrough and McDonnell 1998 163) Thesoil survey program was simply not designed to furnish data for such applications Theincreasingly sophisticated use of soil data has led to a greater demand for data aboutsoil properties than the conventional soil map can accommodate (Cook et al 1996)Traditional soil survey concepts are based on qualitative recognition of soil propertiesin relation to landscape and environmental variables Although these methodsimplicitly incorporate the expertise of the soil scientist they do not make use of geo-computational technologies that are now widely available

Technological advances during the last few decades have created a tremendouspotential for improvement in the way that soil maps are produced (McKensie et al2000) Remote sensing and photogrammetric techniques provide spatially explicitdigital data representations of the Earthrsquos surface that can be combined with digitizedpaper maps in geographic information systems (GIS) to allow efficient characterizationand analysis of vast amounts of data The future of soil survey lies in using GIS tomodel spatial soil variation from more easily mapped environmental variablesPredictive soil mapping (PSM) begins with the development of a numerical or statisticalmodel of the relationship among environmental variables and soil properties which isthen applied to a geographic data base to create a predictive map (Franklin 1995) Threemain goals of PSM are to (1) exploit the relationship between environmental variablesand soil properties in order to more efficiently collect soil data (2) produce and presentdata that better represent soil landscape continuity and (3) explicitly incorporate expertknowledge in model design PSM can also potentially advance pedology and soilgeography by providing insights into soil forming processes

The State Factor soil-forming model forms the theoretical basis of PSM as well as oftraditional soil survey and has been used in other many other disciplines Hans Jennypopularized the State Factor theory of soil formation in the USA through his publicationof Factors of soil formation (1941) The ideas expressed in this classic text represent a for-malization of Dokuchaievrsquos ideas of soil formation (Simonson 1997) Since that timethe theory has provided a paradigm through which soil genesis and distribution can bestudied (Schaetzl 1991) The theory states that soil profile character is a function ofclimate organisms relief parent material and time implying that if the spatial distrib-ution of the soil-forming factors is known soil character may be inferred Thistheoretical framework has been used by many authors in pedological research andremains the most popular theory of soil genesis (for a recapitulation of Jennyrsquoscontinuing influence on pedology see McSweeney et al 1994)

Until recently predictive soil mapping efforts have been constrained by twoproblems associated with Jennyrsquos State Factor theory of soil genesis the mathematicalequations derived from theory proved virtually unsolvable (Huggett 1975) and datadescribing the soil-forming factors were unavailable or not widely available Recentadvances in mathematical theory (fuzzy logic) and statistical methods (includingmachine learning techniques) have helped researchers to better understand the natureof the equations implied by the theory and to solve them Relative to the problem ofsoil-forming factor knowledge Jenny (1941 262) wrote lsquothe conversion of suchfundamental knowledge (soilndashenvironment relationships) to specific field conditions isimpossible unless the areal distribution of the soil formers is knownrsquo Half a centurylater remote sensing and photogrammetry along with GIS can now be used to charac-terize the spatial distribution of soil-forming factors The foundation has been laid forthe emergence of predictive soil modeling as an active area of research

Several articles have been published outside of geographic literature to trackprogress in the field of PSM over the years Two earlier articles (McBratney 1992Hewitt 1993) called for an overall change in the philosophy of soil survey and outlinedresearch challenges in the field at that time In the mid 1990s the term pedometrics wascoined (Webster 1994) to refer to quantitative research in the field of pedology Thisterm would appear synonymous with PSM as defined above More recently reviewshave been published on a subset of PSM techniques (McBratney et al 2000) and generalmethods of modelling soil variation (Heuvelink and Webster 2001)

Predictive soil mapping has only recently emerged as a research niche but the paceof research breadth of knowledge and variety of techniques have expanded rapidly(pedometrics have accounted for 18 of the subject matter in articles published inrecent years in Geoderma Hartemink et al 2001) A review of recent achievements in thefield of PSM is needed because all methods have not been reviewed in past papersFurther a geographic perspective is needed because of the inherently geographicnature of PSM First we review several topics critical to PSM (Section II) including thenature of soil variability (Section II1) how the soil survey has defined soil mapping(Section II2) and soil taxonomy (Section II3) We then briefly describe the componentsof PSM (Section III) geographic data models of soil variability and soil mapping(Section III1) digital terrain modelling (Section III2) fuzzy logic (Section III3) andremote sensing (Section III4) Lastly we discuss the present state of knowledge in PSM(Section IV) Our focus in this review is to evaluate literature in the field of PSM and toreintroduce the subject matter to the geographic audience

P Scull et al 173

174 Predictive soil mapping a review

II Background

Prediction of soil properties based on knowledge of the effect of environmentalvariables on soil formation has always been the basis for all soil mappingUnfortunately the traditional methods do not yield quantifiable soilndashlandscapeinformation that robustly describes actual soil variation In this section we describe soilvariation and explain why traditional methods for defining soil distribution inlandscapes are inadequate It is also our purpose to demonstrate that there is a logicalflow from these approaches into the new PSM techniques All approaches to soilmapping rest on our ability to use knowledge of the process of soil genesis to predictthe properties of soils at any point in the landscape

1 The nature of soil variability

The complex and highly variable nature of soil patterns in landscapes complicates thealready labour-intensive process of collecting and presenting soil survey data (Wrightand Wilson 1979) In 1941 Jenny listed three different definitions of soil before heconceded that lsquoit is problematic whether any definition of soil could be formulatedrsquo(Jenny 1941 2) In the half-century that has passed since Jenny wrote those words therehave been numerous attempts to define soil (reviewed by Birkeland 1999) Mosttextbook authors describe soil as extraordinarily complex For example McKnight(1993 336) defines soil as lsquoan infinitely varying mixture of weathered mineral particlesdecaying organic matter living organisms gases and liquid solutionsrsquo (emphasisadded) This definition illustrates how complex soil can be It further demonstrates thatthe soil landscape is continuous and is not composed of distinct individual soil typesThis point has been made repeatedly over the years (Simonson 1959 Webster andBeckett 1968 Cambell 1977 and Moore et al 1993)

A variety of soil genesis models have been proposed in order to account for the highvariability of soil and collectively they help further illustrate the difficulty of charac-terizing the soil landscape (for a review see Huggett 1975 Birkeland 1999) Threedistinctive approaches have been employed factor models (eg Jenny 1941) wherefactors affecting soil development are identified process models (eg Simonson 1959)where soil-forming processes are emphasized and energy models (eg Runge 1973)where the focus is upon process-driving mechanisms A host of hybrid models havealso been employed Still no real consensus exists today as to exactly how to model soildevelopment partially because of the recent emergence of pedology as an academicdiscipline (Johnson and Watson-Stegner 1987) Existing data collection methods do notyield adequate soil information in part because many of the processes that shape thesoil landscape are still poorly understood

2 How the soil survey has defined soil mapping

Traditional soil survey persists as the most popular form of soil mapping and inventoryand in many cases is the only manner in which the highly variable nature of the soillandscape is catalogued The method consists of three steps (Cook et al 1996) The firstis direct observation of ancillary data (aerial photography geology vegetation etc) and

soil profile characteristics In the second step the observations of soil attributes areincorporated into an implicit conceptual model that is used to infer soil variation Thethird step involves applying the conceptual model to the survey area to predict soilvariation at unobserved sites Usually less than 0001 of the survey area is actuallyobserved (Burrough et al 1971) a fact reflecting the high cost of field sampling Theconceptual model of soil variation is then transformed into a cartographic model thechoropleth map by drawing map unit boundaries on aerial photographs In effect pho-tographic scale determines the resolution of the soil map

This process has been severely criticized in the scientific literature for two reasonsFirst the conceptual model developed by the soil surveyor is primarily implicit beingconstructed in a heuristic manner This results in an excessive dependence upon tacitknowledge and as such incomplete information exists relative to the derivation of theultimate soil survey product (Hudson 1992) This aspect of soil survey is especiallyfrustrating because it fails to document most of the knowledge that the soil surveyoraccumulates during the expensive field mapping process In essence the soil survey isunfalsifiable and therefore untestable (Hewitt 1993) The final product of soil survey isa soil map that has unknown assumptions limitations and accuracy (Burrough et al1971 Dijkerman 1974) PSM techniques are similar in theory to soil survey (they bothuse knowledge of soilndashenvironment relations to make inferences) but the methodsemployed often yield quantitative expressions of soil variability with measured levelsof accuracy Clearly one direction of innovation in soil survey is to add objectivity tomodel development which will allow more explicit scientific communication (Gessler1996) The second major criticism of soil survey concerns the role of soil classification

3 Soil classification

The evolution of PSM techniques has been directly impacted and continues to beinfluenced by the process of soil classification This is especially true in the USAwherethe primary focus of the National Cooperative Soil Survey is to develop and map Soiltaxonomy (the official classification developed by the US Department of AgricultureSoil Survey Staff 1975) The system was influenced by nineteenth century biologicaltaxonomy and the practice of geological survey (Heuvelink and Webster 2001) Thepurpose of Soil taxonomy was to provide an objective manner to systematically classifysoil and was adopted at a time when soil information had to be abstracted to the levelof the modal profile (classified in Soil taxonomy) because it was impossible to catalogueand present the full amount of soil variability (Cambell and Edmonds 1984) In orderto map soil taxa the soil must be perceived as a spatial entity a lsquopedonrsquo (a term used inSoil taxonomy to refer to the smallest recognizable unit that can be called lsquoa soilrsquo) Inpractice this spatial perception of soil results in a map whose classes are homogenousunits with unknown variability and sharply defined boundaries (Burrough andMcDonnell 1998) Since the initial development of Soil taxonomy the perception of thesoil landscape has changed from a collection of individual soil types to a continuallyvarying mixture of soil components (Dmitriev 1983) Field observations have shownthat in concert with the environmental variables soil properties vary continuouslyacross landscapes exhibiting different and complex scales of variation (Simonson1959) Therefore soil distribution is not well represented by choropleth maps

P Scull et al 175

176 Predictive soil mapping a review

(McBratney 1992 Gessler et al 1995) While this change has not yet been recognized inSoil taxonomy PSM research has developed methods of soil inventory that moreaccurately describe the soil landscape Such methods are at odds with the traditionalapproaches because they often involve dismissing the concept of soil as a spatial entityRather the new methods focus on mapping continuously varying soil properties Theimplementation of these techniques has been difficult because of the entrenchment ofSoil taxonomy

III Components of predictive soil mapping

1 Geographic data models of soil variability and soil mapping

Traditionally soil maps have been digitized to fulfil the need for soil data within GIS-based environmental modelling research This information from the paper mapdigitized for the computer is fraught with the same problems of the original choroplethmaps ndash assumed homogeneous units with unknown variability and sharply definedboundaries GIS allows for a more robust characterization of spatial variability (relativeto the cartographic generalizations of the past) by allowing information to be analysedand stored using a variety of data models As defined in the GIS literature datamodelling is the process of discretizing spatial variation which entails abstracting gen-eralizing or approximating geographic reality (defined as empirically verifiable factsabout the real world) Unfortunately data modelling is often confused with issues ofdata structure and limited by software selection (Kemp 1992) The process is of crucialimportance because it controls the manner in which the data can be processed oranalysed (Goodchild 1994) as well as the view of the data the end user ultimatelyreceives (Goodchild 1992a Kemp 1992) Some data models are more accurate thanothers at portraying geographical reality (Goodchild 1992b) In the case of soil howwell do the data including the data model represent the highly variable continuousnature of soil How can digital computers be used to manage spatial data to bestrepresent the soil landscape

The choice of data models is partially dependent upon how the soil is perceived ingeographic space Objects can be thought of as existing as independent entities inempty space (object view) or as one of an infinite set of tuples (the foundation ofgeographic information ndash xyz where z is a measured value and xy are its location inspace) approximated by regions and segments (field view) (Goodchild 1994) The fieldview better represents continuous surfaces such as the soil landscape but soil datahave traditionally been modelled using the feature model of geographic space thechoropleth map Soil data have probably been managed this way because of theinfluence of Soil taxonomy which defines individual soil lsquotypesrsquo and because soil datacollection began during a time when few alternatives existed

Using PSM techniques a fundamental change in the soil data model from thechoropleth map to the raster grid allows better characterization of actual soilndashlandscapevariability A raster data model accommodates a lsquofield viewrsquo representation of thelandscape and is defined as a regular rectangular array of cells with some aggregatevalue of the field recorded for each cell (Goodchild 1994) The resolution of the datastored in this format is a function of the grid cell size which can be made small enoughto simulate continuous variation at the landscape scale The raster has become the most

widely used data model for PSM and is also routinely used to manage other environ-mental information such as elevation (DEMs) and remotely sensed data Spatialanalysis and integration with other types of raster-based environmental data can beeasily performed with soil data stored using a raster data structure (Burrough andMcDonnell 1998)

2 Digital terrain modelling

Terrain analysis quantifies the relief component of models characterizing soilformation Soil development and its associated profile characteristics often occurs inresponse to the way in which water moves through and over the landscape which iscontrolled by local relief Accordingly terrain analysis will be most useful in environ-ments where topographic shape is strongly related to the processes driving soilformation (McKensie et al 2000) Digital terrain modelling is a technique for derivingspatially explicit quantitative measures of the shape character of topography (Weibeland Heller 1991 Wilson and Gallant 2000) The spatial distribution of the resultingterrain attributes (characterizing local water flow paths) can also capture the spatialvariability of soil attributes Moore ID et al (1991) reviewed the analysis of digitalelevation data (including DEMs) for hydrological geomorphological and biologicalapplications They provided a table that summarized the significance and physicalmeaning of various terrain attributes to landscape processes Building on their workmany authors have used terrain attributes derived from digital elevation models(DEMs) as explanatory variables in predictive soil models (Odeh et al 1991 Moore etal 1993 Gessler et al 1995 Skidmore et al 1996 and others) Methods used to deriveterrain attributes have been greatly refined over the last 15 years and future satellitesaiding in the development of more accurate DEMs will make terrain analysis anincreasingly important component of predictive soils mapping (Moore et al 1993Mackay and Band 1998) Several recent review articles have been specifically devotedto the role of terrain analysis in soil mapping (Ventura and Irvin 1996 Irvin et al 1996McKensie et al 2000)

3 Remote sensing

Remote sensing data are an important component of PSM because they provide aspatially contiguous quantitative measure of surface reflectance which is related tosome soil properties (Agbu et al 1990) Both physical factors (eg particle size andsurface roughness) and chemical factors (eg surface mineralogy organic mattercontent and moisture) control soil spectral reflectance (Irons et al 1989) Surfacemineralogy can be derived by wavelength specific charge transfer and crystal fieldabsorptions associated with the presence of iron and iron-oxides (Fe2+ and Fe3+) andvibrational absorptions associated with hydroxyl bonds in clays adsorbed water andthe carbonate ion (Goetz 1989 Irons et al 1989) The presence and strength of theseabsorption features can be used to identify and quantify concentrations of mixed suitesof minerals in soil (Johnson et al 1983 Shipman and Adams 1987) Organic matterparticle size and moisture content in contrast influence soil reflectance primarilythrough a change in average surface reflectance and produce only broad spectral

P Scull et al 177

178 Predictive soil mapping a review

expression (Irons et al 1989) Adecrease in particle size tends to increase surface albedoand decrease spectral contrast of absorption features while an increase in organicmatter or soil moisture decreases average reflectance or albedo

Numerous studies have shown the potential benefits of using remote sensing for soilidentification and mapping Comprehensive surveys of soil spectral reflectance includestudies by Stoner et al (1980) Henderson et al (1992) and Csillag et al (1993) Remotesensing studies based on broad-band sensors such as Landsat TM include Agbu et al(1990) Coleman et al (1993) Seyler et al (1998) and Oliveira (2000) Traditionallyremote sensing has been used to classify soil units through photo-interpretation ordigital image processing Combining remotely sensed information with ancillaryinformation such as thematic maps or vegetation cover can yield significant improve-ments (Wilcox et al 1994 Cialella et al 1997 Wanchang et al 2000)

Recent developments in hyperspectral remote sensing offer the potential of signifi-cantly improving data input to predictive soil models Hyperspectral sensors such asthe Airborne VisibleInfrared Imaging Spectrometer (AVIRIS) measure a contiguousspectrum in the visible and NIR and thereby better characterize atmospheric andsurface properties (Goetz et al 1985) The large number of spectral bands permits directidentification of minerals in surface soils For example Clark and Swayze (1996)mapped over 30 minerals using AVIRIS at Cuprite Nevada Palacios-Orueta and Ustin(1996) showed that enhanced spectral information was suitable for discriminating evensubtle spectral changes associated with differences in organic matter and iron contentOther examples of the application of AVIRIS to aid soil mapping include Palacios-Orueta et al (1998) Okin et al (1998) and Roberts et al (1998)

Sensors that operate in the microwave portion of the electromagnetic spectrum havealso shown promise in soil mapping research Microwave sensing can be broadlydivided into active (eg radar) systems and passive systems and are capable ofpenetrating the atmosphere under virtually all conditions offering a significantadvantage over visible and near-infrared spectroscopy (for a general overview ofmicrowave sensing see Lillesand and Kiefer 1994 chapter 8) Synthetic aperature radar(SAR) is one example of an active system SAR has been used to aid soil propertymapping such as soil salinity (Metternicht 1998) and soil moisture (Engman andChauhan 1995 Narayanan and Hirsave 2001) Active radar systems can also bedesigned to collect data at varying look angles providing the opportunity for theacquisition of stereo radar images Such images can be used to produce high resolutionand extremely accurate DEMs (eg Fang 2000) A similar active sensing system isLiDAR (light detection and ranging) which uses pulses of laser light rather thanmicrowave energy to illuminate the surface (see Bunkin and Bunkin 2000 for a reviewof applications to soil mapping research) While passive microwave systems haveseemed to receive less attention in the literature a few examples of soil mapping appli-cations can be found (see Kleshchenko et al 2000 and Laymon et al 2001) Regardlessof the type of system remote sensing data and derived products are potentially usefulexplanatory variables in predictive soil mapping models

4 Fuzzy logic

Fuzzy set theory or fuzzy logic provides an alternative conceptual paradigm withinPSM research The use of this theory has increased greatly in the last few years making

it an important component of PSM Fuzzy logic is an alternative to Boolean logic thatattempts to recognize the concept of partial truth (Brule 1996) Dr Lotfi Zadeh (1965)introduced the concept and accompanying mathematics in his seminal work lsquoFuzzysetsrsquo The theory permits partial class membership in contrast to traditional set theorywhere set memberships are crisp and binary (ie a soil sample is either completelyType A or it is not at all Type A) Central to the fuzzy concept is the idea that objects innature rarely fit exactly the classification types to which they are assigned (Zadeh1965) Rather they show varying signs of similarity to multiple classes (ie an observedsoil pedon often resembles more than one of the defined soil series within the area) Byusing fuzzy membership values (ranging from 0 nonmembership to 1 totalmembership) within predictive soil models to express degrees of similarity generaliza-tion problems associated with classification schemes (filtering of information) areminimized and the complex nature of soil data is allowed to propagate through themodelling process Similarity values between 0 and 1 are not comparable to proportionsand need not add up to 1 Within Boolean logic probability statements refer to thelikelihood of an outcome the soil sample is either one series or another With fuzzylogic a given sample is not definitively a member of the subset of any one particularseries Fuzzy logic is especially useful in soil research because of the continuous andcomplex nature of the soil landscape It serves as an important alternative to thesubjective rigidity imposed on soils data by Soil taxonomy Several recent articlesprovide a thorough review of the use of fuzzy sets in soil science (Burrough 1989McBratney and Odeh 1997 Burrough et al 1997 De Gruijter et al 1997)

Within PSM research two different approaches to creating continuous classes usingfuzzy logic exist The first is based on the fuzzy-k-means classifier which partitionsobservations in multivariate space into natural classes This approach is similar tocluster analysis and numerical taxonomy but the resulting classes are continuous witheach observation assigned a fuzzy membership value that characterizes its degree ofsimilarity to each individual class The concept has been integrated into geostatisticalmethods and will be discussed in more detail below (see Section IV1) The secondapproach is known as the Semantic Import model (SI) and is used in situations whenclassification schemes are pre-defined and class limits are relatively well understoodThe SI model is commonly used in concert with expert knowledge and will bediscussed in the expert systems section (see Section IV4)

IV Recent advances in predictive soil mapping

Within the last decade many authors have sought to model the soil landscape using avariety of methods Literature in this field could be summarized many different waysbut we concentrate on the literature that directly addresses the goals of predictive soilmapping stated in the introduction (see Table 1) Therefore we will review research thatattempts to exploit the relationship between quantifiable landscape indices and soilcharacter in order to model the soil landscape in a more continuous and thereforerealistic manner

The research reviewed here is distinguished from decades of previous researchdocumenting the correlation between landscape position and soil attributes (reviewedby Hall and Olsen 1991) That body of research is informative but not useful for

P Scull et al 179

180 Predictive soil mapping a review

Table 1

Selected

recen

t literature on pred

ictive

soil m

apping an

d m

apping

(cited in

this article) d

escribing the mod

ellin

g metho

dused the

dep

enden

t variables used a

nd the

env

iron

men

tal v

ariables (ex

plan

atory) used in

the

mod

els

Stud

yMod

ellin

g metho

dDep

ende

nt variables

Environm

ental variab

les

Goa

ls atta

ined

a

Bell et

al 20

00Line

ar and

exp

onen

tial

Total soil orga

nic

Slop

e curvature aspe

ct

ECx

regression

carbon

hillslope

position

Burge

ss and

Web

ster

Punc

tual and

block

Na co

nten

t co

ver

Non

eC

1980

ab

kriging

loam

thickn

ess

ston

e co

nten

t

Burroug

h 198

9Fu

zzy mathe

matical

ndashndash

rpmetho

ds

Burroug

h et

al 19

97Con

tinuo

us classificatio

nndash

ndashrp

Castrigna

no eta

l 200

0Factorial kriging

CEC

pH N

P K

Non

e ex

plicitly used

Cx

Na

Cialella

eta

l 199

7Classifica

tion tree

Drainag

e class

Elev aspe

ct NDVI

Ec

Coo

k et

al 19

96Baysian

rule-ba

sed metho

dsOrgan

ic m

atter

Slop

e aspec

t wetne

ssEcX

inde

x

Dale et

al 19

89 (a review

)Ex

pert systems

ndashndash

ndash

Ellis 19

96Decision tree

ana

lysis

Soil erosion class

Slop

e aspec

t wetne

ssEc

neutral ne

tworks

inde

x flow le

ngth and

accu

mulation Lan

dsat

TM tree

cov

er

Gessler 1

996

A large

variety of statistic

Field an

d labo

ratory

A variety of digital

EC

metho

dsco

llected

phy

sica

len

vironm

ental da

tach

emical and

morph

olog

ical soil

prop

ertie

s

Gessler eta

l 199

5Line

ar and

logit reg

ression

A horizon

and

solum

Curvature CTI topo

EC

depth E horizon

positio

npresen

ce

Goo

vaerts 19

92Factorial kriging

Total carbon

Non

eC

P Scull et al 181pH

N CEC

extractab

lecatio

ns (K Ca M

g)Hartemink et

al 20

01ndash

ndashndash

rp

Heu

velin

k an

d W

ebster 20

01ndash

ndashndash

rp

Hew

itt 19

93ndash

ndashndash

rp

Indo

rante et

al 19

96ndash

ndashndash

rp

King et

al 19

99Lo

gistic regression

Presen

ceabsen

ceSlop

e aspec

t po

t solar

Ec

Non

calc c

lay-loam

energy

Kno

tters eta

l 199

5Kriging

co

-kriging

So

ft layer de

pth

Hillslop

e po

sitio

nCx

regression

kriging

Laga

cherie and

Holmes 19

97Classifica

tion tree

Map

ping

unit

Geo

logy variou

s topo

Ec

indices

Laslett e

tal 198

7Kriging

splin

es tren

dpH

Non

ec

surfac

e nea

rest neigh

bor

McB

ratney 1

992

ndashndash

ndashrp

McB

ratney

eta

l 199

1Block kriging

Clay co

nten

tNon

eC

McB

ratney

eta

l 200

0ndash

ndashndash

rp

McB

ratney

and

Ode

h 199

7Fu

zzy sets in

soil scienc

endash

ndashRp

McB

ratney

and

de Gruijter

Fuzz

y-k-mea

ns w

ithFu

zzy classes

Field co

llected

phy

sical

C19

92ex

tragrade

sch

emical and

morph

olog

ical soil

prop

ertie

s

McC

rack

en and

Cate 198

6Artificial intellig

ence

ndash

ndashrp

expe

rt systems

McK

ensie an

d Austin

19

93Gen

eralized

line

ar m

odels

Clay co

nten

t CEC

Slop

e relief land

form

ec

(logit)

pH EC

COLE

slop

e po

sitio

nbu

lk den

sity and

othe

rs

Moo

re eta

l 199

3Line

ar reg

ression

A horizon

dep

th O

MSlop

e w

etne

ss and

strea

mEC

and P co

nten

t pH

power in

dices aspect

curvature

182 Predictive soil mapping a reviewTa

ble 1

Con

tinu

ed

Stud

yMod

ellin

g metho

dDep

ende

nt variables

Environm

ental v

ariables

Goa

ls atta

ined

a

McK

ensie an

d Ryan 199

9Reg

ression tree and

linea

rSo

lum dep

th P an

dElevation slope

CE

regression

N con

tent

curvature CTI

contribu

ting area do

wn-

slop

e mea

ns for slope

clim

ate da

ta Presco

ttInde

x G

amma

Rad

iometry La

ndsat T

M

and Geo

logy unit

Ode

h et

al 19

92ab

Fuzz

y-c-mea

ns and

Fuzzy classes

Field-co

llected

phy

sica

lC

kriging

chem

ical and

morph

olog

ical soil

prop

ertie

s

Ode

h et

al 19

94

Reg

ression kriging

co

-So

lum dep

th de

pth

Slop

e aspect cu

rvature

eC

1995

kriging regression

to bed

rock gravel

kriging

and clay con

tent

Skidmore et

al 19

91Bayesian expe

rt system

Soil land

scap

e un

itVeg type

wetne

ss in

dex

EcX

1996

grad

ient terrain po

sitio

n

Voltz and

Web

ster 19

90Kriging

cu

bic splin

eClay co

nten

tNon

eC

Web

ster 1

994

Dev

elop

men

t of

ndashndash

rppe

dometrics

Zhu

199

7ab

Fuzz

y logic expe

rt system

Soil series A

horizon

Elev pm aspe

ct c

anop

yECX

Zhu

and

Ban

d 199

4(SoL

IM)

depth in

dividu

alco

verage

grad

ient

Zhu

eta

l 199

7series m

aps

curvature

Not

es

a Letters refer to the de

gree

to w

hich

the

goa

ls of PS

M defined

in the

introdu

ction are achiev

ed

Soilndash

e nvironm

ent relatio

ns utilized

(letter E) be

tter represen

tatio

n of soil c o

ntinuity (C

) an

d ex

pert kno

wledg

e utilize

d (X) c

orrespon

d to goa

ls 1 2 and

3respec

tively from

the

introdu

ction and

cap

ital lette

rs (E C X) ind

icate the metho

d is relatively more successful th

an th

ose metho

ds den

oted

by lower

case le

tters (e c x

) rp ind

icates rev

iew pap

ers

P Scull et al 183

predictive mapping because landscape position is never quantified (rather positionwas often qualitatively defined eg toe-slope) and thus the documented relationshipscannot be generalized using environmental data and digital elevation models to predictsoil character at unvisited sites

Table 1 documents modelling methods model variables and the extent to which thereferenced research satisfies the previously defined goals of PSM (lsquoGoals attainedrsquocolumn) Soil-environment relations utilized (letter E) better representation of soilcontinuity (C) and expert knowledge utilized (X) correspond to goals 1 2 and 3respectively from the introduction and capital letters (E C X) indicate the method isrelatively more successful than those methods denoted by lower case letters (e c x)Citations with no letters present within the lsquoGoals attainedrsquo column do not address theaforementioned goals For example the Cialella et al (1997) received a rating of lsquoEcrsquomeaning that the methods employed successfully utilized environmentalndashsoil characterrelations (E) and somewhat successfully presented a better method of representing soilcontinuity (c) The ratings are provided simply to help organize the literature that wasreviewed Review papers are included within the table denoted by lsquorprsquo Geostatisticalmethods are not included in the table because they have been comprehensivelysurveyed elsewhere (Odeh et al 1994 Burrough et al 1997 McBratney et al 2000Heuvelink and Webster 2001) and because the objectives and assumptions of geostatis-tical methods differ slightly from other PSM research We briefly outline thesedifferences in the following section

1 Geostatistical methods

Geostatistics are a subset of traditional statistics that deal primarily with spatial dataand account for spatial autocorrelation using kriging as the spatial interpolator Theconcept is based upon the theory of regionalized variables which was mainlydeveloped by Matheron (1963) and Krige (1963) Kriging is a form of weighted localaveraging that uses a measure of spatial dependence the variogram to determine theweights applied to the data when computing the averages Geostatistical methods havebeen used in predictive soil mapping research to spatially interpolate soil propertyvalues at unmeasured sites from field-collected data

Burgess and Webster (1980a b) were the first to introduce ordinary kriging to the soilcommunity and since that time an enormous amount of work has been published Forexample ordinary kriging has been used to interpolate many different soil propertiesincluding pollution trace element deficiencies salinity and fertility (Heuvelink andWebster 2001) Ordinary kriging has been criticized for a variety of reasons Forexample Laslett et al (1987) reported that several authors had criticized geostatisticsbecause kriging is a global rather than local technique failing to take into accountknowledge of soil materials and processes Other authors have criticized geostatisticsbecause they are excessively data dependent requiring a large number of closelyspaced data points (Zhu 1997a) As Webster and Oliver (1992) suggest in excess of ahundred samples may be needed to use geostatistics at the field scale because of highspatial variability of soil in some areas Geostatistics also assume spatial autocorrela-tion which sometimes may be a poor assumption in complex terrain where abruptchanges in soil-forming factors occur (McBratney et al 2000) Ordinary kriging by itself

184 Predictive soil mapping a review

does not satisfy two of the three goals of PSM presented in the introduction ndash it does notadequately incorporate expert knowledge and it does not exploit the relationshipbetween environmental variables and soil properties

Ordinary kriging has been modified in a variety of ways to better incorporateancillary data and known soilndashlandscape relationships Block kriging involvesdetermining estimates over meaningful areas rather than at specific points (Burgess andWebster 1980a McBratney et al 1991) Using this method a study area can be stratifiedinto different regions that are reflective of the pedogenetic processes at work In orderto accommodate a trend within a dependent soil variable universal kriging has beenused (Webster 1994) Kriging with external drift is similar to universal kriging but ituses an ancillary variable to represent the trend (McBratney et al 2000) Co-krigingtakes advantage of correlation that may exist between the variable of interest and othermore easily measured variables (Odeh et al 1995) Regression kriging involvesspatially interpolating the residuals from a non-spatial model by kriging and addingthe result to the prediction obtained from that model (Goovaerts 1997 Castrignano etal 2000) Factorial kriging is another method to integrate multivariate data into thestandard kriging routine to extrapolate soil data (Goovaerts 1992) Many authors havecompared these various methods (Laslett et al 1987 Voltz and Webster 1990 Odeh etal 1994 Knotters et al 1995)

Fuzzy logic has been used with geostatistics by various authors to produce new kindsof fuzzy soil maps with continuous classes (Burrough 1989 McBratney and DeGruijter1992 Odeh et al 1992a and reviewed by McBratney and Odeh 1997) The processentails kriging the matrix of membership values determined by the fuzzy k-meansclassifier resulting in a continuous soil surface where individual locations are allowedto belong to more than one class and no rigid boundaries are designated to separate thesoil into discrete units or entities The results of such analysis can be used to assess thepedologic process validity of soil taxonomy by determining whether soils grouptogether into classes that are similar to taxonomic types Fuzzy classes wouldpresumably reflect the main pedologic features within a given area In this sense theclassification is quantitative whereas soil taxonomy is rooted in qualitative discrimina-tion A comparison between the two could provide insightful and help assist soiltaxonomy in making a classification that is more indicative of underlying soil processes

Geostatistics in soil research were originally introduced to quantitatively assess soilvariability within soil mapping units (McBratney et al 1991) in response to criticisms inthe early 1970s that soil unit composition was not well quantified (Beckett and Webster1971) In this regard geostatistics have been very useful having served well the originalgoals set forth by Burgess and Webster (1980b) when they drew kriging to the attentionof soil scientists as a means of spatial prediction At the field scale soil variation islargely due to the effect that topography has on soil genesis Geostatistics have beensuccessfully applied in such environments by using terrain attributes as ancillary datawithin many of the kriging routines described above Such quantitative within-unitvariability of soil properties is very useful in the field of precision agriculture and othersituations (eg pollutants) where very detailed soil attribute information is needed atthe field scale (Heuvelink and Webster 2001)

However geostatistics have not been applied in a wide variety of environments or atlarger scales In order to be successfully applied in different environments geostatisticswill likely require a different suite of ancillary data For example remote sensing data

P Scull et al 185

could be used in arid regions where soil toposequences are less well expressed Atlarger scales of prediction selection of different sets of ancillary variables is requiredbecause different processes define soil character at different scales The most obviousexample is that of climate which may control soil distribution at large scales(continental) but has little explanatory power at the field level Regardless of whetheradequate ancillary data exist the amount of data required to use geostatistics forlandscape-level prediction would be extremely difficult and costly to collect given thestrict sampling protocol required to characterize spatial dependence It is also unclearat what landscape scale soils exhibit spatial autocorrelation

Geostatistical approaches do provide a means of creating continuous soil attributesurfaces to better represent soil continuity (Goal 2) and they can be used to exploit therelationship between environmental variables and soil properties in order to moreefficiently collect soil data (Goal 1) However they do not sufficiently utilize expertknowledge (Goal 3) as no attempt has been made in geostatistical approaches todirectly integrate expert knowledge Fundamentally kriging is a process of interpola-tion designed to predict attribute values in between locations of measured samples Inthis sense geostatistics represent a middle ground between pure interpolation (egnearest neighbour type classifier) in which only measured points for the variable ofinterest are used to determine unknown values and other predictive models thatprimarily use soilndashenvironment correlation to create predictive maps

2 Statistical methods

Statistical methods can be used to exploit the relationship between quantifiablelandscape indices and soil properties to create predictive soil maps For exampleMcKensie and Austin (1993) used a regression to account for a large percentage ofvariation for many soil characteristics (A horizon clay content CEC EC pH bulkdensity and COLE B horizon clay content CEC ESP EC pH bulk density and COLE)using a variety of predictor variables (slope presence or absence of impeding layerrelief landform topographic position) Their results confirm the hypothesis of MooreID et al (1991) that soil character is related to quantifiable landscape indices Howevertheir methods do not provide inference of soil properties at unmeasured sites frommapped environment data because the topographic variables were measured in thefield Linear regression has also been used with terrain variables derived from a 15-mDEM in northeastern Colorado to predict soil attributes (organic matter contentextractable phosphorous pH and texture) at unvisited sites (Moore et al 1993) In thatparticular study 50 of the variance of A-horizon thickness was explained by slope andthe wetness index Gessler et al (1995) also used regression to model A-horizonthickness from topographic variables in southeastern Australia (plan curvature andwetness index r2 = 063 P = 0001) They modeled solum depth and used logisticregression to model E horizon presenceabsence Elsewhere logistic regression hasbeen used to model the presenceabsence of noncalcareous clay loam horizon in centralFrance using terrain attributes from a 20-m DEM (King et al 1999) Exponentialregression has been used to model soil organic carbon using terrain variables (Bell et al2000) in glacial outwash soils in east-central Minnesota Generalized additive models(GAM) have been used less frequently in PSM research Gessler (1996) used a GAM

186 Predictive soil mapping a review

model to predict total soil carbon A horizon depth and solum depth using a variety ofenvironmental predictors (slope elevation wetness index mean annual temperatureprecipitation and radiation)

This small body of research opened the door to more complex methods by demon-strating the existence of quantifiable relationships These authors were able to producesoil attribute maps using raster data models whose scale was dependent upon the gridcell resolution of the environmental data They were successful at exploiting the rela-tionship between quantifiable topographic attributes and soil profile character (Goal 1)The continuous soil attribute surfaces they produce also better represent soil continuitythan the choropleth soil maps produced by traditional soil survey (Goal 2) Howeverthe bulk of these methods (excepting GAMs) are limited by their assumed linear rela-tionship between soil and topographic attributes their assumptions of normallydistributed data and their high data requirements Standard statistical procedures arealso not flexible enough to allow robust integration with a variety of potential datasources such as expert knowledge (Goal 1)

Statistical methods do demonstrate in a quantitative manner that terrain analysis canbe used to predict soil attributes in relatively small areas with homogeneous parentmaterial A large proportion of the research using statistical methods was conducted insemi-arid landscapes at small scales (the largest study area of the entire group was~2000 ha) Obviously for statistical approaches to be most effective they need to bemore universal As such they need to be tested andor developed at larger scales andin more diverse landscapes

3 Decision tree analysis (DTA)

The use of decision tree analysis is just beginning to be explored in predictive soilmapping research although it has been used successfully in the related field ofpredictive vegetation mapping since the early 1990s (Lees and Ritman 1991 MooreDM et al 1991 Franklin 1998) DTA is a form of divisive classification The process oftree modelling involves successively partitioning data (called recursive partitioning inthe tree modelling literature) into increasingly homogeneous subsets which once thepartitioning has ceased are called terminal nodes (Lees and Ritman 1991) Splits orrules defining how to partition the data are selected based on information statistics thatdefine how well the split decreases impurity within the data set (Clark and Pregibon1992) Splits are based on threshold values of an explanatory variable selected bycomparing the increase in resulting purity of node membership for all possiblethresholds and variables The process is iterative growing from the root node (thecomplete data set) to the terminal nodes in a dendritic fashion (Friedl and Brodley1997) Once the tree has been constructed (or grown) it encodes a set of decision rulesthat describe the data partitioning process These rules can be used to classify or predictother data sets (Moore DM et al 1991) Pruning the tree is often necessary to preventthe tree from being overfit to the sample data and to reduce tree complexity Pruningentails combining pairs of terminal nodes into single nodes and can be accomplishedusing cross-validation which yields an initial indication of how large a tree makesrobust predictions (Safavian and Norvig 1991) Cross-validation involves systematical-ly removing portions of the data set and running the remaining sample through the tree

in an iterative manner eventually yielding estimates of the misclassification rates foreach class each node and the whole model (Breiman et al 1984) In this mannerdifferent sized trees can be compared in terms of parsimony

The term DTA is used to collectively refer to all types of tree-based modelling (theword lsquodecisionrsquo is used because it is descriptive indicating that the analysis eventuallyleads to a set of decision rules defining data partitions) The term should be distin-guished from classification tree analysis because the latter refers specifically to DTAwhere the response variable is categorical The term CART (classification andregression trees ndash Breiman et al 1984) is sometimes used but strictly speaking refers tospecific software Friedl and Brodley (1997) provide a review of the decision treealgorithms They divide types of DTA into two classes (1) homogeneous decision treesfor which a single algorithm is used to estimate each split (eg CART) and (2) hybriddecision trees (HDT) for which different splitting methods can be used at differentpoints in the tree (eg Quinlan 1993) They further divide homogeneous decision treesinto univariate (UDT) where single features of the input data define splits and multi-variate decision trees (MDT) where multiple features of the input data can define splitsAccording to this naming convention no distinction is made between methods withdifferent types of response variables although all methods can be used with bothcategorical and continuous response variables In a comparison of these various typesof algorithms on a variety of data sets Friedl and Brodley (1997) found that HDT hadthe highest classification accuracy

The overall aim of DTA is to design a set of predictive rules (eg if geology type Athen soil type B) developed from training data which can then be applied to ageographic data base to predict the value of a response variable (Michaelsen et al 1994)Therefore DTA explicitly uses soilndashlandscape correlation in model development (Goal2) The technique appears promising in soil research but needs to be further tested asnot many of the above types of DTA have been tested in the PSM literature in fact onlyunivariate approaches have been employed For example Lagacherie and Holmes(1997) successfully used univariate DTA to model a categorical response variable soiltype within a training set and then assuming that the training set was representativeextrapolated the model to a much larger region Their work is interesting becausealthough not a single sample came from the area they eventually mapped they wereable to produce a soil map that was more accurate (74 versus 69) than the existingmap produced from traditional methods Cialella et al (1997) also used univariate DTAto predict soil drainage class from a variety of terrain attributes and remotely senseddata They predicted soil drainage class with an average of 78 accuracy ndash impressivegiven that the variation accounted for by a typical soil survey ranges from about halfthe total variance for physical attributes to less than one-tenth for some soil chemicalattributes (Gessler et al 1995)

DTA has been compared with other approaches by several authors In the applicationof erosion modelling DTA results were similar to Artificial Neural Networks (Ellis1996) Both methods achieved high training accuracy (as measured by the KappaStatistic) but in terms of prediction accuracy both methods performed poorly Gessleret al (1995) compared DTA with generalized linear models (GLM) and generalizedadditive models (GAM) to predict A-horizon thickness and concluded that GLM waspreferable to both DTA and GAM McKensie and Ryan (1999) compared regressiontrees and standard linear regression to predict soil properties (total solum depth soil

P Scull et al 187

188 Predictive soil mapping a review

phophorus and soil carbon) using a large variety of predictors (elevation slopecurvature wetness index climate variables geology unit Landsat TM data andGamma radiometric data) and found that different methods work best for differentproperties and overall success hinged on the strength of the relationship between soiland environmental variables Their study of a 50 000 ha forested landscape in southernNew South Wales Australia is one of the few studies at such a small map scale (largearea) or in forested landscape The most extensive comparison of techniques involvingDTA was conducted by McBratney et al (2000) They compared regression techniques(GLM GAM DTA-regression tree) geostatistical techniques (kriging and heterotopicco-kriging) and a hybrid technique (regression kriging) DTA was found to be thepoorest performing of the regression techniques because of the unrealistic predictionsurface generated by the DTA model DTA has also been criticized by other authorsbecause of the stepped prediction surface (Gessler 1996) This phenomenon can beespecially apparent in situations where predictor variables have different resolutionsThe extent to which DTA yields a better spatial representation of soil continuity is afunction of the scale and type of predictor variables used and therefore varies withindividual models

DTA is gaining widespread popularity as a means to develop prediction rules thatcan be rapidly and repeatedly evaluated (Cialella et al 1997 Franklin et al 2000) DTAprovides the following advantages over standard statistical techniques (1) it is easier tointerpret when explanatory variables are both nominal and continuous (2) it isinvariant to monotone re-expressions (transformations) of predictor variables (3) itdeals more satisfactorily with missing data values and outliers (4) it is more adept atcapturing nonadditive and nonlinear behaviour (5) it doesnrsquot make any assumptionsabout data distribution and (6) it is easily updateable as more data are collected(Moore DM et al 1991) The DTA model framework is especially appealing becauseof its capability to integrate a wide range of data sets as explanatory variables

DTA offers a unique opportunity for interaction between soil experts and soilmodellers because the output of the model is a set of rules that can be pedologicallyinterpreted by the soil expert In this sense expert knowledge is used in an implicitmanner in DTA (somewhat effectively achieving Goal 3) While these rules can often beexceedingly complex at minimum the expert can decide whether initial splits makesense given their understanding of the landscape The success of DTA results oftenhinges on the ability of the modeller to make key decisions during the model buildingprocess there is unfortunately no definitive way to determine the most optimal treeThe aid of the soil expert can potentially elucidate this problem

4 Expert systems

A variety of expert system approaches to PSM have been developed to utilize expertknowledge The purpose of such methods is to exploit the information the soil surveyoraccumulates while working in the field by integrating such knowledge into thepredictive model (McCracken and Cate 1986) Unlike the majority of the researchreviewed thus far the dependent variable in many expert systems models is often soiltaxa or mapping unit This apparent disadvantage of expert systems (using classifica-tion to characterize soil continuity) does make them easier to integrate into traditional

soil survey In addition several authors have developed methods to developcontinuous soil property maps from the output of expert system models designed topredict soil unit occurrence

Expert systems are composed of data (information on spatial environmentalvariables eg topography climate etc) a knowledge base (rules and facts related tosoil variation supplied from the soil surveyor) and an inference engine (whichcombines data and the knowledge base to infer logically valid conclusions) (Skidmoreet al 1996) Expert systems differ from conventional models in two ways (1) they storeand manipulate qualitative information (allowing them access to information thatcannot normally be used in other modelling frameworks) and (2) they are structuredas meta-models (the knowledge is separated from the model) (Davis 1993) This allowsthe model to selectively choose which information is relevant at various stages of themodelling process and it allows for information to be easily updated Davis (1993)reviews the application of expert systems to environmental modelling research conclu-ding that the technique is becoming more widely accepted He further states that theapplication of expert systems is constrained by an absence of fundamental knowledgefor rule generation a problem that would appear less relevant to soil mapping giventhe amount of untapped expert knowledge accumulated by the soil surveyor

The first mention of the use of expert systems in pedology was in a paper presenta-tion at the Northeast Committee Soil Survey Conference 1984 (Flach 1985) In herpaper Flach (1985) hinted that recent developments in computer science especiallyexpert systems and artificial intelligence could make modelling a practical mappingtool for soil scientists in the near future A year later McCracken and Cate (1986) hopedto encourage soil scientists to explore expert systems and its application to soil sciencethrough an optimistic article they wrote on the potential use of expert systemsHowever little research was actually conducted in the late 1980s and expert systemshad not yet begun to fulfil the lofty goals set forth by early practitioners (Dale et al1989)

In the early 1990s expert systems approaches to predictive soil mapping began toappear in the literature Skidmore et al (1991) used a Bayesian expert system to mapforest soil into different classes and their results compared favourably with availablesoil maps and actual field-collected data Their methods successfully incorporatedsurveyor knowledge and remotely sensed and digital terrain attributes but failed tobetter characterize continuous soilndashlandscape variability because their final productwas a choropleth map Skidmore et al (1996) revisited their earlier research andassessed the mapping accuracy of their results They found that the soil map producedby the expert system achieved an overall accuracy of 698 (sample size = 53) while themap derived from conventional methods had an accuracy of 736 In addition toknowledge provided by a soil scientist Skidmore et alrsquos (1996) expert system usedvegetation derived for aerial photography and topographic variables derived from a 10-m DEM (wetness index topographic position and slope)

Using an expert system Cook et al (1996) successfully produced a continuous soilproperty map for organic matter content using wetness index aspect and slope asexplanatory variables Their methods were somewhat inefficient because they requireda separate expert system for each soil property of interest However their researchrepresents the only example of expert systems used to directly predict the spatial dis-tribution of a soil property All of the other examples of expert systems in the literature

P Scull et al 189

190 Predictive soil mapping a review

deal with soil type or class The use of expert systems to map soil properties needs tobe explored further

The expert systems discussed thus far have all used Boolean logic within theirtheoretical framework whereby an observation can belong to one and only one classand the soil properties of that class are assigned to the observation With thedevelopment of fuzzy logic and the semantic import model (the second fuzzy logicapproach to creating continuous classes) the opportunity exists to develop fuzzy logic-based expert systems Such systems can be used in conjunction with expert knowledgein situations where experts have a good qualitative idea of how to group data but havetrouble dealing with observations that are not well represented by rigid classificationschemes (Burrough and McDonnell 1998) This method can be particularly useful insituations when taxonomic schemes have been previously developed as is the case insoil taxonomy Several examples of this type of approach were published in the mid-1990s by A Zhu and colleagues (Zhu and Band 1994 Zhu et al 1996 1997 Zhu1997ab) Such systems proved useful for mapping soil at unvisited locations usingsurveyor knowledge and were also capable of producing continuous soil propertymaps The use of fuzzy logic within the theoretical framework of the expert systemallows the complex nature of soil to propagate through the modelling process neversubjected to classification schemes that filter out potentially useful lsquonoisersquo The use offuzzy logic also gives the soil surveyor more latitude during the interview processwhen the knowledge base is defined for the expert system Zhu et al (1997) developeda fuzzy logic-based expert system called SoLIM to determine the similarity of eachgrid cell in a study area to the various taxonomic mapping units delimited by the soilsurvey Continuous soil attribute maps were calculated using the similarity values andtheir relative soil survey determined attribute values (effectively achieving Goal 2) Theresulting data (which proved more accurate than soil survey data once field checked)consists of a raster grid whose resolution is determined by the resolution of the inputenvironmental and digital elevation data As noted previously this type of data modelis more applicable to environmental modelling than the choropleth map (Burrough andMcDonnell 1998)

Expert system approaches to PSM are capable of exploiting soil surveyor knowledgeby developing rule-based systems that imitate the surveyorrsquos conceptual model of soilvariability (the primary focus of Goal 3) The method would appear extremely usefulfor mapping projects (such as those conducted by the NRCS-NCSS) where fieldwork isinitially conducted to determine soilndashlandscape relations Expert system developmentcould be directly inserted into the traditional soil survey mapping approach as asubstitute for the step where the surveyor converts hisher conceptual model into achoropleth map Rather that knowledge could be incorporated into the expert systemwhich could be used to predictively map soil The resulting raster map would be morescientifically based and explicit than the hand-drawn choropleth maps of the past Itsscale would not be limited to that of the aerial photography but rather to the scale ofthe environmental data

Despite these advantages the expert system approach has some drawbacks Becausethe method is dependent upon expert knowledge it cannot be applied whereenvironmentndashsoil relations are poorly understood Of course this criticism can be madeof all PSM models as inductive or deductive knowledge of soilndashenvironment relationsis a prerequisite for PSM However expert systems are deductive models and as such

are not driven by any specific field-collected data (although presumably the soil experthas field experience in the mapping area) Expert systems donrsquot afford the opportunityto first statistically document landscapendashsoil relations and then extrapolate the resultsbecause expert systems do not directly use sample soil data (lsquohard datarsquo) to determinesoilndashlandscape relations Thus the utilization of the relationship between environmen-tal variables and soil properties (Goal 1) is only indirectly achieved Expert systemsapproaches have been demonstrated to be extremely effective in a small number of casestudies The possibility of satisfying all three goals discussed in this review makesexpert systems a predictive soil mapping method that needs to be further tested ndash bothgeographically and across different scales of analysis

V Conclusion

Most of the predictive soil mapping research outlined in this review was conducted atvery large map scales (over small areas) In fact the majority of the research wasconcerned with assessing the spatial variability of soil character within individual fieldsor across soil toposequences The primary driving force behind this type of research hasbeen the need to provide accurate soils information for agriculture and ecologicalmodels It is clear that terrain attributes are powerful predictors at the local scaleGeostatistical tools have been successful at using terrain attributes and the spatialdependence of soil properties to interpolate between existing data points withinindividual fields Across soil toposequences statistical approaches provide a usefulmeans of predicting soil character PSM research has been most successful at the fieldscale because many of the soil-forming factors are held constant For example the neteffect of four of Jennyrsquos five soil-forming factors (climate organism parent material andtime) was minimal within many of the studies reviewed While some research hastackled larger areas of study there exists a deficiency in our ability to predictively mapsoils at smaller map scales Since the distribution of soil is scale-dependent differentPSM methods and predictors are likely to work better at different scales Focus in thefuture must continue to move toward working over larger spatial extents of study inorder to produce landscape-scale soil information

Further a large proportion of the research was conducted in semi-arid gentlysloping agricultural landscapes Humid forestlands mountainous regions and desertshave received little attention As such PSM methods need to continue to be testedandor developed in a wider variety of landscapes where spatial soil distributions canbe more complex Different methods will likely be successful to different degrees indifferent environments Whereas terrain attributes are dominant predictors of soilcharacter across toposequences in prairie lands other predictors are likely necessary inareas where soils do not develop into clear toposequences For example remote sensingdata has been a useful predictor of soil chronosequences in desert landscapes wheresurface appearance is often related to soil character Surficial geology is often reflectiveof soil character in mountain regions where soils are thin and significant bedrock existsat the surface Focus in the future must continue to determine which methods andpredictors work best in which environments

Since the most useful PSM approach will vary across spatial scales and environmen-tal gradients the method used should be driven by the mapping objectives of the

P Scull et al 191

192 Predictive soil mapping a review

project Whereas the traditional soil survey of the past was expected to meet the needsof a diverse group of end users PSM methods will vary given the objectives of thesurvey As a result the end users of the soil data need to play a more active role in thesurvey process Because precision agriculture and large-scale mapping has been theprimary focus of PSM there exists a deficiency in our ability to predictively map soilsfor the purpose of general land use planning and management Even though lessdetailed soil information is needed for these purposes the development of PSMmethods is complicated by the fact that soils exhibit complex spatial variability atsmall map scales where soil-forming environments vary greatly from one location tothe next

A number of alternative methods of characterizing the continuous nature of the soillandscape have been developed Thus far most of the PSM research has provided soilinformation in a nonobject form (as opposed to defining soil types as independententities) Soil data have been generated organized and presented in the form of eitherisorithmic maps or fine-scale raster grids Both of these data models are field-viewmodels of geographic space which allow the soil to be perceived as a constantlyvarying surface Two distinct approaches have been employed (1) mapping individualsoil properties and (2) mapping continuously varying (fuzzy) soil classes Mappingindividual properties is the most common approach and will likely continue todominate PSM research The use of fuzzy soil classes in the literature is less commonand more difficult to be integrated into standard mapping procedures (such as use ofsoil taxa) because the concept radically differs from the traditional view of the soillandscape However fuzzy soil classes (generated using the fuzzy-k means approach) dohave the potential to help further advance Soil taxonomy by identifying taxonomicclasses that are more reflective of pedologic processes at work

Expert systems have been greatly underutilized in PSM research especiallyconsidering how effective a small number of case studies have been Expert systemshave the potential to satisfy successfully all three goals discussed throughout thereview They also have the potential to bridge the gap between traditional approachesand PSM methods because field soil scientists do not have to change their conceptualapproach to mapping They can still conduct field reconnaissance to determinesoilndashlandscape relationships Afterwards though they can use expert systems to betterexploit the knowledge they have garnered Such an approach could help ease thetransition from traditional soil survey to more scientifically explicit methods Expertsystems are also the most fruitful approach to utilizing a wealth of data that has alreadybeen collected in a nonexplicit manner the wealth of expert knowledge (in the form ofthe conceptual model) that senior soil mappers have accumulated

Focus in the future must also move in the direction of operability The utility of manyPSM approaches has been clearly demonstrated yet changes in how soils are mappedand perceived has been slow In the USA various members of the National CooperativeSoil Survey have called for a change in the philosophy of survey (from mappingdiscrete soil types to mapping continuous varying soil properties McSweeny et al1994) but little change has actually taken place To realize the potential of PSM in con-tributing to an overall change in standard soil mapping procedures practitionersworking in the field of PSM need to form working relationships with field soilscientists PSM methods need also to be presented in a manner that is comprehensibleto the soil science trained field mapper Predictive soil mapping is a relatively recent

phenomenon yet much progress has been made In the process of mapping soilknowledge will continue to advance

P Scull et al 193

References

Agbu PA Fehrenbacher DJ and Jansen IJ1990 Statistical comparison of SPOT spectralmaps with field soil maps Soil Science Society ofAmerica Journal 54 818ndash18

Barrett L 1999 Particulars in contextmaintaining a balance in soil geography Annalsof the Association of America Geographers 89707ndash13

Beckett PHT and Webster R 1971 Soilvariability a review Soils and Fertilizers 341ndash15

Bell JC Grigal DF and Bates PC 2000 A soil-terrain model for estimating spatial patterns ofsoil organic carbon In Wilson J and Gallant Jeditors Terrain analysis principles and applica-tions New York City NY John Wiley and Sons295ndash310

Birkeland PW 1999 Soil and geomorphologyThird edition New York NY Oxford UniversityPress

Breiman L Friedman JH Olshen RA andStone CJ 1984 Classification and regressiontrees Belmont CA Wadsworth

Brule FJ 1996 Fuzzy systems ndash a tutorialhttp newsgroup compai httpwwwquadralaycom (last accessed 30 August 2001)

Bunkin FV and Bunkin AF 2000 Lidarsounding of water soil and plants Atmosphericand Oceanic Optics 13 54ndash72

Buol SW Hole FD McCracken RJ andSouthard RJ 1997 Soil genesis and classifica-tion Ames IA Iowa State University Press

Burgess TM and Webster R 1980a Optimalinterpolation and isarithmic mapping of soilproperties the semi-variogram and punctualkriging Journal of Soil Science 31 315ndash31

ndashndashndashndash 1980b Optimal interpolation and isarithmicmapping of soil properties block kringingJournal of Soil Science 31 331ndash41

Burrough PA 1989 Fuzzy mathematicalmethods for soil survey and land evaluationJournal of Soil Science 40 477ndash92

Burrough PA and McDonnell RA 1998Principles of geographic information systems(Revised edition) Oxford Clarendon Press

Burrough PA Beckett PHT and Jarvis MG1971 The relation between cost and utility insoil survey Journal of Soil Science 22 368ndash81

Burrough PA Van Gaans PMF and

Hootsman R 1997 Continuous classificationin soil survey spatial correlation confusionand boundaries Geoderma 77 115ndash35

Cambell JB 1977 Variation of selectedproperties across a soil boundary Soil ScienceSociety of America Journal 41 578ndash82

Cambell JB and Edmonds WJ 1984 Themissing geographic dimension to soiltaxonomy Annals of the Association of AmericanGeographers 74 83ndash97

Castrignano A Giugliarini L Risaliti R andMartinelli N 2000 Study of spatial relation-ships among some soil physico-chemicalproperties of a field in central Italy using multi-variate geostatistics Geoderma 97 39ndash60

Cialella AT Dubayah R Lawrence W andLevine E 1997 Predicting soil drainage classusing remotely sensed and digital elevationdata Journal of Soil Science 62(2) 171ndash78

Clark RN and Swayze GA 1996 Evolution inimaging spectroscopy analysis and sensorsignal-to-noise an examination of how far wehave come Summaries of the sixth annual JPLairborne Earth science workshop 4ndash8 March 1996AVIRIS Workshop Vol 1 5

Clarke LA and Pregibon D 1992 Tree-basedmodels In Chambers J and Hastie J editorsStatistical models in S Pacific Grove Wadsworthand Brooks 377ndash419

Coleman TL Agbu PA and MontgomeryOL 1993 Spectral differentiation of surfacesoils and soil properties ndash is it possible fromspace platforms Soil Science 155 283ndash93

Cook SE Corner RJ Grealish G GesslerPE and Chartres CJ 1996 A rule-basedsystem to map soil properties Soil ScienceSociety of America Journal 60 1893ndash900

Cruickshank JG 1972 Soil geography New YorkNY John Wiley amp Sons

Csillag F Pasztor L Biehl LL 1993 Spectralband selection for the characterization ofsalinity status of soils Remote Sensing ofEnvironment 43 231ndash42

Dale MB McBratney AB and Russell JS1989 On the role of expert systems andnumerical taxonomy in soil classificationJournal of Soil Science 40 223ndash34

Davis JR 1993 Expert systems and environ-

172 Predictive soil mapping a review

spatial distribution and management of soil is critical to maintain a productive societyand to understand the complex balance of chemical and physical processes that makelife possible on Earth

While primarily referred to as pedology and studied in soil science departments suchthemes are the focus of many different academic disciplines For example soilgeography is the study of the location distribution and pattern of soils on the landscape(Buol et al 1997) Geography and pedology share a common history in this centuryCF Marbut a geographer by training was responsible for communicating pedologicalideas developed by the Russian school of pedology within the USA (Cruickshank1972) Marbut was a student of WM Davis the first president of the AmericanAssociation of Geographers (AAG) and later became president himself in 1924 As aresult soil was an active area of research within geography during the early years of theAAG While soil geography is still occasionally published in geography journals (seeBarrett 1999 for a recent perspective) a great deal of geographic soil research hasappeared over the last 20 years outside of geographic journals

From an applied perspective pedology and soil geography form the basis of soilsurvey which remains the primary means by which information on the spatialproperties of soil is collected presented and archived in the USA and throughout theworld The United States Department of Agriculture (USDA) in cooperation with otherfederal and state agencies began soil survey work in 1896 The combined state andfederal government effort became known as the National Cooperative Soil Survey(Indorante et al 1996) Its mission was to furnish private landowners land managersand consultants with soil maps to aid in land-use decision making More recentlyhowever soil maps have been used to provide chemical and physical data input withinecological and hydrological process models (Burrough and McDonnell 1998 163) Thesoil survey program was simply not designed to furnish data for such applications Theincreasingly sophisticated use of soil data has led to a greater demand for data aboutsoil properties than the conventional soil map can accommodate (Cook et al 1996)Traditional soil survey concepts are based on qualitative recognition of soil propertiesin relation to landscape and environmental variables Although these methodsimplicitly incorporate the expertise of the soil scientist they do not make use of geo-computational technologies that are now widely available

Technological advances during the last few decades have created a tremendouspotential for improvement in the way that soil maps are produced (McKensie et al2000) Remote sensing and photogrammetric techniques provide spatially explicitdigital data representations of the Earthrsquos surface that can be combined with digitizedpaper maps in geographic information systems (GIS) to allow efficient characterizationand analysis of vast amounts of data The future of soil survey lies in using GIS tomodel spatial soil variation from more easily mapped environmental variablesPredictive soil mapping (PSM) begins with the development of a numerical or statisticalmodel of the relationship among environmental variables and soil properties which isthen applied to a geographic data base to create a predictive map (Franklin 1995) Threemain goals of PSM are to (1) exploit the relationship between environmental variablesand soil properties in order to more efficiently collect soil data (2) produce and presentdata that better represent soil landscape continuity and (3) explicitly incorporate expertknowledge in model design PSM can also potentially advance pedology and soilgeography by providing insights into soil forming processes

The State Factor soil-forming model forms the theoretical basis of PSM as well as oftraditional soil survey and has been used in other many other disciplines Hans Jennypopularized the State Factor theory of soil formation in the USA through his publicationof Factors of soil formation (1941) The ideas expressed in this classic text represent a for-malization of Dokuchaievrsquos ideas of soil formation (Simonson 1997) Since that timethe theory has provided a paradigm through which soil genesis and distribution can bestudied (Schaetzl 1991) The theory states that soil profile character is a function ofclimate organisms relief parent material and time implying that if the spatial distrib-ution of the soil-forming factors is known soil character may be inferred Thistheoretical framework has been used by many authors in pedological research andremains the most popular theory of soil genesis (for a recapitulation of Jennyrsquoscontinuing influence on pedology see McSweeney et al 1994)

Until recently predictive soil mapping efforts have been constrained by twoproblems associated with Jennyrsquos State Factor theory of soil genesis the mathematicalequations derived from theory proved virtually unsolvable (Huggett 1975) and datadescribing the soil-forming factors were unavailable or not widely available Recentadvances in mathematical theory (fuzzy logic) and statistical methods (includingmachine learning techniques) have helped researchers to better understand the natureof the equations implied by the theory and to solve them Relative to the problem ofsoil-forming factor knowledge Jenny (1941 262) wrote lsquothe conversion of suchfundamental knowledge (soilndashenvironment relationships) to specific field conditions isimpossible unless the areal distribution of the soil formers is knownrsquo Half a centurylater remote sensing and photogrammetry along with GIS can now be used to charac-terize the spatial distribution of soil-forming factors The foundation has been laid forthe emergence of predictive soil modeling as an active area of research

Several articles have been published outside of geographic literature to trackprogress in the field of PSM over the years Two earlier articles (McBratney 1992Hewitt 1993) called for an overall change in the philosophy of soil survey and outlinedresearch challenges in the field at that time In the mid 1990s the term pedometrics wascoined (Webster 1994) to refer to quantitative research in the field of pedology Thisterm would appear synonymous with PSM as defined above More recently reviewshave been published on a subset of PSM techniques (McBratney et al 2000) and generalmethods of modelling soil variation (Heuvelink and Webster 2001)

Predictive soil mapping has only recently emerged as a research niche but the paceof research breadth of knowledge and variety of techniques have expanded rapidly(pedometrics have accounted for 18 of the subject matter in articles published inrecent years in Geoderma Hartemink et al 2001) A review of recent achievements in thefield of PSM is needed because all methods have not been reviewed in past papersFurther a geographic perspective is needed because of the inherently geographicnature of PSM First we review several topics critical to PSM (Section II) including thenature of soil variability (Section II1) how the soil survey has defined soil mapping(Section II2) and soil taxonomy (Section II3) We then briefly describe the componentsof PSM (Section III) geographic data models of soil variability and soil mapping(Section III1) digital terrain modelling (Section III2) fuzzy logic (Section III3) andremote sensing (Section III4) Lastly we discuss the present state of knowledge in PSM(Section IV) Our focus in this review is to evaluate literature in the field of PSM and toreintroduce the subject matter to the geographic audience

P Scull et al 173

174 Predictive soil mapping a review

II Background

Prediction of soil properties based on knowledge of the effect of environmentalvariables on soil formation has always been the basis for all soil mappingUnfortunately the traditional methods do not yield quantifiable soilndashlandscapeinformation that robustly describes actual soil variation In this section we describe soilvariation and explain why traditional methods for defining soil distribution inlandscapes are inadequate It is also our purpose to demonstrate that there is a logicalflow from these approaches into the new PSM techniques All approaches to soilmapping rest on our ability to use knowledge of the process of soil genesis to predictthe properties of soils at any point in the landscape

1 The nature of soil variability

The complex and highly variable nature of soil patterns in landscapes complicates thealready labour-intensive process of collecting and presenting soil survey data (Wrightand Wilson 1979) In 1941 Jenny listed three different definitions of soil before heconceded that lsquoit is problematic whether any definition of soil could be formulatedrsquo(Jenny 1941 2) In the half-century that has passed since Jenny wrote those words therehave been numerous attempts to define soil (reviewed by Birkeland 1999) Mosttextbook authors describe soil as extraordinarily complex For example McKnight(1993 336) defines soil as lsquoan infinitely varying mixture of weathered mineral particlesdecaying organic matter living organisms gases and liquid solutionsrsquo (emphasisadded) This definition illustrates how complex soil can be It further demonstrates thatthe soil landscape is continuous and is not composed of distinct individual soil typesThis point has been made repeatedly over the years (Simonson 1959 Webster andBeckett 1968 Cambell 1977 and Moore et al 1993)

A variety of soil genesis models have been proposed in order to account for the highvariability of soil and collectively they help further illustrate the difficulty of charac-terizing the soil landscape (for a review see Huggett 1975 Birkeland 1999) Threedistinctive approaches have been employed factor models (eg Jenny 1941) wherefactors affecting soil development are identified process models (eg Simonson 1959)where soil-forming processes are emphasized and energy models (eg Runge 1973)where the focus is upon process-driving mechanisms A host of hybrid models havealso been employed Still no real consensus exists today as to exactly how to model soildevelopment partially because of the recent emergence of pedology as an academicdiscipline (Johnson and Watson-Stegner 1987) Existing data collection methods do notyield adequate soil information in part because many of the processes that shape thesoil landscape are still poorly understood

2 How the soil survey has defined soil mapping

Traditional soil survey persists as the most popular form of soil mapping and inventoryand in many cases is the only manner in which the highly variable nature of the soillandscape is catalogued The method consists of three steps (Cook et al 1996) The firstis direct observation of ancillary data (aerial photography geology vegetation etc) and

soil profile characteristics In the second step the observations of soil attributes areincorporated into an implicit conceptual model that is used to infer soil variation Thethird step involves applying the conceptual model to the survey area to predict soilvariation at unobserved sites Usually less than 0001 of the survey area is actuallyobserved (Burrough et al 1971) a fact reflecting the high cost of field sampling Theconceptual model of soil variation is then transformed into a cartographic model thechoropleth map by drawing map unit boundaries on aerial photographs In effect pho-tographic scale determines the resolution of the soil map

This process has been severely criticized in the scientific literature for two reasonsFirst the conceptual model developed by the soil surveyor is primarily implicit beingconstructed in a heuristic manner This results in an excessive dependence upon tacitknowledge and as such incomplete information exists relative to the derivation of theultimate soil survey product (Hudson 1992) This aspect of soil survey is especiallyfrustrating because it fails to document most of the knowledge that the soil surveyoraccumulates during the expensive field mapping process In essence the soil survey isunfalsifiable and therefore untestable (Hewitt 1993) The final product of soil survey isa soil map that has unknown assumptions limitations and accuracy (Burrough et al1971 Dijkerman 1974) PSM techniques are similar in theory to soil survey (they bothuse knowledge of soilndashenvironment relations to make inferences) but the methodsemployed often yield quantitative expressions of soil variability with measured levelsof accuracy Clearly one direction of innovation in soil survey is to add objectivity tomodel development which will allow more explicit scientific communication (Gessler1996) The second major criticism of soil survey concerns the role of soil classification

3 Soil classification

The evolution of PSM techniques has been directly impacted and continues to beinfluenced by the process of soil classification This is especially true in the USAwherethe primary focus of the National Cooperative Soil Survey is to develop and map Soiltaxonomy (the official classification developed by the US Department of AgricultureSoil Survey Staff 1975) The system was influenced by nineteenth century biologicaltaxonomy and the practice of geological survey (Heuvelink and Webster 2001) Thepurpose of Soil taxonomy was to provide an objective manner to systematically classifysoil and was adopted at a time when soil information had to be abstracted to the levelof the modal profile (classified in Soil taxonomy) because it was impossible to catalogueand present the full amount of soil variability (Cambell and Edmonds 1984) In orderto map soil taxa the soil must be perceived as a spatial entity a lsquopedonrsquo (a term used inSoil taxonomy to refer to the smallest recognizable unit that can be called lsquoa soilrsquo) Inpractice this spatial perception of soil results in a map whose classes are homogenousunits with unknown variability and sharply defined boundaries (Burrough andMcDonnell 1998) Since the initial development of Soil taxonomy the perception of thesoil landscape has changed from a collection of individual soil types to a continuallyvarying mixture of soil components (Dmitriev 1983) Field observations have shownthat in concert with the environmental variables soil properties vary continuouslyacross landscapes exhibiting different and complex scales of variation (Simonson1959) Therefore soil distribution is not well represented by choropleth maps

P Scull et al 175

176 Predictive soil mapping a review

(McBratney 1992 Gessler et al 1995) While this change has not yet been recognized inSoil taxonomy PSM research has developed methods of soil inventory that moreaccurately describe the soil landscape Such methods are at odds with the traditionalapproaches because they often involve dismissing the concept of soil as a spatial entityRather the new methods focus on mapping continuously varying soil properties Theimplementation of these techniques has been difficult because of the entrenchment ofSoil taxonomy

III Components of predictive soil mapping

1 Geographic data models of soil variability and soil mapping

Traditionally soil maps have been digitized to fulfil the need for soil data within GIS-based environmental modelling research This information from the paper mapdigitized for the computer is fraught with the same problems of the original choroplethmaps ndash assumed homogeneous units with unknown variability and sharply definedboundaries GIS allows for a more robust characterization of spatial variability (relativeto the cartographic generalizations of the past) by allowing information to be analysedand stored using a variety of data models As defined in the GIS literature datamodelling is the process of discretizing spatial variation which entails abstracting gen-eralizing or approximating geographic reality (defined as empirically verifiable factsabout the real world) Unfortunately data modelling is often confused with issues ofdata structure and limited by software selection (Kemp 1992) The process is of crucialimportance because it controls the manner in which the data can be processed oranalysed (Goodchild 1994) as well as the view of the data the end user ultimatelyreceives (Goodchild 1992a Kemp 1992) Some data models are more accurate thanothers at portraying geographical reality (Goodchild 1992b) In the case of soil howwell do the data including the data model represent the highly variable continuousnature of soil How can digital computers be used to manage spatial data to bestrepresent the soil landscape

The choice of data models is partially dependent upon how the soil is perceived ingeographic space Objects can be thought of as existing as independent entities inempty space (object view) or as one of an infinite set of tuples (the foundation ofgeographic information ndash xyz where z is a measured value and xy are its location inspace) approximated by regions and segments (field view) (Goodchild 1994) The fieldview better represents continuous surfaces such as the soil landscape but soil datahave traditionally been modelled using the feature model of geographic space thechoropleth map Soil data have probably been managed this way because of theinfluence of Soil taxonomy which defines individual soil lsquotypesrsquo and because soil datacollection began during a time when few alternatives existed

Using PSM techniques a fundamental change in the soil data model from thechoropleth map to the raster grid allows better characterization of actual soilndashlandscapevariability A raster data model accommodates a lsquofield viewrsquo representation of thelandscape and is defined as a regular rectangular array of cells with some aggregatevalue of the field recorded for each cell (Goodchild 1994) The resolution of the datastored in this format is a function of the grid cell size which can be made small enoughto simulate continuous variation at the landscape scale The raster has become the most

widely used data model for PSM and is also routinely used to manage other environ-mental information such as elevation (DEMs) and remotely sensed data Spatialanalysis and integration with other types of raster-based environmental data can beeasily performed with soil data stored using a raster data structure (Burrough andMcDonnell 1998)

2 Digital terrain modelling

Terrain analysis quantifies the relief component of models characterizing soilformation Soil development and its associated profile characteristics often occurs inresponse to the way in which water moves through and over the landscape which iscontrolled by local relief Accordingly terrain analysis will be most useful in environ-ments where topographic shape is strongly related to the processes driving soilformation (McKensie et al 2000) Digital terrain modelling is a technique for derivingspatially explicit quantitative measures of the shape character of topography (Weibeland Heller 1991 Wilson and Gallant 2000) The spatial distribution of the resultingterrain attributes (characterizing local water flow paths) can also capture the spatialvariability of soil attributes Moore ID et al (1991) reviewed the analysis of digitalelevation data (including DEMs) for hydrological geomorphological and biologicalapplications They provided a table that summarized the significance and physicalmeaning of various terrain attributes to landscape processes Building on their workmany authors have used terrain attributes derived from digital elevation models(DEMs) as explanatory variables in predictive soil models (Odeh et al 1991 Moore etal 1993 Gessler et al 1995 Skidmore et al 1996 and others) Methods used to deriveterrain attributes have been greatly refined over the last 15 years and future satellitesaiding in the development of more accurate DEMs will make terrain analysis anincreasingly important component of predictive soils mapping (Moore et al 1993Mackay and Band 1998) Several recent review articles have been specifically devotedto the role of terrain analysis in soil mapping (Ventura and Irvin 1996 Irvin et al 1996McKensie et al 2000)

3 Remote sensing

Remote sensing data are an important component of PSM because they provide aspatially contiguous quantitative measure of surface reflectance which is related tosome soil properties (Agbu et al 1990) Both physical factors (eg particle size andsurface roughness) and chemical factors (eg surface mineralogy organic mattercontent and moisture) control soil spectral reflectance (Irons et al 1989) Surfacemineralogy can be derived by wavelength specific charge transfer and crystal fieldabsorptions associated with the presence of iron and iron-oxides (Fe2+ and Fe3+) andvibrational absorptions associated with hydroxyl bonds in clays adsorbed water andthe carbonate ion (Goetz 1989 Irons et al 1989) The presence and strength of theseabsorption features can be used to identify and quantify concentrations of mixed suitesof minerals in soil (Johnson et al 1983 Shipman and Adams 1987) Organic matterparticle size and moisture content in contrast influence soil reflectance primarilythrough a change in average surface reflectance and produce only broad spectral

P Scull et al 177

178 Predictive soil mapping a review

expression (Irons et al 1989) Adecrease in particle size tends to increase surface albedoand decrease spectral contrast of absorption features while an increase in organicmatter or soil moisture decreases average reflectance or albedo

Numerous studies have shown the potential benefits of using remote sensing for soilidentification and mapping Comprehensive surveys of soil spectral reflectance includestudies by Stoner et al (1980) Henderson et al (1992) and Csillag et al (1993) Remotesensing studies based on broad-band sensors such as Landsat TM include Agbu et al(1990) Coleman et al (1993) Seyler et al (1998) and Oliveira (2000) Traditionallyremote sensing has been used to classify soil units through photo-interpretation ordigital image processing Combining remotely sensed information with ancillaryinformation such as thematic maps or vegetation cover can yield significant improve-ments (Wilcox et al 1994 Cialella et al 1997 Wanchang et al 2000)

Recent developments in hyperspectral remote sensing offer the potential of signifi-cantly improving data input to predictive soil models Hyperspectral sensors such asthe Airborne VisibleInfrared Imaging Spectrometer (AVIRIS) measure a contiguousspectrum in the visible and NIR and thereby better characterize atmospheric andsurface properties (Goetz et al 1985) The large number of spectral bands permits directidentification of minerals in surface soils For example Clark and Swayze (1996)mapped over 30 minerals using AVIRIS at Cuprite Nevada Palacios-Orueta and Ustin(1996) showed that enhanced spectral information was suitable for discriminating evensubtle spectral changes associated with differences in organic matter and iron contentOther examples of the application of AVIRIS to aid soil mapping include Palacios-Orueta et al (1998) Okin et al (1998) and Roberts et al (1998)

Sensors that operate in the microwave portion of the electromagnetic spectrum havealso shown promise in soil mapping research Microwave sensing can be broadlydivided into active (eg radar) systems and passive systems and are capable ofpenetrating the atmosphere under virtually all conditions offering a significantadvantage over visible and near-infrared spectroscopy (for a general overview ofmicrowave sensing see Lillesand and Kiefer 1994 chapter 8) Synthetic aperature radar(SAR) is one example of an active system SAR has been used to aid soil propertymapping such as soil salinity (Metternicht 1998) and soil moisture (Engman andChauhan 1995 Narayanan and Hirsave 2001) Active radar systems can also bedesigned to collect data at varying look angles providing the opportunity for theacquisition of stereo radar images Such images can be used to produce high resolutionand extremely accurate DEMs (eg Fang 2000) A similar active sensing system isLiDAR (light detection and ranging) which uses pulses of laser light rather thanmicrowave energy to illuminate the surface (see Bunkin and Bunkin 2000 for a reviewof applications to soil mapping research) While passive microwave systems haveseemed to receive less attention in the literature a few examples of soil mapping appli-cations can be found (see Kleshchenko et al 2000 and Laymon et al 2001) Regardlessof the type of system remote sensing data and derived products are potentially usefulexplanatory variables in predictive soil mapping models

4 Fuzzy logic

Fuzzy set theory or fuzzy logic provides an alternative conceptual paradigm withinPSM research The use of this theory has increased greatly in the last few years making

it an important component of PSM Fuzzy logic is an alternative to Boolean logic thatattempts to recognize the concept of partial truth (Brule 1996) Dr Lotfi Zadeh (1965)introduced the concept and accompanying mathematics in his seminal work lsquoFuzzysetsrsquo The theory permits partial class membership in contrast to traditional set theorywhere set memberships are crisp and binary (ie a soil sample is either completelyType A or it is not at all Type A) Central to the fuzzy concept is the idea that objects innature rarely fit exactly the classification types to which they are assigned (Zadeh1965) Rather they show varying signs of similarity to multiple classes (ie an observedsoil pedon often resembles more than one of the defined soil series within the area) Byusing fuzzy membership values (ranging from 0 nonmembership to 1 totalmembership) within predictive soil models to express degrees of similarity generaliza-tion problems associated with classification schemes (filtering of information) areminimized and the complex nature of soil data is allowed to propagate through themodelling process Similarity values between 0 and 1 are not comparable to proportionsand need not add up to 1 Within Boolean logic probability statements refer to thelikelihood of an outcome the soil sample is either one series or another With fuzzylogic a given sample is not definitively a member of the subset of any one particularseries Fuzzy logic is especially useful in soil research because of the continuous andcomplex nature of the soil landscape It serves as an important alternative to thesubjective rigidity imposed on soils data by Soil taxonomy Several recent articlesprovide a thorough review of the use of fuzzy sets in soil science (Burrough 1989McBratney and Odeh 1997 Burrough et al 1997 De Gruijter et al 1997)

Within PSM research two different approaches to creating continuous classes usingfuzzy logic exist The first is based on the fuzzy-k-means classifier which partitionsobservations in multivariate space into natural classes This approach is similar tocluster analysis and numerical taxonomy but the resulting classes are continuous witheach observation assigned a fuzzy membership value that characterizes its degree ofsimilarity to each individual class The concept has been integrated into geostatisticalmethods and will be discussed in more detail below (see Section IV1) The secondapproach is known as the Semantic Import model (SI) and is used in situations whenclassification schemes are pre-defined and class limits are relatively well understoodThe SI model is commonly used in concert with expert knowledge and will bediscussed in the expert systems section (see Section IV4)

IV Recent advances in predictive soil mapping

Within the last decade many authors have sought to model the soil landscape using avariety of methods Literature in this field could be summarized many different waysbut we concentrate on the literature that directly addresses the goals of predictive soilmapping stated in the introduction (see Table 1) Therefore we will review research thatattempts to exploit the relationship between quantifiable landscape indices and soilcharacter in order to model the soil landscape in a more continuous and thereforerealistic manner

The research reviewed here is distinguished from decades of previous researchdocumenting the correlation between landscape position and soil attributes (reviewedby Hall and Olsen 1991) That body of research is informative but not useful for

P Scull et al 179

180 Predictive soil mapping a review

Table 1

Selected

recen

t literature on pred

ictive

soil m

apping an

d m

apping

(cited in

this article) d

escribing the mod

ellin

g metho

dused the

dep

enden

t variables used a

nd the

env

iron

men

tal v

ariables (ex

plan

atory) used in

the

mod

els

Stud

yMod

ellin

g metho

dDep

ende

nt variables

Environm

ental variab

les

Goa

ls atta

ined

a

Bell et

al 20

00Line

ar and

exp

onen

tial

Total soil orga

nic

Slop

e curvature aspe

ct

ECx

regression

carbon

hillslope

position

Burge

ss and

Web

ster

Punc

tual and

block

Na co

nten

t co

ver

Non

eC

1980

ab

kriging

loam

thickn

ess

ston

e co

nten

t

Burroug

h 198

9Fu

zzy mathe

matical

ndashndash

rpmetho

ds

Burroug

h et

al 19

97Con

tinuo

us classificatio

nndash

ndashrp

Castrigna

no eta

l 200

0Factorial kriging

CEC

pH N

P K

Non

e ex

plicitly used

Cx

Na

Cialella

eta

l 199

7Classifica

tion tree

Drainag

e class

Elev aspe

ct NDVI

Ec

Coo

k et

al 19

96Baysian

rule-ba

sed metho

dsOrgan

ic m

atter

Slop

e aspec

t wetne

ssEcX

inde

x

Dale et

al 19

89 (a review

)Ex

pert systems

ndashndash

ndash

Ellis 19

96Decision tree

ana

lysis

Soil erosion class

Slop

e aspec

t wetne

ssEc

neutral ne

tworks

inde

x flow le

ngth and

accu

mulation Lan

dsat

TM tree

cov

er

Gessler 1

996

A large

variety of statistic

Field an

d labo

ratory

A variety of digital

EC

metho

dsco

llected

phy

sica

len

vironm

ental da

tach

emical and

morph

olog

ical soil

prop

ertie

s

Gessler eta

l 199

5Line

ar and

logit reg

ression

A horizon

and

solum

Curvature CTI topo

EC

depth E horizon

positio

npresen

ce

Goo

vaerts 19

92Factorial kriging

Total carbon

Non

eC

P Scull et al 181pH

N CEC

extractab

lecatio

ns (K Ca M

g)Hartemink et

al 20

01ndash

ndashndash

rp

Heu

velin

k an

d W

ebster 20

01ndash

ndashndash

rp

Hew

itt 19

93ndash

ndashndash

rp

Indo

rante et

al 19

96ndash

ndashndash

rp

King et

al 19

99Lo

gistic regression

Presen

ceabsen

ceSlop

e aspec

t po

t solar

Ec

Non

calc c

lay-loam

energy

Kno

tters eta

l 199

5Kriging

co

-kriging

So

ft layer de

pth

Hillslop

e po

sitio

nCx

regression

kriging

Laga

cherie and

Holmes 19

97Classifica

tion tree

Map

ping

unit

Geo

logy variou

s topo

Ec

indices

Laslett e

tal 198

7Kriging

splin

es tren

dpH

Non

ec

surfac

e nea

rest neigh

bor

McB

ratney 1

992

ndashndash

ndashrp

McB

ratney

eta

l 199

1Block kriging

Clay co

nten

tNon

eC

McB

ratney

eta

l 200

0ndash

ndashndash

rp

McB

ratney

and

Ode

h 199

7Fu

zzy sets in

soil scienc

endash

ndashRp

McB

ratney

and

de Gruijter

Fuzz

y-k-mea

ns w

ithFu

zzy classes

Field co

llected

phy

sical

C19

92ex

tragrade

sch

emical and

morph

olog

ical soil

prop

ertie

s

McC

rack

en and

Cate 198

6Artificial intellig

ence

ndash

ndashrp

expe

rt systems

McK

ensie an

d Austin

19

93Gen

eralized

line

ar m

odels

Clay co

nten

t CEC

Slop

e relief land

form

ec

(logit)

pH EC

COLE

slop

e po

sitio

nbu

lk den

sity and

othe

rs

Moo

re eta

l 199

3Line

ar reg

ression

A horizon

dep

th O

MSlop

e w

etne

ss and

strea

mEC

and P co

nten

t pH

power in

dices aspect

curvature

182 Predictive soil mapping a reviewTa

ble 1

Con

tinu

ed

Stud

yMod

ellin

g metho

dDep

ende

nt variables

Environm

ental v

ariables

Goa

ls atta

ined

a

McK

ensie an

d Ryan 199

9Reg

ression tree and

linea

rSo

lum dep

th P an

dElevation slope

CE

regression

N con

tent

curvature CTI

contribu

ting area do

wn-

slop

e mea

ns for slope

clim

ate da

ta Presco

ttInde

x G

amma

Rad

iometry La

ndsat T

M

and Geo

logy unit

Ode

h et

al 19

92ab

Fuzz

y-c-mea

ns and

Fuzzy classes

Field-co

llected

phy

sica

lC

kriging

chem

ical and

morph

olog

ical soil

prop

ertie

s

Ode

h et

al 19

94

Reg

ression kriging

co

-So

lum dep

th de

pth

Slop

e aspect cu

rvature

eC

1995

kriging regression

to bed

rock gravel

kriging

and clay con

tent

Skidmore et

al 19

91Bayesian expe

rt system

Soil land

scap

e un

itVeg type

wetne

ss in

dex

EcX

1996

grad

ient terrain po

sitio

n

Voltz and

Web

ster 19

90Kriging

cu

bic splin

eClay co

nten

tNon

eC

Web

ster 1

994

Dev

elop

men

t of

ndashndash

rppe

dometrics

Zhu

199

7ab

Fuzz

y logic expe

rt system

Soil series A

horizon

Elev pm aspe

ct c

anop

yECX

Zhu

and

Ban

d 199

4(SoL

IM)

depth in

dividu

alco

verage

grad

ient

Zhu

eta

l 199

7series m

aps

curvature

Not

es

a Letters refer to the de

gree

to w

hich

the

goa

ls of PS

M defined

in the

introdu

ction are achiev

ed

Soilndash

e nvironm

ent relatio

ns utilized

(letter E) be

tter represen

tatio

n of soil c o

ntinuity (C

) an

d ex

pert kno

wledg

e utilize

d (X) c

orrespon

d to goa

ls 1 2 and

3respec

tively from

the

introdu

ction and

cap

ital lette

rs (E C X) ind

icate the metho

d is relatively more successful th

an th

ose metho

ds den

oted

by lower

case le

tters (e c x

) rp ind

icates rev

iew pap

ers

P Scull et al 183

predictive mapping because landscape position is never quantified (rather positionwas often qualitatively defined eg toe-slope) and thus the documented relationshipscannot be generalized using environmental data and digital elevation models to predictsoil character at unvisited sites

Table 1 documents modelling methods model variables and the extent to which thereferenced research satisfies the previously defined goals of PSM (lsquoGoals attainedrsquocolumn) Soil-environment relations utilized (letter E) better representation of soilcontinuity (C) and expert knowledge utilized (X) correspond to goals 1 2 and 3respectively from the introduction and capital letters (E C X) indicate the method isrelatively more successful than those methods denoted by lower case letters (e c x)Citations with no letters present within the lsquoGoals attainedrsquo column do not address theaforementioned goals For example the Cialella et al (1997) received a rating of lsquoEcrsquomeaning that the methods employed successfully utilized environmentalndashsoil characterrelations (E) and somewhat successfully presented a better method of representing soilcontinuity (c) The ratings are provided simply to help organize the literature that wasreviewed Review papers are included within the table denoted by lsquorprsquo Geostatisticalmethods are not included in the table because they have been comprehensivelysurveyed elsewhere (Odeh et al 1994 Burrough et al 1997 McBratney et al 2000Heuvelink and Webster 2001) and because the objectives and assumptions of geostatis-tical methods differ slightly from other PSM research We briefly outline thesedifferences in the following section

1 Geostatistical methods

Geostatistics are a subset of traditional statistics that deal primarily with spatial dataand account for spatial autocorrelation using kriging as the spatial interpolator Theconcept is based upon the theory of regionalized variables which was mainlydeveloped by Matheron (1963) and Krige (1963) Kriging is a form of weighted localaveraging that uses a measure of spatial dependence the variogram to determine theweights applied to the data when computing the averages Geostatistical methods havebeen used in predictive soil mapping research to spatially interpolate soil propertyvalues at unmeasured sites from field-collected data

Burgess and Webster (1980a b) were the first to introduce ordinary kriging to the soilcommunity and since that time an enormous amount of work has been published Forexample ordinary kriging has been used to interpolate many different soil propertiesincluding pollution trace element deficiencies salinity and fertility (Heuvelink andWebster 2001) Ordinary kriging has been criticized for a variety of reasons Forexample Laslett et al (1987) reported that several authors had criticized geostatisticsbecause kriging is a global rather than local technique failing to take into accountknowledge of soil materials and processes Other authors have criticized geostatisticsbecause they are excessively data dependent requiring a large number of closelyspaced data points (Zhu 1997a) As Webster and Oliver (1992) suggest in excess of ahundred samples may be needed to use geostatistics at the field scale because of highspatial variability of soil in some areas Geostatistics also assume spatial autocorrela-tion which sometimes may be a poor assumption in complex terrain where abruptchanges in soil-forming factors occur (McBratney et al 2000) Ordinary kriging by itself

184 Predictive soil mapping a review

does not satisfy two of the three goals of PSM presented in the introduction ndash it does notadequately incorporate expert knowledge and it does not exploit the relationshipbetween environmental variables and soil properties

Ordinary kriging has been modified in a variety of ways to better incorporateancillary data and known soilndashlandscape relationships Block kriging involvesdetermining estimates over meaningful areas rather than at specific points (Burgess andWebster 1980a McBratney et al 1991) Using this method a study area can be stratifiedinto different regions that are reflective of the pedogenetic processes at work In orderto accommodate a trend within a dependent soil variable universal kriging has beenused (Webster 1994) Kriging with external drift is similar to universal kriging but ituses an ancillary variable to represent the trend (McBratney et al 2000) Co-krigingtakes advantage of correlation that may exist between the variable of interest and othermore easily measured variables (Odeh et al 1995) Regression kriging involvesspatially interpolating the residuals from a non-spatial model by kriging and addingthe result to the prediction obtained from that model (Goovaerts 1997 Castrignano etal 2000) Factorial kriging is another method to integrate multivariate data into thestandard kriging routine to extrapolate soil data (Goovaerts 1992) Many authors havecompared these various methods (Laslett et al 1987 Voltz and Webster 1990 Odeh etal 1994 Knotters et al 1995)

Fuzzy logic has been used with geostatistics by various authors to produce new kindsof fuzzy soil maps with continuous classes (Burrough 1989 McBratney and DeGruijter1992 Odeh et al 1992a and reviewed by McBratney and Odeh 1997) The processentails kriging the matrix of membership values determined by the fuzzy k-meansclassifier resulting in a continuous soil surface where individual locations are allowedto belong to more than one class and no rigid boundaries are designated to separate thesoil into discrete units or entities The results of such analysis can be used to assess thepedologic process validity of soil taxonomy by determining whether soils grouptogether into classes that are similar to taxonomic types Fuzzy classes wouldpresumably reflect the main pedologic features within a given area In this sense theclassification is quantitative whereas soil taxonomy is rooted in qualitative discrimina-tion A comparison between the two could provide insightful and help assist soiltaxonomy in making a classification that is more indicative of underlying soil processes

Geostatistics in soil research were originally introduced to quantitatively assess soilvariability within soil mapping units (McBratney et al 1991) in response to criticisms inthe early 1970s that soil unit composition was not well quantified (Beckett and Webster1971) In this regard geostatistics have been very useful having served well the originalgoals set forth by Burgess and Webster (1980b) when they drew kriging to the attentionof soil scientists as a means of spatial prediction At the field scale soil variation islargely due to the effect that topography has on soil genesis Geostatistics have beensuccessfully applied in such environments by using terrain attributes as ancillary datawithin many of the kriging routines described above Such quantitative within-unitvariability of soil properties is very useful in the field of precision agriculture and othersituations (eg pollutants) where very detailed soil attribute information is needed atthe field scale (Heuvelink and Webster 2001)

However geostatistics have not been applied in a wide variety of environments or atlarger scales In order to be successfully applied in different environments geostatisticswill likely require a different suite of ancillary data For example remote sensing data

P Scull et al 185

could be used in arid regions where soil toposequences are less well expressed Atlarger scales of prediction selection of different sets of ancillary variables is requiredbecause different processes define soil character at different scales The most obviousexample is that of climate which may control soil distribution at large scales(continental) but has little explanatory power at the field level Regardless of whetheradequate ancillary data exist the amount of data required to use geostatistics forlandscape-level prediction would be extremely difficult and costly to collect given thestrict sampling protocol required to characterize spatial dependence It is also unclearat what landscape scale soils exhibit spatial autocorrelation

Geostatistical approaches do provide a means of creating continuous soil attributesurfaces to better represent soil continuity (Goal 2) and they can be used to exploit therelationship between environmental variables and soil properties in order to moreefficiently collect soil data (Goal 1) However they do not sufficiently utilize expertknowledge (Goal 3) as no attempt has been made in geostatistical approaches todirectly integrate expert knowledge Fundamentally kriging is a process of interpola-tion designed to predict attribute values in between locations of measured samples Inthis sense geostatistics represent a middle ground between pure interpolation (egnearest neighbour type classifier) in which only measured points for the variable ofinterest are used to determine unknown values and other predictive models thatprimarily use soilndashenvironment correlation to create predictive maps

2 Statistical methods

Statistical methods can be used to exploit the relationship between quantifiablelandscape indices and soil properties to create predictive soil maps For exampleMcKensie and Austin (1993) used a regression to account for a large percentage ofvariation for many soil characteristics (A horizon clay content CEC EC pH bulkdensity and COLE B horizon clay content CEC ESP EC pH bulk density and COLE)using a variety of predictor variables (slope presence or absence of impeding layerrelief landform topographic position) Their results confirm the hypothesis of MooreID et al (1991) that soil character is related to quantifiable landscape indices Howevertheir methods do not provide inference of soil properties at unmeasured sites frommapped environment data because the topographic variables were measured in thefield Linear regression has also been used with terrain variables derived from a 15-mDEM in northeastern Colorado to predict soil attributes (organic matter contentextractable phosphorous pH and texture) at unvisited sites (Moore et al 1993) In thatparticular study 50 of the variance of A-horizon thickness was explained by slope andthe wetness index Gessler et al (1995) also used regression to model A-horizonthickness from topographic variables in southeastern Australia (plan curvature andwetness index r2 = 063 P = 0001) They modeled solum depth and used logisticregression to model E horizon presenceabsence Elsewhere logistic regression hasbeen used to model the presenceabsence of noncalcareous clay loam horizon in centralFrance using terrain attributes from a 20-m DEM (King et al 1999) Exponentialregression has been used to model soil organic carbon using terrain variables (Bell et al2000) in glacial outwash soils in east-central Minnesota Generalized additive models(GAM) have been used less frequently in PSM research Gessler (1996) used a GAM

186 Predictive soil mapping a review

model to predict total soil carbon A horizon depth and solum depth using a variety ofenvironmental predictors (slope elevation wetness index mean annual temperatureprecipitation and radiation)

This small body of research opened the door to more complex methods by demon-strating the existence of quantifiable relationships These authors were able to producesoil attribute maps using raster data models whose scale was dependent upon the gridcell resolution of the environmental data They were successful at exploiting the rela-tionship between quantifiable topographic attributes and soil profile character (Goal 1)The continuous soil attribute surfaces they produce also better represent soil continuitythan the choropleth soil maps produced by traditional soil survey (Goal 2) Howeverthe bulk of these methods (excepting GAMs) are limited by their assumed linear rela-tionship between soil and topographic attributes their assumptions of normallydistributed data and their high data requirements Standard statistical procedures arealso not flexible enough to allow robust integration with a variety of potential datasources such as expert knowledge (Goal 1)

Statistical methods do demonstrate in a quantitative manner that terrain analysis canbe used to predict soil attributes in relatively small areas with homogeneous parentmaterial A large proportion of the research using statistical methods was conducted insemi-arid landscapes at small scales (the largest study area of the entire group was~2000 ha) Obviously for statistical approaches to be most effective they need to bemore universal As such they need to be tested andor developed at larger scales andin more diverse landscapes

3 Decision tree analysis (DTA)

The use of decision tree analysis is just beginning to be explored in predictive soilmapping research although it has been used successfully in the related field ofpredictive vegetation mapping since the early 1990s (Lees and Ritman 1991 MooreDM et al 1991 Franklin 1998) DTA is a form of divisive classification The process oftree modelling involves successively partitioning data (called recursive partitioning inthe tree modelling literature) into increasingly homogeneous subsets which once thepartitioning has ceased are called terminal nodes (Lees and Ritman 1991) Splits orrules defining how to partition the data are selected based on information statistics thatdefine how well the split decreases impurity within the data set (Clark and Pregibon1992) Splits are based on threshold values of an explanatory variable selected bycomparing the increase in resulting purity of node membership for all possiblethresholds and variables The process is iterative growing from the root node (thecomplete data set) to the terminal nodes in a dendritic fashion (Friedl and Brodley1997) Once the tree has been constructed (or grown) it encodes a set of decision rulesthat describe the data partitioning process These rules can be used to classify or predictother data sets (Moore DM et al 1991) Pruning the tree is often necessary to preventthe tree from being overfit to the sample data and to reduce tree complexity Pruningentails combining pairs of terminal nodes into single nodes and can be accomplishedusing cross-validation which yields an initial indication of how large a tree makesrobust predictions (Safavian and Norvig 1991) Cross-validation involves systematical-ly removing portions of the data set and running the remaining sample through the tree

in an iterative manner eventually yielding estimates of the misclassification rates foreach class each node and the whole model (Breiman et al 1984) In this mannerdifferent sized trees can be compared in terms of parsimony

The term DTA is used to collectively refer to all types of tree-based modelling (theword lsquodecisionrsquo is used because it is descriptive indicating that the analysis eventuallyleads to a set of decision rules defining data partitions) The term should be distin-guished from classification tree analysis because the latter refers specifically to DTAwhere the response variable is categorical The term CART (classification andregression trees ndash Breiman et al 1984) is sometimes used but strictly speaking refers tospecific software Friedl and Brodley (1997) provide a review of the decision treealgorithms They divide types of DTA into two classes (1) homogeneous decision treesfor which a single algorithm is used to estimate each split (eg CART) and (2) hybriddecision trees (HDT) for which different splitting methods can be used at differentpoints in the tree (eg Quinlan 1993) They further divide homogeneous decision treesinto univariate (UDT) where single features of the input data define splits and multi-variate decision trees (MDT) where multiple features of the input data can define splitsAccording to this naming convention no distinction is made between methods withdifferent types of response variables although all methods can be used with bothcategorical and continuous response variables In a comparison of these various typesof algorithms on a variety of data sets Friedl and Brodley (1997) found that HDT hadthe highest classification accuracy

The overall aim of DTA is to design a set of predictive rules (eg if geology type Athen soil type B) developed from training data which can then be applied to ageographic data base to predict the value of a response variable (Michaelsen et al 1994)Therefore DTA explicitly uses soilndashlandscape correlation in model development (Goal2) The technique appears promising in soil research but needs to be further tested asnot many of the above types of DTA have been tested in the PSM literature in fact onlyunivariate approaches have been employed For example Lagacherie and Holmes(1997) successfully used univariate DTA to model a categorical response variable soiltype within a training set and then assuming that the training set was representativeextrapolated the model to a much larger region Their work is interesting becausealthough not a single sample came from the area they eventually mapped they wereable to produce a soil map that was more accurate (74 versus 69) than the existingmap produced from traditional methods Cialella et al (1997) also used univariate DTAto predict soil drainage class from a variety of terrain attributes and remotely senseddata They predicted soil drainage class with an average of 78 accuracy ndash impressivegiven that the variation accounted for by a typical soil survey ranges from about halfthe total variance for physical attributes to less than one-tenth for some soil chemicalattributes (Gessler et al 1995)

DTA has been compared with other approaches by several authors In the applicationof erosion modelling DTA results were similar to Artificial Neural Networks (Ellis1996) Both methods achieved high training accuracy (as measured by the KappaStatistic) but in terms of prediction accuracy both methods performed poorly Gessleret al (1995) compared DTA with generalized linear models (GLM) and generalizedadditive models (GAM) to predict A-horizon thickness and concluded that GLM waspreferable to both DTA and GAM McKensie and Ryan (1999) compared regressiontrees and standard linear regression to predict soil properties (total solum depth soil

P Scull et al 187

188 Predictive soil mapping a review

phophorus and soil carbon) using a large variety of predictors (elevation slopecurvature wetness index climate variables geology unit Landsat TM data andGamma radiometric data) and found that different methods work best for differentproperties and overall success hinged on the strength of the relationship between soiland environmental variables Their study of a 50 000 ha forested landscape in southernNew South Wales Australia is one of the few studies at such a small map scale (largearea) or in forested landscape The most extensive comparison of techniques involvingDTA was conducted by McBratney et al (2000) They compared regression techniques(GLM GAM DTA-regression tree) geostatistical techniques (kriging and heterotopicco-kriging) and a hybrid technique (regression kriging) DTA was found to be thepoorest performing of the regression techniques because of the unrealistic predictionsurface generated by the DTA model DTA has also been criticized by other authorsbecause of the stepped prediction surface (Gessler 1996) This phenomenon can beespecially apparent in situations where predictor variables have different resolutionsThe extent to which DTA yields a better spatial representation of soil continuity is afunction of the scale and type of predictor variables used and therefore varies withindividual models

DTA is gaining widespread popularity as a means to develop prediction rules thatcan be rapidly and repeatedly evaluated (Cialella et al 1997 Franklin et al 2000) DTAprovides the following advantages over standard statistical techniques (1) it is easier tointerpret when explanatory variables are both nominal and continuous (2) it isinvariant to monotone re-expressions (transformations) of predictor variables (3) itdeals more satisfactorily with missing data values and outliers (4) it is more adept atcapturing nonadditive and nonlinear behaviour (5) it doesnrsquot make any assumptionsabout data distribution and (6) it is easily updateable as more data are collected(Moore DM et al 1991) The DTA model framework is especially appealing becauseof its capability to integrate a wide range of data sets as explanatory variables

DTA offers a unique opportunity for interaction between soil experts and soilmodellers because the output of the model is a set of rules that can be pedologicallyinterpreted by the soil expert In this sense expert knowledge is used in an implicitmanner in DTA (somewhat effectively achieving Goal 3) While these rules can often beexceedingly complex at minimum the expert can decide whether initial splits makesense given their understanding of the landscape The success of DTA results oftenhinges on the ability of the modeller to make key decisions during the model buildingprocess there is unfortunately no definitive way to determine the most optimal treeThe aid of the soil expert can potentially elucidate this problem

4 Expert systems

A variety of expert system approaches to PSM have been developed to utilize expertknowledge The purpose of such methods is to exploit the information the soil surveyoraccumulates while working in the field by integrating such knowledge into thepredictive model (McCracken and Cate 1986) Unlike the majority of the researchreviewed thus far the dependent variable in many expert systems models is often soiltaxa or mapping unit This apparent disadvantage of expert systems (using classifica-tion to characterize soil continuity) does make them easier to integrate into traditional

soil survey In addition several authors have developed methods to developcontinuous soil property maps from the output of expert system models designed topredict soil unit occurrence

Expert systems are composed of data (information on spatial environmentalvariables eg topography climate etc) a knowledge base (rules and facts related tosoil variation supplied from the soil surveyor) and an inference engine (whichcombines data and the knowledge base to infer logically valid conclusions) (Skidmoreet al 1996) Expert systems differ from conventional models in two ways (1) they storeand manipulate qualitative information (allowing them access to information thatcannot normally be used in other modelling frameworks) and (2) they are structuredas meta-models (the knowledge is separated from the model) (Davis 1993) This allowsthe model to selectively choose which information is relevant at various stages of themodelling process and it allows for information to be easily updated Davis (1993)reviews the application of expert systems to environmental modelling research conclu-ding that the technique is becoming more widely accepted He further states that theapplication of expert systems is constrained by an absence of fundamental knowledgefor rule generation a problem that would appear less relevant to soil mapping giventhe amount of untapped expert knowledge accumulated by the soil surveyor

The first mention of the use of expert systems in pedology was in a paper presenta-tion at the Northeast Committee Soil Survey Conference 1984 (Flach 1985) In herpaper Flach (1985) hinted that recent developments in computer science especiallyexpert systems and artificial intelligence could make modelling a practical mappingtool for soil scientists in the near future A year later McCracken and Cate (1986) hopedto encourage soil scientists to explore expert systems and its application to soil sciencethrough an optimistic article they wrote on the potential use of expert systemsHowever little research was actually conducted in the late 1980s and expert systemshad not yet begun to fulfil the lofty goals set forth by early practitioners (Dale et al1989)

In the early 1990s expert systems approaches to predictive soil mapping began toappear in the literature Skidmore et al (1991) used a Bayesian expert system to mapforest soil into different classes and their results compared favourably with availablesoil maps and actual field-collected data Their methods successfully incorporatedsurveyor knowledge and remotely sensed and digital terrain attributes but failed tobetter characterize continuous soilndashlandscape variability because their final productwas a choropleth map Skidmore et al (1996) revisited their earlier research andassessed the mapping accuracy of their results They found that the soil map producedby the expert system achieved an overall accuracy of 698 (sample size = 53) while themap derived from conventional methods had an accuracy of 736 In addition toknowledge provided by a soil scientist Skidmore et alrsquos (1996) expert system usedvegetation derived for aerial photography and topographic variables derived from a 10-m DEM (wetness index topographic position and slope)

Using an expert system Cook et al (1996) successfully produced a continuous soilproperty map for organic matter content using wetness index aspect and slope asexplanatory variables Their methods were somewhat inefficient because they requireda separate expert system for each soil property of interest However their researchrepresents the only example of expert systems used to directly predict the spatial dis-tribution of a soil property All of the other examples of expert systems in the literature

P Scull et al 189

190 Predictive soil mapping a review

deal with soil type or class The use of expert systems to map soil properties needs tobe explored further

The expert systems discussed thus far have all used Boolean logic within theirtheoretical framework whereby an observation can belong to one and only one classand the soil properties of that class are assigned to the observation With thedevelopment of fuzzy logic and the semantic import model (the second fuzzy logicapproach to creating continuous classes) the opportunity exists to develop fuzzy logic-based expert systems Such systems can be used in conjunction with expert knowledgein situations where experts have a good qualitative idea of how to group data but havetrouble dealing with observations that are not well represented by rigid classificationschemes (Burrough and McDonnell 1998) This method can be particularly useful insituations when taxonomic schemes have been previously developed as is the case insoil taxonomy Several examples of this type of approach were published in the mid-1990s by A Zhu and colleagues (Zhu and Band 1994 Zhu et al 1996 1997 Zhu1997ab) Such systems proved useful for mapping soil at unvisited locations usingsurveyor knowledge and were also capable of producing continuous soil propertymaps The use of fuzzy logic within the theoretical framework of the expert systemallows the complex nature of soil to propagate through the modelling process neversubjected to classification schemes that filter out potentially useful lsquonoisersquo The use offuzzy logic also gives the soil surveyor more latitude during the interview processwhen the knowledge base is defined for the expert system Zhu et al (1997) developeda fuzzy logic-based expert system called SoLIM to determine the similarity of eachgrid cell in a study area to the various taxonomic mapping units delimited by the soilsurvey Continuous soil attribute maps were calculated using the similarity values andtheir relative soil survey determined attribute values (effectively achieving Goal 2) Theresulting data (which proved more accurate than soil survey data once field checked)consists of a raster grid whose resolution is determined by the resolution of the inputenvironmental and digital elevation data As noted previously this type of data modelis more applicable to environmental modelling than the choropleth map (Burrough andMcDonnell 1998)

Expert system approaches to PSM are capable of exploiting soil surveyor knowledgeby developing rule-based systems that imitate the surveyorrsquos conceptual model of soilvariability (the primary focus of Goal 3) The method would appear extremely usefulfor mapping projects (such as those conducted by the NRCS-NCSS) where fieldwork isinitially conducted to determine soilndashlandscape relations Expert system developmentcould be directly inserted into the traditional soil survey mapping approach as asubstitute for the step where the surveyor converts hisher conceptual model into achoropleth map Rather that knowledge could be incorporated into the expert systemwhich could be used to predictively map soil The resulting raster map would be morescientifically based and explicit than the hand-drawn choropleth maps of the past Itsscale would not be limited to that of the aerial photography but rather to the scale ofthe environmental data

Despite these advantages the expert system approach has some drawbacks Becausethe method is dependent upon expert knowledge it cannot be applied whereenvironmentndashsoil relations are poorly understood Of course this criticism can be madeof all PSM models as inductive or deductive knowledge of soilndashenvironment relationsis a prerequisite for PSM However expert systems are deductive models and as such

are not driven by any specific field-collected data (although presumably the soil experthas field experience in the mapping area) Expert systems donrsquot afford the opportunityto first statistically document landscapendashsoil relations and then extrapolate the resultsbecause expert systems do not directly use sample soil data (lsquohard datarsquo) to determinesoilndashlandscape relations Thus the utilization of the relationship between environmen-tal variables and soil properties (Goal 1) is only indirectly achieved Expert systemsapproaches have been demonstrated to be extremely effective in a small number of casestudies The possibility of satisfying all three goals discussed in this review makesexpert systems a predictive soil mapping method that needs to be further tested ndash bothgeographically and across different scales of analysis

V Conclusion

Most of the predictive soil mapping research outlined in this review was conducted atvery large map scales (over small areas) In fact the majority of the research wasconcerned with assessing the spatial variability of soil character within individual fieldsor across soil toposequences The primary driving force behind this type of research hasbeen the need to provide accurate soils information for agriculture and ecologicalmodels It is clear that terrain attributes are powerful predictors at the local scaleGeostatistical tools have been successful at using terrain attributes and the spatialdependence of soil properties to interpolate between existing data points withinindividual fields Across soil toposequences statistical approaches provide a usefulmeans of predicting soil character PSM research has been most successful at the fieldscale because many of the soil-forming factors are held constant For example the neteffect of four of Jennyrsquos five soil-forming factors (climate organism parent material andtime) was minimal within many of the studies reviewed While some research hastackled larger areas of study there exists a deficiency in our ability to predictively mapsoils at smaller map scales Since the distribution of soil is scale-dependent differentPSM methods and predictors are likely to work better at different scales Focus in thefuture must continue to move toward working over larger spatial extents of study inorder to produce landscape-scale soil information

Further a large proportion of the research was conducted in semi-arid gentlysloping agricultural landscapes Humid forestlands mountainous regions and desertshave received little attention As such PSM methods need to continue to be testedandor developed in a wider variety of landscapes where spatial soil distributions canbe more complex Different methods will likely be successful to different degrees indifferent environments Whereas terrain attributes are dominant predictors of soilcharacter across toposequences in prairie lands other predictors are likely necessary inareas where soils do not develop into clear toposequences For example remote sensingdata has been a useful predictor of soil chronosequences in desert landscapes wheresurface appearance is often related to soil character Surficial geology is often reflectiveof soil character in mountain regions where soils are thin and significant bedrock existsat the surface Focus in the future must continue to determine which methods andpredictors work best in which environments

Since the most useful PSM approach will vary across spatial scales and environmen-tal gradients the method used should be driven by the mapping objectives of the

P Scull et al 191

192 Predictive soil mapping a review

project Whereas the traditional soil survey of the past was expected to meet the needsof a diverse group of end users PSM methods will vary given the objectives of thesurvey As a result the end users of the soil data need to play a more active role in thesurvey process Because precision agriculture and large-scale mapping has been theprimary focus of PSM there exists a deficiency in our ability to predictively map soilsfor the purpose of general land use planning and management Even though lessdetailed soil information is needed for these purposes the development of PSMmethods is complicated by the fact that soils exhibit complex spatial variability atsmall map scales where soil-forming environments vary greatly from one location tothe next

A number of alternative methods of characterizing the continuous nature of the soillandscape have been developed Thus far most of the PSM research has provided soilinformation in a nonobject form (as opposed to defining soil types as independententities) Soil data have been generated organized and presented in the form of eitherisorithmic maps or fine-scale raster grids Both of these data models are field-viewmodels of geographic space which allow the soil to be perceived as a constantlyvarying surface Two distinct approaches have been employed (1) mapping individualsoil properties and (2) mapping continuously varying (fuzzy) soil classes Mappingindividual properties is the most common approach and will likely continue todominate PSM research The use of fuzzy soil classes in the literature is less commonand more difficult to be integrated into standard mapping procedures (such as use ofsoil taxa) because the concept radically differs from the traditional view of the soillandscape However fuzzy soil classes (generated using the fuzzy-k means approach) dohave the potential to help further advance Soil taxonomy by identifying taxonomicclasses that are more reflective of pedologic processes at work

Expert systems have been greatly underutilized in PSM research especiallyconsidering how effective a small number of case studies have been Expert systemshave the potential to satisfy successfully all three goals discussed throughout thereview They also have the potential to bridge the gap between traditional approachesand PSM methods because field soil scientists do not have to change their conceptualapproach to mapping They can still conduct field reconnaissance to determinesoilndashlandscape relationships Afterwards though they can use expert systems to betterexploit the knowledge they have garnered Such an approach could help ease thetransition from traditional soil survey to more scientifically explicit methods Expertsystems are also the most fruitful approach to utilizing a wealth of data that has alreadybeen collected in a nonexplicit manner the wealth of expert knowledge (in the form ofthe conceptual model) that senior soil mappers have accumulated

Focus in the future must also move in the direction of operability The utility of manyPSM approaches has been clearly demonstrated yet changes in how soils are mappedand perceived has been slow In the USA various members of the National CooperativeSoil Survey have called for a change in the philosophy of survey (from mappingdiscrete soil types to mapping continuous varying soil properties McSweeny et al1994) but little change has actually taken place To realize the potential of PSM in con-tributing to an overall change in standard soil mapping procedures practitionersworking in the field of PSM need to form working relationships with field soilscientists PSM methods need also to be presented in a manner that is comprehensibleto the soil science trained field mapper Predictive soil mapping is a relatively recent

phenomenon yet much progress has been made In the process of mapping soilknowledge will continue to advance

P Scull et al 193

References

Agbu PA Fehrenbacher DJ and Jansen IJ1990 Statistical comparison of SPOT spectralmaps with field soil maps Soil Science Society ofAmerica Journal 54 818ndash18

Barrett L 1999 Particulars in contextmaintaining a balance in soil geography Annalsof the Association of America Geographers 89707ndash13

Beckett PHT and Webster R 1971 Soilvariability a review Soils and Fertilizers 341ndash15

Bell JC Grigal DF and Bates PC 2000 A soil-terrain model for estimating spatial patterns ofsoil organic carbon In Wilson J and Gallant Jeditors Terrain analysis principles and applica-tions New York City NY John Wiley and Sons295ndash310

Birkeland PW 1999 Soil and geomorphologyThird edition New York NY Oxford UniversityPress

Breiman L Friedman JH Olshen RA andStone CJ 1984 Classification and regressiontrees Belmont CA Wadsworth

Brule FJ 1996 Fuzzy systems ndash a tutorialhttp newsgroup compai httpwwwquadralaycom (last accessed 30 August 2001)

Bunkin FV and Bunkin AF 2000 Lidarsounding of water soil and plants Atmosphericand Oceanic Optics 13 54ndash72

Buol SW Hole FD McCracken RJ andSouthard RJ 1997 Soil genesis and classifica-tion Ames IA Iowa State University Press

Burgess TM and Webster R 1980a Optimalinterpolation and isarithmic mapping of soilproperties the semi-variogram and punctualkriging Journal of Soil Science 31 315ndash31

ndashndashndashndash 1980b Optimal interpolation and isarithmicmapping of soil properties block kringingJournal of Soil Science 31 331ndash41

Burrough PA 1989 Fuzzy mathematicalmethods for soil survey and land evaluationJournal of Soil Science 40 477ndash92

Burrough PA and McDonnell RA 1998Principles of geographic information systems(Revised edition) Oxford Clarendon Press

Burrough PA Beckett PHT and Jarvis MG1971 The relation between cost and utility insoil survey Journal of Soil Science 22 368ndash81

Burrough PA Van Gaans PMF and

Hootsman R 1997 Continuous classificationin soil survey spatial correlation confusionand boundaries Geoderma 77 115ndash35

Cambell JB 1977 Variation of selectedproperties across a soil boundary Soil ScienceSociety of America Journal 41 578ndash82

Cambell JB and Edmonds WJ 1984 Themissing geographic dimension to soiltaxonomy Annals of the Association of AmericanGeographers 74 83ndash97

Castrignano A Giugliarini L Risaliti R andMartinelli N 2000 Study of spatial relation-ships among some soil physico-chemicalproperties of a field in central Italy using multi-variate geostatistics Geoderma 97 39ndash60

Cialella AT Dubayah R Lawrence W andLevine E 1997 Predicting soil drainage classusing remotely sensed and digital elevationdata Journal of Soil Science 62(2) 171ndash78

Clark RN and Swayze GA 1996 Evolution inimaging spectroscopy analysis and sensorsignal-to-noise an examination of how far wehave come Summaries of the sixth annual JPLairborne Earth science workshop 4ndash8 March 1996AVIRIS Workshop Vol 1 5

Clarke LA and Pregibon D 1992 Tree-basedmodels In Chambers J and Hastie J editorsStatistical models in S Pacific Grove Wadsworthand Brooks 377ndash419

Coleman TL Agbu PA and MontgomeryOL 1993 Spectral differentiation of surfacesoils and soil properties ndash is it possible fromspace platforms Soil Science 155 283ndash93

Cook SE Corner RJ Grealish G GesslerPE and Chartres CJ 1996 A rule-basedsystem to map soil properties Soil ScienceSociety of America Journal 60 1893ndash900

Cruickshank JG 1972 Soil geography New YorkNY John Wiley amp Sons

Csillag F Pasztor L Biehl LL 1993 Spectralband selection for the characterization ofsalinity status of soils Remote Sensing ofEnvironment 43 231ndash42

Dale MB McBratney AB and Russell JS1989 On the role of expert systems andnumerical taxonomy in soil classificationJournal of Soil Science 40 223ndash34

Davis JR 1993 Expert systems and environ-

The State Factor soil-forming model forms the theoretical basis of PSM as well as oftraditional soil survey and has been used in other many other disciplines Hans Jennypopularized the State Factor theory of soil formation in the USA through his publicationof Factors of soil formation (1941) The ideas expressed in this classic text represent a for-malization of Dokuchaievrsquos ideas of soil formation (Simonson 1997) Since that timethe theory has provided a paradigm through which soil genesis and distribution can bestudied (Schaetzl 1991) The theory states that soil profile character is a function ofclimate organisms relief parent material and time implying that if the spatial distrib-ution of the soil-forming factors is known soil character may be inferred Thistheoretical framework has been used by many authors in pedological research andremains the most popular theory of soil genesis (for a recapitulation of Jennyrsquoscontinuing influence on pedology see McSweeney et al 1994)

Until recently predictive soil mapping efforts have been constrained by twoproblems associated with Jennyrsquos State Factor theory of soil genesis the mathematicalequations derived from theory proved virtually unsolvable (Huggett 1975) and datadescribing the soil-forming factors were unavailable or not widely available Recentadvances in mathematical theory (fuzzy logic) and statistical methods (includingmachine learning techniques) have helped researchers to better understand the natureof the equations implied by the theory and to solve them Relative to the problem ofsoil-forming factor knowledge Jenny (1941 262) wrote lsquothe conversion of suchfundamental knowledge (soilndashenvironment relationships) to specific field conditions isimpossible unless the areal distribution of the soil formers is knownrsquo Half a centurylater remote sensing and photogrammetry along with GIS can now be used to charac-terize the spatial distribution of soil-forming factors The foundation has been laid forthe emergence of predictive soil modeling as an active area of research

Several articles have been published outside of geographic literature to trackprogress in the field of PSM over the years Two earlier articles (McBratney 1992Hewitt 1993) called for an overall change in the philosophy of soil survey and outlinedresearch challenges in the field at that time In the mid 1990s the term pedometrics wascoined (Webster 1994) to refer to quantitative research in the field of pedology Thisterm would appear synonymous with PSM as defined above More recently reviewshave been published on a subset of PSM techniques (McBratney et al 2000) and generalmethods of modelling soil variation (Heuvelink and Webster 2001)

Predictive soil mapping has only recently emerged as a research niche but the paceof research breadth of knowledge and variety of techniques have expanded rapidly(pedometrics have accounted for 18 of the subject matter in articles published inrecent years in Geoderma Hartemink et al 2001) A review of recent achievements in thefield of PSM is needed because all methods have not been reviewed in past papersFurther a geographic perspective is needed because of the inherently geographicnature of PSM First we review several topics critical to PSM (Section II) including thenature of soil variability (Section II1) how the soil survey has defined soil mapping(Section II2) and soil taxonomy (Section II3) We then briefly describe the componentsof PSM (Section III) geographic data models of soil variability and soil mapping(Section III1) digital terrain modelling (Section III2) fuzzy logic (Section III3) andremote sensing (Section III4) Lastly we discuss the present state of knowledge in PSM(Section IV) Our focus in this review is to evaluate literature in the field of PSM and toreintroduce the subject matter to the geographic audience

P Scull et al 173

174 Predictive soil mapping a review

II Background

Prediction of soil properties based on knowledge of the effect of environmentalvariables on soil formation has always been the basis for all soil mappingUnfortunately the traditional methods do not yield quantifiable soilndashlandscapeinformation that robustly describes actual soil variation In this section we describe soilvariation and explain why traditional methods for defining soil distribution inlandscapes are inadequate It is also our purpose to demonstrate that there is a logicalflow from these approaches into the new PSM techniques All approaches to soilmapping rest on our ability to use knowledge of the process of soil genesis to predictthe properties of soils at any point in the landscape

1 The nature of soil variability

The complex and highly variable nature of soil patterns in landscapes complicates thealready labour-intensive process of collecting and presenting soil survey data (Wrightand Wilson 1979) In 1941 Jenny listed three different definitions of soil before heconceded that lsquoit is problematic whether any definition of soil could be formulatedrsquo(Jenny 1941 2) In the half-century that has passed since Jenny wrote those words therehave been numerous attempts to define soil (reviewed by Birkeland 1999) Mosttextbook authors describe soil as extraordinarily complex For example McKnight(1993 336) defines soil as lsquoan infinitely varying mixture of weathered mineral particlesdecaying organic matter living organisms gases and liquid solutionsrsquo (emphasisadded) This definition illustrates how complex soil can be It further demonstrates thatthe soil landscape is continuous and is not composed of distinct individual soil typesThis point has been made repeatedly over the years (Simonson 1959 Webster andBeckett 1968 Cambell 1977 and Moore et al 1993)

A variety of soil genesis models have been proposed in order to account for the highvariability of soil and collectively they help further illustrate the difficulty of charac-terizing the soil landscape (for a review see Huggett 1975 Birkeland 1999) Threedistinctive approaches have been employed factor models (eg Jenny 1941) wherefactors affecting soil development are identified process models (eg Simonson 1959)where soil-forming processes are emphasized and energy models (eg Runge 1973)where the focus is upon process-driving mechanisms A host of hybrid models havealso been employed Still no real consensus exists today as to exactly how to model soildevelopment partially because of the recent emergence of pedology as an academicdiscipline (Johnson and Watson-Stegner 1987) Existing data collection methods do notyield adequate soil information in part because many of the processes that shape thesoil landscape are still poorly understood

2 How the soil survey has defined soil mapping

Traditional soil survey persists as the most popular form of soil mapping and inventoryand in many cases is the only manner in which the highly variable nature of the soillandscape is catalogued The method consists of three steps (Cook et al 1996) The firstis direct observation of ancillary data (aerial photography geology vegetation etc) and

soil profile characteristics In the second step the observations of soil attributes areincorporated into an implicit conceptual model that is used to infer soil variation Thethird step involves applying the conceptual model to the survey area to predict soilvariation at unobserved sites Usually less than 0001 of the survey area is actuallyobserved (Burrough et al 1971) a fact reflecting the high cost of field sampling Theconceptual model of soil variation is then transformed into a cartographic model thechoropleth map by drawing map unit boundaries on aerial photographs In effect pho-tographic scale determines the resolution of the soil map

This process has been severely criticized in the scientific literature for two reasonsFirst the conceptual model developed by the soil surveyor is primarily implicit beingconstructed in a heuristic manner This results in an excessive dependence upon tacitknowledge and as such incomplete information exists relative to the derivation of theultimate soil survey product (Hudson 1992) This aspect of soil survey is especiallyfrustrating because it fails to document most of the knowledge that the soil surveyoraccumulates during the expensive field mapping process In essence the soil survey isunfalsifiable and therefore untestable (Hewitt 1993) The final product of soil survey isa soil map that has unknown assumptions limitations and accuracy (Burrough et al1971 Dijkerman 1974) PSM techniques are similar in theory to soil survey (they bothuse knowledge of soilndashenvironment relations to make inferences) but the methodsemployed often yield quantitative expressions of soil variability with measured levelsof accuracy Clearly one direction of innovation in soil survey is to add objectivity tomodel development which will allow more explicit scientific communication (Gessler1996) The second major criticism of soil survey concerns the role of soil classification

3 Soil classification

The evolution of PSM techniques has been directly impacted and continues to beinfluenced by the process of soil classification This is especially true in the USAwherethe primary focus of the National Cooperative Soil Survey is to develop and map Soiltaxonomy (the official classification developed by the US Department of AgricultureSoil Survey Staff 1975) The system was influenced by nineteenth century biologicaltaxonomy and the practice of geological survey (Heuvelink and Webster 2001) Thepurpose of Soil taxonomy was to provide an objective manner to systematically classifysoil and was adopted at a time when soil information had to be abstracted to the levelof the modal profile (classified in Soil taxonomy) because it was impossible to catalogueand present the full amount of soil variability (Cambell and Edmonds 1984) In orderto map soil taxa the soil must be perceived as a spatial entity a lsquopedonrsquo (a term used inSoil taxonomy to refer to the smallest recognizable unit that can be called lsquoa soilrsquo) Inpractice this spatial perception of soil results in a map whose classes are homogenousunits with unknown variability and sharply defined boundaries (Burrough andMcDonnell 1998) Since the initial development of Soil taxonomy the perception of thesoil landscape has changed from a collection of individual soil types to a continuallyvarying mixture of soil components (Dmitriev 1983) Field observations have shownthat in concert with the environmental variables soil properties vary continuouslyacross landscapes exhibiting different and complex scales of variation (Simonson1959) Therefore soil distribution is not well represented by choropleth maps

P Scull et al 175

176 Predictive soil mapping a review

(McBratney 1992 Gessler et al 1995) While this change has not yet been recognized inSoil taxonomy PSM research has developed methods of soil inventory that moreaccurately describe the soil landscape Such methods are at odds with the traditionalapproaches because they often involve dismissing the concept of soil as a spatial entityRather the new methods focus on mapping continuously varying soil properties Theimplementation of these techniques has been difficult because of the entrenchment ofSoil taxonomy

III Components of predictive soil mapping

1 Geographic data models of soil variability and soil mapping

Traditionally soil maps have been digitized to fulfil the need for soil data within GIS-based environmental modelling research This information from the paper mapdigitized for the computer is fraught with the same problems of the original choroplethmaps ndash assumed homogeneous units with unknown variability and sharply definedboundaries GIS allows for a more robust characterization of spatial variability (relativeto the cartographic generalizations of the past) by allowing information to be analysedand stored using a variety of data models As defined in the GIS literature datamodelling is the process of discretizing spatial variation which entails abstracting gen-eralizing or approximating geographic reality (defined as empirically verifiable factsabout the real world) Unfortunately data modelling is often confused with issues ofdata structure and limited by software selection (Kemp 1992) The process is of crucialimportance because it controls the manner in which the data can be processed oranalysed (Goodchild 1994) as well as the view of the data the end user ultimatelyreceives (Goodchild 1992a Kemp 1992) Some data models are more accurate thanothers at portraying geographical reality (Goodchild 1992b) In the case of soil howwell do the data including the data model represent the highly variable continuousnature of soil How can digital computers be used to manage spatial data to bestrepresent the soil landscape

The choice of data models is partially dependent upon how the soil is perceived ingeographic space Objects can be thought of as existing as independent entities inempty space (object view) or as one of an infinite set of tuples (the foundation ofgeographic information ndash xyz where z is a measured value and xy are its location inspace) approximated by regions and segments (field view) (Goodchild 1994) The fieldview better represents continuous surfaces such as the soil landscape but soil datahave traditionally been modelled using the feature model of geographic space thechoropleth map Soil data have probably been managed this way because of theinfluence of Soil taxonomy which defines individual soil lsquotypesrsquo and because soil datacollection began during a time when few alternatives existed

Using PSM techniques a fundamental change in the soil data model from thechoropleth map to the raster grid allows better characterization of actual soilndashlandscapevariability A raster data model accommodates a lsquofield viewrsquo representation of thelandscape and is defined as a regular rectangular array of cells with some aggregatevalue of the field recorded for each cell (Goodchild 1994) The resolution of the datastored in this format is a function of the grid cell size which can be made small enoughto simulate continuous variation at the landscape scale The raster has become the most

widely used data model for PSM and is also routinely used to manage other environ-mental information such as elevation (DEMs) and remotely sensed data Spatialanalysis and integration with other types of raster-based environmental data can beeasily performed with soil data stored using a raster data structure (Burrough andMcDonnell 1998)

2 Digital terrain modelling

Terrain analysis quantifies the relief component of models characterizing soilformation Soil development and its associated profile characteristics often occurs inresponse to the way in which water moves through and over the landscape which iscontrolled by local relief Accordingly terrain analysis will be most useful in environ-ments where topographic shape is strongly related to the processes driving soilformation (McKensie et al 2000) Digital terrain modelling is a technique for derivingspatially explicit quantitative measures of the shape character of topography (Weibeland Heller 1991 Wilson and Gallant 2000) The spatial distribution of the resultingterrain attributes (characterizing local water flow paths) can also capture the spatialvariability of soil attributes Moore ID et al (1991) reviewed the analysis of digitalelevation data (including DEMs) for hydrological geomorphological and biologicalapplications They provided a table that summarized the significance and physicalmeaning of various terrain attributes to landscape processes Building on their workmany authors have used terrain attributes derived from digital elevation models(DEMs) as explanatory variables in predictive soil models (Odeh et al 1991 Moore etal 1993 Gessler et al 1995 Skidmore et al 1996 and others) Methods used to deriveterrain attributes have been greatly refined over the last 15 years and future satellitesaiding in the development of more accurate DEMs will make terrain analysis anincreasingly important component of predictive soils mapping (Moore et al 1993Mackay and Band 1998) Several recent review articles have been specifically devotedto the role of terrain analysis in soil mapping (Ventura and Irvin 1996 Irvin et al 1996McKensie et al 2000)

3 Remote sensing

Remote sensing data are an important component of PSM because they provide aspatially contiguous quantitative measure of surface reflectance which is related tosome soil properties (Agbu et al 1990) Both physical factors (eg particle size andsurface roughness) and chemical factors (eg surface mineralogy organic mattercontent and moisture) control soil spectral reflectance (Irons et al 1989) Surfacemineralogy can be derived by wavelength specific charge transfer and crystal fieldabsorptions associated with the presence of iron and iron-oxides (Fe2+ and Fe3+) andvibrational absorptions associated with hydroxyl bonds in clays adsorbed water andthe carbonate ion (Goetz 1989 Irons et al 1989) The presence and strength of theseabsorption features can be used to identify and quantify concentrations of mixed suitesof minerals in soil (Johnson et al 1983 Shipman and Adams 1987) Organic matterparticle size and moisture content in contrast influence soil reflectance primarilythrough a change in average surface reflectance and produce only broad spectral

P Scull et al 177

178 Predictive soil mapping a review

expression (Irons et al 1989) Adecrease in particle size tends to increase surface albedoand decrease spectral contrast of absorption features while an increase in organicmatter or soil moisture decreases average reflectance or albedo

Numerous studies have shown the potential benefits of using remote sensing for soilidentification and mapping Comprehensive surveys of soil spectral reflectance includestudies by Stoner et al (1980) Henderson et al (1992) and Csillag et al (1993) Remotesensing studies based on broad-band sensors such as Landsat TM include Agbu et al(1990) Coleman et al (1993) Seyler et al (1998) and Oliveira (2000) Traditionallyremote sensing has been used to classify soil units through photo-interpretation ordigital image processing Combining remotely sensed information with ancillaryinformation such as thematic maps or vegetation cover can yield significant improve-ments (Wilcox et al 1994 Cialella et al 1997 Wanchang et al 2000)

Recent developments in hyperspectral remote sensing offer the potential of signifi-cantly improving data input to predictive soil models Hyperspectral sensors such asthe Airborne VisibleInfrared Imaging Spectrometer (AVIRIS) measure a contiguousspectrum in the visible and NIR and thereby better characterize atmospheric andsurface properties (Goetz et al 1985) The large number of spectral bands permits directidentification of minerals in surface soils For example Clark and Swayze (1996)mapped over 30 minerals using AVIRIS at Cuprite Nevada Palacios-Orueta and Ustin(1996) showed that enhanced spectral information was suitable for discriminating evensubtle spectral changes associated with differences in organic matter and iron contentOther examples of the application of AVIRIS to aid soil mapping include Palacios-Orueta et al (1998) Okin et al (1998) and Roberts et al (1998)

Sensors that operate in the microwave portion of the electromagnetic spectrum havealso shown promise in soil mapping research Microwave sensing can be broadlydivided into active (eg radar) systems and passive systems and are capable ofpenetrating the atmosphere under virtually all conditions offering a significantadvantage over visible and near-infrared spectroscopy (for a general overview ofmicrowave sensing see Lillesand and Kiefer 1994 chapter 8) Synthetic aperature radar(SAR) is one example of an active system SAR has been used to aid soil propertymapping such as soil salinity (Metternicht 1998) and soil moisture (Engman andChauhan 1995 Narayanan and Hirsave 2001) Active radar systems can also bedesigned to collect data at varying look angles providing the opportunity for theacquisition of stereo radar images Such images can be used to produce high resolutionand extremely accurate DEMs (eg Fang 2000) A similar active sensing system isLiDAR (light detection and ranging) which uses pulses of laser light rather thanmicrowave energy to illuminate the surface (see Bunkin and Bunkin 2000 for a reviewof applications to soil mapping research) While passive microwave systems haveseemed to receive less attention in the literature a few examples of soil mapping appli-cations can be found (see Kleshchenko et al 2000 and Laymon et al 2001) Regardlessof the type of system remote sensing data and derived products are potentially usefulexplanatory variables in predictive soil mapping models

4 Fuzzy logic

Fuzzy set theory or fuzzy logic provides an alternative conceptual paradigm withinPSM research The use of this theory has increased greatly in the last few years making

it an important component of PSM Fuzzy logic is an alternative to Boolean logic thatattempts to recognize the concept of partial truth (Brule 1996) Dr Lotfi Zadeh (1965)introduced the concept and accompanying mathematics in his seminal work lsquoFuzzysetsrsquo The theory permits partial class membership in contrast to traditional set theorywhere set memberships are crisp and binary (ie a soil sample is either completelyType A or it is not at all Type A) Central to the fuzzy concept is the idea that objects innature rarely fit exactly the classification types to which they are assigned (Zadeh1965) Rather they show varying signs of similarity to multiple classes (ie an observedsoil pedon often resembles more than one of the defined soil series within the area) Byusing fuzzy membership values (ranging from 0 nonmembership to 1 totalmembership) within predictive soil models to express degrees of similarity generaliza-tion problems associated with classification schemes (filtering of information) areminimized and the complex nature of soil data is allowed to propagate through themodelling process Similarity values between 0 and 1 are not comparable to proportionsand need not add up to 1 Within Boolean logic probability statements refer to thelikelihood of an outcome the soil sample is either one series or another With fuzzylogic a given sample is not definitively a member of the subset of any one particularseries Fuzzy logic is especially useful in soil research because of the continuous andcomplex nature of the soil landscape It serves as an important alternative to thesubjective rigidity imposed on soils data by Soil taxonomy Several recent articlesprovide a thorough review of the use of fuzzy sets in soil science (Burrough 1989McBratney and Odeh 1997 Burrough et al 1997 De Gruijter et al 1997)

Within PSM research two different approaches to creating continuous classes usingfuzzy logic exist The first is based on the fuzzy-k-means classifier which partitionsobservations in multivariate space into natural classes This approach is similar tocluster analysis and numerical taxonomy but the resulting classes are continuous witheach observation assigned a fuzzy membership value that characterizes its degree ofsimilarity to each individual class The concept has been integrated into geostatisticalmethods and will be discussed in more detail below (see Section IV1) The secondapproach is known as the Semantic Import model (SI) and is used in situations whenclassification schemes are pre-defined and class limits are relatively well understoodThe SI model is commonly used in concert with expert knowledge and will bediscussed in the expert systems section (see Section IV4)

IV Recent advances in predictive soil mapping

Within the last decade many authors have sought to model the soil landscape using avariety of methods Literature in this field could be summarized many different waysbut we concentrate on the literature that directly addresses the goals of predictive soilmapping stated in the introduction (see Table 1) Therefore we will review research thatattempts to exploit the relationship between quantifiable landscape indices and soilcharacter in order to model the soil landscape in a more continuous and thereforerealistic manner

The research reviewed here is distinguished from decades of previous researchdocumenting the correlation between landscape position and soil attributes (reviewedby Hall and Olsen 1991) That body of research is informative but not useful for

P Scull et al 179

180 Predictive soil mapping a review

Table 1

Selected

recen

t literature on pred

ictive

soil m

apping an

d m

apping

(cited in

this article) d

escribing the mod

ellin

g metho

dused the

dep

enden

t variables used a

nd the

env

iron

men

tal v

ariables (ex

plan

atory) used in

the

mod

els

Stud

yMod

ellin

g metho

dDep

ende

nt variables

Environm

ental variab

les

Goa

ls atta

ined

a

Bell et

al 20

00Line

ar and

exp

onen

tial

Total soil orga

nic

Slop

e curvature aspe

ct

ECx

regression

carbon

hillslope

position

Burge

ss and

Web

ster

Punc

tual and

block

Na co

nten

t co

ver

Non

eC

1980

ab

kriging

loam

thickn

ess

ston

e co

nten

t

Burroug

h 198

9Fu

zzy mathe

matical

ndashndash

rpmetho

ds

Burroug

h et

al 19

97Con

tinuo

us classificatio

nndash

ndashrp

Castrigna

no eta

l 200

0Factorial kriging

CEC

pH N

P K

Non

e ex

plicitly used

Cx

Na

Cialella

eta

l 199

7Classifica

tion tree

Drainag

e class

Elev aspe

ct NDVI

Ec

Coo

k et

al 19

96Baysian

rule-ba

sed metho

dsOrgan

ic m

atter

Slop

e aspec

t wetne

ssEcX

inde

x

Dale et

al 19

89 (a review

)Ex

pert systems

ndashndash

ndash

Ellis 19

96Decision tree

ana

lysis

Soil erosion class

Slop

e aspec

t wetne

ssEc

neutral ne

tworks

inde

x flow le

ngth and

accu

mulation Lan

dsat

TM tree

cov

er

Gessler 1

996

A large

variety of statistic

Field an

d labo

ratory

A variety of digital

EC

metho

dsco

llected

phy

sica

len

vironm

ental da

tach

emical and

morph

olog

ical soil

prop

ertie

s

Gessler eta

l 199

5Line

ar and

logit reg

ression

A horizon

and

solum

Curvature CTI topo

EC

depth E horizon

positio

npresen

ce

Goo

vaerts 19

92Factorial kriging

Total carbon

Non

eC

P Scull et al 181pH

N CEC

extractab

lecatio

ns (K Ca M

g)Hartemink et

al 20

01ndash

ndashndash

rp

Heu

velin

k an

d W

ebster 20

01ndash

ndashndash

rp

Hew

itt 19

93ndash

ndashndash

rp

Indo

rante et

al 19

96ndash

ndashndash

rp

King et

al 19

99Lo

gistic regression

Presen

ceabsen

ceSlop

e aspec

t po

t solar

Ec

Non

calc c

lay-loam

energy

Kno

tters eta

l 199

5Kriging

co

-kriging

So

ft layer de

pth

Hillslop

e po

sitio

nCx

regression

kriging

Laga

cherie and

Holmes 19

97Classifica

tion tree

Map

ping

unit

Geo

logy variou

s topo

Ec

indices

Laslett e

tal 198

7Kriging

splin

es tren

dpH

Non

ec

surfac

e nea

rest neigh

bor

McB

ratney 1

992

ndashndash

ndashrp

McB

ratney

eta

l 199

1Block kriging

Clay co

nten

tNon

eC

McB

ratney

eta

l 200

0ndash

ndashndash

rp

McB

ratney

and

Ode

h 199

7Fu

zzy sets in

soil scienc

endash

ndashRp

McB

ratney

and

de Gruijter

Fuzz

y-k-mea

ns w

ithFu

zzy classes

Field co

llected

phy

sical

C19

92ex

tragrade

sch

emical and

morph

olog

ical soil

prop

ertie

s

McC

rack

en and

Cate 198

6Artificial intellig

ence

ndash

ndashrp

expe

rt systems

McK

ensie an

d Austin

19

93Gen

eralized

line

ar m

odels

Clay co

nten

t CEC

Slop

e relief land

form

ec

(logit)

pH EC

COLE

slop

e po

sitio

nbu

lk den

sity and

othe

rs

Moo

re eta

l 199

3Line

ar reg

ression

A horizon

dep

th O

MSlop

e w

etne

ss and

strea

mEC

and P co

nten

t pH

power in

dices aspect

curvature

182 Predictive soil mapping a reviewTa

ble 1

Con

tinu

ed

Stud

yMod

ellin

g metho

dDep

ende

nt variables

Environm

ental v

ariables

Goa

ls atta

ined

a

McK

ensie an

d Ryan 199

9Reg

ression tree and

linea

rSo

lum dep

th P an

dElevation slope

CE

regression

N con

tent

curvature CTI

contribu

ting area do

wn-

slop

e mea

ns for slope

clim

ate da

ta Presco

ttInde

x G

amma

Rad

iometry La

ndsat T

M

and Geo

logy unit

Ode

h et

al 19

92ab

Fuzz

y-c-mea

ns and

Fuzzy classes

Field-co

llected

phy

sica

lC

kriging

chem

ical and

morph

olog

ical soil

prop

ertie

s

Ode

h et

al 19

94

Reg

ression kriging

co

-So

lum dep

th de

pth

Slop

e aspect cu

rvature

eC

1995

kriging regression

to bed

rock gravel

kriging

and clay con

tent

Skidmore et

al 19

91Bayesian expe

rt system

Soil land

scap

e un

itVeg type

wetne

ss in

dex

EcX

1996

grad

ient terrain po

sitio

n

Voltz and

Web

ster 19

90Kriging

cu

bic splin

eClay co

nten

tNon

eC

Web

ster 1

994

Dev

elop

men

t of

ndashndash

rppe

dometrics

Zhu

199

7ab

Fuzz

y logic expe

rt system

Soil series A

horizon

Elev pm aspe

ct c

anop

yECX

Zhu

and

Ban

d 199

4(SoL

IM)

depth in

dividu

alco

verage

grad

ient

Zhu

eta

l 199

7series m

aps

curvature

Not

es

a Letters refer to the de

gree

to w

hich

the

goa

ls of PS

M defined

in the

introdu

ction are achiev

ed

Soilndash

e nvironm

ent relatio

ns utilized

(letter E) be

tter represen

tatio

n of soil c o

ntinuity (C

) an

d ex

pert kno

wledg

e utilize

d (X) c

orrespon

d to goa

ls 1 2 and

3respec

tively from

the

introdu

ction and

cap

ital lette

rs (E C X) ind

icate the metho

d is relatively more successful th

an th

ose metho

ds den

oted

by lower

case le

tters (e c x

) rp ind

icates rev

iew pap

ers

P Scull et al 183

predictive mapping because landscape position is never quantified (rather positionwas often qualitatively defined eg toe-slope) and thus the documented relationshipscannot be generalized using environmental data and digital elevation models to predictsoil character at unvisited sites

Table 1 documents modelling methods model variables and the extent to which thereferenced research satisfies the previously defined goals of PSM (lsquoGoals attainedrsquocolumn) Soil-environment relations utilized (letter E) better representation of soilcontinuity (C) and expert knowledge utilized (X) correspond to goals 1 2 and 3respectively from the introduction and capital letters (E C X) indicate the method isrelatively more successful than those methods denoted by lower case letters (e c x)Citations with no letters present within the lsquoGoals attainedrsquo column do not address theaforementioned goals For example the Cialella et al (1997) received a rating of lsquoEcrsquomeaning that the methods employed successfully utilized environmentalndashsoil characterrelations (E) and somewhat successfully presented a better method of representing soilcontinuity (c) The ratings are provided simply to help organize the literature that wasreviewed Review papers are included within the table denoted by lsquorprsquo Geostatisticalmethods are not included in the table because they have been comprehensivelysurveyed elsewhere (Odeh et al 1994 Burrough et al 1997 McBratney et al 2000Heuvelink and Webster 2001) and because the objectives and assumptions of geostatis-tical methods differ slightly from other PSM research We briefly outline thesedifferences in the following section

1 Geostatistical methods

Geostatistics are a subset of traditional statistics that deal primarily with spatial dataand account for spatial autocorrelation using kriging as the spatial interpolator Theconcept is based upon the theory of regionalized variables which was mainlydeveloped by Matheron (1963) and Krige (1963) Kriging is a form of weighted localaveraging that uses a measure of spatial dependence the variogram to determine theweights applied to the data when computing the averages Geostatistical methods havebeen used in predictive soil mapping research to spatially interpolate soil propertyvalues at unmeasured sites from field-collected data

Burgess and Webster (1980a b) were the first to introduce ordinary kriging to the soilcommunity and since that time an enormous amount of work has been published Forexample ordinary kriging has been used to interpolate many different soil propertiesincluding pollution trace element deficiencies salinity and fertility (Heuvelink andWebster 2001) Ordinary kriging has been criticized for a variety of reasons Forexample Laslett et al (1987) reported that several authors had criticized geostatisticsbecause kriging is a global rather than local technique failing to take into accountknowledge of soil materials and processes Other authors have criticized geostatisticsbecause they are excessively data dependent requiring a large number of closelyspaced data points (Zhu 1997a) As Webster and Oliver (1992) suggest in excess of ahundred samples may be needed to use geostatistics at the field scale because of highspatial variability of soil in some areas Geostatistics also assume spatial autocorrela-tion which sometimes may be a poor assumption in complex terrain where abruptchanges in soil-forming factors occur (McBratney et al 2000) Ordinary kriging by itself

184 Predictive soil mapping a review

does not satisfy two of the three goals of PSM presented in the introduction ndash it does notadequately incorporate expert knowledge and it does not exploit the relationshipbetween environmental variables and soil properties

Ordinary kriging has been modified in a variety of ways to better incorporateancillary data and known soilndashlandscape relationships Block kriging involvesdetermining estimates over meaningful areas rather than at specific points (Burgess andWebster 1980a McBratney et al 1991) Using this method a study area can be stratifiedinto different regions that are reflective of the pedogenetic processes at work In orderto accommodate a trend within a dependent soil variable universal kriging has beenused (Webster 1994) Kriging with external drift is similar to universal kriging but ituses an ancillary variable to represent the trend (McBratney et al 2000) Co-krigingtakes advantage of correlation that may exist between the variable of interest and othermore easily measured variables (Odeh et al 1995) Regression kriging involvesspatially interpolating the residuals from a non-spatial model by kriging and addingthe result to the prediction obtained from that model (Goovaerts 1997 Castrignano etal 2000) Factorial kriging is another method to integrate multivariate data into thestandard kriging routine to extrapolate soil data (Goovaerts 1992) Many authors havecompared these various methods (Laslett et al 1987 Voltz and Webster 1990 Odeh etal 1994 Knotters et al 1995)

Fuzzy logic has been used with geostatistics by various authors to produce new kindsof fuzzy soil maps with continuous classes (Burrough 1989 McBratney and DeGruijter1992 Odeh et al 1992a and reviewed by McBratney and Odeh 1997) The processentails kriging the matrix of membership values determined by the fuzzy k-meansclassifier resulting in a continuous soil surface where individual locations are allowedto belong to more than one class and no rigid boundaries are designated to separate thesoil into discrete units or entities The results of such analysis can be used to assess thepedologic process validity of soil taxonomy by determining whether soils grouptogether into classes that are similar to taxonomic types Fuzzy classes wouldpresumably reflect the main pedologic features within a given area In this sense theclassification is quantitative whereas soil taxonomy is rooted in qualitative discrimina-tion A comparison between the two could provide insightful and help assist soiltaxonomy in making a classification that is more indicative of underlying soil processes

Geostatistics in soil research were originally introduced to quantitatively assess soilvariability within soil mapping units (McBratney et al 1991) in response to criticisms inthe early 1970s that soil unit composition was not well quantified (Beckett and Webster1971) In this regard geostatistics have been very useful having served well the originalgoals set forth by Burgess and Webster (1980b) when they drew kriging to the attentionof soil scientists as a means of spatial prediction At the field scale soil variation islargely due to the effect that topography has on soil genesis Geostatistics have beensuccessfully applied in such environments by using terrain attributes as ancillary datawithin many of the kriging routines described above Such quantitative within-unitvariability of soil properties is very useful in the field of precision agriculture and othersituations (eg pollutants) where very detailed soil attribute information is needed atthe field scale (Heuvelink and Webster 2001)

However geostatistics have not been applied in a wide variety of environments or atlarger scales In order to be successfully applied in different environments geostatisticswill likely require a different suite of ancillary data For example remote sensing data

P Scull et al 185

could be used in arid regions where soil toposequences are less well expressed Atlarger scales of prediction selection of different sets of ancillary variables is requiredbecause different processes define soil character at different scales The most obviousexample is that of climate which may control soil distribution at large scales(continental) but has little explanatory power at the field level Regardless of whetheradequate ancillary data exist the amount of data required to use geostatistics forlandscape-level prediction would be extremely difficult and costly to collect given thestrict sampling protocol required to characterize spatial dependence It is also unclearat what landscape scale soils exhibit spatial autocorrelation

Geostatistical approaches do provide a means of creating continuous soil attributesurfaces to better represent soil continuity (Goal 2) and they can be used to exploit therelationship between environmental variables and soil properties in order to moreefficiently collect soil data (Goal 1) However they do not sufficiently utilize expertknowledge (Goal 3) as no attempt has been made in geostatistical approaches todirectly integrate expert knowledge Fundamentally kriging is a process of interpola-tion designed to predict attribute values in between locations of measured samples Inthis sense geostatistics represent a middle ground between pure interpolation (egnearest neighbour type classifier) in which only measured points for the variable ofinterest are used to determine unknown values and other predictive models thatprimarily use soilndashenvironment correlation to create predictive maps

2 Statistical methods

Statistical methods can be used to exploit the relationship between quantifiablelandscape indices and soil properties to create predictive soil maps For exampleMcKensie and Austin (1993) used a regression to account for a large percentage ofvariation for many soil characteristics (A horizon clay content CEC EC pH bulkdensity and COLE B horizon clay content CEC ESP EC pH bulk density and COLE)using a variety of predictor variables (slope presence or absence of impeding layerrelief landform topographic position) Their results confirm the hypothesis of MooreID et al (1991) that soil character is related to quantifiable landscape indices Howevertheir methods do not provide inference of soil properties at unmeasured sites frommapped environment data because the topographic variables were measured in thefield Linear regression has also been used with terrain variables derived from a 15-mDEM in northeastern Colorado to predict soil attributes (organic matter contentextractable phosphorous pH and texture) at unvisited sites (Moore et al 1993) In thatparticular study 50 of the variance of A-horizon thickness was explained by slope andthe wetness index Gessler et al (1995) also used regression to model A-horizonthickness from topographic variables in southeastern Australia (plan curvature andwetness index r2 = 063 P = 0001) They modeled solum depth and used logisticregression to model E horizon presenceabsence Elsewhere logistic regression hasbeen used to model the presenceabsence of noncalcareous clay loam horizon in centralFrance using terrain attributes from a 20-m DEM (King et al 1999) Exponentialregression has been used to model soil organic carbon using terrain variables (Bell et al2000) in glacial outwash soils in east-central Minnesota Generalized additive models(GAM) have been used less frequently in PSM research Gessler (1996) used a GAM

186 Predictive soil mapping a review

model to predict total soil carbon A horizon depth and solum depth using a variety ofenvironmental predictors (slope elevation wetness index mean annual temperatureprecipitation and radiation)

This small body of research opened the door to more complex methods by demon-strating the existence of quantifiable relationships These authors were able to producesoil attribute maps using raster data models whose scale was dependent upon the gridcell resolution of the environmental data They were successful at exploiting the rela-tionship between quantifiable topographic attributes and soil profile character (Goal 1)The continuous soil attribute surfaces they produce also better represent soil continuitythan the choropleth soil maps produced by traditional soil survey (Goal 2) Howeverthe bulk of these methods (excepting GAMs) are limited by their assumed linear rela-tionship between soil and topographic attributes their assumptions of normallydistributed data and their high data requirements Standard statistical procedures arealso not flexible enough to allow robust integration with a variety of potential datasources such as expert knowledge (Goal 1)

Statistical methods do demonstrate in a quantitative manner that terrain analysis canbe used to predict soil attributes in relatively small areas with homogeneous parentmaterial A large proportion of the research using statistical methods was conducted insemi-arid landscapes at small scales (the largest study area of the entire group was~2000 ha) Obviously for statistical approaches to be most effective they need to bemore universal As such they need to be tested andor developed at larger scales andin more diverse landscapes

3 Decision tree analysis (DTA)

The use of decision tree analysis is just beginning to be explored in predictive soilmapping research although it has been used successfully in the related field ofpredictive vegetation mapping since the early 1990s (Lees and Ritman 1991 MooreDM et al 1991 Franklin 1998) DTA is a form of divisive classification The process oftree modelling involves successively partitioning data (called recursive partitioning inthe tree modelling literature) into increasingly homogeneous subsets which once thepartitioning has ceased are called terminal nodes (Lees and Ritman 1991) Splits orrules defining how to partition the data are selected based on information statistics thatdefine how well the split decreases impurity within the data set (Clark and Pregibon1992) Splits are based on threshold values of an explanatory variable selected bycomparing the increase in resulting purity of node membership for all possiblethresholds and variables The process is iterative growing from the root node (thecomplete data set) to the terminal nodes in a dendritic fashion (Friedl and Brodley1997) Once the tree has been constructed (or grown) it encodes a set of decision rulesthat describe the data partitioning process These rules can be used to classify or predictother data sets (Moore DM et al 1991) Pruning the tree is often necessary to preventthe tree from being overfit to the sample data and to reduce tree complexity Pruningentails combining pairs of terminal nodes into single nodes and can be accomplishedusing cross-validation which yields an initial indication of how large a tree makesrobust predictions (Safavian and Norvig 1991) Cross-validation involves systematical-ly removing portions of the data set and running the remaining sample through the tree

in an iterative manner eventually yielding estimates of the misclassification rates foreach class each node and the whole model (Breiman et al 1984) In this mannerdifferent sized trees can be compared in terms of parsimony

The term DTA is used to collectively refer to all types of tree-based modelling (theword lsquodecisionrsquo is used because it is descriptive indicating that the analysis eventuallyleads to a set of decision rules defining data partitions) The term should be distin-guished from classification tree analysis because the latter refers specifically to DTAwhere the response variable is categorical The term CART (classification andregression trees ndash Breiman et al 1984) is sometimes used but strictly speaking refers tospecific software Friedl and Brodley (1997) provide a review of the decision treealgorithms They divide types of DTA into two classes (1) homogeneous decision treesfor which a single algorithm is used to estimate each split (eg CART) and (2) hybriddecision trees (HDT) for which different splitting methods can be used at differentpoints in the tree (eg Quinlan 1993) They further divide homogeneous decision treesinto univariate (UDT) where single features of the input data define splits and multi-variate decision trees (MDT) where multiple features of the input data can define splitsAccording to this naming convention no distinction is made between methods withdifferent types of response variables although all methods can be used with bothcategorical and continuous response variables In a comparison of these various typesof algorithms on a variety of data sets Friedl and Brodley (1997) found that HDT hadthe highest classification accuracy

The overall aim of DTA is to design a set of predictive rules (eg if geology type Athen soil type B) developed from training data which can then be applied to ageographic data base to predict the value of a response variable (Michaelsen et al 1994)Therefore DTA explicitly uses soilndashlandscape correlation in model development (Goal2) The technique appears promising in soil research but needs to be further tested asnot many of the above types of DTA have been tested in the PSM literature in fact onlyunivariate approaches have been employed For example Lagacherie and Holmes(1997) successfully used univariate DTA to model a categorical response variable soiltype within a training set and then assuming that the training set was representativeextrapolated the model to a much larger region Their work is interesting becausealthough not a single sample came from the area they eventually mapped they wereable to produce a soil map that was more accurate (74 versus 69) than the existingmap produced from traditional methods Cialella et al (1997) also used univariate DTAto predict soil drainage class from a variety of terrain attributes and remotely senseddata They predicted soil drainage class with an average of 78 accuracy ndash impressivegiven that the variation accounted for by a typical soil survey ranges from about halfthe total variance for physical attributes to less than one-tenth for some soil chemicalattributes (Gessler et al 1995)

DTA has been compared with other approaches by several authors In the applicationof erosion modelling DTA results were similar to Artificial Neural Networks (Ellis1996) Both methods achieved high training accuracy (as measured by the KappaStatistic) but in terms of prediction accuracy both methods performed poorly Gessleret al (1995) compared DTA with generalized linear models (GLM) and generalizedadditive models (GAM) to predict A-horizon thickness and concluded that GLM waspreferable to both DTA and GAM McKensie and Ryan (1999) compared regressiontrees and standard linear regression to predict soil properties (total solum depth soil

P Scull et al 187

188 Predictive soil mapping a review

phophorus and soil carbon) using a large variety of predictors (elevation slopecurvature wetness index climate variables geology unit Landsat TM data andGamma radiometric data) and found that different methods work best for differentproperties and overall success hinged on the strength of the relationship between soiland environmental variables Their study of a 50 000 ha forested landscape in southernNew South Wales Australia is one of the few studies at such a small map scale (largearea) or in forested landscape The most extensive comparison of techniques involvingDTA was conducted by McBratney et al (2000) They compared regression techniques(GLM GAM DTA-regression tree) geostatistical techniques (kriging and heterotopicco-kriging) and a hybrid technique (regression kriging) DTA was found to be thepoorest performing of the regression techniques because of the unrealistic predictionsurface generated by the DTA model DTA has also been criticized by other authorsbecause of the stepped prediction surface (Gessler 1996) This phenomenon can beespecially apparent in situations where predictor variables have different resolutionsThe extent to which DTA yields a better spatial representation of soil continuity is afunction of the scale and type of predictor variables used and therefore varies withindividual models

DTA is gaining widespread popularity as a means to develop prediction rules thatcan be rapidly and repeatedly evaluated (Cialella et al 1997 Franklin et al 2000) DTAprovides the following advantages over standard statistical techniques (1) it is easier tointerpret when explanatory variables are both nominal and continuous (2) it isinvariant to monotone re-expressions (transformations) of predictor variables (3) itdeals more satisfactorily with missing data values and outliers (4) it is more adept atcapturing nonadditive and nonlinear behaviour (5) it doesnrsquot make any assumptionsabout data distribution and (6) it is easily updateable as more data are collected(Moore DM et al 1991) The DTA model framework is especially appealing becauseof its capability to integrate a wide range of data sets as explanatory variables

DTA offers a unique opportunity for interaction between soil experts and soilmodellers because the output of the model is a set of rules that can be pedologicallyinterpreted by the soil expert In this sense expert knowledge is used in an implicitmanner in DTA (somewhat effectively achieving Goal 3) While these rules can often beexceedingly complex at minimum the expert can decide whether initial splits makesense given their understanding of the landscape The success of DTA results oftenhinges on the ability of the modeller to make key decisions during the model buildingprocess there is unfortunately no definitive way to determine the most optimal treeThe aid of the soil expert can potentially elucidate this problem

4 Expert systems

A variety of expert system approaches to PSM have been developed to utilize expertknowledge The purpose of such methods is to exploit the information the soil surveyoraccumulates while working in the field by integrating such knowledge into thepredictive model (McCracken and Cate 1986) Unlike the majority of the researchreviewed thus far the dependent variable in many expert systems models is often soiltaxa or mapping unit This apparent disadvantage of expert systems (using classifica-tion to characterize soil continuity) does make them easier to integrate into traditional

soil survey In addition several authors have developed methods to developcontinuous soil property maps from the output of expert system models designed topredict soil unit occurrence

Expert systems are composed of data (information on spatial environmentalvariables eg topography climate etc) a knowledge base (rules and facts related tosoil variation supplied from the soil surveyor) and an inference engine (whichcombines data and the knowledge base to infer logically valid conclusions) (Skidmoreet al 1996) Expert systems differ from conventional models in two ways (1) they storeand manipulate qualitative information (allowing them access to information thatcannot normally be used in other modelling frameworks) and (2) they are structuredas meta-models (the knowledge is separated from the model) (Davis 1993) This allowsthe model to selectively choose which information is relevant at various stages of themodelling process and it allows for information to be easily updated Davis (1993)reviews the application of expert systems to environmental modelling research conclu-ding that the technique is becoming more widely accepted He further states that theapplication of expert systems is constrained by an absence of fundamental knowledgefor rule generation a problem that would appear less relevant to soil mapping giventhe amount of untapped expert knowledge accumulated by the soil surveyor

The first mention of the use of expert systems in pedology was in a paper presenta-tion at the Northeast Committee Soil Survey Conference 1984 (Flach 1985) In herpaper Flach (1985) hinted that recent developments in computer science especiallyexpert systems and artificial intelligence could make modelling a practical mappingtool for soil scientists in the near future A year later McCracken and Cate (1986) hopedto encourage soil scientists to explore expert systems and its application to soil sciencethrough an optimistic article they wrote on the potential use of expert systemsHowever little research was actually conducted in the late 1980s and expert systemshad not yet begun to fulfil the lofty goals set forth by early practitioners (Dale et al1989)

In the early 1990s expert systems approaches to predictive soil mapping began toappear in the literature Skidmore et al (1991) used a Bayesian expert system to mapforest soil into different classes and their results compared favourably with availablesoil maps and actual field-collected data Their methods successfully incorporatedsurveyor knowledge and remotely sensed and digital terrain attributes but failed tobetter characterize continuous soilndashlandscape variability because their final productwas a choropleth map Skidmore et al (1996) revisited their earlier research andassessed the mapping accuracy of their results They found that the soil map producedby the expert system achieved an overall accuracy of 698 (sample size = 53) while themap derived from conventional methods had an accuracy of 736 In addition toknowledge provided by a soil scientist Skidmore et alrsquos (1996) expert system usedvegetation derived for aerial photography and topographic variables derived from a 10-m DEM (wetness index topographic position and slope)

Using an expert system Cook et al (1996) successfully produced a continuous soilproperty map for organic matter content using wetness index aspect and slope asexplanatory variables Their methods were somewhat inefficient because they requireda separate expert system for each soil property of interest However their researchrepresents the only example of expert systems used to directly predict the spatial dis-tribution of a soil property All of the other examples of expert systems in the literature

P Scull et al 189

190 Predictive soil mapping a review

deal with soil type or class The use of expert systems to map soil properties needs tobe explored further

The expert systems discussed thus far have all used Boolean logic within theirtheoretical framework whereby an observation can belong to one and only one classand the soil properties of that class are assigned to the observation With thedevelopment of fuzzy logic and the semantic import model (the second fuzzy logicapproach to creating continuous classes) the opportunity exists to develop fuzzy logic-based expert systems Such systems can be used in conjunction with expert knowledgein situations where experts have a good qualitative idea of how to group data but havetrouble dealing with observations that are not well represented by rigid classificationschemes (Burrough and McDonnell 1998) This method can be particularly useful insituations when taxonomic schemes have been previously developed as is the case insoil taxonomy Several examples of this type of approach were published in the mid-1990s by A Zhu and colleagues (Zhu and Band 1994 Zhu et al 1996 1997 Zhu1997ab) Such systems proved useful for mapping soil at unvisited locations usingsurveyor knowledge and were also capable of producing continuous soil propertymaps The use of fuzzy logic within the theoretical framework of the expert systemallows the complex nature of soil to propagate through the modelling process neversubjected to classification schemes that filter out potentially useful lsquonoisersquo The use offuzzy logic also gives the soil surveyor more latitude during the interview processwhen the knowledge base is defined for the expert system Zhu et al (1997) developeda fuzzy logic-based expert system called SoLIM to determine the similarity of eachgrid cell in a study area to the various taxonomic mapping units delimited by the soilsurvey Continuous soil attribute maps were calculated using the similarity values andtheir relative soil survey determined attribute values (effectively achieving Goal 2) Theresulting data (which proved more accurate than soil survey data once field checked)consists of a raster grid whose resolution is determined by the resolution of the inputenvironmental and digital elevation data As noted previously this type of data modelis more applicable to environmental modelling than the choropleth map (Burrough andMcDonnell 1998)

Expert system approaches to PSM are capable of exploiting soil surveyor knowledgeby developing rule-based systems that imitate the surveyorrsquos conceptual model of soilvariability (the primary focus of Goal 3) The method would appear extremely usefulfor mapping projects (such as those conducted by the NRCS-NCSS) where fieldwork isinitially conducted to determine soilndashlandscape relations Expert system developmentcould be directly inserted into the traditional soil survey mapping approach as asubstitute for the step where the surveyor converts hisher conceptual model into achoropleth map Rather that knowledge could be incorporated into the expert systemwhich could be used to predictively map soil The resulting raster map would be morescientifically based and explicit than the hand-drawn choropleth maps of the past Itsscale would not be limited to that of the aerial photography but rather to the scale ofthe environmental data

Despite these advantages the expert system approach has some drawbacks Becausethe method is dependent upon expert knowledge it cannot be applied whereenvironmentndashsoil relations are poorly understood Of course this criticism can be madeof all PSM models as inductive or deductive knowledge of soilndashenvironment relationsis a prerequisite for PSM However expert systems are deductive models and as such

are not driven by any specific field-collected data (although presumably the soil experthas field experience in the mapping area) Expert systems donrsquot afford the opportunityto first statistically document landscapendashsoil relations and then extrapolate the resultsbecause expert systems do not directly use sample soil data (lsquohard datarsquo) to determinesoilndashlandscape relations Thus the utilization of the relationship between environmen-tal variables and soil properties (Goal 1) is only indirectly achieved Expert systemsapproaches have been demonstrated to be extremely effective in a small number of casestudies The possibility of satisfying all three goals discussed in this review makesexpert systems a predictive soil mapping method that needs to be further tested ndash bothgeographically and across different scales of analysis

V Conclusion

Most of the predictive soil mapping research outlined in this review was conducted atvery large map scales (over small areas) In fact the majority of the research wasconcerned with assessing the spatial variability of soil character within individual fieldsor across soil toposequences The primary driving force behind this type of research hasbeen the need to provide accurate soils information for agriculture and ecologicalmodels It is clear that terrain attributes are powerful predictors at the local scaleGeostatistical tools have been successful at using terrain attributes and the spatialdependence of soil properties to interpolate between existing data points withinindividual fields Across soil toposequences statistical approaches provide a usefulmeans of predicting soil character PSM research has been most successful at the fieldscale because many of the soil-forming factors are held constant For example the neteffect of four of Jennyrsquos five soil-forming factors (climate organism parent material andtime) was minimal within many of the studies reviewed While some research hastackled larger areas of study there exists a deficiency in our ability to predictively mapsoils at smaller map scales Since the distribution of soil is scale-dependent differentPSM methods and predictors are likely to work better at different scales Focus in thefuture must continue to move toward working over larger spatial extents of study inorder to produce landscape-scale soil information

Further a large proportion of the research was conducted in semi-arid gentlysloping agricultural landscapes Humid forestlands mountainous regions and desertshave received little attention As such PSM methods need to continue to be testedandor developed in a wider variety of landscapes where spatial soil distributions canbe more complex Different methods will likely be successful to different degrees indifferent environments Whereas terrain attributes are dominant predictors of soilcharacter across toposequences in prairie lands other predictors are likely necessary inareas where soils do not develop into clear toposequences For example remote sensingdata has been a useful predictor of soil chronosequences in desert landscapes wheresurface appearance is often related to soil character Surficial geology is often reflectiveof soil character in mountain regions where soils are thin and significant bedrock existsat the surface Focus in the future must continue to determine which methods andpredictors work best in which environments

Since the most useful PSM approach will vary across spatial scales and environmen-tal gradients the method used should be driven by the mapping objectives of the

P Scull et al 191

192 Predictive soil mapping a review

project Whereas the traditional soil survey of the past was expected to meet the needsof a diverse group of end users PSM methods will vary given the objectives of thesurvey As a result the end users of the soil data need to play a more active role in thesurvey process Because precision agriculture and large-scale mapping has been theprimary focus of PSM there exists a deficiency in our ability to predictively map soilsfor the purpose of general land use planning and management Even though lessdetailed soil information is needed for these purposes the development of PSMmethods is complicated by the fact that soils exhibit complex spatial variability atsmall map scales where soil-forming environments vary greatly from one location tothe next

A number of alternative methods of characterizing the continuous nature of the soillandscape have been developed Thus far most of the PSM research has provided soilinformation in a nonobject form (as opposed to defining soil types as independententities) Soil data have been generated organized and presented in the form of eitherisorithmic maps or fine-scale raster grids Both of these data models are field-viewmodels of geographic space which allow the soil to be perceived as a constantlyvarying surface Two distinct approaches have been employed (1) mapping individualsoil properties and (2) mapping continuously varying (fuzzy) soil classes Mappingindividual properties is the most common approach and will likely continue todominate PSM research The use of fuzzy soil classes in the literature is less commonand more difficult to be integrated into standard mapping procedures (such as use ofsoil taxa) because the concept radically differs from the traditional view of the soillandscape However fuzzy soil classes (generated using the fuzzy-k means approach) dohave the potential to help further advance Soil taxonomy by identifying taxonomicclasses that are more reflective of pedologic processes at work

Expert systems have been greatly underutilized in PSM research especiallyconsidering how effective a small number of case studies have been Expert systemshave the potential to satisfy successfully all three goals discussed throughout thereview They also have the potential to bridge the gap between traditional approachesand PSM methods because field soil scientists do not have to change their conceptualapproach to mapping They can still conduct field reconnaissance to determinesoilndashlandscape relationships Afterwards though they can use expert systems to betterexploit the knowledge they have garnered Such an approach could help ease thetransition from traditional soil survey to more scientifically explicit methods Expertsystems are also the most fruitful approach to utilizing a wealth of data that has alreadybeen collected in a nonexplicit manner the wealth of expert knowledge (in the form ofthe conceptual model) that senior soil mappers have accumulated

Focus in the future must also move in the direction of operability The utility of manyPSM approaches has been clearly demonstrated yet changes in how soils are mappedand perceived has been slow In the USA various members of the National CooperativeSoil Survey have called for a change in the philosophy of survey (from mappingdiscrete soil types to mapping continuous varying soil properties McSweeny et al1994) but little change has actually taken place To realize the potential of PSM in con-tributing to an overall change in standard soil mapping procedures practitionersworking in the field of PSM need to form working relationships with field soilscientists PSM methods need also to be presented in a manner that is comprehensibleto the soil science trained field mapper Predictive soil mapping is a relatively recent

phenomenon yet much progress has been made In the process of mapping soilknowledge will continue to advance

P Scull et al 193

References

Agbu PA Fehrenbacher DJ and Jansen IJ1990 Statistical comparison of SPOT spectralmaps with field soil maps Soil Science Society ofAmerica Journal 54 818ndash18

Barrett L 1999 Particulars in contextmaintaining a balance in soil geography Annalsof the Association of America Geographers 89707ndash13

Beckett PHT and Webster R 1971 Soilvariability a review Soils and Fertilizers 341ndash15

Bell JC Grigal DF and Bates PC 2000 A soil-terrain model for estimating spatial patterns ofsoil organic carbon In Wilson J and Gallant Jeditors Terrain analysis principles and applica-tions New York City NY John Wiley and Sons295ndash310

Birkeland PW 1999 Soil and geomorphologyThird edition New York NY Oxford UniversityPress

Breiman L Friedman JH Olshen RA andStone CJ 1984 Classification and regressiontrees Belmont CA Wadsworth

Brule FJ 1996 Fuzzy systems ndash a tutorialhttp newsgroup compai httpwwwquadralaycom (last accessed 30 August 2001)

Bunkin FV and Bunkin AF 2000 Lidarsounding of water soil and plants Atmosphericand Oceanic Optics 13 54ndash72

Buol SW Hole FD McCracken RJ andSouthard RJ 1997 Soil genesis and classifica-tion Ames IA Iowa State University Press

Burgess TM and Webster R 1980a Optimalinterpolation and isarithmic mapping of soilproperties the semi-variogram and punctualkriging Journal of Soil Science 31 315ndash31

ndashndashndashndash 1980b Optimal interpolation and isarithmicmapping of soil properties block kringingJournal of Soil Science 31 331ndash41

Burrough PA 1989 Fuzzy mathematicalmethods for soil survey and land evaluationJournal of Soil Science 40 477ndash92

Burrough PA and McDonnell RA 1998Principles of geographic information systems(Revised edition) Oxford Clarendon Press

Burrough PA Beckett PHT and Jarvis MG1971 The relation between cost and utility insoil survey Journal of Soil Science 22 368ndash81

Burrough PA Van Gaans PMF and

Hootsman R 1997 Continuous classificationin soil survey spatial correlation confusionand boundaries Geoderma 77 115ndash35

Cambell JB 1977 Variation of selectedproperties across a soil boundary Soil ScienceSociety of America Journal 41 578ndash82

Cambell JB and Edmonds WJ 1984 Themissing geographic dimension to soiltaxonomy Annals of the Association of AmericanGeographers 74 83ndash97

Castrignano A Giugliarini L Risaliti R andMartinelli N 2000 Study of spatial relation-ships among some soil physico-chemicalproperties of a field in central Italy using multi-variate geostatistics Geoderma 97 39ndash60

Cialella AT Dubayah R Lawrence W andLevine E 1997 Predicting soil drainage classusing remotely sensed and digital elevationdata Journal of Soil Science 62(2) 171ndash78

Clark RN and Swayze GA 1996 Evolution inimaging spectroscopy analysis and sensorsignal-to-noise an examination of how far wehave come Summaries of the sixth annual JPLairborne Earth science workshop 4ndash8 March 1996AVIRIS Workshop Vol 1 5

Clarke LA and Pregibon D 1992 Tree-basedmodels In Chambers J and Hastie J editorsStatistical models in S Pacific Grove Wadsworthand Brooks 377ndash419

Coleman TL Agbu PA and MontgomeryOL 1993 Spectral differentiation of surfacesoils and soil properties ndash is it possible fromspace platforms Soil Science 155 283ndash93

Cook SE Corner RJ Grealish G GesslerPE and Chartres CJ 1996 A rule-basedsystem to map soil properties Soil ScienceSociety of America Journal 60 1893ndash900

Cruickshank JG 1972 Soil geography New YorkNY John Wiley amp Sons

Csillag F Pasztor L Biehl LL 1993 Spectralband selection for the characterization ofsalinity status of soils Remote Sensing ofEnvironment 43 231ndash42

Dale MB McBratney AB and Russell JS1989 On the role of expert systems andnumerical taxonomy in soil classificationJournal of Soil Science 40 223ndash34

Davis JR 1993 Expert systems and environ-

174 Predictive soil mapping a review

II Background

Prediction of soil properties based on knowledge of the effect of environmentalvariables on soil formation has always been the basis for all soil mappingUnfortunately the traditional methods do not yield quantifiable soilndashlandscapeinformation that robustly describes actual soil variation In this section we describe soilvariation and explain why traditional methods for defining soil distribution inlandscapes are inadequate It is also our purpose to demonstrate that there is a logicalflow from these approaches into the new PSM techniques All approaches to soilmapping rest on our ability to use knowledge of the process of soil genesis to predictthe properties of soils at any point in the landscape

1 The nature of soil variability

The complex and highly variable nature of soil patterns in landscapes complicates thealready labour-intensive process of collecting and presenting soil survey data (Wrightand Wilson 1979) In 1941 Jenny listed three different definitions of soil before heconceded that lsquoit is problematic whether any definition of soil could be formulatedrsquo(Jenny 1941 2) In the half-century that has passed since Jenny wrote those words therehave been numerous attempts to define soil (reviewed by Birkeland 1999) Mosttextbook authors describe soil as extraordinarily complex For example McKnight(1993 336) defines soil as lsquoan infinitely varying mixture of weathered mineral particlesdecaying organic matter living organisms gases and liquid solutionsrsquo (emphasisadded) This definition illustrates how complex soil can be It further demonstrates thatthe soil landscape is continuous and is not composed of distinct individual soil typesThis point has been made repeatedly over the years (Simonson 1959 Webster andBeckett 1968 Cambell 1977 and Moore et al 1993)

A variety of soil genesis models have been proposed in order to account for the highvariability of soil and collectively they help further illustrate the difficulty of charac-terizing the soil landscape (for a review see Huggett 1975 Birkeland 1999) Threedistinctive approaches have been employed factor models (eg Jenny 1941) wherefactors affecting soil development are identified process models (eg Simonson 1959)where soil-forming processes are emphasized and energy models (eg Runge 1973)where the focus is upon process-driving mechanisms A host of hybrid models havealso been employed Still no real consensus exists today as to exactly how to model soildevelopment partially because of the recent emergence of pedology as an academicdiscipline (Johnson and Watson-Stegner 1987) Existing data collection methods do notyield adequate soil information in part because many of the processes that shape thesoil landscape are still poorly understood

2 How the soil survey has defined soil mapping

Traditional soil survey persists as the most popular form of soil mapping and inventoryand in many cases is the only manner in which the highly variable nature of the soillandscape is catalogued The method consists of three steps (Cook et al 1996) The firstis direct observation of ancillary data (aerial photography geology vegetation etc) and

soil profile characteristics In the second step the observations of soil attributes areincorporated into an implicit conceptual model that is used to infer soil variation Thethird step involves applying the conceptual model to the survey area to predict soilvariation at unobserved sites Usually less than 0001 of the survey area is actuallyobserved (Burrough et al 1971) a fact reflecting the high cost of field sampling Theconceptual model of soil variation is then transformed into a cartographic model thechoropleth map by drawing map unit boundaries on aerial photographs In effect pho-tographic scale determines the resolution of the soil map

This process has been severely criticized in the scientific literature for two reasonsFirst the conceptual model developed by the soil surveyor is primarily implicit beingconstructed in a heuristic manner This results in an excessive dependence upon tacitknowledge and as such incomplete information exists relative to the derivation of theultimate soil survey product (Hudson 1992) This aspect of soil survey is especiallyfrustrating because it fails to document most of the knowledge that the soil surveyoraccumulates during the expensive field mapping process In essence the soil survey isunfalsifiable and therefore untestable (Hewitt 1993) The final product of soil survey isa soil map that has unknown assumptions limitations and accuracy (Burrough et al1971 Dijkerman 1974) PSM techniques are similar in theory to soil survey (they bothuse knowledge of soilndashenvironment relations to make inferences) but the methodsemployed often yield quantitative expressions of soil variability with measured levelsof accuracy Clearly one direction of innovation in soil survey is to add objectivity tomodel development which will allow more explicit scientific communication (Gessler1996) The second major criticism of soil survey concerns the role of soil classification

3 Soil classification

The evolution of PSM techniques has been directly impacted and continues to beinfluenced by the process of soil classification This is especially true in the USAwherethe primary focus of the National Cooperative Soil Survey is to develop and map Soiltaxonomy (the official classification developed by the US Department of AgricultureSoil Survey Staff 1975) The system was influenced by nineteenth century biologicaltaxonomy and the practice of geological survey (Heuvelink and Webster 2001) Thepurpose of Soil taxonomy was to provide an objective manner to systematically classifysoil and was adopted at a time when soil information had to be abstracted to the levelof the modal profile (classified in Soil taxonomy) because it was impossible to catalogueand present the full amount of soil variability (Cambell and Edmonds 1984) In orderto map soil taxa the soil must be perceived as a spatial entity a lsquopedonrsquo (a term used inSoil taxonomy to refer to the smallest recognizable unit that can be called lsquoa soilrsquo) Inpractice this spatial perception of soil results in a map whose classes are homogenousunits with unknown variability and sharply defined boundaries (Burrough andMcDonnell 1998) Since the initial development of Soil taxonomy the perception of thesoil landscape has changed from a collection of individual soil types to a continuallyvarying mixture of soil components (Dmitriev 1983) Field observations have shownthat in concert with the environmental variables soil properties vary continuouslyacross landscapes exhibiting different and complex scales of variation (Simonson1959) Therefore soil distribution is not well represented by choropleth maps

P Scull et al 175

176 Predictive soil mapping a review

(McBratney 1992 Gessler et al 1995) While this change has not yet been recognized inSoil taxonomy PSM research has developed methods of soil inventory that moreaccurately describe the soil landscape Such methods are at odds with the traditionalapproaches because they often involve dismissing the concept of soil as a spatial entityRather the new methods focus on mapping continuously varying soil properties Theimplementation of these techniques has been difficult because of the entrenchment ofSoil taxonomy

III Components of predictive soil mapping

1 Geographic data models of soil variability and soil mapping

Traditionally soil maps have been digitized to fulfil the need for soil data within GIS-based environmental modelling research This information from the paper mapdigitized for the computer is fraught with the same problems of the original choroplethmaps ndash assumed homogeneous units with unknown variability and sharply definedboundaries GIS allows for a more robust characterization of spatial variability (relativeto the cartographic generalizations of the past) by allowing information to be analysedand stored using a variety of data models As defined in the GIS literature datamodelling is the process of discretizing spatial variation which entails abstracting gen-eralizing or approximating geographic reality (defined as empirically verifiable factsabout the real world) Unfortunately data modelling is often confused with issues ofdata structure and limited by software selection (Kemp 1992) The process is of crucialimportance because it controls the manner in which the data can be processed oranalysed (Goodchild 1994) as well as the view of the data the end user ultimatelyreceives (Goodchild 1992a Kemp 1992) Some data models are more accurate thanothers at portraying geographical reality (Goodchild 1992b) In the case of soil howwell do the data including the data model represent the highly variable continuousnature of soil How can digital computers be used to manage spatial data to bestrepresent the soil landscape

The choice of data models is partially dependent upon how the soil is perceived ingeographic space Objects can be thought of as existing as independent entities inempty space (object view) or as one of an infinite set of tuples (the foundation ofgeographic information ndash xyz where z is a measured value and xy are its location inspace) approximated by regions and segments (field view) (Goodchild 1994) The fieldview better represents continuous surfaces such as the soil landscape but soil datahave traditionally been modelled using the feature model of geographic space thechoropleth map Soil data have probably been managed this way because of theinfluence of Soil taxonomy which defines individual soil lsquotypesrsquo and because soil datacollection began during a time when few alternatives existed

Using PSM techniques a fundamental change in the soil data model from thechoropleth map to the raster grid allows better characterization of actual soilndashlandscapevariability A raster data model accommodates a lsquofield viewrsquo representation of thelandscape and is defined as a regular rectangular array of cells with some aggregatevalue of the field recorded for each cell (Goodchild 1994) The resolution of the datastored in this format is a function of the grid cell size which can be made small enoughto simulate continuous variation at the landscape scale The raster has become the most

widely used data model for PSM and is also routinely used to manage other environ-mental information such as elevation (DEMs) and remotely sensed data Spatialanalysis and integration with other types of raster-based environmental data can beeasily performed with soil data stored using a raster data structure (Burrough andMcDonnell 1998)

2 Digital terrain modelling

Terrain analysis quantifies the relief component of models characterizing soilformation Soil development and its associated profile characteristics often occurs inresponse to the way in which water moves through and over the landscape which iscontrolled by local relief Accordingly terrain analysis will be most useful in environ-ments where topographic shape is strongly related to the processes driving soilformation (McKensie et al 2000) Digital terrain modelling is a technique for derivingspatially explicit quantitative measures of the shape character of topography (Weibeland Heller 1991 Wilson and Gallant 2000) The spatial distribution of the resultingterrain attributes (characterizing local water flow paths) can also capture the spatialvariability of soil attributes Moore ID et al (1991) reviewed the analysis of digitalelevation data (including DEMs) for hydrological geomorphological and biologicalapplications They provided a table that summarized the significance and physicalmeaning of various terrain attributes to landscape processes Building on their workmany authors have used terrain attributes derived from digital elevation models(DEMs) as explanatory variables in predictive soil models (Odeh et al 1991 Moore etal 1993 Gessler et al 1995 Skidmore et al 1996 and others) Methods used to deriveterrain attributes have been greatly refined over the last 15 years and future satellitesaiding in the development of more accurate DEMs will make terrain analysis anincreasingly important component of predictive soils mapping (Moore et al 1993Mackay and Band 1998) Several recent review articles have been specifically devotedto the role of terrain analysis in soil mapping (Ventura and Irvin 1996 Irvin et al 1996McKensie et al 2000)

3 Remote sensing

Remote sensing data are an important component of PSM because they provide aspatially contiguous quantitative measure of surface reflectance which is related tosome soil properties (Agbu et al 1990) Both physical factors (eg particle size andsurface roughness) and chemical factors (eg surface mineralogy organic mattercontent and moisture) control soil spectral reflectance (Irons et al 1989) Surfacemineralogy can be derived by wavelength specific charge transfer and crystal fieldabsorptions associated with the presence of iron and iron-oxides (Fe2+ and Fe3+) andvibrational absorptions associated with hydroxyl bonds in clays adsorbed water andthe carbonate ion (Goetz 1989 Irons et al 1989) The presence and strength of theseabsorption features can be used to identify and quantify concentrations of mixed suitesof minerals in soil (Johnson et al 1983 Shipman and Adams 1987) Organic matterparticle size and moisture content in contrast influence soil reflectance primarilythrough a change in average surface reflectance and produce only broad spectral

P Scull et al 177

178 Predictive soil mapping a review

expression (Irons et al 1989) Adecrease in particle size tends to increase surface albedoand decrease spectral contrast of absorption features while an increase in organicmatter or soil moisture decreases average reflectance or albedo

Numerous studies have shown the potential benefits of using remote sensing for soilidentification and mapping Comprehensive surveys of soil spectral reflectance includestudies by Stoner et al (1980) Henderson et al (1992) and Csillag et al (1993) Remotesensing studies based on broad-band sensors such as Landsat TM include Agbu et al(1990) Coleman et al (1993) Seyler et al (1998) and Oliveira (2000) Traditionallyremote sensing has been used to classify soil units through photo-interpretation ordigital image processing Combining remotely sensed information with ancillaryinformation such as thematic maps or vegetation cover can yield significant improve-ments (Wilcox et al 1994 Cialella et al 1997 Wanchang et al 2000)

Recent developments in hyperspectral remote sensing offer the potential of signifi-cantly improving data input to predictive soil models Hyperspectral sensors such asthe Airborne VisibleInfrared Imaging Spectrometer (AVIRIS) measure a contiguousspectrum in the visible and NIR and thereby better characterize atmospheric andsurface properties (Goetz et al 1985) The large number of spectral bands permits directidentification of minerals in surface soils For example Clark and Swayze (1996)mapped over 30 minerals using AVIRIS at Cuprite Nevada Palacios-Orueta and Ustin(1996) showed that enhanced spectral information was suitable for discriminating evensubtle spectral changes associated with differences in organic matter and iron contentOther examples of the application of AVIRIS to aid soil mapping include Palacios-Orueta et al (1998) Okin et al (1998) and Roberts et al (1998)

Sensors that operate in the microwave portion of the electromagnetic spectrum havealso shown promise in soil mapping research Microwave sensing can be broadlydivided into active (eg radar) systems and passive systems and are capable ofpenetrating the atmosphere under virtually all conditions offering a significantadvantage over visible and near-infrared spectroscopy (for a general overview ofmicrowave sensing see Lillesand and Kiefer 1994 chapter 8) Synthetic aperature radar(SAR) is one example of an active system SAR has been used to aid soil propertymapping such as soil salinity (Metternicht 1998) and soil moisture (Engman andChauhan 1995 Narayanan and Hirsave 2001) Active radar systems can also bedesigned to collect data at varying look angles providing the opportunity for theacquisition of stereo radar images Such images can be used to produce high resolutionand extremely accurate DEMs (eg Fang 2000) A similar active sensing system isLiDAR (light detection and ranging) which uses pulses of laser light rather thanmicrowave energy to illuminate the surface (see Bunkin and Bunkin 2000 for a reviewof applications to soil mapping research) While passive microwave systems haveseemed to receive less attention in the literature a few examples of soil mapping appli-cations can be found (see Kleshchenko et al 2000 and Laymon et al 2001) Regardlessof the type of system remote sensing data and derived products are potentially usefulexplanatory variables in predictive soil mapping models

4 Fuzzy logic

Fuzzy set theory or fuzzy logic provides an alternative conceptual paradigm withinPSM research The use of this theory has increased greatly in the last few years making

it an important component of PSM Fuzzy logic is an alternative to Boolean logic thatattempts to recognize the concept of partial truth (Brule 1996) Dr Lotfi Zadeh (1965)introduced the concept and accompanying mathematics in his seminal work lsquoFuzzysetsrsquo The theory permits partial class membership in contrast to traditional set theorywhere set memberships are crisp and binary (ie a soil sample is either completelyType A or it is not at all Type A) Central to the fuzzy concept is the idea that objects innature rarely fit exactly the classification types to which they are assigned (Zadeh1965) Rather they show varying signs of similarity to multiple classes (ie an observedsoil pedon often resembles more than one of the defined soil series within the area) Byusing fuzzy membership values (ranging from 0 nonmembership to 1 totalmembership) within predictive soil models to express degrees of similarity generaliza-tion problems associated with classification schemes (filtering of information) areminimized and the complex nature of soil data is allowed to propagate through themodelling process Similarity values between 0 and 1 are not comparable to proportionsand need not add up to 1 Within Boolean logic probability statements refer to thelikelihood of an outcome the soil sample is either one series or another With fuzzylogic a given sample is not definitively a member of the subset of any one particularseries Fuzzy logic is especially useful in soil research because of the continuous andcomplex nature of the soil landscape It serves as an important alternative to thesubjective rigidity imposed on soils data by Soil taxonomy Several recent articlesprovide a thorough review of the use of fuzzy sets in soil science (Burrough 1989McBratney and Odeh 1997 Burrough et al 1997 De Gruijter et al 1997)

Within PSM research two different approaches to creating continuous classes usingfuzzy logic exist The first is based on the fuzzy-k-means classifier which partitionsobservations in multivariate space into natural classes This approach is similar tocluster analysis and numerical taxonomy but the resulting classes are continuous witheach observation assigned a fuzzy membership value that characterizes its degree ofsimilarity to each individual class The concept has been integrated into geostatisticalmethods and will be discussed in more detail below (see Section IV1) The secondapproach is known as the Semantic Import model (SI) and is used in situations whenclassification schemes are pre-defined and class limits are relatively well understoodThe SI model is commonly used in concert with expert knowledge and will bediscussed in the expert systems section (see Section IV4)

IV Recent advances in predictive soil mapping

Within the last decade many authors have sought to model the soil landscape using avariety of methods Literature in this field could be summarized many different waysbut we concentrate on the literature that directly addresses the goals of predictive soilmapping stated in the introduction (see Table 1) Therefore we will review research thatattempts to exploit the relationship between quantifiable landscape indices and soilcharacter in order to model the soil landscape in a more continuous and thereforerealistic manner

The research reviewed here is distinguished from decades of previous researchdocumenting the correlation between landscape position and soil attributes (reviewedby Hall and Olsen 1991) That body of research is informative but not useful for

P Scull et al 179

180 Predictive soil mapping a review

Table 1

Selected

recen

t literature on pred

ictive

soil m

apping an

d m

apping

(cited in

this article) d

escribing the mod

ellin

g metho

dused the

dep

enden

t variables used a

nd the

env

iron

men

tal v

ariables (ex

plan

atory) used in

the

mod

els

Stud

yMod

ellin

g metho

dDep

ende

nt variables

Environm

ental variab

les

Goa

ls atta

ined

a

Bell et

al 20

00Line

ar and

exp

onen

tial

Total soil orga

nic

Slop

e curvature aspe

ct

ECx

regression

carbon

hillslope

position

Burge

ss and

Web

ster

Punc

tual and

block

Na co

nten

t co

ver

Non

eC

1980

ab

kriging

loam

thickn

ess

ston

e co

nten

t

Burroug

h 198

9Fu

zzy mathe

matical

ndashndash

rpmetho

ds

Burroug

h et

al 19

97Con

tinuo

us classificatio

nndash

ndashrp

Castrigna

no eta

l 200

0Factorial kriging

CEC

pH N

P K

Non

e ex

plicitly used

Cx

Na

Cialella

eta

l 199

7Classifica

tion tree

Drainag

e class

Elev aspe

ct NDVI

Ec

Coo

k et

al 19

96Baysian

rule-ba

sed metho

dsOrgan

ic m

atter

Slop

e aspec

t wetne

ssEcX

inde

x

Dale et

al 19

89 (a review

)Ex

pert systems

ndashndash

ndash

Ellis 19

96Decision tree

ana

lysis

Soil erosion class

Slop

e aspec

t wetne

ssEc

neutral ne

tworks

inde

x flow le

ngth and

accu

mulation Lan

dsat

TM tree

cov

er

Gessler 1

996

A large

variety of statistic

Field an

d labo

ratory

A variety of digital

EC

metho

dsco

llected

phy

sica

len

vironm

ental da

tach

emical and

morph

olog

ical soil

prop

ertie

s

Gessler eta

l 199

5Line

ar and

logit reg

ression

A horizon

and

solum

Curvature CTI topo

EC

depth E horizon

positio

npresen

ce

Goo

vaerts 19

92Factorial kriging

Total carbon

Non

eC

P Scull et al 181pH

N CEC

extractab

lecatio

ns (K Ca M

g)Hartemink et

al 20

01ndash

ndashndash

rp

Heu

velin

k an

d W

ebster 20

01ndash

ndashndash

rp

Hew

itt 19

93ndash

ndashndash

rp

Indo

rante et

al 19

96ndash

ndashndash

rp

King et

al 19

99Lo

gistic regression

Presen

ceabsen

ceSlop

e aspec

t po

t solar

Ec

Non

calc c

lay-loam

energy

Kno

tters eta

l 199

5Kriging

co

-kriging

So

ft layer de

pth

Hillslop

e po

sitio

nCx

regression

kriging

Laga

cherie and

Holmes 19

97Classifica

tion tree

Map

ping

unit

Geo

logy variou

s topo

Ec

indices

Laslett e

tal 198

7Kriging

splin

es tren

dpH

Non

ec

surfac

e nea

rest neigh

bor

McB

ratney 1

992

ndashndash

ndashrp

McB

ratney

eta

l 199

1Block kriging

Clay co

nten

tNon

eC

McB

ratney

eta

l 200

0ndash

ndashndash

rp

McB

ratney

and

Ode

h 199

7Fu

zzy sets in

soil scienc

endash

ndashRp

McB

ratney

and

de Gruijter

Fuzz

y-k-mea

ns w

ithFu

zzy classes

Field co

llected

phy

sical

C19

92ex

tragrade

sch

emical and

morph

olog

ical soil

prop

ertie

s

McC

rack

en and

Cate 198

6Artificial intellig

ence

ndash

ndashrp

expe

rt systems

McK

ensie an

d Austin

19

93Gen

eralized

line

ar m

odels

Clay co

nten

t CEC

Slop

e relief land

form

ec

(logit)

pH EC

COLE

slop

e po

sitio

nbu

lk den

sity and

othe

rs

Moo

re eta

l 199

3Line

ar reg

ression

A horizon

dep

th O

MSlop

e w

etne

ss and

strea

mEC

and P co

nten

t pH

power in

dices aspect

curvature

182 Predictive soil mapping a reviewTa

ble 1

Con

tinu

ed

Stud

yMod

ellin

g metho

dDep

ende

nt variables

Environm

ental v

ariables

Goa

ls atta

ined

a

McK

ensie an

d Ryan 199

9Reg

ression tree and

linea

rSo

lum dep

th P an

dElevation slope

CE

regression

N con

tent

curvature CTI

contribu

ting area do

wn-

slop

e mea

ns for slope

clim

ate da

ta Presco

ttInde

x G

amma

Rad

iometry La

ndsat T

M

and Geo

logy unit

Ode

h et

al 19

92ab

Fuzz

y-c-mea

ns and

Fuzzy classes

Field-co

llected

phy

sica

lC

kriging

chem

ical and

morph

olog

ical soil

prop

ertie

s

Ode

h et

al 19

94

Reg

ression kriging

co

-So

lum dep

th de

pth

Slop

e aspect cu

rvature

eC

1995

kriging regression

to bed

rock gravel

kriging

and clay con

tent

Skidmore et

al 19

91Bayesian expe

rt system

Soil land

scap

e un

itVeg type

wetne

ss in

dex

EcX

1996

grad

ient terrain po

sitio

n

Voltz and

Web

ster 19

90Kriging

cu

bic splin

eClay co

nten

tNon

eC

Web

ster 1

994

Dev

elop

men

t of

ndashndash

rppe

dometrics

Zhu

199

7ab

Fuzz

y logic expe

rt system

Soil series A

horizon

Elev pm aspe

ct c

anop

yECX

Zhu

and

Ban

d 199

4(SoL

IM)

depth in

dividu

alco

verage

grad

ient

Zhu

eta

l 199

7series m

aps

curvature

Not

es

a Letters refer to the de

gree

to w

hich

the

goa

ls of PS

M defined

in the

introdu

ction are achiev

ed

Soilndash

e nvironm

ent relatio

ns utilized

(letter E) be

tter represen

tatio

n of soil c o

ntinuity (C

) an

d ex

pert kno

wledg

e utilize

d (X) c

orrespon

d to goa

ls 1 2 and

3respec

tively from

the

introdu

ction and

cap

ital lette

rs (E C X) ind

icate the metho

d is relatively more successful th

an th

ose metho

ds den

oted

by lower

case le

tters (e c x

) rp ind

icates rev

iew pap

ers

P Scull et al 183

predictive mapping because landscape position is never quantified (rather positionwas often qualitatively defined eg toe-slope) and thus the documented relationshipscannot be generalized using environmental data and digital elevation models to predictsoil character at unvisited sites

Table 1 documents modelling methods model variables and the extent to which thereferenced research satisfies the previously defined goals of PSM (lsquoGoals attainedrsquocolumn) Soil-environment relations utilized (letter E) better representation of soilcontinuity (C) and expert knowledge utilized (X) correspond to goals 1 2 and 3respectively from the introduction and capital letters (E C X) indicate the method isrelatively more successful than those methods denoted by lower case letters (e c x)Citations with no letters present within the lsquoGoals attainedrsquo column do not address theaforementioned goals For example the Cialella et al (1997) received a rating of lsquoEcrsquomeaning that the methods employed successfully utilized environmentalndashsoil characterrelations (E) and somewhat successfully presented a better method of representing soilcontinuity (c) The ratings are provided simply to help organize the literature that wasreviewed Review papers are included within the table denoted by lsquorprsquo Geostatisticalmethods are not included in the table because they have been comprehensivelysurveyed elsewhere (Odeh et al 1994 Burrough et al 1997 McBratney et al 2000Heuvelink and Webster 2001) and because the objectives and assumptions of geostatis-tical methods differ slightly from other PSM research We briefly outline thesedifferences in the following section

1 Geostatistical methods

Geostatistics are a subset of traditional statistics that deal primarily with spatial dataand account for spatial autocorrelation using kriging as the spatial interpolator Theconcept is based upon the theory of regionalized variables which was mainlydeveloped by Matheron (1963) and Krige (1963) Kriging is a form of weighted localaveraging that uses a measure of spatial dependence the variogram to determine theweights applied to the data when computing the averages Geostatistical methods havebeen used in predictive soil mapping research to spatially interpolate soil propertyvalues at unmeasured sites from field-collected data

Burgess and Webster (1980a b) were the first to introduce ordinary kriging to the soilcommunity and since that time an enormous amount of work has been published Forexample ordinary kriging has been used to interpolate many different soil propertiesincluding pollution trace element deficiencies salinity and fertility (Heuvelink andWebster 2001) Ordinary kriging has been criticized for a variety of reasons Forexample Laslett et al (1987) reported that several authors had criticized geostatisticsbecause kriging is a global rather than local technique failing to take into accountknowledge of soil materials and processes Other authors have criticized geostatisticsbecause they are excessively data dependent requiring a large number of closelyspaced data points (Zhu 1997a) As Webster and Oliver (1992) suggest in excess of ahundred samples may be needed to use geostatistics at the field scale because of highspatial variability of soil in some areas Geostatistics also assume spatial autocorrela-tion which sometimes may be a poor assumption in complex terrain where abruptchanges in soil-forming factors occur (McBratney et al 2000) Ordinary kriging by itself

184 Predictive soil mapping a review

does not satisfy two of the three goals of PSM presented in the introduction ndash it does notadequately incorporate expert knowledge and it does not exploit the relationshipbetween environmental variables and soil properties

Ordinary kriging has been modified in a variety of ways to better incorporateancillary data and known soilndashlandscape relationships Block kriging involvesdetermining estimates over meaningful areas rather than at specific points (Burgess andWebster 1980a McBratney et al 1991) Using this method a study area can be stratifiedinto different regions that are reflective of the pedogenetic processes at work In orderto accommodate a trend within a dependent soil variable universal kriging has beenused (Webster 1994) Kriging with external drift is similar to universal kriging but ituses an ancillary variable to represent the trend (McBratney et al 2000) Co-krigingtakes advantage of correlation that may exist between the variable of interest and othermore easily measured variables (Odeh et al 1995) Regression kriging involvesspatially interpolating the residuals from a non-spatial model by kriging and addingthe result to the prediction obtained from that model (Goovaerts 1997 Castrignano etal 2000) Factorial kriging is another method to integrate multivariate data into thestandard kriging routine to extrapolate soil data (Goovaerts 1992) Many authors havecompared these various methods (Laslett et al 1987 Voltz and Webster 1990 Odeh etal 1994 Knotters et al 1995)

Fuzzy logic has been used with geostatistics by various authors to produce new kindsof fuzzy soil maps with continuous classes (Burrough 1989 McBratney and DeGruijter1992 Odeh et al 1992a and reviewed by McBratney and Odeh 1997) The processentails kriging the matrix of membership values determined by the fuzzy k-meansclassifier resulting in a continuous soil surface where individual locations are allowedto belong to more than one class and no rigid boundaries are designated to separate thesoil into discrete units or entities The results of such analysis can be used to assess thepedologic process validity of soil taxonomy by determining whether soils grouptogether into classes that are similar to taxonomic types Fuzzy classes wouldpresumably reflect the main pedologic features within a given area In this sense theclassification is quantitative whereas soil taxonomy is rooted in qualitative discrimina-tion A comparison between the two could provide insightful and help assist soiltaxonomy in making a classification that is more indicative of underlying soil processes

Geostatistics in soil research were originally introduced to quantitatively assess soilvariability within soil mapping units (McBratney et al 1991) in response to criticisms inthe early 1970s that soil unit composition was not well quantified (Beckett and Webster1971) In this regard geostatistics have been very useful having served well the originalgoals set forth by Burgess and Webster (1980b) when they drew kriging to the attentionof soil scientists as a means of spatial prediction At the field scale soil variation islargely due to the effect that topography has on soil genesis Geostatistics have beensuccessfully applied in such environments by using terrain attributes as ancillary datawithin many of the kriging routines described above Such quantitative within-unitvariability of soil properties is very useful in the field of precision agriculture and othersituations (eg pollutants) where very detailed soil attribute information is needed atthe field scale (Heuvelink and Webster 2001)

However geostatistics have not been applied in a wide variety of environments or atlarger scales In order to be successfully applied in different environments geostatisticswill likely require a different suite of ancillary data For example remote sensing data

P Scull et al 185

could be used in arid regions where soil toposequences are less well expressed Atlarger scales of prediction selection of different sets of ancillary variables is requiredbecause different processes define soil character at different scales The most obviousexample is that of climate which may control soil distribution at large scales(continental) but has little explanatory power at the field level Regardless of whetheradequate ancillary data exist the amount of data required to use geostatistics forlandscape-level prediction would be extremely difficult and costly to collect given thestrict sampling protocol required to characterize spatial dependence It is also unclearat what landscape scale soils exhibit spatial autocorrelation

Geostatistical approaches do provide a means of creating continuous soil attributesurfaces to better represent soil continuity (Goal 2) and they can be used to exploit therelationship between environmental variables and soil properties in order to moreefficiently collect soil data (Goal 1) However they do not sufficiently utilize expertknowledge (Goal 3) as no attempt has been made in geostatistical approaches todirectly integrate expert knowledge Fundamentally kriging is a process of interpola-tion designed to predict attribute values in between locations of measured samples Inthis sense geostatistics represent a middle ground between pure interpolation (egnearest neighbour type classifier) in which only measured points for the variable ofinterest are used to determine unknown values and other predictive models thatprimarily use soilndashenvironment correlation to create predictive maps

2 Statistical methods

Statistical methods can be used to exploit the relationship between quantifiablelandscape indices and soil properties to create predictive soil maps For exampleMcKensie and Austin (1993) used a regression to account for a large percentage ofvariation for many soil characteristics (A horizon clay content CEC EC pH bulkdensity and COLE B horizon clay content CEC ESP EC pH bulk density and COLE)using a variety of predictor variables (slope presence or absence of impeding layerrelief landform topographic position) Their results confirm the hypothesis of MooreID et al (1991) that soil character is related to quantifiable landscape indices Howevertheir methods do not provide inference of soil properties at unmeasured sites frommapped environment data because the topographic variables were measured in thefield Linear regression has also been used with terrain variables derived from a 15-mDEM in northeastern Colorado to predict soil attributes (organic matter contentextractable phosphorous pH and texture) at unvisited sites (Moore et al 1993) In thatparticular study 50 of the variance of A-horizon thickness was explained by slope andthe wetness index Gessler et al (1995) also used regression to model A-horizonthickness from topographic variables in southeastern Australia (plan curvature andwetness index r2 = 063 P = 0001) They modeled solum depth and used logisticregression to model E horizon presenceabsence Elsewhere logistic regression hasbeen used to model the presenceabsence of noncalcareous clay loam horizon in centralFrance using terrain attributes from a 20-m DEM (King et al 1999) Exponentialregression has been used to model soil organic carbon using terrain variables (Bell et al2000) in glacial outwash soils in east-central Minnesota Generalized additive models(GAM) have been used less frequently in PSM research Gessler (1996) used a GAM

186 Predictive soil mapping a review

model to predict total soil carbon A horizon depth and solum depth using a variety ofenvironmental predictors (slope elevation wetness index mean annual temperatureprecipitation and radiation)

This small body of research opened the door to more complex methods by demon-strating the existence of quantifiable relationships These authors were able to producesoil attribute maps using raster data models whose scale was dependent upon the gridcell resolution of the environmental data They were successful at exploiting the rela-tionship between quantifiable topographic attributes and soil profile character (Goal 1)The continuous soil attribute surfaces they produce also better represent soil continuitythan the choropleth soil maps produced by traditional soil survey (Goal 2) Howeverthe bulk of these methods (excepting GAMs) are limited by their assumed linear rela-tionship between soil and topographic attributes their assumptions of normallydistributed data and their high data requirements Standard statistical procedures arealso not flexible enough to allow robust integration with a variety of potential datasources such as expert knowledge (Goal 1)

Statistical methods do demonstrate in a quantitative manner that terrain analysis canbe used to predict soil attributes in relatively small areas with homogeneous parentmaterial A large proportion of the research using statistical methods was conducted insemi-arid landscapes at small scales (the largest study area of the entire group was~2000 ha) Obviously for statistical approaches to be most effective they need to bemore universal As such they need to be tested andor developed at larger scales andin more diverse landscapes

3 Decision tree analysis (DTA)

The use of decision tree analysis is just beginning to be explored in predictive soilmapping research although it has been used successfully in the related field ofpredictive vegetation mapping since the early 1990s (Lees and Ritman 1991 MooreDM et al 1991 Franklin 1998) DTA is a form of divisive classification The process oftree modelling involves successively partitioning data (called recursive partitioning inthe tree modelling literature) into increasingly homogeneous subsets which once thepartitioning has ceased are called terminal nodes (Lees and Ritman 1991) Splits orrules defining how to partition the data are selected based on information statistics thatdefine how well the split decreases impurity within the data set (Clark and Pregibon1992) Splits are based on threshold values of an explanatory variable selected bycomparing the increase in resulting purity of node membership for all possiblethresholds and variables The process is iterative growing from the root node (thecomplete data set) to the terminal nodes in a dendritic fashion (Friedl and Brodley1997) Once the tree has been constructed (or grown) it encodes a set of decision rulesthat describe the data partitioning process These rules can be used to classify or predictother data sets (Moore DM et al 1991) Pruning the tree is often necessary to preventthe tree from being overfit to the sample data and to reduce tree complexity Pruningentails combining pairs of terminal nodes into single nodes and can be accomplishedusing cross-validation which yields an initial indication of how large a tree makesrobust predictions (Safavian and Norvig 1991) Cross-validation involves systematical-ly removing portions of the data set and running the remaining sample through the tree

in an iterative manner eventually yielding estimates of the misclassification rates foreach class each node and the whole model (Breiman et al 1984) In this mannerdifferent sized trees can be compared in terms of parsimony

The term DTA is used to collectively refer to all types of tree-based modelling (theword lsquodecisionrsquo is used because it is descriptive indicating that the analysis eventuallyleads to a set of decision rules defining data partitions) The term should be distin-guished from classification tree analysis because the latter refers specifically to DTAwhere the response variable is categorical The term CART (classification andregression trees ndash Breiman et al 1984) is sometimes used but strictly speaking refers tospecific software Friedl and Brodley (1997) provide a review of the decision treealgorithms They divide types of DTA into two classes (1) homogeneous decision treesfor which a single algorithm is used to estimate each split (eg CART) and (2) hybriddecision trees (HDT) for which different splitting methods can be used at differentpoints in the tree (eg Quinlan 1993) They further divide homogeneous decision treesinto univariate (UDT) where single features of the input data define splits and multi-variate decision trees (MDT) where multiple features of the input data can define splitsAccording to this naming convention no distinction is made between methods withdifferent types of response variables although all methods can be used with bothcategorical and continuous response variables In a comparison of these various typesof algorithms on a variety of data sets Friedl and Brodley (1997) found that HDT hadthe highest classification accuracy

The overall aim of DTA is to design a set of predictive rules (eg if geology type Athen soil type B) developed from training data which can then be applied to ageographic data base to predict the value of a response variable (Michaelsen et al 1994)Therefore DTA explicitly uses soilndashlandscape correlation in model development (Goal2) The technique appears promising in soil research but needs to be further tested asnot many of the above types of DTA have been tested in the PSM literature in fact onlyunivariate approaches have been employed For example Lagacherie and Holmes(1997) successfully used univariate DTA to model a categorical response variable soiltype within a training set and then assuming that the training set was representativeextrapolated the model to a much larger region Their work is interesting becausealthough not a single sample came from the area they eventually mapped they wereable to produce a soil map that was more accurate (74 versus 69) than the existingmap produced from traditional methods Cialella et al (1997) also used univariate DTAto predict soil drainage class from a variety of terrain attributes and remotely senseddata They predicted soil drainage class with an average of 78 accuracy ndash impressivegiven that the variation accounted for by a typical soil survey ranges from about halfthe total variance for physical attributes to less than one-tenth for some soil chemicalattributes (Gessler et al 1995)

DTA has been compared with other approaches by several authors In the applicationof erosion modelling DTA results were similar to Artificial Neural Networks (Ellis1996) Both methods achieved high training accuracy (as measured by the KappaStatistic) but in terms of prediction accuracy both methods performed poorly Gessleret al (1995) compared DTA with generalized linear models (GLM) and generalizedadditive models (GAM) to predict A-horizon thickness and concluded that GLM waspreferable to both DTA and GAM McKensie and Ryan (1999) compared regressiontrees and standard linear regression to predict soil properties (total solum depth soil

P Scull et al 187

188 Predictive soil mapping a review

phophorus and soil carbon) using a large variety of predictors (elevation slopecurvature wetness index climate variables geology unit Landsat TM data andGamma radiometric data) and found that different methods work best for differentproperties and overall success hinged on the strength of the relationship between soiland environmental variables Their study of a 50 000 ha forested landscape in southernNew South Wales Australia is one of the few studies at such a small map scale (largearea) or in forested landscape The most extensive comparison of techniques involvingDTA was conducted by McBratney et al (2000) They compared regression techniques(GLM GAM DTA-regression tree) geostatistical techniques (kriging and heterotopicco-kriging) and a hybrid technique (regression kriging) DTA was found to be thepoorest performing of the regression techniques because of the unrealistic predictionsurface generated by the DTA model DTA has also been criticized by other authorsbecause of the stepped prediction surface (Gessler 1996) This phenomenon can beespecially apparent in situations where predictor variables have different resolutionsThe extent to which DTA yields a better spatial representation of soil continuity is afunction of the scale and type of predictor variables used and therefore varies withindividual models

DTA is gaining widespread popularity as a means to develop prediction rules thatcan be rapidly and repeatedly evaluated (Cialella et al 1997 Franklin et al 2000) DTAprovides the following advantages over standard statistical techniques (1) it is easier tointerpret when explanatory variables are both nominal and continuous (2) it isinvariant to monotone re-expressions (transformations) of predictor variables (3) itdeals more satisfactorily with missing data values and outliers (4) it is more adept atcapturing nonadditive and nonlinear behaviour (5) it doesnrsquot make any assumptionsabout data distribution and (6) it is easily updateable as more data are collected(Moore DM et al 1991) The DTA model framework is especially appealing becauseof its capability to integrate a wide range of data sets as explanatory variables

DTA offers a unique opportunity for interaction between soil experts and soilmodellers because the output of the model is a set of rules that can be pedologicallyinterpreted by the soil expert In this sense expert knowledge is used in an implicitmanner in DTA (somewhat effectively achieving Goal 3) While these rules can often beexceedingly complex at minimum the expert can decide whether initial splits makesense given their understanding of the landscape The success of DTA results oftenhinges on the ability of the modeller to make key decisions during the model buildingprocess there is unfortunately no definitive way to determine the most optimal treeThe aid of the soil expert can potentially elucidate this problem

4 Expert systems

A variety of expert system approaches to PSM have been developed to utilize expertknowledge The purpose of such methods is to exploit the information the soil surveyoraccumulates while working in the field by integrating such knowledge into thepredictive model (McCracken and Cate 1986) Unlike the majority of the researchreviewed thus far the dependent variable in many expert systems models is often soiltaxa or mapping unit This apparent disadvantage of expert systems (using classifica-tion to characterize soil continuity) does make them easier to integrate into traditional

soil survey In addition several authors have developed methods to developcontinuous soil property maps from the output of expert system models designed topredict soil unit occurrence

Expert systems are composed of data (information on spatial environmentalvariables eg topography climate etc) a knowledge base (rules and facts related tosoil variation supplied from the soil surveyor) and an inference engine (whichcombines data and the knowledge base to infer logically valid conclusions) (Skidmoreet al 1996) Expert systems differ from conventional models in two ways (1) they storeand manipulate qualitative information (allowing them access to information thatcannot normally be used in other modelling frameworks) and (2) they are structuredas meta-models (the knowledge is separated from the model) (Davis 1993) This allowsthe model to selectively choose which information is relevant at various stages of themodelling process and it allows for information to be easily updated Davis (1993)reviews the application of expert systems to environmental modelling research conclu-ding that the technique is becoming more widely accepted He further states that theapplication of expert systems is constrained by an absence of fundamental knowledgefor rule generation a problem that would appear less relevant to soil mapping giventhe amount of untapped expert knowledge accumulated by the soil surveyor

The first mention of the use of expert systems in pedology was in a paper presenta-tion at the Northeast Committee Soil Survey Conference 1984 (Flach 1985) In herpaper Flach (1985) hinted that recent developments in computer science especiallyexpert systems and artificial intelligence could make modelling a practical mappingtool for soil scientists in the near future A year later McCracken and Cate (1986) hopedto encourage soil scientists to explore expert systems and its application to soil sciencethrough an optimistic article they wrote on the potential use of expert systemsHowever little research was actually conducted in the late 1980s and expert systemshad not yet begun to fulfil the lofty goals set forth by early practitioners (Dale et al1989)

In the early 1990s expert systems approaches to predictive soil mapping began toappear in the literature Skidmore et al (1991) used a Bayesian expert system to mapforest soil into different classes and their results compared favourably with availablesoil maps and actual field-collected data Their methods successfully incorporatedsurveyor knowledge and remotely sensed and digital terrain attributes but failed tobetter characterize continuous soilndashlandscape variability because their final productwas a choropleth map Skidmore et al (1996) revisited their earlier research andassessed the mapping accuracy of their results They found that the soil map producedby the expert system achieved an overall accuracy of 698 (sample size = 53) while themap derived from conventional methods had an accuracy of 736 In addition toknowledge provided by a soil scientist Skidmore et alrsquos (1996) expert system usedvegetation derived for aerial photography and topographic variables derived from a 10-m DEM (wetness index topographic position and slope)

Using an expert system Cook et al (1996) successfully produced a continuous soilproperty map for organic matter content using wetness index aspect and slope asexplanatory variables Their methods were somewhat inefficient because they requireda separate expert system for each soil property of interest However their researchrepresents the only example of expert systems used to directly predict the spatial dis-tribution of a soil property All of the other examples of expert systems in the literature

P Scull et al 189

190 Predictive soil mapping a review

deal with soil type or class The use of expert systems to map soil properties needs tobe explored further

The expert systems discussed thus far have all used Boolean logic within theirtheoretical framework whereby an observation can belong to one and only one classand the soil properties of that class are assigned to the observation With thedevelopment of fuzzy logic and the semantic import model (the second fuzzy logicapproach to creating continuous classes) the opportunity exists to develop fuzzy logic-based expert systems Such systems can be used in conjunction with expert knowledgein situations where experts have a good qualitative idea of how to group data but havetrouble dealing with observations that are not well represented by rigid classificationschemes (Burrough and McDonnell 1998) This method can be particularly useful insituations when taxonomic schemes have been previously developed as is the case insoil taxonomy Several examples of this type of approach were published in the mid-1990s by A Zhu and colleagues (Zhu and Band 1994 Zhu et al 1996 1997 Zhu1997ab) Such systems proved useful for mapping soil at unvisited locations usingsurveyor knowledge and were also capable of producing continuous soil propertymaps The use of fuzzy logic within the theoretical framework of the expert systemallows the complex nature of soil to propagate through the modelling process neversubjected to classification schemes that filter out potentially useful lsquonoisersquo The use offuzzy logic also gives the soil surveyor more latitude during the interview processwhen the knowledge base is defined for the expert system Zhu et al (1997) developeda fuzzy logic-based expert system called SoLIM to determine the similarity of eachgrid cell in a study area to the various taxonomic mapping units delimited by the soilsurvey Continuous soil attribute maps were calculated using the similarity values andtheir relative soil survey determined attribute values (effectively achieving Goal 2) Theresulting data (which proved more accurate than soil survey data once field checked)consists of a raster grid whose resolution is determined by the resolution of the inputenvironmental and digital elevation data As noted previously this type of data modelis more applicable to environmental modelling than the choropleth map (Burrough andMcDonnell 1998)

Expert system approaches to PSM are capable of exploiting soil surveyor knowledgeby developing rule-based systems that imitate the surveyorrsquos conceptual model of soilvariability (the primary focus of Goal 3) The method would appear extremely usefulfor mapping projects (such as those conducted by the NRCS-NCSS) where fieldwork isinitially conducted to determine soilndashlandscape relations Expert system developmentcould be directly inserted into the traditional soil survey mapping approach as asubstitute for the step where the surveyor converts hisher conceptual model into achoropleth map Rather that knowledge could be incorporated into the expert systemwhich could be used to predictively map soil The resulting raster map would be morescientifically based and explicit than the hand-drawn choropleth maps of the past Itsscale would not be limited to that of the aerial photography but rather to the scale ofthe environmental data

Despite these advantages the expert system approach has some drawbacks Becausethe method is dependent upon expert knowledge it cannot be applied whereenvironmentndashsoil relations are poorly understood Of course this criticism can be madeof all PSM models as inductive or deductive knowledge of soilndashenvironment relationsis a prerequisite for PSM However expert systems are deductive models and as such

are not driven by any specific field-collected data (although presumably the soil experthas field experience in the mapping area) Expert systems donrsquot afford the opportunityto first statistically document landscapendashsoil relations and then extrapolate the resultsbecause expert systems do not directly use sample soil data (lsquohard datarsquo) to determinesoilndashlandscape relations Thus the utilization of the relationship between environmen-tal variables and soil properties (Goal 1) is only indirectly achieved Expert systemsapproaches have been demonstrated to be extremely effective in a small number of casestudies The possibility of satisfying all three goals discussed in this review makesexpert systems a predictive soil mapping method that needs to be further tested ndash bothgeographically and across different scales of analysis

V Conclusion

Most of the predictive soil mapping research outlined in this review was conducted atvery large map scales (over small areas) In fact the majority of the research wasconcerned with assessing the spatial variability of soil character within individual fieldsor across soil toposequences The primary driving force behind this type of research hasbeen the need to provide accurate soils information for agriculture and ecologicalmodels It is clear that terrain attributes are powerful predictors at the local scaleGeostatistical tools have been successful at using terrain attributes and the spatialdependence of soil properties to interpolate between existing data points withinindividual fields Across soil toposequences statistical approaches provide a usefulmeans of predicting soil character PSM research has been most successful at the fieldscale because many of the soil-forming factors are held constant For example the neteffect of four of Jennyrsquos five soil-forming factors (climate organism parent material andtime) was minimal within many of the studies reviewed While some research hastackled larger areas of study there exists a deficiency in our ability to predictively mapsoils at smaller map scales Since the distribution of soil is scale-dependent differentPSM methods and predictors are likely to work better at different scales Focus in thefuture must continue to move toward working over larger spatial extents of study inorder to produce landscape-scale soil information

Further a large proportion of the research was conducted in semi-arid gentlysloping agricultural landscapes Humid forestlands mountainous regions and desertshave received little attention As such PSM methods need to continue to be testedandor developed in a wider variety of landscapes where spatial soil distributions canbe more complex Different methods will likely be successful to different degrees indifferent environments Whereas terrain attributes are dominant predictors of soilcharacter across toposequences in prairie lands other predictors are likely necessary inareas where soils do not develop into clear toposequences For example remote sensingdata has been a useful predictor of soil chronosequences in desert landscapes wheresurface appearance is often related to soil character Surficial geology is often reflectiveof soil character in mountain regions where soils are thin and significant bedrock existsat the surface Focus in the future must continue to determine which methods andpredictors work best in which environments

Since the most useful PSM approach will vary across spatial scales and environmen-tal gradients the method used should be driven by the mapping objectives of the

P Scull et al 191

192 Predictive soil mapping a review

project Whereas the traditional soil survey of the past was expected to meet the needsof a diverse group of end users PSM methods will vary given the objectives of thesurvey As a result the end users of the soil data need to play a more active role in thesurvey process Because precision agriculture and large-scale mapping has been theprimary focus of PSM there exists a deficiency in our ability to predictively map soilsfor the purpose of general land use planning and management Even though lessdetailed soil information is needed for these purposes the development of PSMmethods is complicated by the fact that soils exhibit complex spatial variability atsmall map scales where soil-forming environments vary greatly from one location tothe next

A number of alternative methods of characterizing the continuous nature of the soillandscape have been developed Thus far most of the PSM research has provided soilinformation in a nonobject form (as opposed to defining soil types as independententities) Soil data have been generated organized and presented in the form of eitherisorithmic maps or fine-scale raster grids Both of these data models are field-viewmodels of geographic space which allow the soil to be perceived as a constantlyvarying surface Two distinct approaches have been employed (1) mapping individualsoil properties and (2) mapping continuously varying (fuzzy) soil classes Mappingindividual properties is the most common approach and will likely continue todominate PSM research The use of fuzzy soil classes in the literature is less commonand more difficult to be integrated into standard mapping procedures (such as use ofsoil taxa) because the concept radically differs from the traditional view of the soillandscape However fuzzy soil classes (generated using the fuzzy-k means approach) dohave the potential to help further advance Soil taxonomy by identifying taxonomicclasses that are more reflective of pedologic processes at work

Expert systems have been greatly underutilized in PSM research especiallyconsidering how effective a small number of case studies have been Expert systemshave the potential to satisfy successfully all three goals discussed throughout thereview They also have the potential to bridge the gap between traditional approachesand PSM methods because field soil scientists do not have to change their conceptualapproach to mapping They can still conduct field reconnaissance to determinesoilndashlandscape relationships Afterwards though they can use expert systems to betterexploit the knowledge they have garnered Such an approach could help ease thetransition from traditional soil survey to more scientifically explicit methods Expertsystems are also the most fruitful approach to utilizing a wealth of data that has alreadybeen collected in a nonexplicit manner the wealth of expert knowledge (in the form ofthe conceptual model) that senior soil mappers have accumulated

Focus in the future must also move in the direction of operability The utility of manyPSM approaches has been clearly demonstrated yet changes in how soils are mappedand perceived has been slow In the USA various members of the National CooperativeSoil Survey have called for a change in the philosophy of survey (from mappingdiscrete soil types to mapping continuous varying soil properties McSweeny et al1994) but little change has actually taken place To realize the potential of PSM in con-tributing to an overall change in standard soil mapping procedures practitionersworking in the field of PSM need to form working relationships with field soilscientists PSM methods need also to be presented in a manner that is comprehensibleto the soil science trained field mapper Predictive soil mapping is a relatively recent

phenomenon yet much progress has been made In the process of mapping soilknowledge will continue to advance

P Scull et al 193

References

Agbu PA Fehrenbacher DJ and Jansen IJ1990 Statistical comparison of SPOT spectralmaps with field soil maps Soil Science Society ofAmerica Journal 54 818ndash18

Barrett L 1999 Particulars in contextmaintaining a balance in soil geography Annalsof the Association of America Geographers 89707ndash13

Beckett PHT and Webster R 1971 Soilvariability a review Soils and Fertilizers 341ndash15

Bell JC Grigal DF and Bates PC 2000 A soil-terrain model for estimating spatial patterns ofsoil organic carbon In Wilson J and Gallant Jeditors Terrain analysis principles and applica-tions New York City NY John Wiley and Sons295ndash310

Birkeland PW 1999 Soil and geomorphologyThird edition New York NY Oxford UniversityPress

Breiman L Friedman JH Olshen RA andStone CJ 1984 Classification and regressiontrees Belmont CA Wadsworth

Brule FJ 1996 Fuzzy systems ndash a tutorialhttp newsgroup compai httpwwwquadralaycom (last accessed 30 August 2001)

Bunkin FV and Bunkin AF 2000 Lidarsounding of water soil and plants Atmosphericand Oceanic Optics 13 54ndash72

Buol SW Hole FD McCracken RJ andSouthard RJ 1997 Soil genesis and classifica-tion Ames IA Iowa State University Press

Burgess TM and Webster R 1980a Optimalinterpolation and isarithmic mapping of soilproperties the semi-variogram and punctualkriging Journal of Soil Science 31 315ndash31

ndashndashndashndash 1980b Optimal interpolation and isarithmicmapping of soil properties block kringingJournal of Soil Science 31 331ndash41

Burrough PA 1989 Fuzzy mathematicalmethods for soil survey and land evaluationJournal of Soil Science 40 477ndash92

Burrough PA and McDonnell RA 1998Principles of geographic information systems(Revised edition) Oxford Clarendon Press

Burrough PA Beckett PHT and Jarvis MG1971 The relation between cost and utility insoil survey Journal of Soil Science 22 368ndash81

Burrough PA Van Gaans PMF and

Hootsman R 1997 Continuous classificationin soil survey spatial correlation confusionand boundaries Geoderma 77 115ndash35

Cambell JB 1977 Variation of selectedproperties across a soil boundary Soil ScienceSociety of America Journal 41 578ndash82

Cambell JB and Edmonds WJ 1984 Themissing geographic dimension to soiltaxonomy Annals of the Association of AmericanGeographers 74 83ndash97

Castrignano A Giugliarini L Risaliti R andMartinelli N 2000 Study of spatial relation-ships among some soil physico-chemicalproperties of a field in central Italy using multi-variate geostatistics Geoderma 97 39ndash60

Cialella AT Dubayah R Lawrence W andLevine E 1997 Predicting soil drainage classusing remotely sensed and digital elevationdata Journal of Soil Science 62(2) 171ndash78

Clark RN and Swayze GA 1996 Evolution inimaging spectroscopy analysis and sensorsignal-to-noise an examination of how far wehave come Summaries of the sixth annual JPLairborne Earth science workshop 4ndash8 March 1996AVIRIS Workshop Vol 1 5

Clarke LA and Pregibon D 1992 Tree-basedmodels In Chambers J and Hastie J editorsStatistical models in S Pacific Grove Wadsworthand Brooks 377ndash419

Coleman TL Agbu PA and MontgomeryOL 1993 Spectral differentiation of surfacesoils and soil properties ndash is it possible fromspace platforms Soil Science 155 283ndash93

Cook SE Corner RJ Grealish G GesslerPE and Chartres CJ 1996 A rule-basedsystem to map soil properties Soil ScienceSociety of America Journal 60 1893ndash900

Cruickshank JG 1972 Soil geography New YorkNY John Wiley amp Sons

Csillag F Pasztor L Biehl LL 1993 Spectralband selection for the characterization ofsalinity status of soils Remote Sensing ofEnvironment 43 231ndash42

Dale MB McBratney AB and Russell JS1989 On the role of expert systems andnumerical taxonomy in soil classificationJournal of Soil Science 40 223ndash34

Davis JR 1993 Expert systems and environ-

soil profile characteristics In the second step the observations of soil attributes areincorporated into an implicit conceptual model that is used to infer soil variation Thethird step involves applying the conceptual model to the survey area to predict soilvariation at unobserved sites Usually less than 0001 of the survey area is actuallyobserved (Burrough et al 1971) a fact reflecting the high cost of field sampling Theconceptual model of soil variation is then transformed into a cartographic model thechoropleth map by drawing map unit boundaries on aerial photographs In effect pho-tographic scale determines the resolution of the soil map

This process has been severely criticized in the scientific literature for two reasonsFirst the conceptual model developed by the soil surveyor is primarily implicit beingconstructed in a heuristic manner This results in an excessive dependence upon tacitknowledge and as such incomplete information exists relative to the derivation of theultimate soil survey product (Hudson 1992) This aspect of soil survey is especiallyfrustrating because it fails to document most of the knowledge that the soil surveyoraccumulates during the expensive field mapping process In essence the soil survey isunfalsifiable and therefore untestable (Hewitt 1993) The final product of soil survey isa soil map that has unknown assumptions limitations and accuracy (Burrough et al1971 Dijkerman 1974) PSM techniques are similar in theory to soil survey (they bothuse knowledge of soilndashenvironment relations to make inferences) but the methodsemployed often yield quantitative expressions of soil variability with measured levelsof accuracy Clearly one direction of innovation in soil survey is to add objectivity tomodel development which will allow more explicit scientific communication (Gessler1996) The second major criticism of soil survey concerns the role of soil classification

3 Soil classification

The evolution of PSM techniques has been directly impacted and continues to beinfluenced by the process of soil classification This is especially true in the USAwherethe primary focus of the National Cooperative Soil Survey is to develop and map Soiltaxonomy (the official classification developed by the US Department of AgricultureSoil Survey Staff 1975) The system was influenced by nineteenth century biologicaltaxonomy and the practice of geological survey (Heuvelink and Webster 2001) Thepurpose of Soil taxonomy was to provide an objective manner to systematically classifysoil and was adopted at a time when soil information had to be abstracted to the levelof the modal profile (classified in Soil taxonomy) because it was impossible to catalogueand present the full amount of soil variability (Cambell and Edmonds 1984) In orderto map soil taxa the soil must be perceived as a spatial entity a lsquopedonrsquo (a term used inSoil taxonomy to refer to the smallest recognizable unit that can be called lsquoa soilrsquo) Inpractice this spatial perception of soil results in a map whose classes are homogenousunits with unknown variability and sharply defined boundaries (Burrough andMcDonnell 1998) Since the initial development of Soil taxonomy the perception of thesoil landscape has changed from a collection of individual soil types to a continuallyvarying mixture of soil components (Dmitriev 1983) Field observations have shownthat in concert with the environmental variables soil properties vary continuouslyacross landscapes exhibiting different and complex scales of variation (Simonson1959) Therefore soil distribution is not well represented by choropleth maps

P Scull et al 175

176 Predictive soil mapping a review

(McBratney 1992 Gessler et al 1995) While this change has not yet been recognized inSoil taxonomy PSM research has developed methods of soil inventory that moreaccurately describe the soil landscape Such methods are at odds with the traditionalapproaches because they often involve dismissing the concept of soil as a spatial entityRather the new methods focus on mapping continuously varying soil properties Theimplementation of these techniques has been difficult because of the entrenchment ofSoil taxonomy

III Components of predictive soil mapping

1 Geographic data models of soil variability and soil mapping

Traditionally soil maps have been digitized to fulfil the need for soil data within GIS-based environmental modelling research This information from the paper mapdigitized for the computer is fraught with the same problems of the original choroplethmaps ndash assumed homogeneous units with unknown variability and sharply definedboundaries GIS allows for a more robust characterization of spatial variability (relativeto the cartographic generalizations of the past) by allowing information to be analysedand stored using a variety of data models As defined in the GIS literature datamodelling is the process of discretizing spatial variation which entails abstracting gen-eralizing or approximating geographic reality (defined as empirically verifiable factsabout the real world) Unfortunately data modelling is often confused with issues ofdata structure and limited by software selection (Kemp 1992) The process is of crucialimportance because it controls the manner in which the data can be processed oranalysed (Goodchild 1994) as well as the view of the data the end user ultimatelyreceives (Goodchild 1992a Kemp 1992) Some data models are more accurate thanothers at portraying geographical reality (Goodchild 1992b) In the case of soil howwell do the data including the data model represent the highly variable continuousnature of soil How can digital computers be used to manage spatial data to bestrepresent the soil landscape

The choice of data models is partially dependent upon how the soil is perceived ingeographic space Objects can be thought of as existing as independent entities inempty space (object view) or as one of an infinite set of tuples (the foundation ofgeographic information ndash xyz where z is a measured value and xy are its location inspace) approximated by regions and segments (field view) (Goodchild 1994) The fieldview better represents continuous surfaces such as the soil landscape but soil datahave traditionally been modelled using the feature model of geographic space thechoropleth map Soil data have probably been managed this way because of theinfluence of Soil taxonomy which defines individual soil lsquotypesrsquo and because soil datacollection began during a time when few alternatives existed

Using PSM techniques a fundamental change in the soil data model from thechoropleth map to the raster grid allows better characterization of actual soilndashlandscapevariability A raster data model accommodates a lsquofield viewrsquo representation of thelandscape and is defined as a regular rectangular array of cells with some aggregatevalue of the field recorded for each cell (Goodchild 1994) The resolution of the datastored in this format is a function of the grid cell size which can be made small enoughto simulate continuous variation at the landscape scale The raster has become the most

widely used data model for PSM and is also routinely used to manage other environ-mental information such as elevation (DEMs) and remotely sensed data Spatialanalysis and integration with other types of raster-based environmental data can beeasily performed with soil data stored using a raster data structure (Burrough andMcDonnell 1998)

2 Digital terrain modelling

Terrain analysis quantifies the relief component of models characterizing soilformation Soil development and its associated profile characteristics often occurs inresponse to the way in which water moves through and over the landscape which iscontrolled by local relief Accordingly terrain analysis will be most useful in environ-ments where topographic shape is strongly related to the processes driving soilformation (McKensie et al 2000) Digital terrain modelling is a technique for derivingspatially explicit quantitative measures of the shape character of topography (Weibeland Heller 1991 Wilson and Gallant 2000) The spatial distribution of the resultingterrain attributes (characterizing local water flow paths) can also capture the spatialvariability of soil attributes Moore ID et al (1991) reviewed the analysis of digitalelevation data (including DEMs) for hydrological geomorphological and biologicalapplications They provided a table that summarized the significance and physicalmeaning of various terrain attributes to landscape processes Building on their workmany authors have used terrain attributes derived from digital elevation models(DEMs) as explanatory variables in predictive soil models (Odeh et al 1991 Moore etal 1993 Gessler et al 1995 Skidmore et al 1996 and others) Methods used to deriveterrain attributes have been greatly refined over the last 15 years and future satellitesaiding in the development of more accurate DEMs will make terrain analysis anincreasingly important component of predictive soils mapping (Moore et al 1993Mackay and Band 1998) Several recent review articles have been specifically devotedto the role of terrain analysis in soil mapping (Ventura and Irvin 1996 Irvin et al 1996McKensie et al 2000)

3 Remote sensing

Remote sensing data are an important component of PSM because they provide aspatially contiguous quantitative measure of surface reflectance which is related tosome soil properties (Agbu et al 1990) Both physical factors (eg particle size andsurface roughness) and chemical factors (eg surface mineralogy organic mattercontent and moisture) control soil spectral reflectance (Irons et al 1989) Surfacemineralogy can be derived by wavelength specific charge transfer and crystal fieldabsorptions associated with the presence of iron and iron-oxides (Fe2+ and Fe3+) andvibrational absorptions associated with hydroxyl bonds in clays adsorbed water andthe carbonate ion (Goetz 1989 Irons et al 1989) The presence and strength of theseabsorption features can be used to identify and quantify concentrations of mixed suitesof minerals in soil (Johnson et al 1983 Shipman and Adams 1987) Organic matterparticle size and moisture content in contrast influence soil reflectance primarilythrough a change in average surface reflectance and produce only broad spectral

P Scull et al 177

178 Predictive soil mapping a review

expression (Irons et al 1989) Adecrease in particle size tends to increase surface albedoand decrease spectral contrast of absorption features while an increase in organicmatter or soil moisture decreases average reflectance or albedo

Numerous studies have shown the potential benefits of using remote sensing for soilidentification and mapping Comprehensive surveys of soil spectral reflectance includestudies by Stoner et al (1980) Henderson et al (1992) and Csillag et al (1993) Remotesensing studies based on broad-band sensors such as Landsat TM include Agbu et al(1990) Coleman et al (1993) Seyler et al (1998) and Oliveira (2000) Traditionallyremote sensing has been used to classify soil units through photo-interpretation ordigital image processing Combining remotely sensed information with ancillaryinformation such as thematic maps or vegetation cover can yield significant improve-ments (Wilcox et al 1994 Cialella et al 1997 Wanchang et al 2000)

Recent developments in hyperspectral remote sensing offer the potential of signifi-cantly improving data input to predictive soil models Hyperspectral sensors such asthe Airborne VisibleInfrared Imaging Spectrometer (AVIRIS) measure a contiguousspectrum in the visible and NIR and thereby better characterize atmospheric andsurface properties (Goetz et al 1985) The large number of spectral bands permits directidentification of minerals in surface soils For example Clark and Swayze (1996)mapped over 30 minerals using AVIRIS at Cuprite Nevada Palacios-Orueta and Ustin(1996) showed that enhanced spectral information was suitable for discriminating evensubtle spectral changes associated with differences in organic matter and iron contentOther examples of the application of AVIRIS to aid soil mapping include Palacios-Orueta et al (1998) Okin et al (1998) and Roberts et al (1998)

Sensors that operate in the microwave portion of the electromagnetic spectrum havealso shown promise in soil mapping research Microwave sensing can be broadlydivided into active (eg radar) systems and passive systems and are capable ofpenetrating the atmosphere under virtually all conditions offering a significantadvantage over visible and near-infrared spectroscopy (for a general overview ofmicrowave sensing see Lillesand and Kiefer 1994 chapter 8) Synthetic aperature radar(SAR) is one example of an active system SAR has been used to aid soil propertymapping such as soil salinity (Metternicht 1998) and soil moisture (Engman andChauhan 1995 Narayanan and Hirsave 2001) Active radar systems can also bedesigned to collect data at varying look angles providing the opportunity for theacquisition of stereo radar images Such images can be used to produce high resolutionand extremely accurate DEMs (eg Fang 2000) A similar active sensing system isLiDAR (light detection and ranging) which uses pulses of laser light rather thanmicrowave energy to illuminate the surface (see Bunkin and Bunkin 2000 for a reviewof applications to soil mapping research) While passive microwave systems haveseemed to receive less attention in the literature a few examples of soil mapping appli-cations can be found (see Kleshchenko et al 2000 and Laymon et al 2001) Regardlessof the type of system remote sensing data and derived products are potentially usefulexplanatory variables in predictive soil mapping models

4 Fuzzy logic

Fuzzy set theory or fuzzy logic provides an alternative conceptual paradigm withinPSM research The use of this theory has increased greatly in the last few years making

it an important component of PSM Fuzzy logic is an alternative to Boolean logic thatattempts to recognize the concept of partial truth (Brule 1996) Dr Lotfi Zadeh (1965)introduced the concept and accompanying mathematics in his seminal work lsquoFuzzysetsrsquo The theory permits partial class membership in contrast to traditional set theorywhere set memberships are crisp and binary (ie a soil sample is either completelyType A or it is not at all Type A) Central to the fuzzy concept is the idea that objects innature rarely fit exactly the classification types to which they are assigned (Zadeh1965) Rather they show varying signs of similarity to multiple classes (ie an observedsoil pedon often resembles more than one of the defined soil series within the area) Byusing fuzzy membership values (ranging from 0 nonmembership to 1 totalmembership) within predictive soil models to express degrees of similarity generaliza-tion problems associated with classification schemes (filtering of information) areminimized and the complex nature of soil data is allowed to propagate through themodelling process Similarity values between 0 and 1 are not comparable to proportionsand need not add up to 1 Within Boolean logic probability statements refer to thelikelihood of an outcome the soil sample is either one series or another With fuzzylogic a given sample is not definitively a member of the subset of any one particularseries Fuzzy logic is especially useful in soil research because of the continuous andcomplex nature of the soil landscape It serves as an important alternative to thesubjective rigidity imposed on soils data by Soil taxonomy Several recent articlesprovide a thorough review of the use of fuzzy sets in soil science (Burrough 1989McBratney and Odeh 1997 Burrough et al 1997 De Gruijter et al 1997)

Within PSM research two different approaches to creating continuous classes usingfuzzy logic exist The first is based on the fuzzy-k-means classifier which partitionsobservations in multivariate space into natural classes This approach is similar tocluster analysis and numerical taxonomy but the resulting classes are continuous witheach observation assigned a fuzzy membership value that characterizes its degree ofsimilarity to each individual class The concept has been integrated into geostatisticalmethods and will be discussed in more detail below (see Section IV1) The secondapproach is known as the Semantic Import model (SI) and is used in situations whenclassification schemes are pre-defined and class limits are relatively well understoodThe SI model is commonly used in concert with expert knowledge and will bediscussed in the expert systems section (see Section IV4)

IV Recent advances in predictive soil mapping

Within the last decade many authors have sought to model the soil landscape using avariety of methods Literature in this field could be summarized many different waysbut we concentrate on the literature that directly addresses the goals of predictive soilmapping stated in the introduction (see Table 1) Therefore we will review research thatattempts to exploit the relationship between quantifiable landscape indices and soilcharacter in order to model the soil landscape in a more continuous and thereforerealistic manner

The research reviewed here is distinguished from decades of previous researchdocumenting the correlation between landscape position and soil attributes (reviewedby Hall and Olsen 1991) That body of research is informative but not useful for

P Scull et al 179

180 Predictive soil mapping a review

Table 1

Selected

recen

t literature on pred

ictive

soil m

apping an

d m

apping

(cited in

this article) d

escribing the mod

ellin

g metho

dused the

dep

enden

t variables used a

nd the

env

iron

men

tal v

ariables (ex

plan

atory) used in

the

mod

els

Stud

yMod

ellin

g metho

dDep

ende

nt variables

Environm

ental variab

les

Goa

ls atta

ined

a

Bell et

al 20

00Line

ar and

exp

onen

tial

Total soil orga

nic

Slop

e curvature aspe

ct

ECx

regression

carbon

hillslope

position

Burge

ss and

Web

ster

Punc

tual and

block

Na co

nten

t co

ver

Non

eC

1980

ab

kriging

loam

thickn

ess

ston

e co

nten

t

Burroug

h 198

9Fu

zzy mathe

matical

ndashndash

rpmetho

ds

Burroug

h et

al 19

97Con

tinuo

us classificatio

nndash

ndashrp

Castrigna

no eta

l 200

0Factorial kriging

CEC

pH N

P K

Non

e ex

plicitly used

Cx

Na

Cialella

eta

l 199

7Classifica

tion tree

Drainag

e class

Elev aspe

ct NDVI

Ec

Coo

k et

al 19

96Baysian

rule-ba

sed metho

dsOrgan

ic m

atter

Slop

e aspec

t wetne

ssEcX

inde

x

Dale et

al 19

89 (a review

)Ex

pert systems

ndashndash

ndash

Ellis 19

96Decision tree

ana

lysis

Soil erosion class

Slop

e aspec

t wetne

ssEc

neutral ne

tworks

inde

x flow le

ngth and

accu

mulation Lan

dsat

TM tree

cov

er

Gessler 1

996

A large

variety of statistic

Field an

d labo

ratory

A variety of digital

EC

metho

dsco

llected

phy

sica

len

vironm

ental da

tach

emical and

morph

olog

ical soil

prop

ertie

s

Gessler eta

l 199

5Line

ar and

logit reg

ression

A horizon

and

solum

Curvature CTI topo

EC

depth E horizon

positio

npresen

ce

Goo

vaerts 19

92Factorial kriging

Total carbon

Non

eC

P Scull et al 181pH

N CEC

extractab

lecatio

ns (K Ca M

g)Hartemink et

al 20

01ndash

ndashndash

rp

Heu

velin

k an

d W

ebster 20

01ndash

ndashndash

rp

Hew

itt 19

93ndash

ndashndash

rp

Indo

rante et

al 19

96ndash

ndashndash

rp

King et

al 19

99Lo

gistic regression

Presen

ceabsen

ceSlop

e aspec

t po

t solar

Ec

Non

calc c

lay-loam

energy

Kno

tters eta

l 199

5Kriging

co

-kriging

So

ft layer de

pth

Hillslop

e po

sitio

nCx

regression

kriging

Laga

cherie and

Holmes 19

97Classifica

tion tree

Map

ping

unit

Geo

logy variou

s topo

Ec

indices

Laslett e

tal 198

7Kriging

splin

es tren

dpH

Non

ec

surfac

e nea

rest neigh

bor

McB

ratney 1

992

ndashndash

ndashrp

McB

ratney

eta

l 199

1Block kriging

Clay co

nten

tNon

eC

McB

ratney

eta

l 200

0ndash

ndashndash

rp

McB

ratney

and

Ode

h 199

7Fu

zzy sets in

soil scienc

endash

ndashRp

McB

ratney

and

de Gruijter

Fuzz

y-k-mea

ns w

ithFu

zzy classes

Field co

llected

phy

sical

C19

92ex

tragrade

sch

emical and

morph

olog

ical soil

prop

ertie

s

McC

rack

en and

Cate 198

6Artificial intellig

ence

ndash

ndashrp

expe

rt systems

McK

ensie an

d Austin

19

93Gen

eralized

line

ar m

odels

Clay co

nten

t CEC

Slop

e relief land

form

ec

(logit)

pH EC

COLE

slop

e po

sitio

nbu

lk den

sity and

othe

rs

Moo

re eta

l 199

3Line

ar reg

ression

A horizon

dep

th O

MSlop

e w

etne

ss and

strea

mEC

and P co

nten

t pH

power in

dices aspect

curvature

182 Predictive soil mapping a reviewTa

ble 1

Con

tinu

ed

Stud

yMod

ellin

g metho

dDep

ende

nt variables

Environm

ental v

ariables

Goa

ls atta

ined

a

McK

ensie an

d Ryan 199

9Reg

ression tree and

linea

rSo

lum dep

th P an

dElevation slope

CE

regression

N con

tent

curvature CTI

contribu

ting area do

wn-

slop

e mea

ns for slope

clim

ate da

ta Presco

ttInde

x G

amma

Rad

iometry La

ndsat T

M

and Geo

logy unit

Ode

h et

al 19

92ab

Fuzz

y-c-mea

ns and

Fuzzy classes

Field-co

llected

phy

sica

lC

kriging

chem

ical and

morph

olog

ical soil

prop

ertie

s

Ode

h et

al 19

94

Reg

ression kriging

co

-So

lum dep

th de

pth

Slop

e aspect cu

rvature

eC

1995

kriging regression

to bed

rock gravel

kriging

and clay con

tent

Skidmore et

al 19

91Bayesian expe

rt system

Soil land

scap

e un

itVeg type

wetne

ss in

dex

EcX

1996

grad

ient terrain po

sitio

n

Voltz and

Web

ster 19

90Kriging

cu

bic splin

eClay co

nten

tNon

eC

Web

ster 1

994

Dev

elop

men

t of

ndashndash

rppe

dometrics

Zhu

199

7ab

Fuzz

y logic expe

rt system

Soil series A

horizon

Elev pm aspe

ct c

anop

yECX

Zhu

and

Ban

d 199

4(SoL

IM)

depth in

dividu

alco

verage

grad

ient

Zhu

eta

l 199

7series m

aps

curvature

Not

es

a Letters refer to the de

gree

to w

hich

the

goa

ls of PS

M defined

in the

introdu

ction are achiev

ed

Soilndash

e nvironm

ent relatio

ns utilized

(letter E) be

tter represen

tatio

n of soil c o

ntinuity (C

) an

d ex

pert kno

wledg

e utilize

d (X) c

orrespon

d to goa

ls 1 2 and

3respec

tively from

the

introdu

ction and

cap

ital lette

rs (E C X) ind

icate the metho

d is relatively more successful th

an th

ose metho

ds den

oted

by lower

case le

tters (e c x

) rp ind

icates rev

iew pap

ers

P Scull et al 183

predictive mapping because landscape position is never quantified (rather positionwas often qualitatively defined eg toe-slope) and thus the documented relationshipscannot be generalized using environmental data and digital elevation models to predictsoil character at unvisited sites

Table 1 documents modelling methods model variables and the extent to which thereferenced research satisfies the previously defined goals of PSM (lsquoGoals attainedrsquocolumn) Soil-environment relations utilized (letter E) better representation of soilcontinuity (C) and expert knowledge utilized (X) correspond to goals 1 2 and 3respectively from the introduction and capital letters (E C X) indicate the method isrelatively more successful than those methods denoted by lower case letters (e c x)Citations with no letters present within the lsquoGoals attainedrsquo column do not address theaforementioned goals For example the Cialella et al (1997) received a rating of lsquoEcrsquomeaning that the methods employed successfully utilized environmentalndashsoil characterrelations (E) and somewhat successfully presented a better method of representing soilcontinuity (c) The ratings are provided simply to help organize the literature that wasreviewed Review papers are included within the table denoted by lsquorprsquo Geostatisticalmethods are not included in the table because they have been comprehensivelysurveyed elsewhere (Odeh et al 1994 Burrough et al 1997 McBratney et al 2000Heuvelink and Webster 2001) and because the objectives and assumptions of geostatis-tical methods differ slightly from other PSM research We briefly outline thesedifferences in the following section

1 Geostatistical methods

Geostatistics are a subset of traditional statistics that deal primarily with spatial dataand account for spatial autocorrelation using kriging as the spatial interpolator Theconcept is based upon the theory of regionalized variables which was mainlydeveloped by Matheron (1963) and Krige (1963) Kriging is a form of weighted localaveraging that uses a measure of spatial dependence the variogram to determine theweights applied to the data when computing the averages Geostatistical methods havebeen used in predictive soil mapping research to spatially interpolate soil propertyvalues at unmeasured sites from field-collected data

Burgess and Webster (1980a b) were the first to introduce ordinary kriging to the soilcommunity and since that time an enormous amount of work has been published Forexample ordinary kriging has been used to interpolate many different soil propertiesincluding pollution trace element deficiencies salinity and fertility (Heuvelink andWebster 2001) Ordinary kriging has been criticized for a variety of reasons Forexample Laslett et al (1987) reported that several authors had criticized geostatisticsbecause kriging is a global rather than local technique failing to take into accountknowledge of soil materials and processes Other authors have criticized geostatisticsbecause they are excessively data dependent requiring a large number of closelyspaced data points (Zhu 1997a) As Webster and Oliver (1992) suggest in excess of ahundred samples may be needed to use geostatistics at the field scale because of highspatial variability of soil in some areas Geostatistics also assume spatial autocorrela-tion which sometimes may be a poor assumption in complex terrain where abruptchanges in soil-forming factors occur (McBratney et al 2000) Ordinary kriging by itself

184 Predictive soil mapping a review

does not satisfy two of the three goals of PSM presented in the introduction ndash it does notadequately incorporate expert knowledge and it does not exploit the relationshipbetween environmental variables and soil properties

Ordinary kriging has been modified in a variety of ways to better incorporateancillary data and known soilndashlandscape relationships Block kriging involvesdetermining estimates over meaningful areas rather than at specific points (Burgess andWebster 1980a McBratney et al 1991) Using this method a study area can be stratifiedinto different regions that are reflective of the pedogenetic processes at work In orderto accommodate a trend within a dependent soil variable universal kriging has beenused (Webster 1994) Kriging with external drift is similar to universal kriging but ituses an ancillary variable to represent the trend (McBratney et al 2000) Co-krigingtakes advantage of correlation that may exist between the variable of interest and othermore easily measured variables (Odeh et al 1995) Regression kriging involvesspatially interpolating the residuals from a non-spatial model by kriging and addingthe result to the prediction obtained from that model (Goovaerts 1997 Castrignano etal 2000) Factorial kriging is another method to integrate multivariate data into thestandard kriging routine to extrapolate soil data (Goovaerts 1992) Many authors havecompared these various methods (Laslett et al 1987 Voltz and Webster 1990 Odeh etal 1994 Knotters et al 1995)

Fuzzy logic has been used with geostatistics by various authors to produce new kindsof fuzzy soil maps with continuous classes (Burrough 1989 McBratney and DeGruijter1992 Odeh et al 1992a and reviewed by McBratney and Odeh 1997) The processentails kriging the matrix of membership values determined by the fuzzy k-meansclassifier resulting in a continuous soil surface where individual locations are allowedto belong to more than one class and no rigid boundaries are designated to separate thesoil into discrete units or entities The results of such analysis can be used to assess thepedologic process validity of soil taxonomy by determining whether soils grouptogether into classes that are similar to taxonomic types Fuzzy classes wouldpresumably reflect the main pedologic features within a given area In this sense theclassification is quantitative whereas soil taxonomy is rooted in qualitative discrimina-tion A comparison between the two could provide insightful and help assist soiltaxonomy in making a classification that is more indicative of underlying soil processes

Geostatistics in soil research were originally introduced to quantitatively assess soilvariability within soil mapping units (McBratney et al 1991) in response to criticisms inthe early 1970s that soil unit composition was not well quantified (Beckett and Webster1971) In this regard geostatistics have been very useful having served well the originalgoals set forth by Burgess and Webster (1980b) when they drew kriging to the attentionof soil scientists as a means of spatial prediction At the field scale soil variation islargely due to the effect that topography has on soil genesis Geostatistics have beensuccessfully applied in such environments by using terrain attributes as ancillary datawithin many of the kriging routines described above Such quantitative within-unitvariability of soil properties is very useful in the field of precision agriculture and othersituations (eg pollutants) where very detailed soil attribute information is needed atthe field scale (Heuvelink and Webster 2001)

However geostatistics have not been applied in a wide variety of environments or atlarger scales In order to be successfully applied in different environments geostatisticswill likely require a different suite of ancillary data For example remote sensing data

P Scull et al 185

could be used in arid regions where soil toposequences are less well expressed Atlarger scales of prediction selection of different sets of ancillary variables is requiredbecause different processes define soil character at different scales The most obviousexample is that of climate which may control soil distribution at large scales(continental) but has little explanatory power at the field level Regardless of whetheradequate ancillary data exist the amount of data required to use geostatistics forlandscape-level prediction would be extremely difficult and costly to collect given thestrict sampling protocol required to characterize spatial dependence It is also unclearat what landscape scale soils exhibit spatial autocorrelation

Geostatistical approaches do provide a means of creating continuous soil attributesurfaces to better represent soil continuity (Goal 2) and they can be used to exploit therelationship between environmental variables and soil properties in order to moreefficiently collect soil data (Goal 1) However they do not sufficiently utilize expertknowledge (Goal 3) as no attempt has been made in geostatistical approaches todirectly integrate expert knowledge Fundamentally kriging is a process of interpola-tion designed to predict attribute values in between locations of measured samples Inthis sense geostatistics represent a middle ground between pure interpolation (egnearest neighbour type classifier) in which only measured points for the variable ofinterest are used to determine unknown values and other predictive models thatprimarily use soilndashenvironment correlation to create predictive maps

2 Statistical methods

Statistical methods can be used to exploit the relationship between quantifiablelandscape indices and soil properties to create predictive soil maps For exampleMcKensie and Austin (1993) used a regression to account for a large percentage ofvariation for many soil characteristics (A horizon clay content CEC EC pH bulkdensity and COLE B horizon clay content CEC ESP EC pH bulk density and COLE)using a variety of predictor variables (slope presence or absence of impeding layerrelief landform topographic position) Their results confirm the hypothesis of MooreID et al (1991) that soil character is related to quantifiable landscape indices Howevertheir methods do not provide inference of soil properties at unmeasured sites frommapped environment data because the topographic variables were measured in thefield Linear regression has also been used with terrain variables derived from a 15-mDEM in northeastern Colorado to predict soil attributes (organic matter contentextractable phosphorous pH and texture) at unvisited sites (Moore et al 1993) In thatparticular study 50 of the variance of A-horizon thickness was explained by slope andthe wetness index Gessler et al (1995) also used regression to model A-horizonthickness from topographic variables in southeastern Australia (plan curvature andwetness index r2 = 063 P = 0001) They modeled solum depth and used logisticregression to model E horizon presenceabsence Elsewhere logistic regression hasbeen used to model the presenceabsence of noncalcareous clay loam horizon in centralFrance using terrain attributes from a 20-m DEM (King et al 1999) Exponentialregression has been used to model soil organic carbon using terrain variables (Bell et al2000) in glacial outwash soils in east-central Minnesota Generalized additive models(GAM) have been used less frequently in PSM research Gessler (1996) used a GAM

186 Predictive soil mapping a review

model to predict total soil carbon A horizon depth and solum depth using a variety ofenvironmental predictors (slope elevation wetness index mean annual temperatureprecipitation and radiation)

This small body of research opened the door to more complex methods by demon-strating the existence of quantifiable relationships These authors were able to producesoil attribute maps using raster data models whose scale was dependent upon the gridcell resolution of the environmental data They were successful at exploiting the rela-tionship between quantifiable topographic attributes and soil profile character (Goal 1)The continuous soil attribute surfaces they produce also better represent soil continuitythan the choropleth soil maps produced by traditional soil survey (Goal 2) Howeverthe bulk of these methods (excepting GAMs) are limited by their assumed linear rela-tionship between soil and topographic attributes their assumptions of normallydistributed data and their high data requirements Standard statistical procedures arealso not flexible enough to allow robust integration with a variety of potential datasources such as expert knowledge (Goal 1)

Statistical methods do demonstrate in a quantitative manner that terrain analysis canbe used to predict soil attributes in relatively small areas with homogeneous parentmaterial A large proportion of the research using statistical methods was conducted insemi-arid landscapes at small scales (the largest study area of the entire group was~2000 ha) Obviously for statistical approaches to be most effective they need to bemore universal As such they need to be tested andor developed at larger scales andin more diverse landscapes

3 Decision tree analysis (DTA)

The use of decision tree analysis is just beginning to be explored in predictive soilmapping research although it has been used successfully in the related field ofpredictive vegetation mapping since the early 1990s (Lees and Ritman 1991 MooreDM et al 1991 Franklin 1998) DTA is a form of divisive classification The process oftree modelling involves successively partitioning data (called recursive partitioning inthe tree modelling literature) into increasingly homogeneous subsets which once thepartitioning has ceased are called terminal nodes (Lees and Ritman 1991) Splits orrules defining how to partition the data are selected based on information statistics thatdefine how well the split decreases impurity within the data set (Clark and Pregibon1992) Splits are based on threshold values of an explanatory variable selected bycomparing the increase in resulting purity of node membership for all possiblethresholds and variables The process is iterative growing from the root node (thecomplete data set) to the terminal nodes in a dendritic fashion (Friedl and Brodley1997) Once the tree has been constructed (or grown) it encodes a set of decision rulesthat describe the data partitioning process These rules can be used to classify or predictother data sets (Moore DM et al 1991) Pruning the tree is often necessary to preventthe tree from being overfit to the sample data and to reduce tree complexity Pruningentails combining pairs of terminal nodes into single nodes and can be accomplishedusing cross-validation which yields an initial indication of how large a tree makesrobust predictions (Safavian and Norvig 1991) Cross-validation involves systematical-ly removing portions of the data set and running the remaining sample through the tree

in an iterative manner eventually yielding estimates of the misclassification rates foreach class each node and the whole model (Breiman et al 1984) In this mannerdifferent sized trees can be compared in terms of parsimony

The term DTA is used to collectively refer to all types of tree-based modelling (theword lsquodecisionrsquo is used because it is descriptive indicating that the analysis eventuallyleads to a set of decision rules defining data partitions) The term should be distin-guished from classification tree analysis because the latter refers specifically to DTAwhere the response variable is categorical The term CART (classification andregression trees ndash Breiman et al 1984) is sometimes used but strictly speaking refers tospecific software Friedl and Brodley (1997) provide a review of the decision treealgorithms They divide types of DTA into two classes (1) homogeneous decision treesfor which a single algorithm is used to estimate each split (eg CART) and (2) hybriddecision trees (HDT) for which different splitting methods can be used at differentpoints in the tree (eg Quinlan 1993) They further divide homogeneous decision treesinto univariate (UDT) where single features of the input data define splits and multi-variate decision trees (MDT) where multiple features of the input data can define splitsAccording to this naming convention no distinction is made between methods withdifferent types of response variables although all methods can be used with bothcategorical and continuous response variables In a comparison of these various typesof algorithms on a variety of data sets Friedl and Brodley (1997) found that HDT hadthe highest classification accuracy

The overall aim of DTA is to design a set of predictive rules (eg if geology type Athen soil type B) developed from training data which can then be applied to ageographic data base to predict the value of a response variable (Michaelsen et al 1994)Therefore DTA explicitly uses soilndashlandscape correlation in model development (Goal2) The technique appears promising in soil research but needs to be further tested asnot many of the above types of DTA have been tested in the PSM literature in fact onlyunivariate approaches have been employed For example Lagacherie and Holmes(1997) successfully used univariate DTA to model a categorical response variable soiltype within a training set and then assuming that the training set was representativeextrapolated the model to a much larger region Their work is interesting becausealthough not a single sample came from the area they eventually mapped they wereable to produce a soil map that was more accurate (74 versus 69) than the existingmap produced from traditional methods Cialella et al (1997) also used univariate DTAto predict soil drainage class from a variety of terrain attributes and remotely senseddata They predicted soil drainage class with an average of 78 accuracy ndash impressivegiven that the variation accounted for by a typical soil survey ranges from about halfthe total variance for physical attributes to less than one-tenth for some soil chemicalattributes (Gessler et al 1995)

DTA has been compared with other approaches by several authors In the applicationof erosion modelling DTA results were similar to Artificial Neural Networks (Ellis1996) Both methods achieved high training accuracy (as measured by the KappaStatistic) but in terms of prediction accuracy both methods performed poorly Gessleret al (1995) compared DTA with generalized linear models (GLM) and generalizedadditive models (GAM) to predict A-horizon thickness and concluded that GLM waspreferable to both DTA and GAM McKensie and Ryan (1999) compared regressiontrees and standard linear regression to predict soil properties (total solum depth soil

P Scull et al 187

188 Predictive soil mapping a review

phophorus and soil carbon) using a large variety of predictors (elevation slopecurvature wetness index climate variables geology unit Landsat TM data andGamma radiometric data) and found that different methods work best for differentproperties and overall success hinged on the strength of the relationship between soiland environmental variables Their study of a 50 000 ha forested landscape in southernNew South Wales Australia is one of the few studies at such a small map scale (largearea) or in forested landscape The most extensive comparison of techniques involvingDTA was conducted by McBratney et al (2000) They compared regression techniques(GLM GAM DTA-regression tree) geostatistical techniques (kriging and heterotopicco-kriging) and a hybrid technique (regression kriging) DTA was found to be thepoorest performing of the regression techniques because of the unrealistic predictionsurface generated by the DTA model DTA has also been criticized by other authorsbecause of the stepped prediction surface (Gessler 1996) This phenomenon can beespecially apparent in situations where predictor variables have different resolutionsThe extent to which DTA yields a better spatial representation of soil continuity is afunction of the scale and type of predictor variables used and therefore varies withindividual models

DTA is gaining widespread popularity as a means to develop prediction rules thatcan be rapidly and repeatedly evaluated (Cialella et al 1997 Franklin et al 2000) DTAprovides the following advantages over standard statistical techniques (1) it is easier tointerpret when explanatory variables are both nominal and continuous (2) it isinvariant to monotone re-expressions (transformations) of predictor variables (3) itdeals more satisfactorily with missing data values and outliers (4) it is more adept atcapturing nonadditive and nonlinear behaviour (5) it doesnrsquot make any assumptionsabout data distribution and (6) it is easily updateable as more data are collected(Moore DM et al 1991) The DTA model framework is especially appealing becauseof its capability to integrate a wide range of data sets as explanatory variables

DTA offers a unique opportunity for interaction between soil experts and soilmodellers because the output of the model is a set of rules that can be pedologicallyinterpreted by the soil expert In this sense expert knowledge is used in an implicitmanner in DTA (somewhat effectively achieving Goal 3) While these rules can often beexceedingly complex at minimum the expert can decide whether initial splits makesense given their understanding of the landscape The success of DTA results oftenhinges on the ability of the modeller to make key decisions during the model buildingprocess there is unfortunately no definitive way to determine the most optimal treeThe aid of the soil expert can potentially elucidate this problem

4 Expert systems

A variety of expert system approaches to PSM have been developed to utilize expertknowledge The purpose of such methods is to exploit the information the soil surveyoraccumulates while working in the field by integrating such knowledge into thepredictive model (McCracken and Cate 1986) Unlike the majority of the researchreviewed thus far the dependent variable in many expert systems models is often soiltaxa or mapping unit This apparent disadvantage of expert systems (using classifica-tion to characterize soil continuity) does make them easier to integrate into traditional

soil survey In addition several authors have developed methods to developcontinuous soil property maps from the output of expert system models designed topredict soil unit occurrence

Expert systems are composed of data (information on spatial environmentalvariables eg topography climate etc) a knowledge base (rules and facts related tosoil variation supplied from the soil surveyor) and an inference engine (whichcombines data and the knowledge base to infer logically valid conclusions) (Skidmoreet al 1996) Expert systems differ from conventional models in two ways (1) they storeand manipulate qualitative information (allowing them access to information thatcannot normally be used in other modelling frameworks) and (2) they are structuredas meta-models (the knowledge is separated from the model) (Davis 1993) This allowsthe model to selectively choose which information is relevant at various stages of themodelling process and it allows for information to be easily updated Davis (1993)reviews the application of expert systems to environmental modelling research conclu-ding that the technique is becoming more widely accepted He further states that theapplication of expert systems is constrained by an absence of fundamental knowledgefor rule generation a problem that would appear less relevant to soil mapping giventhe amount of untapped expert knowledge accumulated by the soil surveyor

The first mention of the use of expert systems in pedology was in a paper presenta-tion at the Northeast Committee Soil Survey Conference 1984 (Flach 1985) In herpaper Flach (1985) hinted that recent developments in computer science especiallyexpert systems and artificial intelligence could make modelling a practical mappingtool for soil scientists in the near future A year later McCracken and Cate (1986) hopedto encourage soil scientists to explore expert systems and its application to soil sciencethrough an optimistic article they wrote on the potential use of expert systemsHowever little research was actually conducted in the late 1980s and expert systemshad not yet begun to fulfil the lofty goals set forth by early practitioners (Dale et al1989)

In the early 1990s expert systems approaches to predictive soil mapping began toappear in the literature Skidmore et al (1991) used a Bayesian expert system to mapforest soil into different classes and their results compared favourably with availablesoil maps and actual field-collected data Their methods successfully incorporatedsurveyor knowledge and remotely sensed and digital terrain attributes but failed tobetter characterize continuous soilndashlandscape variability because their final productwas a choropleth map Skidmore et al (1996) revisited their earlier research andassessed the mapping accuracy of their results They found that the soil map producedby the expert system achieved an overall accuracy of 698 (sample size = 53) while themap derived from conventional methods had an accuracy of 736 In addition toknowledge provided by a soil scientist Skidmore et alrsquos (1996) expert system usedvegetation derived for aerial photography and topographic variables derived from a 10-m DEM (wetness index topographic position and slope)

Using an expert system Cook et al (1996) successfully produced a continuous soilproperty map for organic matter content using wetness index aspect and slope asexplanatory variables Their methods were somewhat inefficient because they requireda separate expert system for each soil property of interest However their researchrepresents the only example of expert systems used to directly predict the spatial dis-tribution of a soil property All of the other examples of expert systems in the literature

P Scull et al 189

190 Predictive soil mapping a review

deal with soil type or class The use of expert systems to map soil properties needs tobe explored further

The expert systems discussed thus far have all used Boolean logic within theirtheoretical framework whereby an observation can belong to one and only one classand the soil properties of that class are assigned to the observation With thedevelopment of fuzzy logic and the semantic import model (the second fuzzy logicapproach to creating continuous classes) the opportunity exists to develop fuzzy logic-based expert systems Such systems can be used in conjunction with expert knowledgein situations where experts have a good qualitative idea of how to group data but havetrouble dealing with observations that are not well represented by rigid classificationschemes (Burrough and McDonnell 1998) This method can be particularly useful insituations when taxonomic schemes have been previously developed as is the case insoil taxonomy Several examples of this type of approach were published in the mid-1990s by A Zhu and colleagues (Zhu and Band 1994 Zhu et al 1996 1997 Zhu1997ab) Such systems proved useful for mapping soil at unvisited locations usingsurveyor knowledge and were also capable of producing continuous soil propertymaps The use of fuzzy logic within the theoretical framework of the expert systemallows the complex nature of soil to propagate through the modelling process neversubjected to classification schemes that filter out potentially useful lsquonoisersquo The use offuzzy logic also gives the soil surveyor more latitude during the interview processwhen the knowledge base is defined for the expert system Zhu et al (1997) developeda fuzzy logic-based expert system called SoLIM to determine the similarity of eachgrid cell in a study area to the various taxonomic mapping units delimited by the soilsurvey Continuous soil attribute maps were calculated using the similarity values andtheir relative soil survey determined attribute values (effectively achieving Goal 2) Theresulting data (which proved more accurate than soil survey data once field checked)consists of a raster grid whose resolution is determined by the resolution of the inputenvironmental and digital elevation data As noted previously this type of data modelis more applicable to environmental modelling than the choropleth map (Burrough andMcDonnell 1998)

Expert system approaches to PSM are capable of exploiting soil surveyor knowledgeby developing rule-based systems that imitate the surveyorrsquos conceptual model of soilvariability (the primary focus of Goal 3) The method would appear extremely usefulfor mapping projects (such as those conducted by the NRCS-NCSS) where fieldwork isinitially conducted to determine soilndashlandscape relations Expert system developmentcould be directly inserted into the traditional soil survey mapping approach as asubstitute for the step where the surveyor converts hisher conceptual model into achoropleth map Rather that knowledge could be incorporated into the expert systemwhich could be used to predictively map soil The resulting raster map would be morescientifically based and explicit than the hand-drawn choropleth maps of the past Itsscale would not be limited to that of the aerial photography but rather to the scale ofthe environmental data

Despite these advantages the expert system approach has some drawbacks Becausethe method is dependent upon expert knowledge it cannot be applied whereenvironmentndashsoil relations are poorly understood Of course this criticism can be madeof all PSM models as inductive or deductive knowledge of soilndashenvironment relationsis a prerequisite for PSM However expert systems are deductive models and as such

are not driven by any specific field-collected data (although presumably the soil experthas field experience in the mapping area) Expert systems donrsquot afford the opportunityto first statistically document landscapendashsoil relations and then extrapolate the resultsbecause expert systems do not directly use sample soil data (lsquohard datarsquo) to determinesoilndashlandscape relations Thus the utilization of the relationship between environmen-tal variables and soil properties (Goal 1) is only indirectly achieved Expert systemsapproaches have been demonstrated to be extremely effective in a small number of casestudies The possibility of satisfying all three goals discussed in this review makesexpert systems a predictive soil mapping method that needs to be further tested ndash bothgeographically and across different scales of analysis

V Conclusion

Most of the predictive soil mapping research outlined in this review was conducted atvery large map scales (over small areas) In fact the majority of the research wasconcerned with assessing the spatial variability of soil character within individual fieldsor across soil toposequences The primary driving force behind this type of research hasbeen the need to provide accurate soils information for agriculture and ecologicalmodels It is clear that terrain attributes are powerful predictors at the local scaleGeostatistical tools have been successful at using terrain attributes and the spatialdependence of soil properties to interpolate between existing data points withinindividual fields Across soil toposequences statistical approaches provide a usefulmeans of predicting soil character PSM research has been most successful at the fieldscale because many of the soil-forming factors are held constant For example the neteffect of four of Jennyrsquos five soil-forming factors (climate organism parent material andtime) was minimal within many of the studies reviewed While some research hastackled larger areas of study there exists a deficiency in our ability to predictively mapsoils at smaller map scales Since the distribution of soil is scale-dependent differentPSM methods and predictors are likely to work better at different scales Focus in thefuture must continue to move toward working over larger spatial extents of study inorder to produce landscape-scale soil information

Further a large proportion of the research was conducted in semi-arid gentlysloping agricultural landscapes Humid forestlands mountainous regions and desertshave received little attention As such PSM methods need to continue to be testedandor developed in a wider variety of landscapes where spatial soil distributions canbe more complex Different methods will likely be successful to different degrees indifferent environments Whereas terrain attributes are dominant predictors of soilcharacter across toposequences in prairie lands other predictors are likely necessary inareas where soils do not develop into clear toposequences For example remote sensingdata has been a useful predictor of soil chronosequences in desert landscapes wheresurface appearance is often related to soil character Surficial geology is often reflectiveof soil character in mountain regions where soils are thin and significant bedrock existsat the surface Focus in the future must continue to determine which methods andpredictors work best in which environments

Since the most useful PSM approach will vary across spatial scales and environmen-tal gradients the method used should be driven by the mapping objectives of the

P Scull et al 191

192 Predictive soil mapping a review

project Whereas the traditional soil survey of the past was expected to meet the needsof a diverse group of end users PSM methods will vary given the objectives of thesurvey As a result the end users of the soil data need to play a more active role in thesurvey process Because precision agriculture and large-scale mapping has been theprimary focus of PSM there exists a deficiency in our ability to predictively map soilsfor the purpose of general land use planning and management Even though lessdetailed soil information is needed for these purposes the development of PSMmethods is complicated by the fact that soils exhibit complex spatial variability atsmall map scales where soil-forming environments vary greatly from one location tothe next

A number of alternative methods of characterizing the continuous nature of the soillandscape have been developed Thus far most of the PSM research has provided soilinformation in a nonobject form (as opposed to defining soil types as independententities) Soil data have been generated organized and presented in the form of eitherisorithmic maps or fine-scale raster grids Both of these data models are field-viewmodels of geographic space which allow the soil to be perceived as a constantlyvarying surface Two distinct approaches have been employed (1) mapping individualsoil properties and (2) mapping continuously varying (fuzzy) soil classes Mappingindividual properties is the most common approach and will likely continue todominate PSM research The use of fuzzy soil classes in the literature is less commonand more difficult to be integrated into standard mapping procedures (such as use ofsoil taxa) because the concept radically differs from the traditional view of the soillandscape However fuzzy soil classes (generated using the fuzzy-k means approach) dohave the potential to help further advance Soil taxonomy by identifying taxonomicclasses that are more reflective of pedologic processes at work

Expert systems have been greatly underutilized in PSM research especiallyconsidering how effective a small number of case studies have been Expert systemshave the potential to satisfy successfully all three goals discussed throughout thereview They also have the potential to bridge the gap between traditional approachesand PSM methods because field soil scientists do not have to change their conceptualapproach to mapping They can still conduct field reconnaissance to determinesoilndashlandscape relationships Afterwards though they can use expert systems to betterexploit the knowledge they have garnered Such an approach could help ease thetransition from traditional soil survey to more scientifically explicit methods Expertsystems are also the most fruitful approach to utilizing a wealth of data that has alreadybeen collected in a nonexplicit manner the wealth of expert knowledge (in the form ofthe conceptual model) that senior soil mappers have accumulated

Focus in the future must also move in the direction of operability The utility of manyPSM approaches has been clearly demonstrated yet changes in how soils are mappedand perceived has been slow In the USA various members of the National CooperativeSoil Survey have called for a change in the philosophy of survey (from mappingdiscrete soil types to mapping continuous varying soil properties McSweeny et al1994) but little change has actually taken place To realize the potential of PSM in con-tributing to an overall change in standard soil mapping procedures practitionersworking in the field of PSM need to form working relationships with field soilscientists PSM methods need also to be presented in a manner that is comprehensibleto the soil science trained field mapper Predictive soil mapping is a relatively recent

phenomenon yet much progress has been made In the process of mapping soilknowledge will continue to advance

P Scull et al 193

References

Agbu PA Fehrenbacher DJ and Jansen IJ1990 Statistical comparison of SPOT spectralmaps with field soil maps Soil Science Society ofAmerica Journal 54 818ndash18

Barrett L 1999 Particulars in contextmaintaining a balance in soil geography Annalsof the Association of America Geographers 89707ndash13

Beckett PHT and Webster R 1971 Soilvariability a review Soils and Fertilizers 341ndash15

Bell JC Grigal DF and Bates PC 2000 A soil-terrain model for estimating spatial patterns ofsoil organic carbon In Wilson J and Gallant Jeditors Terrain analysis principles and applica-tions New York City NY John Wiley and Sons295ndash310

Birkeland PW 1999 Soil and geomorphologyThird edition New York NY Oxford UniversityPress

Breiman L Friedman JH Olshen RA andStone CJ 1984 Classification and regressiontrees Belmont CA Wadsworth

Brule FJ 1996 Fuzzy systems ndash a tutorialhttp newsgroup compai httpwwwquadralaycom (last accessed 30 August 2001)

Bunkin FV and Bunkin AF 2000 Lidarsounding of water soil and plants Atmosphericand Oceanic Optics 13 54ndash72

Buol SW Hole FD McCracken RJ andSouthard RJ 1997 Soil genesis and classifica-tion Ames IA Iowa State University Press

Burgess TM and Webster R 1980a Optimalinterpolation and isarithmic mapping of soilproperties the semi-variogram and punctualkriging Journal of Soil Science 31 315ndash31

ndashndashndashndash 1980b Optimal interpolation and isarithmicmapping of soil properties block kringingJournal of Soil Science 31 331ndash41

Burrough PA 1989 Fuzzy mathematicalmethods for soil survey and land evaluationJournal of Soil Science 40 477ndash92

Burrough PA and McDonnell RA 1998Principles of geographic information systems(Revised edition) Oxford Clarendon Press

Burrough PA Beckett PHT and Jarvis MG1971 The relation between cost and utility insoil survey Journal of Soil Science 22 368ndash81

Burrough PA Van Gaans PMF and

Hootsman R 1997 Continuous classificationin soil survey spatial correlation confusionand boundaries Geoderma 77 115ndash35

Cambell JB 1977 Variation of selectedproperties across a soil boundary Soil ScienceSociety of America Journal 41 578ndash82

Cambell JB and Edmonds WJ 1984 Themissing geographic dimension to soiltaxonomy Annals of the Association of AmericanGeographers 74 83ndash97

Castrignano A Giugliarini L Risaliti R andMartinelli N 2000 Study of spatial relation-ships among some soil physico-chemicalproperties of a field in central Italy using multi-variate geostatistics Geoderma 97 39ndash60

Cialella AT Dubayah R Lawrence W andLevine E 1997 Predicting soil drainage classusing remotely sensed and digital elevationdata Journal of Soil Science 62(2) 171ndash78

Clark RN and Swayze GA 1996 Evolution inimaging spectroscopy analysis and sensorsignal-to-noise an examination of how far wehave come Summaries of the sixth annual JPLairborne Earth science workshop 4ndash8 March 1996AVIRIS Workshop Vol 1 5

Clarke LA and Pregibon D 1992 Tree-basedmodels In Chambers J and Hastie J editorsStatistical models in S Pacific Grove Wadsworthand Brooks 377ndash419

Coleman TL Agbu PA and MontgomeryOL 1993 Spectral differentiation of surfacesoils and soil properties ndash is it possible fromspace platforms Soil Science 155 283ndash93

Cook SE Corner RJ Grealish G GesslerPE and Chartres CJ 1996 A rule-basedsystem to map soil properties Soil ScienceSociety of America Journal 60 1893ndash900

Cruickshank JG 1972 Soil geography New YorkNY John Wiley amp Sons

Csillag F Pasztor L Biehl LL 1993 Spectralband selection for the characterization ofsalinity status of soils Remote Sensing ofEnvironment 43 231ndash42

Dale MB McBratney AB and Russell JS1989 On the role of expert systems andnumerical taxonomy in soil classificationJournal of Soil Science 40 223ndash34

Davis JR 1993 Expert systems and environ-

176 Predictive soil mapping a review

(McBratney 1992 Gessler et al 1995) While this change has not yet been recognized inSoil taxonomy PSM research has developed methods of soil inventory that moreaccurately describe the soil landscape Such methods are at odds with the traditionalapproaches because they often involve dismissing the concept of soil as a spatial entityRather the new methods focus on mapping continuously varying soil properties Theimplementation of these techniques has been difficult because of the entrenchment ofSoil taxonomy

III Components of predictive soil mapping

1 Geographic data models of soil variability and soil mapping

Traditionally soil maps have been digitized to fulfil the need for soil data within GIS-based environmental modelling research This information from the paper mapdigitized for the computer is fraught with the same problems of the original choroplethmaps ndash assumed homogeneous units with unknown variability and sharply definedboundaries GIS allows for a more robust characterization of spatial variability (relativeto the cartographic generalizations of the past) by allowing information to be analysedand stored using a variety of data models As defined in the GIS literature datamodelling is the process of discretizing spatial variation which entails abstracting gen-eralizing or approximating geographic reality (defined as empirically verifiable factsabout the real world) Unfortunately data modelling is often confused with issues ofdata structure and limited by software selection (Kemp 1992) The process is of crucialimportance because it controls the manner in which the data can be processed oranalysed (Goodchild 1994) as well as the view of the data the end user ultimatelyreceives (Goodchild 1992a Kemp 1992) Some data models are more accurate thanothers at portraying geographical reality (Goodchild 1992b) In the case of soil howwell do the data including the data model represent the highly variable continuousnature of soil How can digital computers be used to manage spatial data to bestrepresent the soil landscape

The choice of data models is partially dependent upon how the soil is perceived ingeographic space Objects can be thought of as existing as independent entities inempty space (object view) or as one of an infinite set of tuples (the foundation ofgeographic information ndash xyz where z is a measured value and xy are its location inspace) approximated by regions and segments (field view) (Goodchild 1994) The fieldview better represents continuous surfaces such as the soil landscape but soil datahave traditionally been modelled using the feature model of geographic space thechoropleth map Soil data have probably been managed this way because of theinfluence of Soil taxonomy which defines individual soil lsquotypesrsquo and because soil datacollection began during a time when few alternatives existed

Using PSM techniques a fundamental change in the soil data model from thechoropleth map to the raster grid allows better characterization of actual soilndashlandscapevariability A raster data model accommodates a lsquofield viewrsquo representation of thelandscape and is defined as a regular rectangular array of cells with some aggregatevalue of the field recorded for each cell (Goodchild 1994) The resolution of the datastored in this format is a function of the grid cell size which can be made small enoughto simulate continuous variation at the landscape scale The raster has become the most

widely used data model for PSM and is also routinely used to manage other environ-mental information such as elevation (DEMs) and remotely sensed data Spatialanalysis and integration with other types of raster-based environmental data can beeasily performed with soil data stored using a raster data structure (Burrough andMcDonnell 1998)

2 Digital terrain modelling

Terrain analysis quantifies the relief component of models characterizing soilformation Soil development and its associated profile characteristics often occurs inresponse to the way in which water moves through and over the landscape which iscontrolled by local relief Accordingly terrain analysis will be most useful in environ-ments where topographic shape is strongly related to the processes driving soilformation (McKensie et al 2000) Digital terrain modelling is a technique for derivingspatially explicit quantitative measures of the shape character of topography (Weibeland Heller 1991 Wilson and Gallant 2000) The spatial distribution of the resultingterrain attributes (characterizing local water flow paths) can also capture the spatialvariability of soil attributes Moore ID et al (1991) reviewed the analysis of digitalelevation data (including DEMs) for hydrological geomorphological and biologicalapplications They provided a table that summarized the significance and physicalmeaning of various terrain attributes to landscape processes Building on their workmany authors have used terrain attributes derived from digital elevation models(DEMs) as explanatory variables in predictive soil models (Odeh et al 1991 Moore etal 1993 Gessler et al 1995 Skidmore et al 1996 and others) Methods used to deriveterrain attributes have been greatly refined over the last 15 years and future satellitesaiding in the development of more accurate DEMs will make terrain analysis anincreasingly important component of predictive soils mapping (Moore et al 1993Mackay and Band 1998) Several recent review articles have been specifically devotedto the role of terrain analysis in soil mapping (Ventura and Irvin 1996 Irvin et al 1996McKensie et al 2000)

3 Remote sensing

Remote sensing data are an important component of PSM because they provide aspatially contiguous quantitative measure of surface reflectance which is related tosome soil properties (Agbu et al 1990) Both physical factors (eg particle size andsurface roughness) and chemical factors (eg surface mineralogy organic mattercontent and moisture) control soil spectral reflectance (Irons et al 1989) Surfacemineralogy can be derived by wavelength specific charge transfer and crystal fieldabsorptions associated with the presence of iron and iron-oxides (Fe2+ and Fe3+) andvibrational absorptions associated with hydroxyl bonds in clays adsorbed water andthe carbonate ion (Goetz 1989 Irons et al 1989) The presence and strength of theseabsorption features can be used to identify and quantify concentrations of mixed suitesof minerals in soil (Johnson et al 1983 Shipman and Adams 1987) Organic matterparticle size and moisture content in contrast influence soil reflectance primarilythrough a change in average surface reflectance and produce only broad spectral

P Scull et al 177

178 Predictive soil mapping a review

expression (Irons et al 1989) Adecrease in particle size tends to increase surface albedoand decrease spectral contrast of absorption features while an increase in organicmatter or soil moisture decreases average reflectance or albedo

Numerous studies have shown the potential benefits of using remote sensing for soilidentification and mapping Comprehensive surveys of soil spectral reflectance includestudies by Stoner et al (1980) Henderson et al (1992) and Csillag et al (1993) Remotesensing studies based on broad-band sensors such as Landsat TM include Agbu et al(1990) Coleman et al (1993) Seyler et al (1998) and Oliveira (2000) Traditionallyremote sensing has been used to classify soil units through photo-interpretation ordigital image processing Combining remotely sensed information with ancillaryinformation such as thematic maps or vegetation cover can yield significant improve-ments (Wilcox et al 1994 Cialella et al 1997 Wanchang et al 2000)

Recent developments in hyperspectral remote sensing offer the potential of signifi-cantly improving data input to predictive soil models Hyperspectral sensors such asthe Airborne VisibleInfrared Imaging Spectrometer (AVIRIS) measure a contiguousspectrum in the visible and NIR and thereby better characterize atmospheric andsurface properties (Goetz et al 1985) The large number of spectral bands permits directidentification of minerals in surface soils For example Clark and Swayze (1996)mapped over 30 minerals using AVIRIS at Cuprite Nevada Palacios-Orueta and Ustin(1996) showed that enhanced spectral information was suitable for discriminating evensubtle spectral changes associated with differences in organic matter and iron contentOther examples of the application of AVIRIS to aid soil mapping include Palacios-Orueta et al (1998) Okin et al (1998) and Roberts et al (1998)

Sensors that operate in the microwave portion of the electromagnetic spectrum havealso shown promise in soil mapping research Microwave sensing can be broadlydivided into active (eg radar) systems and passive systems and are capable ofpenetrating the atmosphere under virtually all conditions offering a significantadvantage over visible and near-infrared spectroscopy (for a general overview ofmicrowave sensing see Lillesand and Kiefer 1994 chapter 8) Synthetic aperature radar(SAR) is one example of an active system SAR has been used to aid soil propertymapping such as soil salinity (Metternicht 1998) and soil moisture (Engman andChauhan 1995 Narayanan and Hirsave 2001) Active radar systems can also bedesigned to collect data at varying look angles providing the opportunity for theacquisition of stereo radar images Such images can be used to produce high resolutionand extremely accurate DEMs (eg Fang 2000) A similar active sensing system isLiDAR (light detection and ranging) which uses pulses of laser light rather thanmicrowave energy to illuminate the surface (see Bunkin and Bunkin 2000 for a reviewof applications to soil mapping research) While passive microwave systems haveseemed to receive less attention in the literature a few examples of soil mapping appli-cations can be found (see Kleshchenko et al 2000 and Laymon et al 2001) Regardlessof the type of system remote sensing data and derived products are potentially usefulexplanatory variables in predictive soil mapping models

4 Fuzzy logic

Fuzzy set theory or fuzzy logic provides an alternative conceptual paradigm withinPSM research The use of this theory has increased greatly in the last few years making

it an important component of PSM Fuzzy logic is an alternative to Boolean logic thatattempts to recognize the concept of partial truth (Brule 1996) Dr Lotfi Zadeh (1965)introduced the concept and accompanying mathematics in his seminal work lsquoFuzzysetsrsquo The theory permits partial class membership in contrast to traditional set theorywhere set memberships are crisp and binary (ie a soil sample is either completelyType A or it is not at all Type A) Central to the fuzzy concept is the idea that objects innature rarely fit exactly the classification types to which they are assigned (Zadeh1965) Rather they show varying signs of similarity to multiple classes (ie an observedsoil pedon often resembles more than one of the defined soil series within the area) Byusing fuzzy membership values (ranging from 0 nonmembership to 1 totalmembership) within predictive soil models to express degrees of similarity generaliza-tion problems associated with classification schemes (filtering of information) areminimized and the complex nature of soil data is allowed to propagate through themodelling process Similarity values between 0 and 1 are not comparable to proportionsand need not add up to 1 Within Boolean logic probability statements refer to thelikelihood of an outcome the soil sample is either one series or another With fuzzylogic a given sample is not definitively a member of the subset of any one particularseries Fuzzy logic is especially useful in soil research because of the continuous andcomplex nature of the soil landscape It serves as an important alternative to thesubjective rigidity imposed on soils data by Soil taxonomy Several recent articlesprovide a thorough review of the use of fuzzy sets in soil science (Burrough 1989McBratney and Odeh 1997 Burrough et al 1997 De Gruijter et al 1997)

Within PSM research two different approaches to creating continuous classes usingfuzzy logic exist The first is based on the fuzzy-k-means classifier which partitionsobservations in multivariate space into natural classes This approach is similar tocluster analysis and numerical taxonomy but the resulting classes are continuous witheach observation assigned a fuzzy membership value that characterizes its degree ofsimilarity to each individual class The concept has been integrated into geostatisticalmethods and will be discussed in more detail below (see Section IV1) The secondapproach is known as the Semantic Import model (SI) and is used in situations whenclassification schemes are pre-defined and class limits are relatively well understoodThe SI model is commonly used in concert with expert knowledge and will bediscussed in the expert systems section (see Section IV4)

IV Recent advances in predictive soil mapping

Within the last decade many authors have sought to model the soil landscape using avariety of methods Literature in this field could be summarized many different waysbut we concentrate on the literature that directly addresses the goals of predictive soilmapping stated in the introduction (see Table 1) Therefore we will review research thatattempts to exploit the relationship between quantifiable landscape indices and soilcharacter in order to model the soil landscape in a more continuous and thereforerealistic manner

The research reviewed here is distinguished from decades of previous researchdocumenting the correlation between landscape position and soil attributes (reviewedby Hall and Olsen 1991) That body of research is informative but not useful for

P Scull et al 179

180 Predictive soil mapping a review

Table 1

Selected

recen

t literature on pred

ictive

soil m

apping an

d m

apping

(cited in

this article) d

escribing the mod

ellin

g metho

dused the

dep

enden

t variables used a

nd the

env

iron

men

tal v

ariables (ex

plan

atory) used in

the

mod

els

Stud

yMod

ellin

g metho

dDep

ende

nt variables

Environm

ental variab

les

Goa

ls atta

ined

a

Bell et

al 20

00Line

ar and

exp

onen

tial

Total soil orga

nic

Slop

e curvature aspe

ct

ECx

regression

carbon

hillslope

position

Burge

ss and

Web

ster

Punc

tual and

block

Na co

nten

t co

ver

Non

eC

1980

ab

kriging

loam

thickn

ess

ston

e co

nten

t

Burroug

h 198

9Fu

zzy mathe

matical

ndashndash

rpmetho

ds

Burroug

h et

al 19

97Con

tinuo

us classificatio

nndash

ndashrp

Castrigna

no eta

l 200

0Factorial kriging

CEC

pH N

P K

Non

e ex

plicitly used

Cx

Na

Cialella

eta

l 199

7Classifica

tion tree

Drainag

e class

Elev aspe

ct NDVI

Ec

Coo

k et

al 19

96Baysian

rule-ba

sed metho

dsOrgan

ic m

atter

Slop

e aspec

t wetne

ssEcX

inde

x

Dale et

al 19

89 (a review

)Ex

pert systems

ndashndash

ndash

Ellis 19

96Decision tree

ana

lysis

Soil erosion class

Slop

e aspec

t wetne

ssEc

neutral ne

tworks

inde

x flow le

ngth and

accu

mulation Lan

dsat

TM tree

cov

er

Gessler 1

996

A large

variety of statistic

Field an

d labo

ratory

A variety of digital

EC

metho

dsco

llected

phy

sica

len

vironm

ental da

tach

emical and

morph

olog

ical soil

prop

ertie

s

Gessler eta

l 199

5Line

ar and

logit reg

ression

A horizon

and

solum

Curvature CTI topo

EC

depth E horizon

positio

npresen

ce

Goo

vaerts 19

92Factorial kriging

Total carbon

Non

eC

P Scull et al 181pH

N CEC

extractab

lecatio

ns (K Ca M

g)Hartemink et

al 20

01ndash

ndashndash

rp

Heu

velin

k an

d W

ebster 20

01ndash

ndashndash

rp

Hew

itt 19

93ndash

ndashndash

rp

Indo

rante et

al 19

96ndash

ndashndash

rp

King et

al 19

99Lo

gistic regression

Presen

ceabsen

ceSlop

e aspec

t po

t solar

Ec

Non

calc c

lay-loam

energy

Kno

tters eta

l 199

5Kriging

co

-kriging

So

ft layer de

pth

Hillslop

e po

sitio

nCx

regression

kriging

Laga

cherie and

Holmes 19

97Classifica

tion tree

Map

ping

unit

Geo

logy variou

s topo

Ec

indices

Laslett e

tal 198

7Kriging

splin

es tren

dpH

Non

ec

surfac

e nea

rest neigh

bor

McB

ratney 1

992

ndashndash

ndashrp

McB

ratney

eta

l 199

1Block kriging

Clay co

nten

tNon

eC

McB

ratney

eta

l 200

0ndash

ndashndash

rp

McB

ratney

and

Ode

h 199

7Fu

zzy sets in

soil scienc

endash

ndashRp

McB

ratney

and

de Gruijter

Fuzz

y-k-mea

ns w

ithFu

zzy classes

Field co

llected

phy

sical

C19

92ex

tragrade

sch

emical and

morph

olog

ical soil

prop

ertie

s

McC

rack

en and

Cate 198

6Artificial intellig

ence

ndash

ndashrp

expe

rt systems

McK

ensie an

d Austin

19

93Gen

eralized

line

ar m

odels

Clay co

nten

t CEC

Slop

e relief land

form

ec

(logit)

pH EC

COLE

slop

e po

sitio

nbu

lk den

sity and

othe

rs

Moo

re eta

l 199

3Line

ar reg

ression

A horizon

dep

th O

MSlop

e w

etne

ss and

strea

mEC

and P co

nten

t pH

power in

dices aspect

curvature

182 Predictive soil mapping a reviewTa

ble 1

Con

tinu

ed

Stud

yMod

ellin

g metho

dDep

ende

nt variables

Environm

ental v

ariables

Goa

ls atta

ined

a

McK

ensie an

d Ryan 199

9Reg

ression tree and

linea

rSo

lum dep

th P an

dElevation slope

CE

regression

N con

tent

curvature CTI

contribu

ting area do

wn-

slop

e mea

ns for slope

clim

ate da

ta Presco

ttInde

x G

amma

Rad

iometry La

ndsat T

M

and Geo

logy unit

Ode

h et

al 19

92ab

Fuzz

y-c-mea

ns and

Fuzzy classes

Field-co

llected

phy

sica

lC

kriging

chem

ical and

morph

olog

ical soil

prop

ertie

s

Ode

h et

al 19

94

Reg

ression kriging

co

-So

lum dep

th de

pth

Slop

e aspect cu

rvature

eC

1995

kriging regression

to bed

rock gravel

kriging

and clay con

tent

Skidmore et

al 19

91Bayesian expe

rt system

Soil land

scap

e un

itVeg type

wetne

ss in

dex

EcX

1996

grad

ient terrain po

sitio

n

Voltz and

Web

ster 19

90Kriging

cu

bic splin

eClay co

nten

tNon

eC

Web

ster 1

994

Dev

elop

men

t of

ndashndash

rppe

dometrics

Zhu

199

7ab

Fuzz

y logic expe

rt system

Soil series A

horizon

Elev pm aspe

ct c

anop

yECX

Zhu

and

Ban

d 199

4(SoL

IM)

depth in

dividu

alco

verage

grad

ient

Zhu

eta

l 199

7series m

aps

curvature

Not

es

a Letters refer to the de

gree

to w

hich

the

goa

ls of PS

M defined

in the

introdu

ction are achiev

ed

Soilndash

e nvironm

ent relatio

ns utilized

(letter E) be

tter represen

tatio

n of soil c o

ntinuity (C

) an

d ex

pert kno

wledg

e utilize

d (X) c

orrespon

d to goa

ls 1 2 and

3respec

tively from

the

introdu

ction and

cap

ital lette

rs (E C X) ind

icate the metho

d is relatively more successful th

an th

ose metho

ds den

oted

by lower

case le

tters (e c x

) rp ind

icates rev

iew pap

ers

P Scull et al 183

predictive mapping because landscape position is never quantified (rather positionwas often qualitatively defined eg toe-slope) and thus the documented relationshipscannot be generalized using environmental data and digital elevation models to predictsoil character at unvisited sites

Table 1 documents modelling methods model variables and the extent to which thereferenced research satisfies the previously defined goals of PSM (lsquoGoals attainedrsquocolumn) Soil-environment relations utilized (letter E) better representation of soilcontinuity (C) and expert knowledge utilized (X) correspond to goals 1 2 and 3respectively from the introduction and capital letters (E C X) indicate the method isrelatively more successful than those methods denoted by lower case letters (e c x)Citations with no letters present within the lsquoGoals attainedrsquo column do not address theaforementioned goals For example the Cialella et al (1997) received a rating of lsquoEcrsquomeaning that the methods employed successfully utilized environmentalndashsoil characterrelations (E) and somewhat successfully presented a better method of representing soilcontinuity (c) The ratings are provided simply to help organize the literature that wasreviewed Review papers are included within the table denoted by lsquorprsquo Geostatisticalmethods are not included in the table because they have been comprehensivelysurveyed elsewhere (Odeh et al 1994 Burrough et al 1997 McBratney et al 2000Heuvelink and Webster 2001) and because the objectives and assumptions of geostatis-tical methods differ slightly from other PSM research We briefly outline thesedifferences in the following section

1 Geostatistical methods

Geostatistics are a subset of traditional statistics that deal primarily with spatial dataand account for spatial autocorrelation using kriging as the spatial interpolator Theconcept is based upon the theory of regionalized variables which was mainlydeveloped by Matheron (1963) and Krige (1963) Kriging is a form of weighted localaveraging that uses a measure of spatial dependence the variogram to determine theweights applied to the data when computing the averages Geostatistical methods havebeen used in predictive soil mapping research to spatially interpolate soil propertyvalues at unmeasured sites from field-collected data

Burgess and Webster (1980a b) were the first to introduce ordinary kriging to the soilcommunity and since that time an enormous amount of work has been published Forexample ordinary kriging has been used to interpolate many different soil propertiesincluding pollution trace element deficiencies salinity and fertility (Heuvelink andWebster 2001) Ordinary kriging has been criticized for a variety of reasons Forexample Laslett et al (1987) reported that several authors had criticized geostatisticsbecause kriging is a global rather than local technique failing to take into accountknowledge of soil materials and processes Other authors have criticized geostatisticsbecause they are excessively data dependent requiring a large number of closelyspaced data points (Zhu 1997a) As Webster and Oliver (1992) suggest in excess of ahundred samples may be needed to use geostatistics at the field scale because of highspatial variability of soil in some areas Geostatistics also assume spatial autocorrela-tion which sometimes may be a poor assumption in complex terrain where abruptchanges in soil-forming factors occur (McBratney et al 2000) Ordinary kriging by itself

184 Predictive soil mapping a review

does not satisfy two of the three goals of PSM presented in the introduction ndash it does notadequately incorporate expert knowledge and it does not exploit the relationshipbetween environmental variables and soil properties

Ordinary kriging has been modified in a variety of ways to better incorporateancillary data and known soilndashlandscape relationships Block kriging involvesdetermining estimates over meaningful areas rather than at specific points (Burgess andWebster 1980a McBratney et al 1991) Using this method a study area can be stratifiedinto different regions that are reflective of the pedogenetic processes at work In orderto accommodate a trend within a dependent soil variable universal kriging has beenused (Webster 1994) Kriging with external drift is similar to universal kriging but ituses an ancillary variable to represent the trend (McBratney et al 2000) Co-krigingtakes advantage of correlation that may exist between the variable of interest and othermore easily measured variables (Odeh et al 1995) Regression kriging involvesspatially interpolating the residuals from a non-spatial model by kriging and addingthe result to the prediction obtained from that model (Goovaerts 1997 Castrignano etal 2000) Factorial kriging is another method to integrate multivariate data into thestandard kriging routine to extrapolate soil data (Goovaerts 1992) Many authors havecompared these various methods (Laslett et al 1987 Voltz and Webster 1990 Odeh etal 1994 Knotters et al 1995)

Fuzzy logic has been used with geostatistics by various authors to produce new kindsof fuzzy soil maps with continuous classes (Burrough 1989 McBratney and DeGruijter1992 Odeh et al 1992a and reviewed by McBratney and Odeh 1997) The processentails kriging the matrix of membership values determined by the fuzzy k-meansclassifier resulting in a continuous soil surface where individual locations are allowedto belong to more than one class and no rigid boundaries are designated to separate thesoil into discrete units or entities The results of such analysis can be used to assess thepedologic process validity of soil taxonomy by determining whether soils grouptogether into classes that are similar to taxonomic types Fuzzy classes wouldpresumably reflect the main pedologic features within a given area In this sense theclassification is quantitative whereas soil taxonomy is rooted in qualitative discrimina-tion A comparison between the two could provide insightful and help assist soiltaxonomy in making a classification that is more indicative of underlying soil processes

Geostatistics in soil research were originally introduced to quantitatively assess soilvariability within soil mapping units (McBratney et al 1991) in response to criticisms inthe early 1970s that soil unit composition was not well quantified (Beckett and Webster1971) In this regard geostatistics have been very useful having served well the originalgoals set forth by Burgess and Webster (1980b) when they drew kriging to the attentionof soil scientists as a means of spatial prediction At the field scale soil variation islargely due to the effect that topography has on soil genesis Geostatistics have beensuccessfully applied in such environments by using terrain attributes as ancillary datawithin many of the kriging routines described above Such quantitative within-unitvariability of soil properties is very useful in the field of precision agriculture and othersituations (eg pollutants) where very detailed soil attribute information is needed atthe field scale (Heuvelink and Webster 2001)

However geostatistics have not been applied in a wide variety of environments or atlarger scales In order to be successfully applied in different environments geostatisticswill likely require a different suite of ancillary data For example remote sensing data

P Scull et al 185

could be used in arid regions where soil toposequences are less well expressed Atlarger scales of prediction selection of different sets of ancillary variables is requiredbecause different processes define soil character at different scales The most obviousexample is that of climate which may control soil distribution at large scales(continental) but has little explanatory power at the field level Regardless of whetheradequate ancillary data exist the amount of data required to use geostatistics forlandscape-level prediction would be extremely difficult and costly to collect given thestrict sampling protocol required to characterize spatial dependence It is also unclearat what landscape scale soils exhibit spatial autocorrelation

Geostatistical approaches do provide a means of creating continuous soil attributesurfaces to better represent soil continuity (Goal 2) and they can be used to exploit therelationship between environmental variables and soil properties in order to moreefficiently collect soil data (Goal 1) However they do not sufficiently utilize expertknowledge (Goal 3) as no attempt has been made in geostatistical approaches todirectly integrate expert knowledge Fundamentally kriging is a process of interpola-tion designed to predict attribute values in between locations of measured samples Inthis sense geostatistics represent a middle ground between pure interpolation (egnearest neighbour type classifier) in which only measured points for the variable ofinterest are used to determine unknown values and other predictive models thatprimarily use soilndashenvironment correlation to create predictive maps

2 Statistical methods

Statistical methods can be used to exploit the relationship between quantifiablelandscape indices and soil properties to create predictive soil maps For exampleMcKensie and Austin (1993) used a regression to account for a large percentage ofvariation for many soil characteristics (A horizon clay content CEC EC pH bulkdensity and COLE B horizon clay content CEC ESP EC pH bulk density and COLE)using a variety of predictor variables (slope presence or absence of impeding layerrelief landform topographic position) Their results confirm the hypothesis of MooreID et al (1991) that soil character is related to quantifiable landscape indices Howevertheir methods do not provide inference of soil properties at unmeasured sites frommapped environment data because the topographic variables were measured in thefield Linear regression has also been used with terrain variables derived from a 15-mDEM in northeastern Colorado to predict soil attributes (organic matter contentextractable phosphorous pH and texture) at unvisited sites (Moore et al 1993) In thatparticular study 50 of the variance of A-horizon thickness was explained by slope andthe wetness index Gessler et al (1995) also used regression to model A-horizonthickness from topographic variables in southeastern Australia (plan curvature andwetness index r2 = 063 P = 0001) They modeled solum depth and used logisticregression to model E horizon presenceabsence Elsewhere logistic regression hasbeen used to model the presenceabsence of noncalcareous clay loam horizon in centralFrance using terrain attributes from a 20-m DEM (King et al 1999) Exponentialregression has been used to model soil organic carbon using terrain variables (Bell et al2000) in glacial outwash soils in east-central Minnesota Generalized additive models(GAM) have been used less frequently in PSM research Gessler (1996) used a GAM

186 Predictive soil mapping a review

model to predict total soil carbon A horizon depth and solum depth using a variety ofenvironmental predictors (slope elevation wetness index mean annual temperatureprecipitation and radiation)

This small body of research opened the door to more complex methods by demon-strating the existence of quantifiable relationships These authors were able to producesoil attribute maps using raster data models whose scale was dependent upon the gridcell resolution of the environmental data They were successful at exploiting the rela-tionship between quantifiable topographic attributes and soil profile character (Goal 1)The continuous soil attribute surfaces they produce also better represent soil continuitythan the choropleth soil maps produced by traditional soil survey (Goal 2) Howeverthe bulk of these methods (excepting GAMs) are limited by their assumed linear rela-tionship between soil and topographic attributes their assumptions of normallydistributed data and their high data requirements Standard statistical procedures arealso not flexible enough to allow robust integration with a variety of potential datasources such as expert knowledge (Goal 1)

Statistical methods do demonstrate in a quantitative manner that terrain analysis canbe used to predict soil attributes in relatively small areas with homogeneous parentmaterial A large proportion of the research using statistical methods was conducted insemi-arid landscapes at small scales (the largest study area of the entire group was~2000 ha) Obviously for statistical approaches to be most effective they need to bemore universal As such they need to be tested andor developed at larger scales andin more diverse landscapes

3 Decision tree analysis (DTA)

The use of decision tree analysis is just beginning to be explored in predictive soilmapping research although it has been used successfully in the related field ofpredictive vegetation mapping since the early 1990s (Lees and Ritman 1991 MooreDM et al 1991 Franklin 1998) DTA is a form of divisive classification The process oftree modelling involves successively partitioning data (called recursive partitioning inthe tree modelling literature) into increasingly homogeneous subsets which once thepartitioning has ceased are called terminal nodes (Lees and Ritman 1991) Splits orrules defining how to partition the data are selected based on information statistics thatdefine how well the split decreases impurity within the data set (Clark and Pregibon1992) Splits are based on threshold values of an explanatory variable selected bycomparing the increase in resulting purity of node membership for all possiblethresholds and variables The process is iterative growing from the root node (thecomplete data set) to the terminal nodes in a dendritic fashion (Friedl and Brodley1997) Once the tree has been constructed (or grown) it encodes a set of decision rulesthat describe the data partitioning process These rules can be used to classify or predictother data sets (Moore DM et al 1991) Pruning the tree is often necessary to preventthe tree from being overfit to the sample data and to reduce tree complexity Pruningentails combining pairs of terminal nodes into single nodes and can be accomplishedusing cross-validation which yields an initial indication of how large a tree makesrobust predictions (Safavian and Norvig 1991) Cross-validation involves systematical-ly removing portions of the data set and running the remaining sample through the tree

in an iterative manner eventually yielding estimates of the misclassification rates foreach class each node and the whole model (Breiman et al 1984) In this mannerdifferent sized trees can be compared in terms of parsimony

The term DTA is used to collectively refer to all types of tree-based modelling (theword lsquodecisionrsquo is used because it is descriptive indicating that the analysis eventuallyleads to a set of decision rules defining data partitions) The term should be distin-guished from classification tree analysis because the latter refers specifically to DTAwhere the response variable is categorical The term CART (classification andregression trees ndash Breiman et al 1984) is sometimes used but strictly speaking refers tospecific software Friedl and Brodley (1997) provide a review of the decision treealgorithms They divide types of DTA into two classes (1) homogeneous decision treesfor which a single algorithm is used to estimate each split (eg CART) and (2) hybriddecision trees (HDT) for which different splitting methods can be used at differentpoints in the tree (eg Quinlan 1993) They further divide homogeneous decision treesinto univariate (UDT) where single features of the input data define splits and multi-variate decision trees (MDT) where multiple features of the input data can define splitsAccording to this naming convention no distinction is made between methods withdifferent types of response variables although all methods can be used with bothcategorical and continuous response variables In a comparison of these various typesof algorithms on a variety of data sets Friedl and Brodley (1997) found that HDT hadthe highest classification accuracy

The overall aim of DTA is to design a set of predictive rules (eg if geology type Athen soil type B) developed from training data which can then be applied to ageographic data base to predict the value of a response variable (Michaelsen et al 1994)Therefore DTA explicitly uses soilndashlandscape correlation in model development (Goal2) The technique appears promising in soil research but needs to be further tested asnot many of the above types of DTA have been tested in the PSM literature in fact onlyunivariate approaches have been employed For example Lagacherie and Holmes(1997) successfully used univariate DTA to model a categorical response variable soiltype within a training set and then assuming that the training set was representativeextrapolated the model to a much larger region Their work is interesting becausealthough not a single sample came from the area they eventually mapped they wereable to produce a soil map that was more accurate (74 versus 69) than the existingmap produced from traditional methods Cialella et al (1997) also used univariate DTAto predict soil drainage class from a variety of terrain attributes and remotely senseddata They predicted soil drainage class with an average of 78 accuracy ndash impressivegiven that the variation accounted for by a typical soil survey ranges from about halfthe total variance for physical attributes to less than one-tenth for some soil chemicalattributes (Gessler et al 1995)

DTA has been compared with other approaches by several authors In the applicationof erosion modelling DTA results were similar to Artificial Neural Networks (Ellis1996) Both methods achieved high training accuracy (as measured by the KappaStatistic) but in terms of prediction accuracy both methods performed poorly Gessleret al (1995) compared DTA with generalized linear models (GLM) and generalizedadditive models (GAM) to predict A-horizon thickness and concluded that GLM waspreferable to both DTA and GAM McKensie and Ryan (1999) compared regressiontrees and standard linear regression to predict soil properties (total solum depth soil

P Scull et al 187

188 Predictive soil mapping a review

phophorus and soil carbon) using a large variety of predictors (elevation slopecurvature wetness index climate variables geology unit Landsat TM data andGamma radiometric data) and found that different methods work best for differentproperties and overall success hinged on the strength of the relationship between soiland environmental variables Their study of a 50 000 ha forested landscape in southernNew South Wales Australia is one of the few studies at such a small map scale (largearea) or in forested landscape The most extensive comparison of techniques involvingDTA was conducted by McBratney et al (2000) They compared regression techniques(GLM GAM DTA-regression tree) geostatistical techniques (kriging and heterotopicco-kriging) and a hybrid technique (regression kriging) DTA was found to be thepoorest performing of the regression techniques because of the unrealistic predictionsurface generated by the DTA model DTA has also been criticized by other authorsbecause of the stepped prediction surface (Gessler 1996) This phenomenon can beespecially apparent in situations where predictor variables have different resolutionsThe extent to which DTA yields a better spatial representation of soil continuity is afunction of the scale and type of predictor variables used and therefore varies withindividual models

DTA is gaining widespread popularity as a means to develop prediction rules thatcan be rapidly and repeatedly evaluated (Cialella et al 1997 Franklin et al 2000) DTAprovides the following advantages over standard statistical techniques (1) it is easier tointerpret when explanatory variables are both nominal and continuous (2) it isinvariant to monotone re-expressions (transformations) of predictor variables (3) itdeals more satisfactorily with missing data values and outliers (4) it is more adept atcapturing nonadditive and nonlinear behaviour (5) it doesnrsquot make any assumptionsabout data distribution and (6) it is easily updateable as more data are collected(Moore DM et al 1991) The DTA model framework is especially appealing becauseof its capability to integrate a wide range of data sets as explanatory variables

DTA offers a unique opportunity for interaction between soil experts and soilmodellers because the output of the model is a set of rules that can be pedologicallyinterpreted by the soil expert In this sense expert knowledge is used in an implicitmanner in DTA (somewhat effectively achieving Goal 3) While these rules can often beexceedingly complex at minimum the expert can decide whether initial splits makesense given their understanding of the landscape The success of DTA results oftenhinges on the ability of the modeller to make key decisions during the model buildingprocess there is unfortunately no definitive way to determine the most optimal treeThe aid of the soil expert can potentially elucidate this problem

4 Expert systems

A variety of expert system approaches to PSM have been developed to utilize expertknowledge The purpose of such methods is to exploit the information the soil surveyoraccumulates while working in the field by integrating such knowledge into thepredictive model (McCracken and Cate 1986) Unlike the majority of the researchreviewed thus far the dependent variable in many expert systems models is often soiltaxa or mapping unit This apparent disadvantage of expert systems (using classifica-tion to characterize soil continuity) does make them easier to integrate into traditional

soil survey In addition several authors have developed methods to developcontinuous soil property maps from the output of expert system models designed topredict soil unit occurrence

Expert systems are composed of data (information on spatial environmentalvariables eg topography climate etc) a knowledge base (rules and facts related tosoil variation supplied from the soil surveyor) and an inference engine (whichcombines data and the knowledge base to infer logically valid conclusions) (Skidmoreet al 1996) Expert systems differ from conventional models in two ways (1) they storeand manipulate qualitative information (allowing them access to information thatcannot normally be used in other modelling frameworks) and (2) they are structuredas meta-models (the knowledge is separated from the model) (Davis 1993) This allowsthe model to selectively choose which information is relevant at various stages of themodelling process and it allows for information to be easily updated Davis (1993)reviews the application of expert systems to environmental modelling research conclu-ding that the technique is becoming more widely accepted He further states that theapplication of expert systems is constrained by an absence of fundamental knowledgefor rule generation a problem that would appear less relevant to soil mapping giventhe amount of untapped expert knowledge accumulated by the soil surveyor

The first mention of the use of expert systems in pedology was in a paper presenta-tion at the Northeast Committee Soil Survey Conference 1984 (Flach 1985) In herpaper Flach (1985) hinted that recent developments in computer science especiallyexpert systems and artificial intelligence could make modelling a practical mappingtool for soil scientists in the near future A year later McCracken and Cate (1986) hopedto encourage soil scientists to explore expert systems and its application to soil sciencethrough an optimistic article they wrote on the potential use of expert systemsHowever little research was actually conducted in the late 1980s and expert systemshad not yet begun to fulfil the lofty goals set forth by early practitioners (Dale et al1989)

In the early 1990s expert systems approaches to predictive soil mapping began toappear in the literature Skidmore et al (1991) used a Bayesian expert system to mapforest soil into different classes and their results compared favourably with availablesoil maps and actual field-collected data Their methods successfully incorporatedsurveyor knowledge and remotely sensed and digital terrain attributes but failed tobetter characterize continuous soilndashlandscape variability because their final productwas a choropleth map Skidmore et al (1996) revisited their earlier research andassessed the mapping accuracy of their results They found that the soil map producedby the expert system achieved an overall accuracy of 698 (sample size = 53) while themap derived from conventional methods had an accuracy of 736 In addition toknowledge provided by a soil scientist Skidmore et alrsquos (1996) expert system usedvegetation derived for aerial photography and topographic variables derived from a 10-m DEM (wetness index topographic position and slope)

Using an expert system Cook et al (1996) successfully produced a continuous soilproperty map for organic matter content using wetness index aspect and slope asexplanatory variables Their methods were somewhat inefficient because they requireda separate expert system for each soil property of interest However their researchrepresents the only example of expert systems used to directly predict the spatial dis-tribution of a soil property All of the other examples of expert systems in the literature

P Scull et al 189

190 Predictive soil mapping a review

deal with soil type or class The use of expert systems to map soil properties needs tobe explored further

The expert systems discussed thus far have all used Boolean logic within theirtheoretical framework whereby an observation can belong to one and only one classand the soil properties of that class are assigned to the observation With thedevelopment of fuzzy logic and the semantic import model (the second fuzzy logicapproach to creating continuous classes) the opportunity exists to develop fuzzy logic-based expert systems Such systems can be used in conjunction with expert knowledgein situations where experts have a good qualitative idea of how to group data but havetrouble dealing with observations that are not well represented by rigid classificationschemes (Burrough and McDonnell 1998) This method can be particularly useful insituations when taxonomic schemes have been previously developed as is the case insoil taxonomy Several examples of this type of approach were published in the mid-1990s by A Zhu and colleagues (Zhu and Band 1994 Zhu et al 1996 1997 Zhu1997ab) Such systems proved useful for mapping soil at unvisited locations usingsurveyor knowledge and were also capable of producing continuous soil propertymaps The use of fuzzy logic within the theoretical framework of the expert systemallows the complex nature of soil to propagate through the modelling process neversubjected to classification schemes that filter out potentially useful lsquonoisersquo The use offuzzy logic also gives the soil surveyor more latitude during the interview processwhen the knowledge base is defined for the expert system Zhu et al (1997) developeda fuzzy logic-based expert system called SoLIM to determine the similarity of eachgrid cell in a study area to the various taxonomic mapping units delimited by the soilsurvey Continuous soil attribute maps were calculated using the similarity values andtheir relative soil survey determined attribute values (effectively achieving Goal 2) Theresulting data (which proved more accurate than soil survey data once field checked)consists of a raster grid whose resolution is determined by the resolution of the inputenvironmental and digital elevation data As noted previously this type of data modelis more applicable to environmental modelling than the choropleth map (Burrough andMcDonnell 1998)

Expert system approaches to PSM are capable of exploiting soil surveyor knowledgeby developing rule-based systems that imitate the surveyorrsquos conceptual model of soilvariability (the primary focus of Goal 3) The method would appear extremely usefulfor mapping projects (such as those conducted by the NRCS-NCSS) where fieldwork isinitially conducted to determine soilndashlandscape relations Expert system developmentcould be directly inserted into the traditional soil survey mapping approach as asubstitute for the step where the surveyor converts hisher conceptual model into achoropleth map Rather that knowledge could be incorporated into the expert systemwhich could be used to predictively map soil The resulting raster map would be morescientifically based and explicit than the hand-drawn choropleth maps of the past Itsscale would not be limited to that of the aerial photography but rather to the scale ofthe environmental data

Despite these advantages the expert system approach has some drawbacks Becausethe method is dependent upon expert knowledge it cannot be applied whereenvironmentndashsoil relations are poorly understood Of course this criticism can be madeof all PSM models as inductive or deductive knowledge of soilndashenvironment relationsis a prerequisite for PSM However expert systems are deductive models and as such

are not driven by any specific field-collected data (although presumably the soil experthas field experience in the mapping area) Expert systems donrsquot afford the opportunityto first statistically document landscapendashsoil relations and then extrapolate the resultsbecause expert systems do not directly use sample soil data (lsquohard datarsquo) to determinesoilndashlandscape relations Thus the utilization of the relationship between environmen-tal variables and soil properties (Goal 1) is only indirectly achieved Expert systemsapproaches have been demonstrated to be extremely effective in a small number of casestudies The possibility of satisfying all three goals discussed in this review makesexpert systems a predictive soil mapping method that needs to be further tested ndash bothgeographically and across different scales of analysis

V Conclusion

Most of the predictive soil mapping research outlined in this review was conducted atvery large map scales (over small areas) In fact the majority of the research wasconcerned with assessing the spatial variability of soil character within individual fieldsor across soil toposequences The primary driving force behind this type of research hasbeen the need to provide accurate soils information for agriculture and ecologicalmodels It is clear that terrain attributes are powerful predictors at the local scaleGeostatistical tools have been successful at using terrain attributes and the spatialdependence of soil properties to interpolate between existing data points withinindividual fields Across soil toposequences statistical approaches provide a usefulmeans of predicting soil character PSM research has been most successful at the fieldscale because many of the soil-forming factors are held constant For example the neteffect of four of Jennyrsquos five soil-forming factors (climate organism parent material andtime) was minimal within many of the studies reviewed While some research hastackled larger areas of study there exists a deficiency in our ability to predictively mapsoils at smaller map scales Since the distribution of soil is scale-dependent differentPSM methods and predictors are likely to work better at different scales Focus in thefuture must continue to move toward working over larger spatial extents of study inorder to produce landscape-scale soil information

Further a large proportion of the research was conducted in semi-arid gentlysloping agricultural landscapes Humid forestlands mountainous regions and desertshave received little attention As such PSM methods need to continue to be testedandor developed in a wider variety of landscapes where spatial soil distributions canbe more complex Different methods will likely be successful to different degrees indifferent environments Whereas terrain attributes are dominant predictors of soilcharacter across toposequences in prairie lands other predictors are likely necessary inareas where soils do not develop into clear toposequences For example remote sensingdata has been a useful predictor of soil chronosequences in desert landscapes wheresurface appearance is often related to soil character Surficial geology is often reflectiveof soil character in mountain regions where soils are thin and significant bedrock existsat the surface Focus in the future must continue to determine which methods andpredictors work best in which environments

Since the most useful PSM approach will vary across spatial scales and environmen-tal gradients the method used should be driven by the mapping objectives of the

P Scull et al 191

192 Predictive soil mapping a review

project Whereas the traditional soil survey of the past was expected to meet the needsof a diverse group of end users PSM methods will vary given the objectives of thesurvey As a result the end users of the soil data need to play a more active role in thesurvey process Because precision agriculture and large-scale mapping has been theprimary focus of PSM there exists a deficiency in our ability to predictively map soilsfor the purpose of general land use planning and management Even though lessdetailed soil information is needed for these purposes the development of PSMmethods is complicated by the fact that soils exhibit complex spatial variability atsmall map scales where soil-forming environments vary greatly from one location tothe next

A number of alternative methods of characterizing the continuous nature of the soillandscape have been developed Thus far most of the PSM research has provided soilinformation in a nonobject form (as opposed to defining soil types as independententities) Soil data have been generated organized and presented in the form of eitherisorithmic maps or fine-scale raster grids Both of these data models are field-viewmodels of geographic space which allow the soil to be perceived as a constantlyvarying surface Two distinct approaches have been employed (1) mapping individualsoil properties and (2) mapping continuously varying (fuzzy) soil classes Mappingindividual properties is the most common approach and will likely continue todominate PSM research The use of fuzzy soil classes in the literature is less commonand more difficult to be integrated into standard mapping procedures (such as use ofsoil taxa) because the concept radically differs from the traditional view of the soillandscape However fuzzy soil classes (generated using the fuzzy-k means approach) dohave the potential to help further advance Soil taxonomy by identifying taxonomicclasses that are more reflective of pedologic processes at work

Expert systems have been greatly underutilized in PSM research especiallyconsidering how effective a small number of case studies have been Expert systemshave the potential to satisfy successfully all three goals discussed throughout thereview They also have the potential to bridge the gap between traditional approachesand PSM methods because field soil scientists do not have to change their conceptualapproach to mapping They can still conduct field reconnaissance to determinesoilndashlandscape relationships Afterwards though they can use expert systems to betterexploit the knowledge they have garnered Such an approach could help ease thetransition from traditional soil survey to more scientifically explicit methods Expertsystems are also the most fruitful approach to utilizing a wealth of data that has alreadybeen collected in a nonexplicit manner the wealth of expert knowledge (in the form ofthe conceptual model) that senior soil mappers have accumulated

Focus in the future must also move in the direction of operability The utility of manyPSM approaches has been clearly demonstrated yet changes in how soils are mappedand perceived has been slow In the USA various members of the National CooperativeSoil Survey have called for a change in the philosophy of survey (from mappingdiscrete soil types to mapping continuous varying soil properties McSweeny et al1994) but little change has actually taken place To realize the potential of PSM in con-tributing to an overall change in standard soil mapping procedures practitionersworking in the field of PSM need to form working relationships with field soilscientists PSM methods need also to be presented in a manner that is comprehensibleto the soil science trained field mapper Predictive soil mapping is a relatively recent

phenomenon yet much progress has been made In the process of mapping soilknowledge will continue to advance

P Scull et al 193

References

Agbu PA Fehrenbacher DJ and Jansen IJ1990 Statistical comparison of SPOT spectralmaps with field soil maps Soil Science Society ofAmerica Journal 54 818ndash18

Barrett L 1999 Particulars in contextmaintaining a balance in soil geography Annalsof the Association of America Geographers 89707ndash13

Beckett PHT and Webster R 1971 Soilvariability a review Soils and Fertilizers 341ndash15

Bell JC Grigal DF and Bates PC 2000 A soil-terrain model for estimating spatial patterns ofsoil organic carbon In Wilson J and Gallant Jeditors Terrain analysis principles and applica-tions New York City NY John Wiley and Sons295ndash310

Birkeland PW 1999 Soil and geomorphologyThird edition New York NY Oxford UniversityPress

Breiman L Friedman JH Olshen RA andStone CJ 1984 Classification and regressiontrees Belmont CA Wadsworth

Brule FJ 1996 Fuzzy systems ndash a tutorialhttp newsgroup compai httpwwwquadralaycom (last accessed 30 August 2001)

Bunkin FV and Bunkin AF 2000 Lidarsounding of water soil and plants Atmosphericand Oceanic Optics 13 54ndash72

Buol SW Hole FD McCracken RJ andSouthard RJ 1997 Soil genesis and classifica-tion Ames IA Iowa State University Press

Burgess TM and Webster R 1980a Optimalinterpolation and isarithmic mapping of soilproperties the semi-variogram and punctualkriging Journal of Soil Science 31 315ndash31

ndashndashndashndash 1980b Optimal interpolation and isarithmicmapping of soil properties block kringingJournal of Soil Science 31 331ndash41

Burrough PA 1989 Fuzzy mathematicalmethods for soil survey and land evaluationJournal of Soil Science 40 477ndash92

Burrough PA and McDonnell RA 1998Principles of geographic information systems(Revised edition) Oxford Clarendon Press

Burrough PA Beckett PHT and Jarvis MG1971 The relation between cost and utility insoil survey Journal of Soil Science 22 368ndash81

Burrough PA Van Gaans PMF and

Hootsman R 1997 Continuous classificationin soil survey spatial correlation confusionand boundaries Geoderma 77 115ndash35

Cambell JB 1977 Variation of selectedproperties across a soil boundary Soil ScienceSociety of America Journal 41 578ndash82

Cambell JB and Edmonds WJ 1984 Themissing geographic dimension to soiltaxonomy Annals of the Association of AmericanGeographers 74 83ndash97

Castrignano A Giugliarini L Risaliti R andMartinelli N 2000 Study of spatial relation-ships among some soil physico-chemicalproperties of a field in central Italy using multi-variate geostatistics Geoderma 97 39ndash60

Cialella AT Dubayah R Lawrence W andLevine E 1997 Predicting soil drainage classusing remotely sensed and digital elevationdata Journal of Soil Science 62(2) 171ndash78

Clark RN and Swayze GA 1996 Evolution inimaging spectroscopy analysis and sensorsignal-to-noise an examination of how far wehave come Summaries of the sixth annual JPLairborne Earth science workshop 4ndash8 March 1996AVIRIS Workshop Vol 1 5

Clarke LA and Pregibon D 1992 Tree-basedmodels In Chambers J and Hastie J editorsStatistical models in S Pacific Grove Wadsworthand Brooks 377ndash419

Coleman TL Agbu PA and MontgomeryOL 1993 Spectral differentiation of surfacesoils and soil properties ndash is it possible fromspace platforms Soil Science 155 283ndash93

Cook SE Corner RJ Grealish G GesslerPE and Chartres CJ 1996 A rule-basedsystem to map soil properties Soil ScienceSociety of America Journal 60 1893ndash900

Cruickshank JG 1972 Soil geography New YorkNY John Wiley amp Sons

Csillag F Pasztor L Biehl LL 1993 Spectralband selection for the characterization ofsalinity status of soils Remote Sensing ofEnvironment 43 231ndash42

Dale MB McBratney AB and Russell JS1989 On the role of expert systems andnumerical taxonomy in soil classificationJournal of Soil Science 40 223ndash34

Davis JR 1993 Expert systems and environ-

widely used data model for PSM and is also routinely used to manage other environ-mental information such as elevation (DEMs) and remotely sensed data Spatialanalysis and integration with other types of raster-based environmental data can beeasily performed with soil data stored using a raster data structure (Burrough andMcDonnell 1998)

2 Digital terrain modelling

Terrain analysis quantifies the relief component of models characterizing soilformation Soil development and its associated profile characteristics often occurs inresponse to the way in which water moves through and over the landscape which iscontrolled by local relief Accordingly terrain analysis will be most useful in environ-ments where topographic shape is strongly related to the processes driving soilformation (McKensie et al 2000) Digital terrain modelling is a technique for derivingspatially explicit quantitative measures of the shape character of topography (Weibeland Heller 1991 Wilson and Gallant 2000) The spatial distribution of the resultingterrain attributes (characterizing local water flow paths) can also capture the spatialvariability of soil attributes Moore ID et al (1991) reviewed the analysis of digitalelevation data (including DEMs) for hydrological geomorphological and biologicalapplications They provided a table that summarized the significance and physicalmeaning of various terrain attributes to landscape processes Building on their workmany authors have used terrain attributes derived from digital elevation models(DEMs) as explanatory variables in predictive soil models (Odeh et al 1991 Moore etal 1993 Gessler et al 1995 Skidmore et al 1996 and others) Methods used to deriveterrain attributes have been greatly refined over the last 15 years and future satellitesaiding in the development of more accurate DEMs will make terrain analysis anincreasingly important component of predictive soils mapping (Moore et al 1993Mackay and Band 1998) Several recent review articles have been specifically devotedto the role of terrain analysis in soil mapping (Ventura and Irvin 1996 Irvin et al 1996McKensie et al 2000)

3 Remote sensing

Remote sensing data are an important component of PSM because they provide aspatially contiguous quantitative measure of surface reflectance which is related tosome soil properties (Agbu et al 1990) Both physical factors (eg particle size andsurface roughness) and chemical factors (eg surface mineralogy organic mattercontent and moisture) control soil spectral reflectance (Irons et al 1989) Surfacemineralogy can be derived by wavelength specific charge transfer and crystal fieldabsorptions associated with the presence of iron and iron-oxides (Fe2+ and Fe3+) andvibrational absorptions associated with hydroxyl bonds in clays adsorbed water andthe carbonate ion (Goetz 1989 Irons et al 1989) The presence and strength of theseabsorption features can be used to identify and quantify concentrations of mixed suitesof minerals in soil (Johnson et al 1983 Shipman and Adams 1987) Organic matterparticle size and moisture content in contrast influence soil reflectance primarilythrough a change in average surface reflectance and produce only broad spectral

P Scull et al 177

178 Predictive soil mapping a review

expression (Irons et al 1989) Adecrease in particle size tends to increase surface albedoand decrease spectral contrast of absorption features while an increase in organicmatter or soil moisture decreases average reflectance or albedo

Numerous studies have shown the potential benefits of using remote sensing for soilidentification and mapping Comprehensive surveys of soil spectral reflectance includestudies by Stoner et al (1980) Henderson et al (1992) and Csillag et al (1993) Remotesensing studies based on broad-band sensors such as Landsat TM include Agbu et al(1990) Coleman et al (1993) Seyler et al (1998) and Oliveira (2000) Traditionallyremote sensing has been used to classify soil units through photo-interpretation ordigital image processing Combining remotely sensed information with ancillaryinformation such as thematic maps or vegetation cover can yield significant improve-ments (Wilcox et al 1994 Cialella et al 1997 Wanchang et al 2000)

Recent developments in hyperspectral remote sensing offer the potential of signifi-cantly improving data input to predictive soil models Hyperspectral sensors such asthe Airborne VisibleInfrared Imaging Spectrometer (AVIRIS) measure a contiguousspectrum in the visible and NIR and thereby better characterize atmospheric andsurface properties (Goetz et al 1985) The large number of spectral bands permits directidentification of minerals in surface soils For example Clark and Swayze (1996)mapped over 30 minerals using AVIRIS at Cuprite Nevada Palacios-Orueta and Ustin(1996) showed that enhanced spectral information was suitable for discriminating evensubtle spectral changes associated with differences in organic matter and iron contentOther examples of the application of AVIRIS to aid soil mapping include Palacios-Orueta et al (1998) Okin et al (1998) and Roberts et al (1998)

Sensors that operate in the microwave portion of the electromagnetic spectrum havealso shown promise in soil mapping research Microwave sensing can be broadlydivided into active (eg radar) systems and passive systems and are capable ofpenetrating the atmosphere under virtually all conditions offering a significantadvantage over visible and near-infrared spectroscopy (for a general overview ofmicrowave sensing see Lillesand and Kiefer 1994 chapter 8) Synthetic aperature radar(SAR) is one example of an active system SAR has been used to aid soil propertymapping such as soil salinity (Metternicht 1998) and soil moisture (Engman andChauhan 1995 Narayanan and Hirsave 2001) Active radar systems can also bedesigned to collect data at varying look angles providing the opportunity for theacquisition of stereo radar images Such images can be used to produce high resolutionand extremely accurate DEMs (eg Fang 2000) A similar active sensing system isLiDAR (light detection and ranging) which uses pulses of laser light rather thanmicrowave energy to illuminate the surface (see Bunkin and Bunkin 2000 for a reviewof applications to soil mapping research) While passive microwave systems haveseemed to receive less attention in the literature a few examples of soil mapping appli-cations can be found (see Kleshchenko et al 2000 and Laymon et al 2001) Regardlessof the type of system remote sensing data and derived products are potentially usefulexplanatory variables in predictive soil mapping models

4 Fuzzy logic

Fuzzy set theory or fuzzy logic provides an alternative conceptual paradigm withinPSM research The use of this theory has increased greatly in the last few years making

it an important component of PSM Fuzzy logic is an alternative to Boolean logic thatattempts to recognize the concept of partial truth (Brule 1996) Dr Lotfi Zadeh (1965)introduced the concept and accompanying mathematics in his seminal work lsquoFuzzysetsrsquo The theory permits partial class membership in contrast to traditional set theorywhere set memberships are crisp and binary (ie a soil sample is either completelyType A or it is not at all Type A) Central to the fuzzy concept is the idea that objects innature rarely fit exactly the classification types to which they are assigned (Zadeh1965) Rather they show varying signs of similarity to multiple classes (ie an observedsoil pedon often resembles more than one of the defined soil series within the area) Byusing fuzzy membership values (ranging from 0 nonmembership to 1 totalmembership) within predictive soil models to express degrees of similarity generaliza-tion problems associated with classification schemes (filtering of information) areminimized and the complex nature of soil data is allowed to propagate through themodelling process Similarity values between 0 and 1 are not comparable to proportionsand need not add up to 1 Within Boolean logic probability statements refer to thelikelihood of an outcome the soil sample is either one series or another With fuzzylogic a given sample is not definitively a member of the subset of any one particularseries Fuzzy logic is especially useful in soil research because of the continuous andcomplex nature of the soil landscape It serves as an important alternative to thesubjective rigidity imposed on soils data by Soil taxonomy Several recent articlesprovide a thorough review of the use of fuzzy sets in soil science (Burrough 1989McBratney and Odeh 1997 Burrough et al 1997 De Gruijter et al 1997)

Within PSM research two different approaches to creating continuous classes usingfuzzy logic exist The first is based on the fuzzy-k-means classifier which partitionsobservations in multivariate space into natural classes This approach is similar tocluster analysis and numerical taxonomy but the resulting classes are continuous witheach observation assigned a fuzzy membership value that characterizes its degree ofsimilarity to each individual class The concept has been integrated into geostatisticalmethods and will be discussed in more detail below (see Section IV1) The secondapproach is known as the Semantic Import model (SI) and is used in situations whenclassification schemes are pre-defined and class limits are relatively well understoodThe SI model is commonly used in concert with expert knowledge and will bediscussed in the expert systems section (see Section IV4)

IV Recent advances in predictive soil mapping

Within the last decade many authors have sought to model the soil landscape using avariety of methods Literature in this field could be summarized many different waysbut we concentrate on the literature that directly addresses the goals of predictive soilmapping stated in the introduction (see Table 1) Therefore we will review research thatattempts to exploit the relationship between quantifiable landscape indices and soilcharacter in order to model the soil landscape in a more continuous and thereforerealistic manner

The research reviewed here is distinguished from decades of previous researchdocumenting the correlation between landscape position and soil attributes (reviewedby Hall and Olsen 1991) That body of research is informative but not useful for

P Scull et al 179

180 Predictive soil mapping a review

Table 1

Selected

recen

t literature on pred

ictive

soil m

apping an

d m

apping

(cited in

this article) d

escribing the mod

ellin

g metho

dused the

dep

enden

t variables used a

nd the

env

iron

men

tal v

ariables (ex

plan

atory) used in

the

mod

els

Stud

yMod

ellin

g metho

dDep

ende

nt variables

Environm

ental variab

les

Goa

ls atta

ined

a

Bell et

al 20

00Line

ar and

exp

onen

tial

Total soil orga

nic

Slop

e curvature aspe

ct

ECx

regression

carbon

hillslope

position

Burge

ss and

Web

ster

Punc

tual and

block

Na co

nten

t co

ver

Non

eC

1980

ab

kriging

loam

thickn

ess

ston

e co

nten

t

Burroug

h 198

9Fu

zzy mathe

matical

ndashndash

rpmetho

ds

Burroug

h et

al 19

97Con

tinuo

us classificatio

nndash

ndashrp

Castrigna

no eta

l 200

0Factorial kriging

CEC

pH N

P K

Non

e ex

plicitly used

Cx

Na

Cialella

eta

l 199

7Classifica

tion tree

Drainag

e class

Elev aspe

ct NDVI

Ec

Coo

k et

al 19

96Baysian

rule-ba

sed metho

dsOrgan

ic m

atter

Slop

e aspec

t wetne

ssEcX

inde

x

Dale et

al 19

89 (a review

)Ex

pert systems

ndashndash

ndash

Ellis 19

96Decision tree

ana

lysis

Soil erosion class

Slop

e aspec

t wetne

ssEc

neutral ne

tworks

inde

x flow le

ngth and

accu

mulation Lan

dsat

TM tree

cov

er

Gessler 1

996

A large

variety of statistic

Field an

d labo

ratory

A variety of digital

EC

metho

dsco

llected

phy

sica

len

vironm

ental da

tach

emical and

morph

olog

ical soil

prop

ertie

s

Gessler eta

l 199

5Line

ar and

logit reg

ression

A horizon

and

solum

Curvature CTI topo

EC

depth E horizon

positio

npresen

ce

Goo

vaerts 19

92Factorial kriging

Total carbon

Non

eC

P Scull et al 181pH

N CEC

extractab

lecatio

ns (K Ca M

g)Hartemink et

al 20

01ndash

ndashndash

rp

Heu

velin

k an

d W

ebster 20

01ndash

ndashndash

rp

Hew

itt 19

93ndash

ndashndash

rp

Indo

rante et

al 19

96ndash

ndashndash

rp

King et

al 19

99Lo

gistic regression

Presen

ceabsen

ceSlop

e aspec

t po

t solar

Ec

Non

calc c

lay-loam

energy

Kno

tters eta

l 199

5Kriging

co

-kriging

So

ft layer de

pth

Hillslop

e po

sitio

nCx

regression

kriging

Laga

cherie and

Holmes 19

97Classifica

tion tree

Map

ping

unit

Geo

logy variou

s topo

Ec

indices

Laslett e

tal 198

7Kriging

splin

es tren

dpH

Non

ec

surfac

e nea

rest neigh

bor

McB

ratney 1

992

ndashndash

ndashrp

McB

ratney

eta

l 199

1Block kriging

Clay co

nten

tNon

eC

McB

ratney

eta

l 200

0ndash

ndashndash

rp

McB

ratney

and

Ode

h 199

7Fu

zzy sets in

soil scienc

endash

ndashRp

McB

ratney

and

de Gruijter

Fuzz

y-k-mea

ns w

ithFu

zzy classes

Field co

llected

phy

sical

C19

92ex

tragrade

sch

emical and

morph

olog

ical soil

prop

ertie

s

McC

rack

en and

Cate 198

6Artificial intellig

ence

ndash

ndashrp

expe

rt systems

McK

ensie an

d Austin

19

93Gen

eralized

line

ar m

odels

Clay co

nten

t CEC

Slop

e relief land

form

ec

(logit)

pH EC

COLE

slop

e po

sitio

nbu

lk den

sity and

othe

rs

Moo

re eta

l 199

3Line

ar reg

ression

A horizon

dep

th O

MSlop

e w

etne

ss and

strea

mEC

and P co

nten

t pH

power in

dices aspect

curvature

182 Predictive soil mapping a reviewTa

ble 1

Con

tinu

ed

Stud

yMod

ellin

g metho

dDep

ende

nt variables

Environm

ental v

ariables

Goa

ls atta

ined

a

McK

ensie an

d Ryan 199

9Reg

ression tree and

linea

rSo

lum dep

th P an

dElevation slope

CE

regression

N con

tent

curvature CTI

contribu

ting area do

wn-

slop

e mea

ns for slope

clim

ate da

ta Presco

ttInde

x G

amma

Rad

iometry La

ndsat T

M

and Geo

logy unit

Ode

h et

al 19

92ab

Fuzz

y-c-mea

ns and

Fuzzy classes

Field-co

llected

phy

sica

lC

kriging

chem

ical and

morph

olog

ical soil

prop

ertie

s

Ode

h et

al 19

94

Reg

ression kriging

co

-So

lum dep

th de

pth

Slop

e aspect cu

rvature

eC

1995

kriging regression

to bed

rock gravel

kriging

and clay con

tent

Skidmore et

al 19

91Bayesian expe

rt system

Soil land

scap

e un

itVeg type

wetne

ss in

dex

EcX

1996

grad

ient terrain po

sitio

n

Voltz and

Web

ster 19

90Kriging

cu

bic splin

eClay co

nten

tNon

eC

Web

ster 1

994

Dev

elop

men

t of

ndashndash

rppe

dometrics

Zhu

199

7ab

Fuzz

y logic expe

rt system

Soil series A

horizon

Elev pm aspe

ct c

anop

yECX

Zhu

and

Ban

d 199

4(SoL

IM)

depth in

dividu

alco

verage

grad

ient

Zhu

eta

l 199

7series m

aps

curvature

Not

es

a Letters refer to the de

gree

to w

hich

the

goa

ls of PS

M defined

in the

introdu

ction are achiev

ed

Soilndash

e nvironm

ent relatio

ns utilized

(letter E) be

tter represen

tatio

n of soil c o

ntinuity (C

) an

d ex

pert kno

wledg

e utilize

d (X) c

orrespon

d to goa

ls 1 2 and

3respec

tively from

the

introdu

ction and

cap

ital lette

rs (E C X) ind

icate the metho

d is relatively more successful th

an th

ose metho

ds den

oted

by lower

case le

tters (e c x

) rp ind

icates rev

iew pap

ers

P Scull et al 183

predictive mapping because landscape position is never quantified (rather positionwas often qualitatively defined eg toe-slope) and thus the documented relationshipscannot be generalized using environmental data and digital elevation models to predictsoil character at unvisited sites

Table 1 documents modelling methods model variables and the extent to which thereferenced research satisfies the previously defined goals of PSM (lsquoGoals attainedrsquocolumn) Soil-environment relations utilized (letter E) better representation of soilcontinuity (C) and expert knowledge utilized (X) correspond to goals 1 2 and 3respectively from the introduction and capital letters (E C X) indicate the method isrelatively more successful than those methods denoted by lower case letters (e c x)Citations with no letters present within the lsquoGoals attainedrsquo column do not address theaforementioned goals For example the Cialella et al (1997) received a rating of lsquoEcrsquomeaning that the methods employed successfully utilized environmentalndashsoil characterrelations (E) and somewhat successfully presented a better method of representing soilcontinuity (c) The ratings are provided simply to help organize the literature that wasreviewed Review papers are included within the table denoted by lsquorprsquo Geostatisticalmethods are not included in the table because they have been comprehensivelysurveyed elsewhere (Odeh et al 1994 Burrough et al 1997 McBratney et al 2000Heuvelink and Webster 2001) and because the objectives and assumptions of geostatis-tical methods differ slightly from other PSM research We briefly outline thesedifferences in the following section

1 Geostatistical methods

Geostatistics are a subset of traditional statistics that deal primarily with spatial dataand account for spatial autocorrelation using kriging as the spatial interpolator Theconcept is based upon the theory of regionalized variables which was mainlydeveloped by Matheron (1963) and Krige (1963) Kriging is a form of weighted localaveraging that uses a measure of spatial dependence the variogram to determine theweights applied to the data when computing the averages Geostatistical methods havebeen used in predictive soil mapping research to spatially interpolate soil propertyvalues at unmeasured sites from field-collected data

Burgess and Webster (1980a b) were the first to introduce ordinary kriging to the soilcommunity and since that time an enormous amount of work has been published Forexample ordinary kriging has been used to interpolate many different soil propertiesincluding pollution trace element deficiencies salinity and fertility (Heuvelink andWebster 2001) Ordinary kriging has been criticized for a variety of reasons Forexample Laslett et al (1987) reported that several authors had criticized geostatisticsbecause kriging is a global rather than local technique failing to take into accountknowledge of soil materials and processes Other authors have criticized geostatisticsbecause they are excessively data dependent requiring a large number of closelyspaced data points (Zhu 1997a) As Webster and Oliver (1992) suggest in excess of ahundred samples may be needed to use geostatistics at the field scale because of highspatial variability of soil in some areas Geostatistics also assume spatial autocorrela-tion which sometimes may be a poor assumption in complex terrain where abruptchanges in soil-forming factors occur (McBratney et al 2000) Ordinary kriging by itself

184 Predictive soil mapping a review

does not satisfy two of the three goals of PSM presented in the introduction ndash it does notadequately incorporate expert knowledge and it does not exploit the relationshipbetween environmental variables and soil properties

Ordinary kriging has been modified in a variety of ways to better incorporateancillary data and known soilndashlandscape relationships Block kriging involvesdetermining estimates over meaningful areas rather than at specific points (Burgess andWebster 1980a McBratney et al 1991) Using this method a study area can be stratifiedinto different regions that are reflective of the pedogenetic processes at work In orderto accommodate a trend within a dependent soil variable universal kriging has beenused (Webster 1994) Kriging with external drift is similar to universal kriging but ituses an ancillary variable to represent the trend (McBratney et al 2000) Co-krigingtakes advantage of correlation that may exist between the variable of interest and othermore easily measured variables (Odeh et al 1995) Regression kriging involvesspatially interpolating the residuals from a non-spatial model by kriging and addingthe result to the prediction obtained from that model (Goovaerts 1997 Castrignano etal 2000) Factorial kriging is another method to integrate multivariate data into thestandard kriging routine to extrapolate soil data (Goovaerts 1992) Many authors havecompared these various methods (Laslett et al 1987 Voltz and Webster 1990 Odeh etal 1994 Knotters et al 1995)

Fuzzy logic has been used with geostatistics by various authors to produce new kindsof fuzzy soil maps with continuous classes (Burrough 1989 McBratney and DeGruijter1992 Odeh et al 1992a and reviewed by McBratney and Odeh 1997) The processentails kriging the matrix of membership values determined by the fuzzy k-meansclassifier resulting in a continuous soil surface where individual locations are allowedto belong to more than one class and no rigid boundaries are designated to separate thesoil into discrete units or entities The results of such analysis can be used to assess thepedologic process validity of soil taxonomy by determining whether soils grouptogether into classes that are similar to taxonomic types Fuzzy classes wouldpresumably reflect the main pedologic features within a given area In this sense theclassification is quantitative whereas soil taxonomy is rooted in qualitative discrimina-tion A comparison between the two could provide insightful and help assist soiltaxonomy in making a classification that is more indicative of underlying soil processes

Geostatistics in soil research were originally introduced to quantitatively assess soilvariability within soil mapping units (McBratney et al 1991) in response to criticisms inthe early 1970s that soil unit composition was not well quantified (Beckett and Webster1971) In this regard geostatistics have been very useful having served well the originalgoals set forth by Burgess and Webster (1980b) when they drew kriging to the attentionof soil scientists as a means of spatial prediction At the field scale soil variation islargely due to the effect that topography has on soil genesis Geostatistics have beensuccessfully applied in such environments by using terrain attributes as ancillary datawithin many of the kriging routines described above Such quantitative within-unitvariability of soil properties is very useful in the field of precision agriculture and othersituations (eg pollutants) where very detailed soil attribute information is needed atthe field scale (Heuvelink and Webster 2001)

However geostatistics have not been applied in a wide variety of environments or atlarger scales In order to be successfully applied in different environments geostatisticswill likely require a different suite of ancillary data For example remote sensing data

P Scull et al 185

could be used in arid regions where soil toposequences are less well expressed Atlarger scales of prediction selection of different sets of ancillary variables is requiredbecause different processes define soil character at different scales The most obviousexample is that of climate which may control soil distribution at large scales(continental) but has little explanatory power at the field level Regardless of whetheradequate ancillary data exist the amount of data required to use geostatistics forlandscape-level prediction would be extremely difficult and costly to collect given thestrict sampling protocol required to characterize spatial dependence It is also unclearat what landscape scale soils exhibit spatial autocorrelation

Geostatistical approaches do provide a means of creating continuous soil attributesurfaces to better represent soil continuity (Goal 2) and they can be used to exploit therelationship between environmental variables and soil properties in order to moreefficiently collect soil data (Goal 1) However they do not sufficiently utilize expertknowledge (Goal 3) as no attempt has been made in geostatistical approaches todirectly integrate expert knowledge Fundamentally kriging is a process of interpola-tion designed to predict attribute values in between locations of measured samples Inthis sense geostatistics represent a middle ground between pure interpolation (egnearest neighbour type classifier) in which only measured points for the variable ofinterest are used to determine unknown values and other predictive models thatprimarily use soilndashenvironment correlation to create predictive maps

2 Statistical methods

Statistical methods can be used to exploit the relationship between quantifiablelandscape indices and soil properties to create predictive soil maps For exampleMcKensie and Austin (1993) used a regression to account for a large percentage ofvariation for many soil characteristics (A horizon clay content CEC EC pH bulkdensity and COLE B horizon clay content CEC ESP EC pH bulk density and COLE)using a variety of predictor variables (slope presence or absence of impeding layerrelief landform topographic position) Their results confirm the hypothesis of MooreID et al (1991) that soil character is related to quantifiable landscape indices Howevertheir methods do not provide inference of soil properties at unmeasured sites frommapped environment data because the topographic variables were measured in thefield Linear regression has also been used with terrain variables derived from a 15-mDEM in northeastern Colorado to predict soil attributes (organic matter contentextractable phosphorous pH and texture) at unvisited sites (Moore et al 1993) In thatparticular study 50 of the variance of A-horizon thickness was explained by slope andthe wetness index Gessler et al (1995) also used regression to model A-horizonthickness from topographic variables in southeastern Australia (plan curvature andwetness index r2 = 063 P = 0001) They modeled solum depth and used logisticregression to model E horizon presenceabsence Elsewhere logistic regression hasbeen used to model the presenceabsence of noncalcareous clay loam horizon in centralFrance using terrain attributes from a 20-m DEM (King et al 1999) Exponentialregression has been used to model soil organic carbon using terrain variables (Bell et al2000) in glacial outwash soils in east-central Minnesota Generalized additive models(GAM) have been used less frequently in PSM research Gessler (1996) used a GAM

186 Predictive soil mapping a review

model to predict total soil carbon A horizon depth and solum depth using a variety ofenvironmental predictors (slope elevation wetness index mean annual temperatureprecipitation and radiation)

This small body of research opened the door to more complex methods by demon-strating the existence of quantifiable relationships These authors were able to producesoil attribute maps using raster data models whose scale was dependent upon the gridcell resolution of the environmental data They were successful at exploiting the rela-tionship between quantifiable topographic attributes and soil profile character (Goal 1)The continuous soil attribute surfaces they produce also better represent soil continuitythan the choropleth soil maps produced by traditional soil survey (Goal 2) Howeverthe bulk of these methods (excepting GAMs) are limited by their assumed linear rela-tionship between soil and topographic attributes their assumptions of normallydistributed data and their high data requirements Standard statistical procedures arealso not flexible enough to allow robust integration with a variety of potential datasources such as expert knowledge (Goal 1)

Statistical methods do demonstrate in a quantitative manner that terrain analysis canbe used to predict soil attributes in relatively small areas with homogeneous parentmaterial A large proportion of the research using statistical methods was conducted insemi-arid landscapes at small scales (the largest study area of the entire group was~2000 ha) Obviously for statistical approaches to be most effective they need to bemore universal As such they need to be tested andor developed at larger scales andin more diverse landscapes

3 Decision tree analysis (DTA)

The use of decision tree analysis is just beginning to be explored in predictive soilmapping research although it has been used successfully in the related field ofpredictive vegetation mapping since the early 1990s (Lees and Ritman 1991 MooreDM et al 1991 Franklin 1998) DTA is a form of divisive classification The process oftree modelling involves successively partitioning data (called recursive partitioning inthe tree modelling literature) into increasingly homogeneous subsets which once thepartitioning has ceased are called terminal nodes (Lees and Ritman 1991) Splits orrules defining how to partition the data are selected based on information statistics thatdefine how well the split decreases impurity within the data set (Clark and Pregibon1992) Splits are based on threshold values of an explanatory variable selected bycomparing the increase in resulting purity of node membership for all possiblethresholds and variables The process is iterative growing from the root node (thecomplete data set) to the terminal nodes in a dendritic fashion (Friedl and Brodley1997) Once the tree has been constructed (or grown) it encodes a set of decision rulesthat describe the data partitioning process These rules can be used to classify or predictother data sets (Moore DM et al 1991) Pruning the tree is often necessary to preventthe tree from being overfit to the sample data and to reduce tree complexity Pruningentails combining pairs of terminal nodes into single nodes and can be accomplishedusing cross-validation which yields an initial indication of how large a tree makesrobust predictions (Safavian and Norvig 1991) Cross-validation involves systematical-ly removing portions of the data set and running the remaining sample through the tree

in an iterative manner eventually yielding estimates of the misclassification rates foreach class each node and the whole model (Breiman et al 1984) In this mannerdifferent sized trees can be compared in terms of parsimony

The term DTA is used to collectively refer to all types of tree-based modelling (theword lsquodecisionrsquo is used because it is descriptive indicating that the analysis eventuallyleads to a set of decision rules defining data partitions) The term should be distin-guished from classification tree analysis because the latter refers specifically to DTAwhere the response variable is categorical The term CART (classification andregression trees ndash Breiman et al 1984) is sometimes used but strictly speaking refers tospecific software Friedl and Brodley (1997) provide a review of the decision treealgorithms They divide types of DTA into two classes (1) homogeneous decision treesfor which a single algorithm is used to estimate each split (eg CART) and (2) hybriddecision trees (HDT) for which different splitting methods can be used at differentpoints in the tree (eg Quinlan 1993) They further divide homogeneous decision treesinto univariate (UDT) where single features of the input data define splits and multi-variate decision trees (MDT) where multiple features of the input data can define splitsAccording to this naming convention no distinction is made between methods withdifferent types of response variables although all methods can be used with bothcategorical and continuous response variables In a comparison of these various typesof algorithms on a variety of data sets Friedl and Brodley (1997) found that HDT hadthe highest classification accuracy

The overall aim of DTA is to design a set of predictive rules (eg if geology type Athen soil type B) developed from training data which can then be applied to ageographic data base to predict the value of a response variable (Michaelsen et al 1994)Therefore DTA explicitly uses soilndashlandscape correlation in model development (Goal2) The technique appears promising in soil research but needs to be further tested asnot many of the above types of DTA have been tested in the PSM literature in fact onlyunivariate approaches have been employed For example Lagacherie and Holmes(1997) successfully used univariate DTA to model a categorical response variable soiltype within a training set and then assuming that the training set was representativeextrapolated the model to a much larger region Their work is interesting becausealthough not a single sample came from the area they eventually mapped they wereable to produce a soil map that was more accurate (74 versus 69) than the existingmap produced from traditional methods Cialella et al (1997) also used univariate DTAto predict soil drainage class from a variety of terrain attributes and remotely senseddata They predicted soil drainage class with an average of 78 accuracy ndash impressivegiven that the variation accounted for by a typical soil survey ranges from about halfthe total variance for physical attributes to less than one-tenth for some soil chemicalattributes (Gessler et al 1995)

DTA has been compared with other approaches by several authors In the applicationof erosion modelling DTA results were similar to Artificial Neural Networks (Ellis1996) Both methods achieved high training accuracy (as measured by the KappaStatistic) but in terms of prediction accuracy both methods performed poorly Gessleret al (1995) compared DTA with generalized linear models (GLM) and generalizedadditive models (GAM) to predict A-horizon thickness and concluded that GLM waspreferable to both DTA and GAM McKensie and Ryan (1999) compared regressiontrees and standard linear regression to predict soil properties (total solum depth soil

P Scull et al 187

188 Predictive soil mapping a review

phophorus and soil carbon) using a large variety of predictors (elevation slopecurvature wetness index climate variables geology unit Landsat TM data andGamma radiometric data) and found that different methods work best for differentproperties and overall success hinged on the strength of the relationship between soiland environmental variables Their study of a 50 000 ha forested landscape in southernNew South Wales Australia is one of the few studies at such a small map scale (largearea) or in forested landscape The most extensive comparison of techniques involvingDTA was conducted by McBratney et al (2000) They compared regression techniques(GLM GAM DTA-regression tree) geostatistical techniques (kriging and heterotopicco-kriging) and a hybrid technique (regression kriging) DTA was found to be thepoorest performing of the regression techniques because of the unrealistic predictionsurface generated by the DTA model DTA has also been criticized by other authorsbecause of the stepped prediction surface (Gessler 1996) This phenomenon can beespecially apparent in situations where predictor variables have different resolutionsThe extent to which DTA yields a better spatial representation of soil continuity is afunction of the scale and type of predictor variables used and therefore varies withindividual models

DTA is gaining widespread popularity as a means to develop prediction rules thatcan be rapidly and repeatedly evaluated (Cialella et al 1997 Franklin et al 2000) DTAprovides the following advantages over standard statistical techniques (1) it is easier tointerpret when explanatory variables are both nominal and continuous (2) it isinvariant to monotone re-expressions (transformations) of predictor variables (3) itdeals more satisfactorily with missing data values and outliers (4) it is more adept atcapturing nonadditive and nonlinear behaviour (5) it doesnrsquot make any assumptionsabout data distribution and (6) it is easily updateable as more data are collected(Moore DM et al 1991) The DTA model framework is especially appealing becauseof its capability to integrate a wide range of data sets as explanatory variables

DTA offers a unique opportunity for interaction between soil experts and soilmodellers because the output of the model is a set of rules that can be pedologicallyinterpreted by the soil expert In this sense expert knowledge is used in an implicitmanner in DTA (somewhat effectively achieving Goal 3) While these rules can often beexceedingly complex at minimum the expert can decide whether initial splits makesense given their understanding of the landscape The success of DTA results oftenhinges on the ability of the modeller to make key decisions during the model buildingprocess there is unfortunately no definitive way to determine the most optimal treeThe aid of the soil expert can potentially elucidate this problem

4 Expert systems

A variety of expert system approaches to PSM have been developed to utilize expertknowledge The purpose of such methods is to exploit the information the soil surveyoraccumulates while working in the field by integrating such knowledge into thepredictive model (McCracken and Cate 1986) Unlike the majority of the researchreviewed thus far the dependent variable in many expert systems models is often soiltaxa or mapping unit This apparent disadvantage of expert systems (using classifica-tion to characterize soil continuity) does make them easier to integrate into traditional

soil survey In addition several authors have developed methods to developcontinuous soil property maps from the output of expert system models designed topredict soil unit occurrence

Expert systems are composed of data (information on spatial environmentalvariables eg topography climate etc) a knowledge base (rules and facts related tosoil variation supplied from the soil surveyor) and an inference engine (whichcombines data and the knowledge base to infer logically valid conclusions) (Skidmoreet al 1996) Expert systems differ from conventional models in two ways (1) they storeand manipulate qualitative information (allowing them access to information thatcannot normally be used in other modelling frameworks) and (2) they are structuredas meta-models (the knowledge is separated from the model) (Davis 1993) This allowsthe model to selectively choose which information is relevant at various stages of themodelling process and it allows for information to be easily updated Davis (1993)reviews the application of expert systems to environmental modelling research conclu-ding that the technique is becoming more widely accepted He further states that theapplication of expert systems is constrained by an absence of fundamental knowledgefor rule generation a problem that would appear less relevant to soil mapping giventhe amount of untapped expert knowledge accumulated by the soil surveyor

The first mention of the use of expert systems in pedology was in a paper presenta-tion at the Northeast Committee Soil Survey Conference 1984 (Flach 1985) In herpaper Flach (1985) hinted that recent developments in computer science especiallyexpert systems and artificial intelligence could make modelling a practical mappingtool for soil scientists in the near future A year later McCracken and Cate (1986) hopedto encourage soil scientists to explore expert systems and its application to soil sciencethrough an optimistic article they wrote on the potential use of expert systemsHowever little research was actually conducted in the late 1980s and expert systemshad not yet begun to fulfil the lofty goals set forth by early practitioners (Dale et al1989)

In the early 1990s expert systems approaches to predictive soil mapping began toappear in the literature Skidmore et al (1991) used a Bayesian expert system to mapforest soil into different classes and their results compared favourably with availablesoil maps and actual field-collected data Their methods successfully incorporatedsurveyor knowledge and remotely sensed and digital terrain attributes but failed tobetter characterize continuous soilndashlandscape variability because their final productwas a choropleth map Skidmore et al (1996) revisited their earlier research andassessed the mapping accuracy of their results They found that the soil map producedby the expert system achieved an overall accuracy of 698 (sample size = 53) while themap derived from conventional methods had an accuracy of 736 In addition toknowledge provided by a soil scientist Skidmore et alrsquos (1996) expert system usedvegetation derived for aerial photography and topographic variables derived from a 10-m DEM (wetness index topographic position and slope)

Using an expert system Cook et al (1996) successfully produced a continuous soilproperty map for organic matter content using wetness index aspect and slope asexplanatory variables Their methods were somewhat inefficient because they requireda separate expert system for each soil property of interest However their researchrepresents the only example of expert systems used to directly predict the spatial dis-tribution of a soil property All of the other examples of expert systems in the literature

P Scull et al 189

190 Predictive soil mapping a review

deal with soil type or class The use of expert systems to map soil properties needs tobe explored further

The expert systems discussed thus far have all used Boolean logic within theirtheoretical framework whereby an observation can belong to one and only one classand the soil properties of that class are assigned to the observation With thedevelopment of fuzzy logic and the semantic import model (the second fuzzy logicapproach to creating continuous classes) the opportunity exists to develop fuzzy logic-based expert systems Such systems can be used in conjunction with expert knowledgein situations where experts have a good qualitative idea of how to group data but havetrouble dealing with observations that are not well represented by rigid classificationschemes (Burrough and McDonnell 1998) This method can be particularly useful insituations when taxonomic schemes have been previously developed as is the case insoil taxonomy Several examples of this type of approach were published in the mid-1990s by A Zhu and colleagues (Zhu and Band 1994 Zhu et al 1996 1997 Zhu1997ab) Such systems proved useful for mapping soil at unvisited locations usingsurveyor knowledge and were also capable of producing continuous soil propertymaps The use of fuzzy logic within the theoretical framework of the expert systemallows the complex nature of soil to propagate through the modelling process neversubjected to classification schemes that filter out potentially useful lsquonoisersquo The use offuzzy logic also gives the soil surveyor more latitude during the interview processwhen the knowledge base is defined for the expert system Zhu et al (1997) developeda fuzzy logic-based expert system called SoLIM to determine the similarity of eachgrid cell in a study area to the various taxonomic mapping units delimited by the soilsurvey Continuous soil attribute maps were calculated using the similarity values andtheir relative soil survey determined attribute values (effectively achieving Goal 2) Theresulting data (which proved more accurate than soil survey data once field checked)consists of a raster grid whose resolution is determined by the resolution of the inputenvironmental and digital elevation data As noted previously this type of data modelis more applicable to environmental modelling than the choropleth map (Burrough andMcDonnell 1998)

Expert system approaches to PSM are capable of exploiting soil surveyor knowledgeby developing rule-based systems that imitate the surveyorrsquos conceptual model of soilvariability (the primary focus of Goal 3) The method would appear extremely usefulfor mapping projects (such as those conducted by the NRCS-NCSS) where fieldwork isinitially conducted to determine soilndashlandscape relations Expert system developmentcould be directly inserted into the traditional soil survey mapping approach as asubstitute for the step where the surveyor converts hisher conceptual model into achoropleth map Rather that knowledge could be incorporated into the expert systemwhich could be used to predictively map soil The resulting raster map would be morescientifically based and explicit than the hand-drawn choropleth maps of the past Itsscale would not be limited to that of the aerial photography but rather to the scale ofthe environmental data

Despite these advantages the expert system approach has some drawbacks Becausethe method is dependent upon expert knowledge it cannot be applied whereenvironmentndashsoil relations are poorly understood Of course this criticism can be madeof all PSM models as inductive or deductive knowledge of soilndashenvironment relationsis a prerequisite for PSM However expert systems are deductive models and as such

are not driven by any specific field-collected data (although presumably the soil experthas field experience in the mapping area) Expert systems donrsquot afford the opportunityto first statistically document landscapendashsoil relations and then extrapolate the resultsbecause expert systems do not directly use sample soil data (lsquohard datarsquo) to determinesoilndashlandscape relations Thus the utilization of the relationship between environmen-tal variables and soil properties (Goal 1) is only indirectly achieved Expert systemsapproaches have been demonstrated to be extremely effective in a small number of casestudies The possibility of satisfying all three goals discussed in this review makesexpert systems a predictive soil mapping method that needs to be further tested ndash bothgeographically and across different scales of analysis

V Conclusion

Most of the predictive soil mapping research outlined in this review was conducted atvery large map scales (over small areas) In fact the majority of the research wasconcerned with assessing the spatial variability of soil character within individual fieldsor across soil toposequences The primary driving force behind this type of research hasbeen the need to provide accurate soils information for agriculture and ecologicalmodels It is clear that terrain attributes are powerful predictors at the local scaleGeostatistical tools have been successful at using terrain attributes and the spatialdependence of soil properties to interpolate between existing data points withinindividual fields Across soil toposequences statistical approaches provide a usefulmeans of predicting soil character PSM research has been most successful at the fieldscale because many of the soil-forming factors are held constant For example the neteffect of four of Jennyrsquos five soil-forming factors (climate organism parent material andtime) was minimal within many of the studies reviewed While some research hastackled larger areas of study there exists a deficiency in our ability to predictively mapsoils at smaller map scales Since the distribution of soil is scale-dependent differentPSM methods and predictors are likely to work better at different scales Focus in thefuture must continue to move toward working over larger spatial extents of study inorder to produce landscape-scale soil information

Further a large proportion of the research was conducted in semi-arid gentlysloping agricultural landscapes Humid forestlands mountainous regions and desertshave received little attention As such PSM methods need to continue to be testedandor developed in a wider variety of landscapes where spatial soil distributions canbe more complex Different methods will likely be successful to different degrees indifferent environments Whereas terrain attributes are dominant predictors of soilcharacter across toposequences in prairie lands other predictors are likely necessary inareas where soils do not develop into clear toposequences For example remote sensingdata has been a useful predictor of soil chronosequences in desert landscapes wheresurface appearance is often related to soil character Surficial geology is often reflectiveof soil character in mountain regions where soils are thin and significant bedrock existsat the surface Focus in the future must continue to determine which methods andpredictors work best in which environments

Since the most useful PSM approach will vary across spatial scales and environmen-tal gradients the method used should be driven by the mapping objectives of the

P Scull et al 191

192 Predictive soil mapping a review

project Whereas the traditional soil survey of the past was expected to meet the needsof a diverse group of end users PSM methods will vary given the objectives of thesurvey As a result the end users of the soil data need to play a more active role in thesurvey process Because precision agriculture and large-scale mapping has been theprimary focus of PSM there exists a deficiency in our ability to predictively map soilsfor the purpose of general land use planning and management Even though lessdetailed soil information is needed for these purposes the development of PSMmethods is complicated by the fact that soils exhibit complex spatial variability atsmall map scales where soil-forming environments vary greatly from one location tothe next

A number of alternative methods of characterizing the continuous nature of the soillandscape have been developed Thus far most of the PSM research has provided soilinformation in a nonobject form (as opposed to defining soil types as independententities) Soil data have been generated organized and presented in the form of eitherisorithmic maps or fine-scale raster grids Both of these data models are field-viewmodels of geographic space which allow the soil to be perceived as a constantlyvarying surface Two distinct approaches have been employed (1) mapping individualsoil properties and (2) mapping continuously varying (fuzzy) soil classes Mappingindividual properties is the most common approach and will likely continue todominate PSM research The use of fuzzy soil classes in the literature is less commonand more difficult to be integrated into standard mapping procedures (such as use ofsoil taxa) because the concept radically differs from the traditional view of the soillandscape However fuzzy soil classes (generated using the fuzzy-k means approach) dohave the potential to help further advance Soil taxonomy by identifying taxonomicclasses that are more reflective of pedologic processes at work

Expert systems have been greatly underutilized in PSM research especiallyconsidering how effective a small number of case studies have been Expert systemshave the potential to satisfy successfully all three goals discussed throughout thereview They also have the potential to bridge the gap between traditional approachesand PSM methods because field soil scientists do not have to change their conceptualapproach to mapping They can still conduct field reconnaissance to determinesoilndashlandscape relationships Afterwards though they can use expert systems to betterexploit the knowledge they have garnered Such an approach could help ease thetransition from traditional soil survey to more scientifically explicit methods Expertsystems are also the most fruitful approach to utilizing a wealth of data that has alreadybeen collected in a nonexplicit manner the wealth of expert knowledge (in the form ofthe conceptual model) that senior soil mappers have accumulated

Focus in the future must also move in the direction of operability The utility of manyPSM approaches has been clearly demonstrated yet changes in how soils are mappedand perceived has been slow In the USA various members of the National CooperativeSoil Survey have called for a change in the philosophy of survey (from mappingdiscrete soil types to mapping continuous varying soil properties McSweeny et al1994) but little change has actually taken place To realize the potential of PSM in con-tributing to an overall change in standard soil mapping procedures practitionersworking in the field of PSM need to form working relationships with field soilscientists PSM methods need also to be presented in a manner that is comprehensibleto the soil science trained field mapper Predictive soil mapping is a relatively recent

phenomenon yet much progress has been made In the process of mapping soilknowledge will continue to advance

P Scull et al 193

References

Agbu PA Fehrenbacher DJ and Jansen IJ1990 Statistical comparison of SPOT spectralmaps with field soil maps Soil Science Society ofAmerica Journal 54 818ndash18

Barrett L 1999 Particulars in contextmaintaining a balance in soil geography Annalsof the Association of America Geographers 89707ndash13

Beckett PHT and Webster R 1971 Soilvariability a review Soils and Fertilizers 341ndash15

Bell JC Grigal DF and Bates PC 2000 A soil-terrain model for estimating spatial patterns ofsoil organic carbon In Wilson J and Gallant Jeditors Terrain analysis principles and applica-tions New York City NY John Wiley and Sons295ndash310

Birkeland PW 1999 Soil and geomorphologyThird edition New York NY Oxford UniversityPress

Breiman L Friedman JH Olshen RA andStone CJ 1984 Classification and regressiontrees Belmont CA Wadsworth

Brule FJ 1996 Fuzzy systems ndash a tutorialhttp newsgroup compai httpwwwquadralaycom (last accessed 30 August 2001)

Bunkin FV and Bunkin AF 2000 Lidarsounding of water soil and plants Atmosphericand Oceanic Optics 13 54ndash72

Buol SW Hole FD McCracken RJ andSouthard RJ 1997 Soil genesis and classifica-tion Ames IA Iowa State University Press

Burgess TM and Webster R 1980a Optimalinterpolation and isarithmic mapping of soilproperties the semi-variogram and punctualkriging Journal of Soil Science 31 315ndash31

ndashndashndashndash 1980b Optimal interpolation and isarithmicmapping of soil properties block kringingJournal of Soil Science 31 331ndash41

Burrough PA 1989 Fuzzy mathematicalmethods for soil survey and land evaluationJournal of Soil Science 40 477ndash92

Burrough PA and McDonnell RA 1998Principles of geographic information systems(Revised edition) Oxford Clarendon Press

Burrough PA Beckett PHT and Jarvis MG1971 The relation between cost and utility insoil survey Journal of Soil Science 22 368ndash81

Burrough PA Van Gaans PMF and

Hootsman R 1997 Continuous classificationin soil survey spatial correlation confusionand boundaries Geoderma 77 115ndash35

Cambell JB 1977 Variation of selectedproperties across a soil boundary Soil ScienceSociety of America Journal 41 578ndash82

Cambell JB and Edmonds WJ 1984 Themissing geographic dimension to soiltaxonomy Annals of the Association of AmericanGeographers 74 83ndash97

Castrignano A Giugliarini L Risaliti R andMartinelli N 2000 Study of spatial relation-ships among some soil physico-chemicalproperties of a field in central Italy using multi-variate geostatistics Geoderma 97 39ndash60

Cialella AT Dubayah R Lawrence W andLevine E 1997 Predicting soil drainage classusing remotely sensed and digital elevationdata Journal of Soil Science 62(2) 171ndash78

Clark RN and Swayze GA 1996 Evolution inimaging spectroscopy analysis and sensorsignal-to-noise an examination of how far wehave come Summaries of the sixth annual JPLairborne Earth science workshop 4ndash8 March 1996AVIRIS Workshop Vol 1 5

Clarke LA and Pregibon D 1992 Tree-basedmodels In Chambers J and Hastie J editorsStatistical models in S Pacific Grove Wadsworthand Brooks 377ndash419

Coleman TL Agbu PA and MontgomeryOL 1993 Spectral differentiation of surfacesoils and soil properties ndash is it possible fromspace platforms Soil Science 155 283ndash93

Cook SE Corner RJ Grealish G GesslerPE and Chartres CJ 1996 A rule-basedsystem to map soil properties Soil ScienceSociety of America Journal 60 1893ndash900

Cruickshank JG 1972 Soil geography New YorkNY John Wiley amp Sons

Csillag F Pasztor L Biehl LL 1993 Spectralband selection for the characterization ofsalinity status of soils Remote Sensing ofEnvironment 43 231ndash42

Dale MB McBratney AB and Russell JS1989 On the role of expert systems andnumerical taxonomy in soil classificationJournal of Soil Science 40 223ndash34

Davis JR 1993 Expert systems and environ-

178 Predictive soil mapping a review

expression (Irons et al 1989) Adecrease in particle size tends to increase surface albedoand decrease spectral contrast of absorption features while an increase in organicmatter or soil moisture decreases average reflectance or albedo

Numerous studies have shown the potential benefits of using remote sensing for soilidentification and mapping Comprehensive surveys of soil spectral reflectance includestudies by Stoner et al (1980) Henderson et al (1992) and Csillag et al (1993) Remotesensing studies based on broad-band sensors such as Landsat TM include Agbu et al(1990) Coleman et al (1993) Seyler et al (1998) and Oliveira (2000) Traditionallyremote sensing has been used to classify soil units through photo-interpretation ordigital image processing Combining remotely sensed information with ancillaryinformation such as thematic maps or vegetation cover can yield significant improve-ments (Wilcox et al 1994 Cialella et al 1997 Wanchang et al 2000)

Recent developments in hyperspectral remote sensing offer the potential of signifi-cantly improving data input to predictive soil models Hyperspectral sensors such asthe Airborne VisibleInfrared Imaging Spectrometer (AVIRIS) measure a contiguousspectrum in the visible and NIR and thereby better characterize atmospheric andsurface properties (Goetz et al 1985) The large number of spectral bands permits directidentification of minerals in surface soils For example Clark and Swayze (1996)mapped over 30 minerals using AVIRIS at Cuprite Nevada Palacios-Orueta and Ustin(1996) showed that enhanced spectral information was suitable for discriminating evensubtle spectral changes associated with differences in organic matter and iron contentOther examples of the application of AVIRIS to aid soil mapping include Palacios-Orueta et al (1998) Okin et al (1998) and Roberts et al (1998)

Sensors that operate in the microwave portion of the electromagnetic spectrum havealso shown promise in soil mapping research Microwave sensing can be broadlydivided into active (eg radar) systems and passive systems and are capable ofpenetrating the atmosphere under virtually all conditions offering a significantadvantage over visible and near-infrared spectroscopy (for a general overview ofmicrowave sensing see Lillesand and Kiefer 1994 chapter 8) Synthetic aperature radar(SAR) is one example of an active system SAR has been used to aid soil propertymapping such as soil salinity (Metternicht 1998) and soil moisture (Engman andChauhan 1995 Narayanan and Hirsave 2001) Active radar systems can also bedesigned to collect data at varying look angles providing the opportunity for theacquisition of stereo radar images Such images can be used to produce high resolutionand extremely accurate DEMs (eg Fang 2000) A similar active sensing system isLiDAR (light detection and ranging) which uses pulses of laser light rather thanmicrowave energy to illuminate the surface (see Bunkin and Bunkin 2000 for a reviewof applications to soil mapping research) While passive microwave systems haveseemed to receive less attention in the literature a few examples of soil mapping appli-cations can be found (see Kleshchenko et al 2000 and Laymon et al 2001) Regardlessof the type of system remote sensing data and derived products are potentially usefulexplanatory variables in predictive soil mapping models

4 Fuzzy logic

Fuzzy set theory or fuzzy logic provides an alternative conceptual paradigm withinPSM research The use of this theory has increased greatly in the last few years making

it an important component of PSM Fuzzy logic is an alternative to Boolean logic thatattempts to recognize the concept of partial truth (Brule 1996) Dr Lotfi Zadeh (1965)introduced the concept and accompanying mathematics in his seminal work lsquoFuzzysetsrsquo The theory permits partial class membership in contrast to traditional set theorywhere set memberships are crisp and binary (ie a soil sample is either completelyType A or it is not at all Type A) Central to the fuzzy concept is the idea that objects innature rarely fit exactly the classification types to which they are assigned (Zadeh1965) Rather they show varying signs of similarity to multiple classes (ie an observedsoil pedon often resembles more than one of the defined soil series within the area) Byusing fuzzy membership values (ranging from 0 nonmembership to 1 totalmembership) within predictive soil models to express degrees of similarity generaliza-tion problems associated with classification schemes (filtering of information) areminimized and the complex nature of soil data is allowed to propagate through themodelling process Similarity values between 0 and 1 are not comparable to proportionsand need not add up to 1 Within Boolean logic probability statements refer to thelikelihood of an outcome the soil sample is either one series or another With fuzzylogic a given sample is not definitively a member of the subset of any one particularseries Fuzzy logic is especially useful in soil research because of the continuous andcomplex nature of the soil landscape It serves as an important alternative to thesubjective rigidity imposed on soils data by Soil taxonomy Several recent articlesprovide a thorough review of the use of fuzzy sets in soil science (Burrough 1989McBratney and Odeh 1997 Burrough et al 1997 De Gruijter et al 1997)

Within PSM research two different approaches to creating continuous classes usingfuzzy logic exist The first is based on the fuzzy-k-means classifier which partitionsobservations in multivariate space into natural classes This approach is similar tocluster analysis and numerical taxonomy but the resulting classes are continuous witheach observation assigned a fuzzy membership value that characterizes its degree ofsimilarity to each individual class The concept has been integrated into geostatisticalmethods and will be discussed in more detail below (see Section IV1) The secondapproach is known as the Semantic Import model (SI) and is used in situations whenclassification schemes are pre-defined and class limits are relatively well understoodThe SI model is commonly used in concert with expert knowledge and will bediscussed in the expert systems section (see Section IV4)

IV Recent advances in predictive soil mapping

Within the last decade many authors have sought to model the soil landscape using avariety of methods Literature in this field could be summarized many different waysbut we concentrate on the literature that directly addresses the goals of predictive soilmapping stated in the introduction (see Table 1) Therefore we will review research thatattempts to exploit the relationship between quantifiable landscape indices and soilcharacter in order to model the soil landscape in a more continuous and thereforerealistic manner

The research reviewed here is distinguished from decades of previous researchdocumenting the correlation between landscape position and soil attributes (reviewedby Hall and Olsen 1991) That body of research is informative but not useful for

P Scull et al 179

180 Predictive soil mapping a review

Table 1

Selected

recen

t literature on pred

ictive

soil m

apping an

d m

apping

(cited in

this article) d

escribing the mod

ellin

g metho

dused the

dep

enden

t variables used a

nd the

env

iron

men

tal v

ariables (ex

plan

atory) used in

the

mod

els

Stud

yMod

ellin

g metho

dDep

ende

nt variables

Environm

ental variab

les

Goa

ls atta

ined

a

Bell et

al 20

00Line

ar and

exp

onen

tial

Total soil orga

nic

Slop

e curvature aspe

ct

ECx

regression

carbon

hillslope

position

Burge

ss and

Web

ster

Punc

tual and

block

Na co

nten

t co

ver

Non

eC

1980

ab

kriging

loam

thickn

ess

ston

e co

nten

t

Burroug

h 198

9Fu

zzy mathe

matical

ndashndash

rpmetho

ds

Burroug

h et

al 19

97Con

tinuo

us classificatio

nndash

ndashrp

Castrigna

no eta

l 200

0Factorial kriging

CEC

pH N

P K

Non

e ex

plicitly used

Cx

Na

Cialella

eta

l 199

7Classifica

tion tree

Drainag

e class

Elev aspe

ct NDVI

Ec

Coo

k et

al 19

96Baysian

rule-ba

sed metho

dsOrgan

ic m

atter

Slop

e aspec

t wetne

ssEcX

inde

x

Dale et

al 19

89 (a review

)Ex

pert systems

ndashndash

ndash

Ellis 19

96Decision tree

ana

lysis

Soil erosion class

Slop

e aspec

t wetne

ssEc

neutral ne

tworks

inde

x flow le

ngth and

accu

mulation Lan

dsat

TM tree

cov

er

Gessler 1

996

A large

variety of statistic

Field an

d labo

ratory

A variety of digital

EC

metho

dsco

llected

phy

sica

len

vironm

ental da

tach

emical and

morph

olog

ical soil

prop

ertie

s

Gessler eta

l 199

5Line

ar and

logit reg

ression

A horizon

and

solum

Curvature CTI topo

EC

depth E horizon

positio

npresen

ce

Goo

vaerts 19

92Factorial kriging

Total carbon

Non

eC

P Scull et al 181pH

N CEC

extractab

lecatio

ns (K Ca M

g)Hartemink et

al 20

01ndash

ndashndash

rp

Heu

velin

k an

d W

ebster 20

01ndash

ndashndash

rp

Hew

itt 19

93ndash

ndashndash

rp

Indo

rante et

al 19

96ndash

ndashndash

rp

King et

al 19

99Lo

gistic regression

Presen

ceabsen

ceSlop

e aspec

t po

t solar

Ec

Non

calc c

lay-loam

energy

Kno

tters eta

l 199

5Kriging

co

-kriging

So

ft layer de

pth

Hillslop

e po

sitio

nCx

regression

kriging

Laga

cherie and

Holmes 19

97Classifica

tion tree

Map

ping

unit

Geo

logy variou

s topo

Ec

indices

Laslett e

tal 198

7Kriging

splin

es tren

dpH

Non

ec

surfac

e nea

rest neigh

bor

McB

ratney 1

992

ndashndash

ndashrp

McB

ratney

eta

l 199

1Block kriging

Clay co

nten

tNon

eC

McB

ratney

eta

l 200

0ndash

ndashndash

rp

McB

ratney

and

Ode

h 199

7Fu

zzy sets in

soil scienc

endash

ndashRp

McB

ratney

and

de Gruijter

Fuzz

y-k-mea

ns w

ithFu

zzy classes

Field co

llected

phy

sical

C19

92ex

tragrade

sch

emical and

morph

olog

ical soil

prop

ertie

s

McC

rack

en and

Cate 198

6Artificial intellig

ence

ndash

ndashrp

expe

rt systems

McK

ensie an

d Austin

19

93Gen

eralized

line

ar m

odels

Clay co

nten

t CEC

Slop

e relief land

form

ec

(logit)

pH EC

COLE

slop

e po

sitio

nbu

lk den

sity and

othe

rs

Moo

re eta

l 199

3Line

ar reg

ression

A horizon

dep

th O

MSlop

e w

etne

ss and

strea

mEC

and P co

nten

t pH

power in

dices aspect

curvature

182 Predictive soil mapping a reviewTa

ble 1

Con

tinu

ed

Stud

yMod

ellin

g metho

dDep

ende

nt variables

Environm

ental v

ariables

Goa

ls atta

ined

a

McK

ensie an

d Ryan 199

9Reg

ression tree and

linea

rSo

lum dep

th P an

dElevation slope

CE

regression

N con

tent

curvature CTI

contribu

ting area do

wn-

slop

e mea

ns for slope

clim

ate da

ta Presco

ttInde

x G

amma

Rad

iometry La

ndsat T

M

and Geo

logy unit

Ode

h et

al 19

92ab

Fuzz

y-c-mea

ns and

Fuzzy classes

Field-co

llected

phy

sica

lC

kriging

chem

ical and

morph

olog

ical soil

prop

ertie

s

Ode

h et

al 19

94

Reg

ression kriging

co

-So

lum dep

th de

pth

Slop

e aspect cu

rvature

eC

1995

kriging regression

to bed

rock gravel

kriging

and clay con

tent

Skidmore et

al 19

91Bayesian expe

rt system

Soil land

scap

e un

itVeg type

wetne

ss in

dex

EcX

1996

grad

ient terrain po

sitio

n

Voltz and

Web

ster 19

90Kriging

cu

bic splin

eClay co

nten

tNon

eC

Web

ster 1

994

Dev

elop

men

t of

ndashndash

rppe

dometrics

Zhu

199

7ab

Fuzz

y logic expe

rt system

Soil series A

horizon

Elev pm aspe

ct c

anop

yECX

Zhu

and

Ban

d 199

4(SoL

IM)

depth in

dividu

alco

verage

grad

ient

Zhu

eta

l 199

7series m

aps

curvature

Not

es

a Letters refer to the de

gree

to w

hich

the

goa

ls of PS

M defined

in the

introdu

ction are achiev

ed

Soilndash

e nvironm

ent relatio

ns utilized

(letter E) be

tter represen

tatio

n of soil c o

ntinuity (C

) an

d ex

pert kno

wledg

e utilize

d (X) c

orrespon

d to goa

ls 1 2 and

3respec

tively from

the

introdu

ction and

cap

ital lette

rs (E C X) ind

icate the metho

d is relatively more successful th

an th

ose metho

ds den

oted

by lower

case le

tters (e c x

) rp ind

icates rev

iew pap

ers

P Scull et al 183

predictive mapping because landscape position is never quantified (rather positionwas often qualitatively defined eg toe-slope) and thus the documented relationshipscannot be generalized using environmental data and digital elevation models to predictsoil character at unvisited sites

Table 1 documents modelling methods model variables and the extent to which thereferenced research satisfies the previously defined goals of PSM (lsquoGoals attainedrsquocolumn) Soil-environment relations utilized (letter E) better representation of soilcontinuity (C) and expert knowledge utilized (X) correspond to goals 1 2 and 3respectively from the introduction and capital letters (E C X) indicate the method isrelatively more successful than those methods denoted by lower case letters (e c x)Citations with no letters present within the lsquoGoals attainedrsquo column do not address theaforementioned goals For example the Cialella et al (1997) received a rating of lsquoEcrsquomeaning that the methods employed successfully utilized environmentalndashsoil characterrelations (E) and somewhat successfully presented a better method of representing soilcontinuity (c) The ratings are provided simply to help organize the literature that wasreviewed Review papers are included within the table denoted by lsquorprsquo Geostatisticalmethods are not included in the table because they have been comprehensivelysurveyed elsewhere (Odeh et al 1994 Burrough et al 1997 McBratney et al 2000Heuvelink and Webster 2001) and because the objectives and assumptions of geostatis-tical methods differ slightly from other PSM research We briefly outline thesedifferences in the following section

1 Geostatistical methods

Geostatistics are a subset of traditional statistics that deal primarily with spatial dataand account for spatial autocorrelation using kriging as the spatial interpolator Theconcept is based upon the theory of regionalized variables which was mainlydeveloped by Matheron (1963) and Krige (1963) Kriging is a form of weighted localaveraging that uses a measure of spatial dependence the variogram to determine theweights applied to the data when computing the averages Geostatistical methods havebeen used in predictive soil mapping research to spatially interpolate soil propertyvalues at unmeasured sites from field-collected data

Burgess and Webster (1980a b) were the first to introduce ordinary kriging to the soilcommunity and since that time an enormous amount of work has been published Forexample ordinary kriging has been used to interpolate many different soil propertiesincluding pollution trace element deficiencies salinity and fertility (Heuvelink andWebster 2001) Ordinary kriging has been criticized for a variety of reasons Forexample Laslett et al (1987) reported that several authors had criticized geostatisticsbecause kriging is a global rather than local technique failing to take into accountknowledge of soil materials and processes Other authors have criticized geostatisticsbecause they are excessively data dependent requiring a large number of closelyspaced data points (Zhu 1997a) As Webster and Oliver (1992) suggest in excess of ahundred samples may be needed to use geostatistics at the field scale because of highspatial variability of soil in some areas Geostatistics also assume spatial autocorrela-tion which sometimes may be a poor assumption in complex terrain where abruptchanges in soil-forming factors occur (McBratney et al 2000) Ordinary kriging by itself

184 Predictive soil mapping a review

does not satisfy two of the three goals of PSM presented in the introduction ndash it does notadequately incorporate expert knowledge and it does not exploit the relationshipbetween environmental variables and soil properties

Ordinary kriging has been modified in a variety of ways to better incorporateancillary data and known soilndashlandscape relationships Block kriging involvesdetermining estimates over meaningful areas rather than at specific points (Burgess andWebster 1980a McBratney et al 1991) Using this method a study area can be stratifiedinto different regions that are reflective of the pedogenetic processes at work In orderto accommodate a trend within a dependent soil variable universal kriging has beenused (Webster 1994) Kriging with external drift is similar to universal kriging but ituses an ancillary variable to represent the trend (McBratney et al 2000) Co-krigingtakes advantage of correlation that may exist between the variable of interest and othermore easily measured variables (Odeh et al 1995) Regression kriging involvesspatially interpolating the residuals from a non-spatial model by kriging and addingthe result to the prediction obtained from that model (Goovaerts 1997 Castrignano etal 2000) Factorial kriging is another method to integrate multivariate data into thestandard kriging routine to extrapolate soil data (Goovaerts 1992) Many authors havecompared these various methods (Laslett et al 1987 Voltz and Webster 1990 Odeh etal 1994 Knotters et al 1995)

Fuzzy logic has been used with geostatistics by various authors to produce new kindsof fuzzy soil maps with continuous classes (Burrough 1989 McBratney and DeGruijter1992 Odeh et al 1992a and reviewed by McBratney and Odeh 1997) The processentails kriging the matrix of membership values determined by the fuzzy k-meansclassifier resulting in a continuous soil surface where individual locations are allowedto belong to more than one class and no rigid boundaries are designated to separate thesoil into discrete units or entities The results of such analysis can be used to assess thepedologic process validity of soil taxonomy by determining whether soils grouptogether into classes that are similar to taxonomic types Fuzzy classes wouldpresumably reflect the main pedologic features within a given area In this sense theclassification is quantitative whereas soil taxonomy is rooted in qualitative discrimina-tion A comparison between the two could provide insightful and help assist soiltaxonomy in making a classification that is more indicative of underlying soil processes

Geostatistics in soil research were originally introduced to quantitatively assess soilvariability within soil mapping units (McBratney et al 1991) in response to criticisms inthe early 1970s that soil unit composition was not well quantified (Beckett and Webster1971) In this regard geostatistics have been very useful having served well the originalgoals set forth by Burgess and Webster (1980b) when they drew kriging to the attentionof soil scientists as a means of spatial prediction At the field scale soil variation islargely due to the effect that topography has on soil genesis Geostatistics have beensuccessfully applied in such environments by using terrain attributes as ancillary datawithin many of the kriging routines described above Such quantitative within-unitvariability of soil properties is very useful in the field of precision agriculture and othersituations (eg pollutants) where very detailed soil attribute information is needed atthe field scale (Heuvelink and Webster 2001)

However geostatistics have not been applied in a wide variety of environments or atlarger scales In order to be successfully applied in different environments geostatisticswill likely require a different suite of ancillary data For example remote sensing data

P Scull et al 185

could be used in arid regions where soil toposequences are less well expressed Atlarger scales of prediction selection of different sets of ancillary variables is requiredbecause different processes define soil character at different scales The most obviousexample is that of climate which may control soil distribution at large scales(continental) but has little explanatory power at the field level Regardless of whetheradequate ancillary data exist the amount of data required to use geostatistics forlandscape-level prediction would be extremely difficult and costly to collect given thestrict sampling protocol required to characterize spatial dependence It is also unclearat what landscape scale soils exhibit spatial autocorrelation

Geostatistical approaches do provide a means of creating continuous soil attributesurfaces to better represent soil continuity (Goal 2) and they can be used to exploit therelationship between environmental variables and soil properties in order to moreefficiently collect soil data (Goal 1) However they do not sufficiently utilize expertknowledge (Goal 3) as no attempt has been made in geostatistical approaches todirectly integrate expert knowledge Fundamentally kriging is a process of interpola-tion designed to predict attribute values in between locations of measured samples Inthis sense geostatistics represent a middle ground between pure interpolation (egnearest neighbour type classifier) in which only measured points for the variable ofinterest are used to determine unknown values and other predictive models thatprimarily use soilndashenvironment correlation to create predictive maps

2 Statistical methods

Statistical methods can be used to exploit the relationship between quantifiablelandscape indices and soil properties to create predictive soil maps For exampleMcKensie and Austin (1993) used a regression to account for a large percentage ofvariation for many soil characteristics (A horizon clay content CEC EC pH bulkdensity and COLE B horizon clay content CEC ESP EC pH bulk density and COLE)using a variety of predictor variables (slope presence or absence of impeding layerrelief landform topographic position) Their results confirm the hypothesis of MooreID et al (1991) that soil character is related to quantifiable landscape indices Howevertheir methods do not provide inference of soil properties at unmeasured sites frommapped environment data because the topographic variables were measured in thefield Linear regression has also been used with terrain variables derived from a 15-mDEM in northeastern Colorado to predict soil attributes (organic matter contentextractable phosphorous pH and texture) at unvisited sites (Moore et al 1993) In thatparticular study 50 of the variance of A-horizon thickness was explained by slope andthe wetness index Gessler et al (1995) also used regression to model A-horizonthickness from topographic variables in southeastern Australia (plan curvature andwetness index r2 = 063 P = 0001) They modeled solum depth and used logisticregression to model E horizon presenceabsence Elsewhere logistic regression hasbeen used to model the presenceabsence of noncalcareous clay loam horizon in centralFrance using terrain attributes from a 20-m DEM (King et al 1999) Exponentialregression has been used to model soil organic carbon using terrain variables (Bell et al2000) in glacial outwash soils in east-central Minnesota Generalized additive models(GAM) have been used less frequently in PSM research Gessler (1996) used a GAM

186 Predictive soil mapping a review

model to predict total soil carbon A horizon depth and solum depth using a variety ofenvironmental predictors (slope elevation wetness index mean annual temperatureprecipitation and radiation)

This small body of research opened the door to more complex methods by demon-strating the existence of quantifiable relationships These authors were able to producesoil attribute maps using raster data models whose scale was dependent upon the gridcell resolution of the environmental data They were successful at exploiting the rela-tionship between quantifiable topographic attributes and soil profile character (Goal 1)The continuous soil attribute surfaces they produce also better represent soil continuitythan the choropleth soil maps produced by traditional soil survey (Goal 2) Howeverthe bulk of these methods (excepting GAMs) are limited by their assumed linear rela-tionship between soil and topographic attributes their assumptions of normallydistributed data and their high data requirements Standard statistical procedures arealso not flexible enough to allow robust integration with a variety of potential datasources such as expert knowledge (Goal 1)

Statistical methods do demonstrate in a quantitative manner that terrain analysis canbe used to predict soil attributes in relatively small areas with homogeneous parentmaterial A large proportion of the research using statistical methods was conducted insemi-arid landscapes at small scales (the largest study area of the entire group was~2000 ha) Obviously for statistical approaches to be most effective they need to bemore universal As such they need to be tested andor developed at larger scales andin more diverse landscapes

3 Decision tree analysis (DTA)

The use of decision tree analysis is just beginning to be explored in predictive soilmapping research although it has been used successfully in the related field ofpredictive vegetation mapping since the early 1990s (Lees and Ritman 1991 MooreDM et al 1991 Franklin 1998) DTA is a form of divisive classification The process oftree modelling involves successively partitioning data (called recursive partitioning inthe tree modelling literature) into increasingly homogeneous subsets which once thepartitioning has ceased are called terminal nodes (Lees and Ritman 1991) Splits orrules defining how to partition the data are selected based on information statistics thatdefine how well the split decreases impurity within the data set (Clark and Pregibon1992) Splits are based on threshold values of an explanatory variable selected bycomparing the increase in resulting purity of node membership for all possiblethresholds and variables The process is iterative growing from the root node (thecomplete data set) to the terminal nodes in a dendritic fashion (Friedl and Brodley1997) Once the tree has been constructed (or grown) it encodes a set of decision rulesthat describe the data partitioning process These rules can be used to classify or predictother data sets (Moore DM et al 1991) Pruning the tree is often necessary to preventthe tree from being overfit to the sample data and to reduce tree complexity Pruningentails combining pairs of terminal nodes into single nodes and can be accomplishedusing cross-validation which yields an initial indication of how large a tree makesrobust predictions (Safavian and Norvig 1991) Cross-validation involves systematical-ly removing portions of the data set and running the remaining sample through the tree

in an iterative manner eventually yielding estimates of the misclassification rates foreach class each node and the whole model (Breiman et al 1984) In this mannerdifferent sized trees can be compared in terms of parsimony

The term DTA is used to collectively refer to all types of tree-based modelling (theword lsquodecisionrsquo is used because it is descriptive indicating that the analysis eventuallyleads to a set of decision rules defining data partitions) The term should be distin-guished from classification tree analysis because the latter refers specifically to DTAwhere the response variable is categorical The term CART (classification andregression trees ndash Breiman et al 1984) is sometimes used but strictly speaking refers tospecific software Friedl and Brodley (1997) provide a review of the decision treealgorithms They divide types of DTA into two classes (1) homogeneous decision treesfor which a single algorithm is used to estimate each split (eg CART) and (2) hybriddecision trees (HDT) for which different splitting methods can be used at differentpoints in the tree (eg Quinlan 1993) They further divide homogeneous decision treesinto univariate (UDT) where single features of the input data define splits and multi-variate decision trees (MDT) where multiple features of the input data can define splitsAccording to this naming convention no distinction is made between methods withdifferent types of response variables although all methods can be used with bothcategorical and continuous response variables In a comparison of these various typesof algorithms on a variety of data sets Friedl and Brodley (1997) found that HDT hadthe highest classification accuracy

The overall aim of DTA is to design a set of predictive rules (eg if geology type Athen soil type B) developed from training data which can then be applied to ageographic data base to predict the value of a response variable (Michaelsen et al 1994)Therefore DTA explicitly uses soilndashlandscape correlation in model development (Goal2) The technique appears promising in soil research but needs to be further tested asnot many of the above types of DTA have been tested in the PSM literature in fact onlyunivariate approaches have been employed For example Lagacherie and Holmes(1997) successfully used univariate DTA to model a categorical response variable soiltype within a training set and then assuming that the training set was representativeextrapolated the model to a much larger region Their work is interesting becausealthough not a single sample came from the area they eventually mapped they wereable to produce a soil map that was more accurate (74 versus 69) than the existingmap produced from traditional methods Cialella et al (1997) also used univariate DTAto predict soil drainage class from a variety of terrain attributes and remotely senseddata They predicted soil drainage class with an average of 78 accuracy ndash impressivegiven that the variation accounted for by a typical soil survey ranges from about halfthe total variance for physical attributes to less than one-tenth for some soil chemicalattributes (Gessler et al 1995)

DTA has been compared with other approaches by several authors In the applicationof erosion modelling DTA results were similar to Artificial Neural Networks (Ellis1996) Both methods achieved high training accuracy (as measured by the KappaStatistic) but in terms of prediction accuracy both methods performed poorly Gessleret al (1995) compared DTA with generalized linear models (GLM) and generalizedadditive models (GAM) to predict A-horizon thickness and concluded that GLM waspreferable to both DTA and GAM McKensie and Ryan (1999) compared regressiontrees and standard linear regression to predict soil properties (total solum depth soil

P Scull et al 187

188 Predictive soil mapping a review

phophorus and soil carbon) using a large variety of predictors (elevation slopecurvature wetness index climate variables geology unit Landsat TM data andGamma radiometric data) and found that different methods work best for differentproperties and overall success hinged on the strength of the relationship between soiland environmental variables Their study of a 50 000 ha forested landscape in southernNew South Wales Australia is one of the few studies at such a small map scale (largearea) or in forested landscape The most extensive comparison of techniques involvingDTA was conducted by McBratney et al (2000) They compared regression techniques(GLM GAM DTA-regression tree) geostatistical techniques (kriging and heterotopicco-kriging) and a hybrid technique (regression kriging) DTA was found to be thepoorest performing of the regression techniques because of the unrealistic predictionsurface generated by the DTA model DTA has also been criticized by other authorsbecause of the stepped prediction surface (Gessler 1996) This phenomenon can beespecially apparent in situations where predictor variables have different resolutionsThe extent to which DTA yields a better spatial representation of soil continuity is afunction of the scale and type of predictor variables used and therefore varies withindividual models

DTA is gaining widespread popularity as a means to develop prediction rules thatcan be rapidly and repeatedly evaluated (Cialella et al 1997 Franklin et al 2000) DTAprovides the following advantages over standard statistical techniques (1) it is easier tointerpret when explanatory variables are both nominal and continuous (2) it isinvariant to monotone re-expressions (transformations) of predictor variables (3) itdeals more satisfactorily with missing data values and outliers (4) it is more adept atcapturing nonadditive and nonlinear behaviour (5) it doesnrsquot make any assumptionsabout data distribution and (6) it is easily updateable as more data are collected(Moore DM et al 1991) The DTA model framework is especially appealing becauseof its capability to integrate a wide range of data sets as explanatory variables

DTA offers a unique opportunity for interaction between soil experts and soilmodellers because the output of the model is a set of rules that can be pedologicallyinterpreted by the soil expert In this sense expert knowledge is used in an implicitmanner in DTA (somewhat effectively achieving Goal 3) While these rules can often beexceedingly complex at minimum the expert can decide whether initial splits makesense given their understanding of the landscape The success of DTA results oftenhinges on the ability of the modeller to make key decisions during the model buildingprocess there is unfortunately no definitive way to determine the most optimal treeThe aid of the soil expert can potentially elucidate this problem

4 Expert systems

A variety of expert system approaches to PSM have been developed to utilize expertknowledge The purpose of such methods is to exploit the information the soil surveyoraccumulates while working in the field by integrating such knowledge into thepredictive model (McCracken and Cate 1986) Unlike the majority of the researchreviewed thus far the dependent variable in many expert systems models is often soiltaxa or mapping unit This apparent disadvantage of expert systems (using classifica-tion to characterize soil continuity) does make them easier to integrate into traditional

soil survey In addition several authors have developed methods to developcontinuous soil property maps from the output of expert system models designed topredict soil unit occurrence

Expert systems are composed of data (information on spatial environmentalvariables eg topography climate etc) a knowledge base (rules and facts related tosoil variation supplied from the soil surveyor) and an inference engine (whichcombines data and the knowledge base to infer logically valid conclusions) (Skidmoreet al 1996) Expert systems differ from conventional models in two ways (1) they storeand manipulate qualitative information (allowing them access to information thatcannot normally be used in other modelling frameworks) and (2) they are structuredas meta-models (the knowledge is separated from the model) (Davis 1993) This allowsthe model to selectively choose which information is relevant at various stages of themodelling process and it allows for information to be easily updated Davis (1993)reviews the application of expert systems to environmental modelling research conclu-ding that the technique is becoming more widely accepted He further states that theapplication of expert systems is constrained by an absence of fundamental knowledgefor rule generation a problem that would appear less relevant to soil mapping giventhe amount of untapped expert knowledge accumulated by the soil surveyor

The first mention of the use of expert systems in pedology was in a paper presenta-tion at the Northeast Committee Soil Survey Conference 1984 (Flach 1985) In herpaper Flach (1985) hinted that recent developments in computer science especiallyexpert systems and artificial intelligence could make modelling a practical mappingtool for soil scientists in the near future A year later McCracken and Cate (1986) hopedto encourage soil scientists to explore expert systems and its application to soil sciencethrough an optimistic article they wrote on the potential use of expert systemsHowever little research was actually conducted in the late 1980s and expert systemshad not yet begun to fulfil the lofty goals set forth by early practitioners (Dale et al1989)

In the early 1990s expert systems approaches to predictive soil mapping began toappear in the literature Skidmore et al (1991) used a Bayesian expert system to mapforest soil into different classes and their results compared favourably with availablesoil maps and actual field-collected data Their methods successfully incorporatedsurveyor knowledge and remotely sensed and digital terrain attributes but failed tobetter characterize continuous soilndashlandscape variability because their final productwas a choropleth map Skidmore et al (1996) revisited their earlier research andassessed the mapping accuracy of their results They found that the soil map producedby the expert system achieved an overall accuracy of 698 (sample size = 53) while themap derived from conventional methods had an accuracy of 736 In addition toknowledge provided by a soil scientist Skidmore et alrsquos (1996) expert system usedvegetation derived for aerial photography and topographic variables derived from a 10-m DEM (wetness index topographic position and slope)

Using an expert system Cook et al (1996) successfully produced a continuous soilproperty map for organic matter content using wetness index aspect and slope asexplanatory variables Their methods were somewhat inefficient because they requireda separate expert system for each soil property of interest However their researchrepresents the only example of expert systems used to directly predict the spatial dis-tribution of a soil property All of the other examples of expert systems in the literature

P Scull et al 189

190 Predictive soil mapping a review

deal with soil type or class The use of expert systems to map soil properties needs tobe explored further

The expert systems discussed thus far have all used Boolean logic within theirtheoretical framework whereby an observation can belong to one and only one classand the soil properties of that class are assigned to the observation With thedevelopment of fuzzy logic and the semantic import model (the second fuzzy logicapproach to creating continuous classes) the opportunity exists to develop fuzzy logic-based expert systems Such systems can be used in conjunction with expert knowledgein situations where experts have a good qualitative idea of how to group data but havetrouble dealing with observations that are not well represented by rigid classificationschemes (Burrough and McDonnell 1998) This method can be particularly useful insituations when taxonomic schemes have been previously developed as is the case insoil taxonomy Several examples of this type of approach were published in the mid-1990s by A Zhu and colleagues (Zhu and Band 1994 Zhu et al 1996 1997 Zhu1997ab) Such systems proved useful for mapping soil at unvisited locations usingsurveyor knowledge and were also capable of producing continuous soil propertymaps The use of fuzzy logic within the theoretical framework of the expert systemallows the complex nature of soil to propagate through the modelling process neversubjected to classification schemes that filter out potentially useful lsquonoisersquo The use offuzzy logic also gives the soil surveyor more latitude during the interview processwhen the knowledge base is defined for the expert system Zhu et al (1997) developeda fuzzy logic-based expert system called SoLIM to determine the similarity of eachgrid cell in a study area to the various taxonomic mapping units delimited by the soilsurvey Continuous soil attribute maps were calculated using the similarity values andtheir relative soil survey determined attribute values (effectively achieving Goal 2) Theresulting data (which proved more accurate than soil survey data once field checked)consists of a raster grid whose resolution is determined by the resolution of the inputenvironmental and digital elevation data As noted previously this type of data modelis more applicable to environmental modelling than the choropleth map (Burrough andMcDonnell 1998)

Expert system approaches to PSM are capable of exploiting soil surveyor knowledgeby developing rule-based systems that imitate the surveyorrsquos conceptual model of soilvariability (the primary focus of Goal 3) The method would appear extremely usefulfor mapping projects (such as those conducted by the NRCS-NCSS) where fieldwork isinitially conducted to determine soilndashlandscape relations Expert system developmentcould be directly inserted into the traditional soil survey mapping approach as asubstitute for the step where the surveyor converts hisher conceptual model into achoropleth map Rather that knowledge could be incorporated into the expert systemwhich could be used to predictively map soil The resulting raster map would be morescientifically based and explicit than the hand-drawn choropleth maps of the past Itsscale would not be limited to that of the aerial photography but rather to the scale ofthe environmental data

Despite these advantages the expert system approach has some drawbacks Becausethe method is dependent upon expert knowledge it cannot be applied whereenvironmentndashsoil relations are poorly understood Of course this criticism can be madeof all PSM models as inductive or deductive knowledge of soilndashenvironment relationsis a prerequisite for PSM However expert systems are deductive models and as such

are not driven by any specific field-collected data (although presumably the soil experthas field experience in the mapping area) Expert systems donrsquot afford the opportunityto first statistically document landscapendashsoil relations and then extrapolate the resultsbecause expert systems do not directly use sample soil data (lsquohard datarsquo) to determinesoilndashlandscape relations Thus the utilization of the relationship between environmen-tal variables and soil properties (Goal 1) is only indirectly achieved Expert systemsapproaches have been demonstrated to be extremely effective in a small number of casestudies The possibility of satisfying all three goals discussed in this review makesexpert systems a predictive soil mapping method that needs to be further tested ndash bothgeographically and across different scales of analysis

V Conclusion

Most of the predictive soil mapping research outlined in this review was conducted atvery large map scales (over small areas) In fact the majority of the research wasconcerned with assessing the spatial variability of soil character within individual fieldsor across soil toposequences The primary driving force behind this type of research hasbeen the need to provide accurate soils information for agriculture and ecologicalmodels It is clear that terrain attributes are powerful predictors at the local scaleGeostatistical tools have been successful at using terrain attributes and the spatialdependence of soil properties to interpolate between existing data points withinindividual fields Across soil toposequences statistical approaches provide a usefulmeans of predicting soil character PSM research has been most successful at the fieldscale because many of the soil-forming factors are held constant For example the neteffect of four of Jennyrsquos five soil-forming factors (climate organism parent material andtime) was minimal within many of the studies reviewed While some research hastackled larger areas of study there exists a deficiency in our ability to predictively mapsoils at smaller map scales Since the distribution of soil is scale-dependent differentPSM methods and predictors are likely to work better at different scales Focus in thefuture must continue to move toward working over larger spatial extents of study inorder to produce landscape-scale soil information

Further a large proportion of the research was conducted in semi-arid gentlysloping agricultural landscapes Humid forestlands mountainous regions and desertshave received little attention As such PSM methods need to continue to be testedandor developed in a wider variety of landscapes where spatial soil distributions canbe more complex Different methods will likely be successful to different degrees indifferent environments Whereas terrain attributes are dominant predictors of soilcharacter across toposequences in prairie lands other predictors are likely necessary inareas where soils do not develop into clear toposequences For example remote sensingdata has been a useful predictor of soil chronosequences in desert landscapes wheresurface appearance is often related to soil character Surficial geology is often reflectiveof soil character in mountain regions where soils are thin and significant bedrock existsat the surface Focus in the future must continue to determine which methods andpredictors work best in which environments

Since the most useful PSM approach will vary across spatial scales and environmen-tal gradients the method used should be driven by the mapping objectives of the

P Scull et al 191

192 Predictive soil mapping a review

project Whereas the traditional soil survey of the past was expected to meet the needsof a diverse group of end users PSM methods will vary given the objectives of thesurvey As a result the end users of the soil data need to play a more active role in thesurvey process Because precision agriculture and large-scale mapping has been theprimary focus of PSM there exists a deficiency in our ability to predictively map soilsfor the purpose of general land use planning and management Even though lessdetailed soil information is needed for these purposes the development of PSMmethods is complicated by the fact that soils exhibit complex spatial variability atsmall map scales where soil-forming environments vary greatly from one location tothe next

A number of alternative methods of characterizing the continuous nature of the soillandscape have been developed Thus far most of the PSM research has provided soilinformation in a nonobject form (as opposed to defining soil types as independententities) Soil data have been generated organized and presented in the form of eitherisorithmic maps or fine-scale raster grids Both of these data models are field-viewmodels of geographic space which allow the soil to be perceived as a constantlyvarying surface Two distinct approaches have been employed (1) mapping individualsoil properties and (2) mapping continuously varying (fuzzy) soil classes Mappingindividual properties is the most common approach and will likely continue todominate PSM research The use of fuzzy soil classes in the literature is less commonand more difficult to be integrated into standard mapping procedures (such as use ofsoil taxa) because the concept radically differs from the traditional view of the soillandscape However fuzzy soil classes (generated using the fuzzy-k means approach) dohave the potential to help further advance Soil taxonomy by identifying taxonomicclasses that are more reflective of pedologic processes at work

Expert systems have been greatly underutilized in PSM research especiallyconsidering how effective a small number of case studies have been Expert systemshave the potential to satisfy successfully all three goals discussed throughout thereview They also have the potential to bridge the gap between traditional approachesand PSM methods because field soil scientists do not have to change their conceptualapproach to mapping They can still conduct field reconnaissance to determinesoilndashlandscape relationships Afterwards though they can use expert systems to betterexploit the knowledge they have garnered Such an approach could help ease thetransition from traditional soil survey to more scientifically explicit methods Expertsystems are also the most fruitful approach to utilizing a wealth of data that has alreadybeen collected in a nonexplicit manner the wealth of expert knowledge (in the form ofthe conceptual model) that senior soil mappers have accumulated

Focus in the future must also move in the direction of operability The utility of manyPSM approaches has been clearly demonstrated yet changes in how soils are mappedand perceived has been slow In the USA various members of the National CooperativeSoil Survey have called for a change in the philosophy of survey (from mappingdiscrete soil types to mapping continuous varying soil properties McSweeny et al1994) but little change has actually taken place To realize the potential of PSM in con-tributing to an overall change in standard soil mapping procedures practitionersworking in the field of PSM need to form working relationships with field soilscientists PSM methods need also to be presented in a manner that is comprehensibleto the soil science trained field mapper Predictive soil mapping is a relatively recent

phenomenon yet much progress has been made In the process of mapping soilknowledge will continue to advance

P Scull et al 193

References

Agbu PA Fehrenbacher DJ and Jansen IJ1990 Statistical comparison of SPOT spectralmaps with field soil maps Soil Science Society ofAmerica Journal 54 818ndash18

Barrett L 1999 Particulars in contextmaintaining a balance in soil geography Annalsof the Association of America Geographers 89707ndash13

Beckett PHT and Webster R 1971 Soilvariability a review Soils and Fertilizers 341ndash15

Bell JC Grigal DF and Bates PC 2000 A soil-terrain model for estimating spatial patterns ofsoil organic carbon In Wilson J and Gallant Jeditors Terrain analysis principles and applica-tions New York City NY John Wiley and Sons295ndash310

Birkeland PW 1999 Soil and geomorphologyThird edition New York NY Oxford UniversityPress

Breiman L Friedman JH Olshen RA andStone CJ 1984 Classification and regressiontrees Belmont CA Wadsworth

Brule FJ 1996 Fuzzy systems ndash a tutorialhttp newsgroup compai httpwwwquadralaycom (last accessed 30 August 2001)

Bunkin FV and Bunkin AF 2000 Lidarsounding of water soil and plants Atmosphericand Oceanic Optics 13 54ndash72

Buol SW Hole FD McCracken RJ andSouthard RJ 1997 Soil genesis and classifica-tion Ames IA Iowa State University Press

Burgess TM and Webster R 1980a Optimalinterpolation and isarithmic mapping of soilproperties the semi-variogram and punctualkriging Journal of Soil Science 31 315ndash31

ndashndashndashndash 1980b Optimal interpolation and isarithmicmapping of soil properties block kringingJournal of Soil Science 31 331ndash41

Burrough PA 1989 Fuzzy mathematicalmethods for soil survey and land evaluationJournal of Soil Science 40 477ndash92

Burrough PA and McDonnell RA 1998Principles of geographic information systems(Revised edition) Oxford Clarendon Press

Burrough PA Beckett PHT and Jarvis MG1971 The relation between cost and utility insoil survey Journal of Soil Science 22 368ndash81

Burrough PA Van Gaans PMF and

Hootsman R 1997 Continuous classificationin soil survey spatial correlation confusionand boundaries Geoderma 77 115ndash35

Cambell JB 1977 Variation of selectedproperties across a soil boundary Soil ScienceSociety of America Journal 41 578ndash82

Cambell JB and Edmonds WJ 1984 Themissing geographic dimension to soiltaxonomy Annals of the Association of AmericanGeographers 74 83ndash97

Castrignano A Giugliarini L Risaliti R andMartinelli N 2000 Study of spatial relation-ships among some soil physico-chemicalproperties of a field in central Italy using multi-variate geostatistics Geoderma 97 39ndash60

Cialella AT Dubayah R Lawrence W andLevine E 1997 Predicting soil drainage classusing remotely sensed and digital elevationdata Journal of Soil Science 62(2) 171ndash78

Clark RN and Swayze GA 1996 Evolution inimaging spectroscopy analysis and sensorsignal-to-noise an examination of how far wehave come Summaries of the sixth annual JPLairborne Earth science workshop 4ndash8 March 1996AVIRIS Workshop Vol 1 5

Clarke LA and Pregibon D 1992 Tree-basedmodels In Chambers J and Hastie J editorsStatistical models in S Pacific Grove Wadsworthand Brooks 377ndash419

Coleman TL Agbu PA and MontgomeryOL 1993 Spectral differentiation of surfacesoils and soil properties ndash is it possible fromspace platforms Soil Science 155 283ndash93

Cook SE Corner RJ Grealish G GesslerPE and Chartres CJ 1996 A rule-basedsystem to map soil properties Soil ScienceSociety of America Journal 60 1893ndash900

Cruickshank JG 1972 Soil geography New YorkNY John Wiley amp Sons

Csillag F Pasztor L Biehl LL 1993 Spectralband selection for the characterization ofsalinity status of soils Remote Sensing ofEnvironment 43 231ndash42

Dale MB McBratney AB and Russell JS1989 On the role of expert systems andnumerical taxonomy in soil classificationJournal of Soil Science 40 223ndash34

Davis JR 1993 Expert systems and environ-

it an important component of PSM Fuzzy logic is an alternative to Boolean logic thatattempts to recognize the concept of partial truth (Brule 1996) Dr Lotfi Zadeh (1965)introduced the concept and accompanying mathematics in his seminal work lsquoFuzzysetsrsquo The theory permits partial class membership in contrast to traditional set theorywhere set memberships are crisp and binary (ie a soil sample is either completelyType A or it is not at all Type A) Central to the fuzzy concept is the idea that objects innature rarely fit exactly the classification types to which they are assigned (Zadeh1965) Rather they show varying signs of similarity to multiple classes (ie an observedsoil pedon often resembles more than one of the defined soil series within the area) Byusing fuzzy membership values (ranging from 0 nonmembership to 1 totalmembership) within predictive soil models to express degrees of similarity generaliza-tion problems associated with classification schemes (filtering of information) areminimized and the complex nature of soil data is allowed to propagate through themodelling process Similarity values between 0 and 1 are not comparable to proportionsand need not add up to 1 Within Boolean logic probability statements refer to thelikelihood of an outcome the soil sample is either one series or another With fuzzylogic a given sample is not definitively a member of the subset of any one particularseries Fuzzy logic is especially useful in soil research because of the continuous andcomplex nature of the soil landscape It serves as an important alternative to thesubjective rigidity imposed on soils data by Soil taxonomy Several recent articlesprovide a thorough review of the use of fuzzy sets in soil science (Burrough 1989McBratney and Odeh 1997 Burrough et al 1997 De Gruijter et al 1997)

Within PSM research two different approaches to creating continuous classes usingfuzzy logic exist The first is based on the fuzzy-k-means classifier which partitionsobservations in multivariate space into natural classes This approach is similar tocluster analysis and numerical taxonomy but the resulting classes are continuous witheach observation assigned a fuzzy membership value that characterizes its degree ofsimilarity to each individual class The concept has been integrated into geostatisticalmethods and will be discussed in more detail below (see Section IV1) The secondapproach is known as the Semantic Import model (SI) and is used in situations whenclassification schemes are pre-defined and class limits are relatively well understoodThe SI model is commonly used in concert with expert knowledge and will bediscussed in the expert systems section (see Section IV4)

IV Recent advances in predictive soil mapping

Within the last decade many authors have sought to model the soil landscape using avariety of methods Literature in this field could be summarized many different waysbut we concentrate on the literature that directly addresses the goals of predictive soilmapping stated in the introduction (see Table 1) Therefore we will review research thatattempts to exploit the relationship between quantifiable landscape indices and soilcharacter in order to model the soil landscape in a more continuous and thereforerealistic manner

The research reviewed here is distinguished from decades of previous researchdocumenting the correlation between landscape position and soil attributes (reviewedby Hall and Olsen 1991) That body of research is informative but not useful for

P Scull et al 179

180 Predictive soil mapping a review

Table 1

Selected

recen

t literature on pred

ictive

soil m

apping an

d m

apping

(cited in

this article) d

escribing the mod

ellin

g metho

dused the

dep

enden

t variables used a

nd the

env

iron

men

tal v

ariables (ex

plan

atory) used in

the

mod

els

Stud

yMod

ellin

g metho

dDep

ende

nt variables

Environm

ental variab

les

Goa

ls atta

ined

a

Bell et

al 20

00Line

ar and

exp

onen

tial

Total soil orga

nic

Slop

e curvature aspe

ct

ECx

regression

carbon

hillslope

position

Burge

ss and

Web

ster

Punc

tual and

block

Na co

nten

t co

ver

Non

eC

1980

ab

kriging

loam

thickn

ess

ston

e co

nten

t

Burroug

h 198

9Fu

zzy mathe

matical

ndashndash

rpmetho

ds

Burroug

h et

al 19

97Con

tinuo

us classificatio

nndash

ndashrp

Castrigna

no eta

l 200

0Factorial kriging

CEC

pH N

P K

Non

e ex

plicitly used

Cx

Na

Cialella

eta

l 199

7Classifica

tion tree

Drainag

e class

Elev aspe

ct NDVI

Ec

Coo

k et

al 19

96Baysian

rule-ba

sed metho

dsOrgan

ic m

atter

Slop

e aspec

t wetne

ssEcX

inde

x

Dale et

al 19

89 (a review

)Ex

pert systems

ndashndash

ndash

Ellis 19

96Decision tree

ana

lysis

Soil erosion class

Slop

e aspec

t wetne

ssEc

neutral ne

tworks

inde

x flow le

ngth and

accu

mulation Lan

dsat

TM tree

cov

er

Gessler 1

996

A large

variety of statistic

Field an

d labo

ratory

A variety of digital

EC

metho

dsco

llected

phy

sica

len

vironm

ental da

tach

emical and

morph

olog

ical soil

prop

ertie

s

Gessler eta

l 199

5Line

ar and

logit reg

ression

A horizon

and

solum

Curvature CTI topo

EC

depth E horizon

positio

npresen

ce

Goo

vaerts 19

92Factorial kriging

Total carbon

Non

eC

P Scull et al 181pH

N CEC

extractab

lecatio

ns (K Ca M

g)Hartemink et

al 20

01ndash

ndashndash

rp

Heu

velin

k an

d W

ebster 20

01ndash

ndashndash

rp

Hew

itt 19

93ndash

ndashndash

rp

Indo

rante et

al 19

96ndash

ndashndash

rp

King et

al 19

99Lo

gistic regression

Presen

ceabsen

ceSlop

e aspec

t po

t solar

Ec

Non

calc c

lay-loam

energy

Kno

tters eta

l 199

5Kriging

co

-kriging

So

ft layer de

pth

Hillslop

e po

sitio

nCx

regression

kriging

Laga

cherie and

Holmes 19

97Classifica

tion tree

Map

ping

unit

Geo

logy variou

s topo

Ec

indices

Laslett e

tal 198

7Kriging

splin

es tren

dpH

Non

ec

surfac

e nea

rest neigh

bor

McB

ratney 1

992

ndashndash

ndashrp

McB

ratney

eta

l 199

1Block kriging

Clay co

nten

tNon

eC

McB

ratney

eta

l 200

0ndash

ndashndash

rp

McB

ratney

and

Ode

h 199

7Fu

zzy sets in

soil scienc

endash

ndashRp

McB

ratney

and

de Gruijter

Fuzz

y-k-mea

ns w

ithFu

zzy classes

Field co

llected

phy

sical

C19

92ex

tragrade

sch

emical and

morph

olog

ical soil

prop

ertie

s

McC

rack

en and

Cate 198

6Artificial intellig

ence

ndash

ndashrp

expe

rt systems

McK

ensie an

d Austin

19

93Gen

eralized

line

ar m

odels

Clay co

nten

t CEC

Slop

e relief land

form

ec

(logit)

pH EC

COLE

slop

e po

sitio

nbu

lk den

sity and

othe

rs

Moo

re eta

l 199

3Line

ar reg

ression

A horizon

dep

th O

MSlop

e w

etne

ss and

strea

mEC

and P co

nten

t pH

power in

dices aspect

curvature

182 Predictive soil mapping a reviewTa

ble 1

Con

tinu

ed

Stud

yMod

ellin

g metho

dDep

ende

nt variables

Environm

ental v

ariables

Goa

ls atta

ined

a

McK

ensie an

d Ryan 199

9Reg

ression tree and

linea

rSo

lum dep

th P an

dElevation slope

CE

regression

N con

tent

curvature CTI

contribu

ting area do

wn-

slop

e mea

ns for slope

clim

ate da

ta Presco

ttInde

x G

amma

Rad

iometry La

ndsat T

M

and Geo

logy unit

Ode

h et

al 19

92ab

Fuzz

y-c-mea

ns and

Fuzzy classes

Field-co

llected

phy

sica

lC

kriging

chem

ical and

morph

olog

ical soil

prop

ertie

s

Ode

h et

al 19

94

Reg

ression kriging

co

-So

lum dep

th de

pth

Slop

e aspect cu

rvature

eC

1995

kriging regression

to bed

rock gravel

kriging

and clay con

tent

Skidmore et

al 19

91Bayesian expe

rt system

Soil land

scap

e un

itVeg type

wetne

ss in

dex

EcX

1996

grad

ient terrain po

sitio

n

Voltz and

Web

ster 19

90Kriging

cu

bic splin

eClay co

nten

tNon

eC

Web

ster 1

994

Dev

elop

men

t of

ndashndash

rppe

dometrics

Zhu

199

7ab

Fuzz

y logic expe

rt system

Soil series A

horizon

Elev pm aspe

ct c

anop

yECX

Zhu

and

Ban

d 199

4(SoL

IM)

depth in

dividu

alco

verage

grad

ient

Zhu

eta

l 199

7series m

aps

curvature

Not

es

a Letters refer to the de

gree

to w

hich

the

goa

ls of PS

M defined

in the

introdu

ction are achiev

ed

Soilndash

e nvironm

ent relatio

ns utilized

(letter E) be

tter represen

tatio

n of soil c o

ntinuity (C

) an

d ex

pert kno

wledg

e utilize

d (X) c

orrespon

d to goa

ls 1 2 and

3respec

tively from

the

introdu

ction and

cap

ital lette

rs (E C X) ind

icate the metho

d is relatively more successful th

an th

ose metho

ds den

oted

by lower

case le

tters (e c x

) rp ind

icates rev

iew pap

ers

P Scull et al 183

predictive mapping because landscape position is never quantified (rather positionwas often qualitatively defined eg toe-slope) and thus the documented relationshipscannot be generalized using environmental data and digital elevation models to predictsoil character at unvisited sites

Table 1 documents modelling methods model variables and the extent to which thereferenced research satisfies the previously defined goals of PSM (lsquoGoals attainedrsquocolumn) Soil-environment relations utilized (letter E) better representation of soilcontinuity (C) and expert knowledge utilized (X) correspond to goals 1 2 and 3respectively from the introduction and capital letters (E C X) indicate the method isrelatively more successful than those methods denoted by lower case letters (e c x)Citations with no letters present within the lsquoGoals attainedrsquo column do not address theaforementioned goals For example the Cialella et al (1997) received a rating of lsquoEcrsquomeaning that the methods employed successfully utilized environmentalndashsoil characterrelations (E) and somewhat successfully presented a better method of representing soilcontinuity (c) The ratings are provided simply to help organize the literature that wasreviewed Review papers are included within the table denoted by lsquorprsquo Geostatisticalmethods are not included in the table because they have been comprehensivelysurveyed elsewhere (Odeh et al 1994 Burrough et al 1997 McBratney et al 2000Heuvelink and Webster 2001) and because the objectives and assumptions of geostatis-tical methods differ slightly from other PSM research We briefly outline thesedifferences in the following section

1 Geostatistical methods

Geostatistics are a subset of traditional statistics that deal primarily with spatial dataand account for spatial autocorrelation using kriging as the spatial interpolator Theconcept is based upon the theory of regionalized variables which was mainlydeveloped by Matheron (1963) and Krige (1963) Kriging is a form of weighted localaveraging that uses a measure of spatial dependence the variogram to determine theweights applied to the data when computing the averages Geostatistical methods havebeen used in predictive soil mapping research to spatially interpolate soil propertyvalues at unmeasured sites from field-collected data

Burgess and Webster (1980a b) were the first to introduce ordinary kriging to the soilcommunity and since that time an enormous amount of work has been published Forexample ordinary kriging has been used to interpolate many different soil propertiesincluding pollution trace element deficiencies salinity and fertility (Heuvelink andWebster 2001) Ordinary kriging has been criticized for a variety of reasons Forexample Laslett et al (1987) reported that several authors had criticized geostatisticsbecause kriging is a global rather than local technique failing to take into accountknowledge of soil materials and processes Other authors have criticized geostatisticsbecause they are excessively data dependent requiring a large number of closelyspaced data points (Zhu 1997a) As Webster and Oliver (1992) suggest in excess of ahundred samples may be needed to use geostatistics at the field scale because of highspatial variability of soil in some areas Geostatistics also assume spatial autocorrela-tion which sometimes may be a poor assumption in complex terrain where abruptchanges in soil-forming factors occur (McBratney et al 2000) Ordinary kriging by itself

184 Predictive soil mapping a review

does not satisfy two of the three goals of PSM presented in the introduction ndash it does notadequately incorporate expert knowledge and it does not exploit the relationshipbetween environmental variables and soil properties

Ordinary kriging has been modified in a variety of ways to better incorporateancillary data and known soilndashlandscape relationships Block kriging involvesdetermining estimates over meaningful areas rather than at specific points (Burgess andWebster 1980a McBratney et al 1991) Using this method a study area can be stratifiedinto different regions that are reflective of the pedogenetic processes at work In orderto accommodate a trend within a dependent soil variable universal kriging has beenused (Webster 1994) Kriging with external drift is similar to universal kriging but ituses an ancillary variable to represent the trend (McBratney et al 2000) Co-krigingtakes advantage of correlation that may exist between the variable of interest and othermore easily measured variables (Odeh et al 1995) Regression kriging involvesspatially interpolating the residuals from a non-spatial model by kriging and addingthe result to the prediction obtained from that model (Goovaerts 1997 Castrignano etal 2000) Factorial kriging is another method to integrate multivariate data into thestandard kriging routine to extrapolate soil data (Goovaerts 1992) Many authors havecompared these various methods (Laslett et al 1987 Voltz and Webster 1990 Odeh etal 1994 Knotters et al 1995)

Fuzzy logic has been used with geostatistics by various authors to produce new kindsof fuzzy soil maps with continuous classes (Burrough 1989 McBratney and DeGruijter1992 Odeh et al 1992a and reviewed by McBratney and Odeh 1997) The processentails kriging the matrix of membership values determined by the fuzzy k-meansclassifier resulting in a continuous soil surface where individual locations are allowedto belong to more than one class and no rigid boundaries are designated to separate thesoil into discrete units or entities The results of such analysis can be used to assess thepedologic process validity of soil taxonomy by determining whether soils grouptogether into classes that are similar to taxonomic types Fuzzy classes wouldpresumably reflect the main pedologic features within a given area In this sense theclassification is quantitative whereas soil taxonomy is rooted in qualitative discrimina-tion A comparison between the two could provide insightful and help assist soiltaxonomy in making a classification that is more indicative of underlying soil processes

Geostatistics in soil research were originally introduced to quantitatively assess soilvariability within soil mapping units (McBratney et al 1991) in response to criticisms inthe early 1970s that soil unit composition was not well quantified (Beckett and Webster1971) In this regard geostatistics have been very useful having served well the originalgoals set forth by Burgess and Webster (1980b) when they drew kriging to the attentionof soil scientists as a means of spatial prediction At the field scale soil variation islargely due to the effect that topography has on soil genesis Geostatistics have beensuccessfully applied in such environments by using terrain attributes as ancillary datawithin many of the kriging routines described above Such quantitative within-unitvariability of soil properties is very useful in the field of precision agriculture and othersituations (eg pollutants) where very detailed soil attribute information is needed atthe field scale (Heuvelink and Webster 2001)

However geostatistics have not been applied in a wide variety of environments or atlarger scales In order to be successfully applied in different environments geostatisticswill likely require a different suite of ancillary data For example remote sensing data

P Scull et al 185

could be used in arid regions where soil toposequences are less well expressed Atlarger scales of prediction selection of different sets of ancillary variables is requiredbecause different processes define soil character at different scales The most obviousexample is that of climate which may control soil distribution at large scales(continental) but has little explanatory power at the field level Regardless of whetheradequate ancillary data exist the amount of data required to use geostatistics forlandscape-level prediction would be extremely difficult and costly to collect given thestrict sampling protocol required to characterize spatial dependence It is also unclearat what landscape scale soils exhibit spatial autocorrelation

Geostatistical approaches do provide a means of creating continuous soil attributesurfaces to better represent soil continuity (Goal 2) and they can be used to exploit therelationship between environmental variables and soil properties in order to moreefficiently collect soil data (Goal 1) However they do not sufficiently utilize expertknowledge (Goal 3) as no attempt has been made in geostatistical approaches todirectly integrate expert knowledge Fundamentally kriging is a process of interpola-tion designed to predict attribute values in between locations of measured samples Inthis sense geostatistics represent a middle ground between pure interpolation (egnearest neighbour type classifier) in which only measured points for the variable ofinterest are used to determine unknown values and other predictive models thatprimarily use soilndashenvironment correlation to create predictive maps

2 Statistical methods

Statistical methods can be used to exploit the relationship between quantifiablelandscape indices and soil properties to create predictive soil maps For exampleMcKensie and Austin (1993) used a regression to account for a large percentage ofvariation for many soil characteristics (A horizon clay content CEC EC pH bulkdensity and COLE B horizon clay content CEC ESP EC pH bulk density and COLE)using a variety of predictor variables (slope presence or absence of impeding layerrelief landform topographic position) Their results confirm the hypothesis of MooreID et al (1991) that soil character is related to quantifiable landscape indices Howevertheir methods do not provide inference of soil properties at unmeasured sites frommapped environment data because the topographic variables were measured in thefield Linear regression has also been used with terrain variables derived from a 15-mDEM in northeastern Colorado to predict soil attributes (organic matter contentextractable phosphorous pH and texture) at unvisited sites (Moore et al 1993) In thatparticular study 50 of the variance of A-horizon thickness was explained by slope andthe wetness index Gessler et al (1995) also used regression to model A-horizonthickness from topographic variables in southeastern Australia (plan curvature andwetness index r2 = 063 P = 0001) They modeled solum depth and used logisticregression to model E horizon presenceabsence Elsewhere logistic regression hasbeen used to model the presenceabsence of noncalcareous clay loam horizon in centralFrance using terrain attributes from a 20-m DEM (King et al 1999) Exponentialregression has been used to model soil organic carbon using terrain variables (Bell et al2000) in glacial outwash soils in east-central Minnesota Generalized additive models(GAM) have been used less frequently in PSM research Gessler (1996) used a GAM

186 Predictive soil mapping a review

model to predict total soil carbon A horizon depth and solum depth using a variety ofenvironmental predictors (slope elevation wetness index mean annual temperatureprecipitation and radiation)

This small body of research opened the door to more complex methods by demon-strating the existence of quantifiable relationships These authors were able to producesoil attribute maps using raster data models whose scale was dependent upon the gridcell resolution of the environmental data They were successful at exploiting the rela-tionship between quantifiable topographic attributes and soil profile character (Goal 1)The continuous soil attribute surfaces they produce also better represent soil continuitythan the choropleth soil maps produced by traditional soil survey (Goal 2) Howeverthe bulk of these methods (excepting GAMs) are limited by their assumed linear rela-tionship between soil and topographic attributes their assumptions of normallydistributed data and their high data requirements Standard statistical procedures arealso not flexible enough to allow robust integration with a variety of potential datasources such as expert knowledge (Goal 1)

Statistical methods do demonstrate in a quantitative manner that terrain analysis canbe used to predict soil attributes in relatively small areas with homogeneous parentmaterial A large proportion of the research using statistical methods was conducted insemi-arid landscapes at small scales (the largest study area of the entire group was~2000 ha) Obviously for statistical approaches to be most effective they need to bemore universal As such they need to be tested andor developed at larger scales andin more diverse landscapes

3 Decision tree analysis (DTA)

The use of decision tree analysis is just beginning to be explored in predictive soilmapping research although it has been used successfully in the related field ofpredictive vegetation mapping since the early 1990s (Lees and Ritman 1991 MooreDM et al 1991 Franklin 1998) DTA is a form of divisive classification The process oftree modelling involves successively partitioning data (called recursive partitioning inthe tree modelling literature) into increasingly homogeneous subsets which once thepartitioning has ceased are called terminal nodes (Lees and Ritman 1991) Splits orrules defining how to partition the data are selected based on information statistics thatdefine how well the split decreases impurity within the data set (Clark and Pregibon1992) Splits are based on threshold values of an explanatory variable selected bycomparing the increase in resulting purity of node membership for all possiblethresholds and variables The process is iterative growing from the root node (thecomplete data set) to the terminal nodes in a dendritic fashion (Friedl and Brodley1997) Once the tree has been constructed (or grown) it encodes a set of decision rulesthat describe the data partitioning process These rules can be used to classify or predictother data sets (Moore DM et al 1991) Pruning the tree is often necessary to preventthe tree from being overfit to the sample data and to reduce tree complexity Pruningentails combining pairs of terminal nodes into single nodes and can be accomplishedusing cross-validation which yields an initial indication of how large a tree makesrobust predictions (Safavian and Norvig 1991) Cross-validation involves systematical-ly removing portions of the data set and running the remaining sample through the tree

in an iterative manner eventually yielding estimates of the misclassification rates foreach class each node and the whole model (Breiman et al 1984) In this mannerdifferent sized trees can be compared in terms of parsimony

The term DTA is used to collectively refer to all types of tree-based modelling (theword lsquodecisionrsquo is used because it is descriptive indicating that the analysis eventuallyleads to a set of decision rules defining data partitions) The term should be distin-guished from classification tree analysis because the latter refers specifically to DTAwhere the response variable is categorical The term CART (classification andregression trees ndash Breiman et al 1984) is sometimes used but strictly speaking refers tospecific software Friedl and Brodley (1997) provide a review of the decision treealgorithms They divide types of DTA into two classes (1) homogeneous decision treesfor which a single algorithm is used to estimate each split (eg CART) and (2) hybriddecision trees (HDT) for which different splitting methods can be used at differentpoints in the tree (eg Quinlan 1993) They further divide homogeneous decision treesinto univariate (UDT) where single features of the input data define splits and multi-variate decision trees (MDT) where multiple features of the input data can define splitsAccording to this naming convention no distinction is made between methods withdifferent types of response variables although all methods can be used with bothcategorical and continuous response variables In a comparison of these various typesof algorithms on a variety of data sets Friedl and Brodley (1997) found that HDT hadthe highest classification accuracy

The overall aim of DTA is to design a set of predictive rules (eg if geology type Athen soil type B) developed from training data which can then be applied to ageographic data base to predict the value of a response variable (Michaelsen et al 1994)Therefore DTA explicitly uses soilndashlandscape correlation in model development (Goal2) The technique appears promising in soil research but needs to be further tested asnot many of the above types of DTA have been tested in the PSM literature in fact onlyunivariate approaches have been employed For example Lagacherie and Holmes(1997) successfully used univariate DTA to model a categorical response variable soiltype within a training set and then assuming that the training set was representativeextrapolated the model to a much larger region Their work is interesting becausealthough not a single sample came from the area they eventually mapped they wereable to produce a soil map that was more accurate (74 versus 69) than the existingmap produced from traditional methods Cialella et al (1997) also used univariate DTAto predict soil drainage class from a variety of terrain attributes and remotely senseddata They predicted soil drainage class with an average of 78 accuracy ndash impressivegiven that the variation accounted for by a typical soil survey ranges from about halfthe total variance for physical attributes to less than one-tenth for some soil chemicalattributes (Gessler et al 1995)

DTA has been compared with other approaches by several authors In the applicationof erosion modelling DTA results were similar to Artificial Neural Networks (Ellis1996) Both methods achieved high training accuracy (as measured by the KappaStatistic) but in terms of prediction accuracy both methods performed poorly Gessleret al (1995) compared DTA with generalized linear models (GLM) and generalizedadditive models (GAM) to predict A-horizon thickness and concluded that GLM waspreferable to both DTA and GAM McKensie and Ryan (1999) compared regressiontrees and standard linear regression to predict soil properties (total solum depth soil

P Scull et al 187

188 Predictive soil mapping a review

phophorus and soil carbon) using a large variety of predictors (elevation slopecurvature wetness index climate variables geology unit Landsat TM data andGamma radiometric data) and found that different methods work best for differentproperties and overall success hinged on the strength of the relationship between soiland environmental variables Their study of a 50 000 ha forested landscape in southernNew South Wales Australia is one of the few studies at such a small map scale (largearea) or in forested landscape The most extensive comparison of techniques involvingDTA was conducted by McBratney et al (2000) They compared regression techniques(GLM GAM DTA-regression tree) geostatistical techniques (kriging and heterotopicco-kriging) and a hybrid technique (regression kriging) DTA was found to be thepoorest performing of the regression techniques because of the unrealistic predictionsurface generated by the DTA model DTA has also been criticized by other authorsbecause of the stepped prediction surface (Gessler 1996) This phenomenon can beespecially apparent in situations where predictor variables have different resolutionsThe extent to which DTA yields a better spatial representation of soil continuity is afunction of the scale and type of predictor variables used and therefore varies withindividual models

DTA is gaining widespread popularity as a means to develop prediction rules thatcan be rapidly and repeatedly evaluated (Cialella et al 1997 Franklin et al 2000) DTAprovides the following advantages over standard statistical techniques (1) it is easier tointerpret when explanatory variables are both nominal and continuous (2) it isinvariant to monotone re-expressions (transformations) of predictor variables (3) itdeals more satisfactorily with missing data values and outliers (4) it is more adept atcapturing nonadditive and nonlinear behaviour (5) it doesnrsquot make any assumptionsabout data distribution and (6) it is easily updateable as more data are collected(Moore DM et al 1991) The DTA model framework is especially appealing becauseof its capability to integrate a wide range of data sets as explanatory variables

DTA offers a unique opportunity for interaction between soil experts and soilmodellers because the output of the model is a set of rules that can be pedologicallyinterpreted by the soil expert In this sense expert knowledge is used in an implicitmanner in DTA (somewhat effectively achieving Goal 3) While these rules can often beexceedingly complex at minimum the expert can decide whether initial splits makesense given their understanding of the landscape The success of DTA results oftenhinges on the ability of the modeller to make key decisions during the model buildingprocess there is unfortunately no definitive way to determine the most optimal treeThe aid of the soil expert can potentially elucidate this problem

4 Expert systems

A variety of expert system approaches to PSM have been developed to utilize expertknowledge The purpose of such methods is to exploit the information the soil surveyoraccumulates while working in the field by integrating such knowledge into thepredictive model (McCracken and Cate 1986) Unlike the majority of the researchreviewed thus far the dependent variable in many expert systems models is often soiltaxa or mapping unit This apparent disadvantage of expert systems (using classifica-tion to characterize soil continuity) does make them easier to integrate into traditional

soil survey In addition several authors have developed methods to developcontinuous soil property maps from the output of expert system models designed topredict soil unit occurrence

Expert systems are composed of data (information on spatial environmentalvariables eg topography climate etc) a knowledge base (rules and facts related tosoil variation supplied from the soil surveyor) and an inference engine (whichcombines data and the knowledge base to infer logically valid conclusions) (Skidmoreet al 1996) Expert systems differ from conventional models in two ways (1) they storeand manipulate qualitative information (allowing them access to information thatcannot normally be used in other modelling frameworks) and (2) they are structuredas meta-models (the knowledge is separated from the model) (Davis 1993) This allowsthe model to selectively choose which information is relevant at various stages of themodelling process and it allows for information to be easily updated Davis (1993)reviews the application of expert systems to environmental modelling research conclu-ding that the technique is becoming more widely accepted He further states that theapplication of expert systems is constrained by an absence of fundamental knowledgefor rule generation a problem that would appear less relevant to soil mapping giventhe amount of untapped expert knowledge accumulated by the soil surveyor

The first mention of the use of expert systems in pedology was in a paper presenta-tion at the Northeast Committee Soil Survey Conference 1984 (Flach 1985) In herpaper Flach (1985) hinted that recent developments in computer science especiallyexpert systems and artificial intelligence could make modelling a practical mappingtool for soil scientists in the near future A year later McCracken and Cate (1986) hopedto encourage soil scientists to explore expert systems and its application to soil sciencethrough an optimistic article they wrote on the potential use of expert systemsHowever little research was actually conducted in the late 1980s and expert systemshad not yet begun to fulfil the lofty goals set forth by early practitioners (Dale et al1989)

In the early 1990s expert systems approaches to predictive soil mapping began toappear in the literature Skidmore et al (1991) used a Bayesian expert system to mapforest soil into different classes and their results compared favourably with availablesoil maps and actual field-collected data Their methods successfully incorporatedsurveyor knowledge and remotely sensed and digital terrain attributes but failed tobetter characterize continuous soilndashlandscape variability because their final productwas a choropleth map Skidmore et al (1996) revisited their earlier research andassessed the mapping accuracy of their results They found that the soil map producedby the expert system achieved an overall accuracy of 698 (sample size = 53) while themap derived from conventional methods had an accuracy of 736 In addition toknowledge provided by a soil scientist Skidmore et alrsquos (1996) expert system usedvegetation derived for aerial photography and topographic variables derived from a 10-m DEM (wetness index topographic position and slope)

Using an expert system Cook et al (1996) successfully produced a continuous soilproperty map for organic matter content using wetness index aspect and slope asexplanatory variables Their methods were somewhat inefficient because they requireda separate expert system for each soil property of interest However their researchrepresents the only example of expert systems used to directly predict the spatial dis-tribution of a soil property All of the other examples of expert systems in the literature

P Scull et al 189

190 Predictive soil mapping a review

deal with soil type or class The use of expert systems to map soil properties needs tobe explored further

The expert systems discussed thus far have all used Boolean logic within theirtheoretical framework whereby an observation can belong to one and only one classand the soil properties of that class are assigned to the observation With thedevelopment of fuzzy logic and the semantic import model (the second fuzzy logicapproach to creating continuous classes) the opportunity exists to develop fuzzy logic-based expert systems Such systems can be used in conjunction with expert knowledgein situations where experts have a good qualitative idea of how to group data but havetrouble dealing with observations that are not well represented by rigid classificationschemes (Burrough and McDonnell 1998) This method can be particularly useful insituations when taxonomic schemes have been previously developed as is the case insoil taxonomy Several examples of this type of approach were published in the mid-1990s by A Zhu and colleagues (Zhu and Band 1994 Zhu et al 1996 1997 Zhu1997ab) Such systems proved useful for mapping soil at unvisited locations usingsurveyor knowledge and were also capable of producing continuous soil propertymaps The use of fuzzy logic within the theoretical framework of the expert systemallows the complex nature of soil to propagate through the modelling process neversubjected to classification schemes that filter out potentially useful lsquonoisersquo The use offuzzy logic also gives the soil surveyor more latitude during the interview processwhen the knowledge base is defined for the expert system Zhu et al (1997) developeda fuzzy logic-based expert system called SoLIM to determine the similarity of eachgrid cell in a study area to the various taxonomic mapping units delimited by the soilsurvey Continuous soil attribute maps were calculated using the similarity values andtheir relative soil survey determined attribute values (effectively achieving Goal 2) Theresulting data (which proved more accurate than soil survey data once field checked)consists of a raster grid whose resolution is determined by the resolution of the inputenvironmental and digital elevation data As noted previously this type of data modelis more applicable to environmental modelling than the choropleth map (Burrough andMcDonnell 1998)

Expert system approaches to PSM are capable of exploiting soil surveyor knowledgeby developing rule-based systems that imitate the surveyorrsquos conceptual model of soilvariability (the primary focus of Goal 3) The method would appear extremely usefulfor mapping projects (such as those conducted by the NRCS-NCSS) where fieldwork isinitially conducted to determine soilndashlandscape relations Expert system developmentcould be directly inserted into the traditional soil survey mapping approach as asubstitute for the step where the surveyor converts hisher conceptual model into achoropleth map Rather that knowledge could be incorporated into the expert systemwhich could be used to predictively map soil The resulting raster map would be morescientifically based and explicit than the hand-drawn choropleth maps of the past Itsscale would not be limited to that of the aerial photography but rather to the scale ofthe environmental data

Despite these advantages the expert system approach has some drawbacks Becausethe method is dependent upon expert knowledge it cannot be applied whereenvironmentndashsoil relations are poorly understood Of course this criticism can be madeof all PSM models as inductive or deductive knowledge of soilndashenvironment relationsis a prerequisite for PSM However expert systems are deductive models and as such

are not driven by any specific field-collected data (although presumably the soil experthas field experience in the mapping area) Expert systems donrsquot afford the opportunityto first statistically document landscapendashsoil relations and then extrapolate the resultsbecause expert systems do not directly use sample soil data (lsquohard datarsquo) to determinesoilndashlandscape relations Thus the utilization of the relationship between environmen-tal variables and soil properties (Goal 1) is only indirectly achieved Expert systemsapproaches have been demonstrated to be extremely effective in a small number of casestudies The possibility of satisfying all three goals discussed in this review makesexpert systems a predictive soil mapping method that needs to be further tested ndash bothgeographically and across different scales of analysis

V Conclusion

Most of the predictive soil mapping research outlined in this review was conducted atvery large map scales (over small areas) In fact the majority of the research wasconcerned with assessing the spatial variability of soil character within individual fieldsor across soil toposequences The primary driving force behind this type of research hasbeen the need to provide accurate soils information for agriculture and ecologicalmodels It is clear that terrain attributes are powerful predictors at the local scaleGeostatistical tools have been successful at using terrain attributes and the spatialdependence of soil properties to interpolate between existing data points withinindividual fields Across soil toposequences statistical approaches provide a usefulmeans of predicting soil character PSM research has been most successful at the fieldscale because many of the soil-forming factors are held constant For example the neteffect of four of Jennyrsquos five soil-forming factors (climate organism parent material andtime) was minimal within many of the studies reviewed While some research hastackled larger areas of study there exists a deficiency in our ability to predictively mapsoils at smaller map scales Since the distribution of soil is scale-dependent differentPSM methods and predictors are likely to work better at different scales Focus in thefuture must continue to move toward working over larger spatial extents of study inorder to produce landscape-scale soil information

Further a large proportion of the research was conducted in semi-arid gentlysloping agricultural landscapes Humid forestlands mountainous regions and desertshave received little attention As such PSM methods need to continue to be testedandor developed in a wider variety of landscapes where spatial soil distributions canbe more complex Different methods will likely be successful to different degrees indifferent environments Whereas terrain attributes are dominant predictors of soilcharacter across toposequences in prairie lands other predictors are likely necessary inareas where soils do not develop into clear toposequences For example remote sensingdata has been a useful predictor of soil chronosequences in desert landscapes wheresurface appearance is often related to soil character Surficial geology is often reflectiveof soil character in mountain regions where soils are thin and significant bedrock existsat the surface Focus in the future must continue to determine which methods andpredictors work best in which environments

Since the most useful PSM approach will vary across spatial scales and environmen-tal gradients the method used should be driven by the mapping objectives of the

P Scull et al 191

192 Predictive soil mapping a review

project Whereas the traditional soil survey of the past was expected to meet the needsof a diverse group of end users PSM methods will vary given the objectives of thesurvey As a result the end users of the soil data need to play a more active role in thesurvey process Because precision agriculture and large-scale mapping has been theprimary focus of PSM there exists a deficiency in our ability to predictively map soilsfor the purpose of general land use planning and management Even though lessdetailed soil information is needed for these purposes the development of PSMmethods is complicated by the fact that soils exhibit complex spatial variability atsmall map scales where soil-forming environments vary greatly from one location tothe next

A number of alternative methods of characterizing the continuous nature of the soillandscape have been developed Thus far most of the PSM research has provided soilinformation in a nonobject form (as opposed to defining soil types as independententities) Soil data have been generated organized and presented in the form of eitherisorithmic maps or fine-scale raster grids Both of these data models are field-viewmodels of geographic space which allow the soil to be perceived as a constantlyvarying surface Two distinct approaches have been employed (1) mapping individualsoil properties and (2) mapping continuously varying (fuzzy) soil classes Mappingindividual properties is the most common approach and will likely continue todominate PSM research The use of fuzzy soil classes in the literature is less commonand more difficult to be integrated into standard mapping procedures (such as use ofsoil taxa) because the concept radically differs from the traditional view of the soillandscape However fuzzy soil classes (generated using the fuzzy-k means approach) dohave the potential to help further advance Soil taxonomy by identifying taxonomicclasses that are more reflective of pedologic processes at work

Expert systems have been greatly underutilized in PSM research especiallyconsidering how effective a small number of case studies have been Expert systemshave the potential to satisfy successfully all three goals discussed throughout thereview They also have the potential to bridge the gap between traditional approachesand PSM methods because field soil scientists do not have to change their conceptualapproach to mapping They can still conduct field reconnaissance to determinesoilndashlandscape relationships Afterwards though they can use expert systems to betterexploit the knowledge they have garnered Such an approach could help ease thetransition from traditional soil survey to more scientifically explicit methods Expertsystems are also the most fruitful approach to utilizing a wealth of data that has alreadybeen collected in a nonexplicit manner the wealth of expert knowledge (in the form ofthe conceptual model) that senior soil mappers have accumulated

Focus in the future must also move in the direction of operability The utility of manyPSM approaches has been clearly demonstrated yet changes in how soils are mappedand perceived has been slow In the USA various members of the National CooperativeSoil Survey have called for a change in the philosophy of survey (from mappingdiscrete soil types to mapping continuous varying soil properties McSweeny et al1994) but little change has actually taken place To realize the potential of PSM in con-tributing to an overall change in standard soil mapping procedures practitionersworking in the field of PSM need to form working relationships with field soilscientists PSM methods need also to be presented in a manner that is comprehensibleto the soil science trained field mapper Predictive soil mapping is a relatively recent

phenomenon yet much progress has been made In the process of mapping soilknowledge will continue to advance

P Scull et al 193

References

Agbu PA Fehrenbacher DJ and Jansen IJ1990 Statistical comparison of SPOT spectralmaps with field soil maps Soil Science Society ofAmerica Journal 54 818ndash18

Barrett L 1999 Particulars in contextmaintaining a balance in soil geography Annalsof the Association of America Geographers 89707ndash13

Beckett PHT and Webster R 1971 Soilvariability a review Soils and Fertilizers 341ndash15

Bell JC Grigal DF and Bates PC 2000 A soil-terrain model for estimating spatial patterns ofsoil organic carbon In Wilson J and Gallant Jeditors Terrain analysis principles and applica-tions New York City NY John Wiley and Sons295ndash310

Birkeland PW 1999 Soil and geomorphologyThird edition New York NY Oxford UniversityPress

Breiman L Friedman JH Olshen RA andStone CJ 1984 Classification and regressiontrees Belmont CA Wadsworth

Brule FJ 1996 Fuzzy systems ndash a tutorialhttp newsgroup compai httpwwwquadralaycom (last accessed 30 August 2001)

Bunkin FV and Bunkin AF 2000 Lidarsounding of water soil and plants Atmosphericand Oceanic Optics 13 54ndash72

Buol SW Hole FD McCracken RJ andSouthard RJ 1997 Soil genesis and classifica-tion Ames IA Iowa State University Press

Burgess TM and Webster R 1980a Optimalinterpolation and isarithmic mapping of soilproperties the semi-variogram and punctualkriging Journal of Soil Science 31 315ndash31

ndashndashndashndash 1980b Optimal interpolation and isarithmicmapping of soil properties block kringingJournal of Soil Science 31 331ndash41

Burrough PA 1989 Fuzzy mathematicalmethods for soil survey and land evaluationJournal of Soil Science 40 477ndash92

Burrough PA and McDonnell RA 1998Principles of geographic information systems(Revised edition) Oxford Clarendon Press

Burrough PA Beckett PHT and Jarvis MG1971 The relation between cost and utility insoil survey Journal of Soil Science 22 368ndash81

Burrough PA Van Gaans PMF and

Hootsman R 1997 Continuous classificationin soil survey spatial correlation confusionand boundaries Geoderma 77 115ndash35

Cambell JB 1977 Variation of selectedproperties across a soil boundary Soil ScienceSociety of America Journal 41 578ndash82

Cambell JB and Edmonds WJ 1984 Themissing geographic dimension to soiltaxonomy Annals of the Association of AmericanGeographers 74 83ndash97

Castrignano A Giugliarini L Risaliti R andMartinelli N 2000 Study of spatial relation-ships among some soil physico-chemicalproperties of a field in central Italy using multi-variate geostatistics Geoderma 97 39ndash60

Cialella AT Dubayah R Lawrence W andLevine E 1997 Predicting soil drainage classusing remotely sensed and digital elevationdata Journal of Soil Science 62(2) 171ndash78

Clark RN and Swayze GA 1996 Evolution inimaging spectroscopy analysis and sensorsignal-to-noise an examination of how far wehave come Summaries of the sixth annual JPLairborne Earth science workshop 4ndash8 March 1996AVIRIS Workshop Vol 1 5

Clarke LA and Pregibon D 1992 Tree-basedmodels In Chambers J and Hastie J editorsStatistical models in S Pacific Grove Wadsworthand Brooks 377ndash419

Coleman TL Agbu PA and MontgomeryOL 1993 Spectral differentiation of surfacesoils and soil properties ndash is it possible fromspace platforms Soil Science 155 283ndash93

Cook SE Corner RJ Grealish G GesslerPE and Chartres CJ 1996 A rule-basedsystem to map soil properties Soil ScienceSociety of America Journal 60 1893ndash900

Cruickshank JG 1972 Soil geography New YorkNY John Wiley amp Sons

Csillag F Pasztor L Biehl LL 1993 Spectralband selection for the characterization ofsalinity status of soils Remote Sensing ofEnvironment 43 231ndash42

Dale MB McBratney AB and Russell JS1989 On the role of expert systems andnumerical taxonomy in soil classificationJournal of Soil Science 40 223ndash34

Davis JR 1993 Expert systems and environ-

180 Predictive soil mapping a review

Table 1

Selected

recen

t literature on pred

ictive

soil m

apping an

d m

apping

(cited in

this article) d

escribing the mod

ellin

g metho

dused the

dep

enden

t variables used a

nd the

env

iron

men

tal v

ariables (ex

plan

atory) used in

the

mod

els

Stud

yMod

ellin

g metho

dDep

ende

nt variables

Environm

ental variab

les

Goa

ls atta

ined

a

Bell et

al 20

00Line

ar and

exp

onen

tial

Total soil orga

nic

Slop

e curvature aspe

ct

ECx

regression

carbon

hillslope

position

Burge

ss and

Web

ster

Punc

tual and

block

Na co

nten

t co

ver

Non

eC

1980

ab

kriging

loam

thickn

ess

ston

e co

nten

t

Burroug

h 198

9Fu

zzy mathe

matical

ndashndash

rpmetho

ds

Burroug

h et

al 19

97Con

tinuo

us classificatio

nndash

ndashrp

Castrigna

no eta

l 200

0Factorial kriging

CEC

pH N

P K

Non

e ex

plicitly used

Cx

Na

Cialella

eta

l 199

7Classifica

tion tree

Drainag

e class

Elev aspe

ct NDVI

Ec

Coo

k et

al 19

96Baysian

rule-ba

sed metho

dsOrgan

ic m

atter

Slop

e aspec

t wetne

ssEcX

inde

x

Dale et

al 19

89 (a review

)Ex

pert systems

ndashndash

ndash

Ellis 19

96Decision tree

ana

lysis

Soil erosion class

Slop

e aspec

t wetne

ssEc

neutral ne

tworks

inde

x flow le

ngth and

accu

mulation Lan

dsat

TM tree

cov

er

Gessler 1

996

A large

variety of statistic

Field an

d labo

ratory

A variety of digital

EC

metho

dsco

llected

phy

sica

len

vironm

ental da

tach

emical and

morph

olog

ical soil

prop

ertie

s

Gessler eta

l 199

5Line

ar and

logit reg

ression

A horizon

and

solum

Curvature CTI topo

EC

depth E horizon

positio

npresen

ce

Goo

vaerts 19

92Factorial kriging

Total carbon

Non

eC

P Scull et al 181pH

N CEC

extractab

lecatio

ns (K Ca M

g)Hartemink et

al 20

01ndash

ndashndash

rp

Heu

velin

k an

d W

ebster 20

01ndash

ndashndash

rp

Hew

itt 19

93ndash

ndashndash

rp

Indo

rante et

al 19

96ndash

ndashndash

rp

King et

al 19

99Lo

gistic regression

Presen

ceabsen

ceSlop

e aspec

t po

t solar

Ec

Non

calc c

lay-loam

energy

Kno

tters eta

l 199

5Kriging

co

-kriging

So

ft layer de

pth

Hillslop

e po

sitio

nCx

regression

kriging

Laga

cherie and

Holmes 19

97Classifica

tion tree

Map

ping

unit

Geo

logy variou

s topo

Ec

indices

Laslett e

tal 198

7Kriging

splin

es tren

dpH

Non

ec

surfac

e nea

rest neigh

bor

McB

ratney 1

992

ndashndash

ndashrp

McB

ratney

eta

l 199

1Block kriging

Clay co

nten

tNon

eC

McB

ratney

eta

l 200

0ndash

ndashndash

rp

McB

ratney

and

Ode

h 199

7Fu

zzy sets in

soil scienc

endash

ndashRp

McB

ratney

and

de Gruijter

Fuzz

y-k-mea

ns w

ithFu

zzy classes

Field co

llected

phy

sical

C19

92ex

tragrade

sch

emical and

morph

olog

ical soil

prop

ertie

s

McC

rack

en and

Cate 198

6Artificial intellig

ence

ndash

ndashrp

expe

rt systems

McK

ensie an

d Austin

19

93Gen

eralized

line

ar m

odels

Clay co

nten

t CEC

Slop

e relief land

form

ec

(logit)

pH EC

COLE

slop

e po

sitio

nbu

lk den

sity and

othe

rs

Moo

re eta

l 199

3Line

ar reg

ression

A horizon

dep

th O

MSlop

e w

etne

ss and

strea

mEC

and P co

nten

t pH

power in

dices aspect

curvature

182 Predictive soil mapping a reviewTa

ble 1

Con

tinu

ed

Stud

yMod

ellin

g metho

dDep

ende

nt variables

Environm

ental v

ariables

Goa

ls atta

ined

a

McK

ensie an

d Ryan 199

9Reg

ression tree and

linea

rSo

lum dep

th P an

dElevation slope

CE

regression

N con

tent

curvature CTI

contribu

ting area do

wn-

slop

e mea

ns for slope

clim

ate da

ta Presco

ttInde

x G

amma

Rad

iometry La

ndsat T

M

and Geo

logy unit

Ode

h et

al 19

92ab

Fuzz

y-c-mea

ns and

Fuzzy classes

Field-co

llected

phy

sica

lC

kriging

chem

ical and

morph

olog

ical soil

prop

ertie

s

Ode

h et

al 19

94

Reg

ression kriging

co

-So

lum dep

th de

pth

Slop

e aspect cu

rvature

eC

1995

kriging regression

to bed

rock gravel

kriging

and clay con

tent

Skidmore et

al 19

91Bayesian expe

rt system

Soil land

scap

e un

itVeg type

wetne

ss in

dex

EcX

1996

grad

ient terrain po

sitio

n

Voltz and

Web

ster 19

90Kriging

cu

bic splin

eClay co

nten

tNon

eC

Web

ster 1

994

Dev

elop

men

t of

ndashndash

rppe

dometrics

Zhu

199

7ab

Fuzz

y logic expe

rt system

Soil series A

horizon

Elev pm aspe

ct c

anop

yECX

Zhu

and

Ban

d 199

4(SoL

IM)

depth in

dividu

alco

verage

grad

ient

Zhu

eta

l 199

7series m

aps

curvature

Not

es

a Letters refer to the de

gree

to w

hich

the

goa

ls of PS

M defined

in the

introdu

ction are achiev

ed

Soilndash

e nvironm

ent relatio

ns utilized

(letter E) be

tter represen

tatio

n of soil c o

ntinuity (C

) an

d ex

pert kno

wledg

e utilize

d (X) c

orrespon

d to goa

ls 1 2 and

3respec

tively from

the

introdu

ction and

cap

ital lette

rs (E C X) ind

icate the metho

d is relatively more successful th

an th

ose metho

ds den

oted

by lower

case le

tters (e c x

) rp ind

icates rev

iew pap

ers

P Scull et al 183

predictive mapping because landscape position is never quantified (rather positionwas often qualitatively defined eg toe-slope) and thus the documented relationshipscannot be generalized using environmental data and digital elevation models to predictsoil character at unvisited sites

Table 1 documents modelling methods model variables and the extent to which thereferenced research satisfies the previously defined goals of PSM (lsquoGoals attainedrsquocolumn) Soil-environment relations utilized (letter E) better representation of soilcontinuity (C) and expert knowledge utilized (X) correspond to goals 1 2 and 3respectively from the introduction and capital letters (E C X) indicate the method isrelatively more successful than those methods denoted by lower case letters (e c x)Citations with no letters present within the lsquoGoals attainedrsquo column do not address theaforementioned goals For example the Cialella et al (1997) received a rating of lsquoEcrsquomeaning that the methods employed successfully utilized environmentalndashsoil characterrelations (E) and somewhat successfully presented a better method of representing soilcontinuity (c) The ratings are provided simply to help organize the literature that wasreviewed Review papers are included within the table denoted by lsquorprsquo Geostatisticalmethods are not included in the table because they have been comprehensivelysurveyed elsewhere (Odeh et al 1994 Burrough et al 1997 McBratney et al 2000Heuvelink and Webster 2001) and because the objectives and assumptions of geostatis-tical methods differ slightly from other PSM research We briefly outline thesedifferences in the following section

1 Geostatistical methods

Geostatistics are a subset of traditional statistics that deal primarily with spatial dataand account for spatial autocorrelation using kriging as the spatial interpolator Theconcept is based upon the theory of regionalized variables which was mainlydeveloped by Matheron (1963) and Krige (1963) Kriging is a form of weighted localaveraging that uses a measure of spatial dependence the variogram to determine theweights applied to the data when computing the averages Geostatistical methods havebeen used in predictive soil mapping research to spatially interpolate soil propertyvalues at unmeasured sites from field-collected data

Burgess and Webster (1980a b) were the first to introduce ordinary kriging to the soilcommunity and since that time an enormous amount of work has been published Forexample ordinary kriging has been used to interpolate many different soil propertiesincluding pollution trace element deficiencies salinity and fertility (Heuvelink andWebster 2001) Ordinary kriging has been criticized for a variety of reasons Forexample Laslett et al (1987) reported that several authors had criticized geostatisticsbecause kriging is a global rather than local technique failing to take into accountknowledge of soil materials and processes Other authors have criticized geostatisticsbecause they are excessively data dependent requiring a large number of closelyspaced data points (Zhu 1997a) As Webster and Oliver (1992) suggest in excess of ahundred samples may be needed to use geostatistics at the field scale because of highspatial variability of soil in some areas Geostatistics also assume spatial autocorrela-tion which sometimes may be a poor assumption in complex terrain where abruptchanges in soil-forming factors occur (McBratney et al 2000) Ordinary kriging by itself

184 Predictive soil mapping a review

does not satisfy two of the three goals of PSM presented in the introduction ndash it does notadequately incorporate expert knowledge and it does not exploit the relationshipbetween environmental variables and soil properties

Ordinary kriging has been modified in a variety of ways to better incorporateancillary data and known soilndashlandscape relationships Block kriging involvesdetermining estimates over meaningful areas rather than at specific points (Burgess andWebster 1980a McBratney et al 1991) Using this method a study area can be stratifiedinto different regions that are reflective of the pedogenetic processes at work In orderto accommodate a trend within a dependent soil variable universal kriging has beenused (Webster 1994) Kriging with external drift is similar to universal kriging but ituses an ancillary variable to represent the trend (McBratney et al 2000) Co-krigingtakes advantage of correlation that may exist between the variable of interest and othermore easily measured variables (Odeh et al 1995) Regression kriging involvesspatially interpolating the residuals from a non-spatial model by kriging and addingthe result to the prediction obtained from that model (Goovaerts 1997 Castrignano etal 2000) Factorial kriging is another method to integrate multivariate data into thestandard kriging routine to extrapolate soil data (Goovaerts 1992) Many authors havecompared these various methods (Laslett et al 1987 Voltz and Webster 1990 Odeh etal 1994 Knotters et al 1995)

Fuzzy logic has been used with geostatistics by various authors to produce new kindsof fuzzy soil maps with continuous classes (Burrough 1989 McBratney and DeGruijter1992 Odeh et al 1992a and reviewed by McBratney and Odeh 1997) The processentails kriging the matrix of membership values determined by the fuzzy k-meansclassifier resulting in a continuous soil surface where individual locations are allowedto belong to more than one class and no rigid boundaries are designated to separate thesoil into discrete units or entities The results of such analysis can be used to assess thepedologic process validity of soil taxonomy by determining whether soils grouptogether into classes that are similar to taxonomic types Fuzzy classes wouldpresumably reflect the main pedologic features within a given area In this sense theclassification is quantitative whereas soil taxonomy is rooted in qualitative discrimina-tion A comparison between the two could provide insightful and help assist soiltaxonomy in making a classification that is more indicative of underlying soil processes

Geostatistics in soil research were originally introduced to quantitatively assess soilvariability within soil mapping units (McBratney et al 1991) in response to criticisms inthe early 1970s that soil unit composition was not well quantified (Beckett and Webster1971) In this regard geostatistics have been very useful having served well the originalgoals set forth by Burgess and Webster (1980b) when they drew kriging to the attentionof soil scientists as a means of spatial prediction At the field scale soil variation islargely due to the effect that topography has on soil genesis Geostatistics have beensuccessfully applied in such environments by using terrain attributes as ancillary datawithin many of the kriging routines described above Such quantitative within-unitvariability of soil properties is very useful in the field of precision agriculture and othersituations (eg pollutants) where very detailed soil attribute information is needed atthe field scale (Heuvelink and Webster 2001)

However geostatistics have not been applied in a wide variety of environments or atlarger scales In order to be successfully applied in different environments geostatisticswill likely require a different suite of ancillary data For example remote sensing data

P Scull et al 185

could be used in arid regions where soil toposequences are less well expressed Atlarger scales of prediction selection of different sets of ancillary variables is requiredbecause different processes define soil character at different scales The most obviousexample is that of climate which may control soil distribution at large scales(continental) but has little explanatory power at the field level Regardless of whetheradequate ancillary data exist the amount of data required to use geostatistics forlandscape-level prediction would be extremely difficult and costly to collect given thestrict sampling protocol required to characterize spatial dependence It is also unclearat what landscape scale soils exhibit spatial autocorrelation

Geostatistical approaches do provide a means of creating continuous soil attributesurfaces to better represent soil continuity (Goal 2) and they can be used to exploit therelationship between environmental variables and soil properties in order to moreefficiently collect soil data (Goal 1) However they do not sufficiently utilize expertknowledge (Goal 3) as no attempt has been made in geostatistical approaches todirectly integrate expert knowledge Fundamentally kriging is a process of interpola-tion designed to predict attribute values in between locations of measured samples Inthis sense geostatistics represent a middle ground between pure interpolation (egnearest neighbour type classifier) in which only measured points for the variable ofinterest are used to determine unknown values and other predictive models thatprimarily use soilndashenvironment correlation to create predictive maps

2 Statistical methods

Statistical methods can be used to exploit the relationship between quantifiablelandscape indices and soil properties to create predictive soil maps For exampleMcKensie and Austin (1993) used a regression to account for a large percentage ofvariation for many soil characteristics (A horizon clay content CEC EC pH bulkdensity and COLE B horizon clay content CEC ESP EC pH bulk density and COLE)using a variety of predictor variables (slope presence or absence of impeding layerrelief landform topographic position) Their results confirm the hypothesis of MooreID et al (1991) that soil character is related to quantifiable landscape indices Howevertheir methods do not provide inference of soil properties at unmeasured sites frommapped environment data because the topographic variables were measured in thefield Linear regression has also been used with terrain variables derived from a 15-mDEM in northeastern Colorado to predict soil attributes (organic matter contentextractable phosphorous pH and texture) at unvisited sites (Moore et al 1993) In thatparticular study 50 of the variance of A-horizon thickness was explained by slope andthe wetness index Gessler et al (1995) also used regression to model A-horizonthickness from topographic variables in southeastern Australia (plan curvature andwetness index r2 = 063 P = 0001) They modeled solum depth and used logisticregression to model E horizon presenceabsence Elsewhere logistic regression hasbeen used to model the presenceabsence of noncalcareous clay loam horizon in centralFrance using terrain attributes from a 20-m DEM (King et al 1999) Exponentialregression has been used to model soil organic carbon using terrain variables (Bell et al2000) in glacial outwash soils in east-central Minnesota Generalized additive models(GAM) have been used less frequently in PSM research Gessler (1996) used a GAM

186 Predictive soil mapping a review

model to predict total soil carbon A horizon depth and solum depth using a variety ofenvironmental predictors (slope elevation wetness index mean annual temperatureprecipitation and radiation)

This small body of research opened the door to more complex methods by demon-strating the existence of quantifiable relationships These authors were able to producesoil attribute maps using raster data models whose scale was dependent upon the gridcell resolution of the environmental data They were successful at exploiting the rela-tionship between quantifiable topographic attributes and soil profile character (Goal 1)The continuous soil attribute surfaces they produce also better represent soil continuitythan the choropleth soil maps produced by traditional soil survey (Goal 2) Howeverthe bulk of these methods (excepting GAMs) are limited by their assumed linear rela-tionship between soil and topographic attributes their assumptions of normallydistributed data and their high data requirements Standard statistical procedures arealso not flexible enough to allow robust integration with a variety of potential datasources such as expert knowledge (Goal 1)

Statistical methods do demonstrate in a quantitative manner that terrain analysis canbe used to predict soil attributes in relatively small areas with homogeneous parentmaterial A large proportion of the research using statistical methods was conducted insemi-arid landscapes at small scales (the largest study area of the entire group was~2000 ha) Obviously for statistical approaches to be most effective they need to bemore universal As such they need to be tested andor developed at larger scales andin more diverse landscapes

3 Decision tree analysis (DTA)

The use of decision tree analysis is just beginning to be explored in predictive soilmapping research although it has been used successfully in the related field ofpredictive vegetation mapping since the early 1990s (Lees and Ritman 1991 MooreDM et al 1991 Franklin 1998) DTA is a form of divisive classification The process oftree modelling involves successively partitioning data (called recursive partitioning inthe tree modelling literature) into increasingly homogeneous subsets which once thepartitioning has ceased are called terminal nodes (Lees and Ritman 1991) Splits orrules defining how to partition the data are selected based on information statistics thatdefine how well the split decreases impurity within the data set (Clark and Pregibon1992) Splits are based on threshold values of an explanatory variable selected bycomparing the increase in resulting purity of node membership for all possiblethresholds and variables The process is iterative growing from the root node (thecomplete data set) to the terminal nodes in a dendritic fashion (Friedl and Brodley1997) Once the tree has been constructed (or grown) it encodes a set of decision rulesthat describe the data partitioning process These rules can be used to classify or predictother data sets (Moore DM et al 1991) Pruning the tree is often necessary to preventthe tree from being overfit to the sample data and to reduce tree complexity Pruningentails combining pairs of terminal nodes into single nodes and can be accomplishedusing cross-validation which yields an initial indication of how large a tree makesrobust predictions (Safavian and Norvig 1991) Cross-validation involves systematical-ly removing portions of the data set and running the remaining sample through the tree

in an iterative manner eventually yielding estimates of the misclassification rates foreach class each node and the whole model (Breiman et al 1984) In this mannerdifferent sized trees can be compared in terms of parsimony

The term DTA is used to collectively refer to all types of tree-based modelling (theword lsquodecisionrsquo is used because it is descriptive indicating that the analysis eventuallyleads to a set of decision rules defining data partitions) The term should be distin-guished from classification tree analysis because the latter refers specifically to DTAwhere the response variable is categorical The term CART (classification andregression trees ndash Breiman et al 1984) is sometimes used but strictly speaking refers tospecific software Friedl and Brodley (1997) provide a review of the decision treealgorithms They divide types of DTA into two classes (1) homogeneous decision treesfor which a single algorithm is used to estimate each split (eg CART) and (2) hybriddecision trees (HDT) for which different splitting methods can be used at differentpoints in the tree (eg Quinlan 1993) They further divide homogeneous decision treesinto univariate (UDT) where single features of the input data define splits and multi-variate decision trees (MDT) where multiple features of the input data can define splitsAccording to this naming convention no distinction is made between methods withdifferent types of response variables although all methods can be used with bothcategorical and continuous response variables In a comparison of these various typesof algorithms on a variety of data sets Friedl and Brodley (1997) found that HDT hadthe highest classification accuracy

The overall aim of DTA is to design a set of predictive rules (eg if geology type Athen soil type B) developed from training data which can then be applied to ageographic data base to predict the value of a response variable (Michaelsen et al 1994)Therefore DTA explicitly uses soilndashlandscape correlation in model development (Goal2) The technique appears promising in soil research but needs to be further tested asnot many of the above types of DTA have been tested in the PSM literature in fact onlyunivariate approaches have been employed For example Lagacherie and Holmes(1997) successfully used univariate DTA to model a categorical response variable soiltype within a training set and then assuming that the training set was representativeextrapolated the model to a much larger region Their work is interesting becausealthough not a single sample came from the area they eventually mapped they wereable to produce a soil map that was more accurate (74 versus 69) than the existingmap produced from traditional methods Cialella et al (1997) also used univariate DTAto predict soil drainage class from a variety of terrain attributes and remotely senseddata They predicted soil drainage class with an average of 78 accuracy ndash impressivegiven that the variation accounted for by a typical soil survey ranges from about halfthe total variance for physical attributes to less than one-tenth for some soil chemicalattributes (Gessler et al 1995)

DTA has been compared with other approaches by several authors In the applicationof erosion modelling DTA results were similar to Artificial Neural Networks (Ellis1996) Both methods achieved high training accuracy (as measured by the KappaStatistic) but in terms of prediction accuracy both methods performed poorly Gessleret al (1995) compared DTA with generalized linear models (GLM) and generalizedadditive models (GAM) to predict A-horizon thickness and concluded that GLM waspreferable to both DTA and GAM McKensie and Ryan (1999) compared regressiontrees and standard linear regression to predict soil properties (total solum depth soil

P Scull et al 187

188 Predictive soil mapping a review

phophorus and soil carbon) using a large variety of predictors (elevation slopecurvature wetness index climate variables geology unit Landsat TM data andGamma radiometric data) and found that different methods work best for differentproperties and overall success hinged on the strength of the relationship between soiland environmental variables Their study of a 50 000 ha forested landscape in southernNew South Wales Australia is one of the few studies at such a small map scale (largearea) or in forested landscape The most extensive comparison of techniques involvingDTA was conducted by McBratney et al (2000) They compared regression techniques(GLM GAM DTA-regression tree) geostatistical techniques (kriging and heterotopicco-kriging) and a hybrid technique (regression kriging) DTA was found to be thepoorest performing of the regression techniques because of the unrealistic predictionsurface generated by the DTA model DTA has also been criticized by other authorsbecause of the stepped prediction surface (Gessler 1996) This phenomenon can beespecially apparent in situations where predictor variables have different resolutionsThe extent to which DTA yields a better spatial representation of soil continuity is afunction of the scale and type of predictor variables used and therefore varies withindividual models

DTA is gaining widespread popularity as a means to develop prediction rules thatcan be rapidly and repeatedly evaluated (Cialella et al 1997 Franklin et al 2000) DTAprovides the following advantages over standard statistical techniques (1) it is easier tointerpret when explanatory variables are both nominal and continuous (2) it isinvariant to monotone re-expressions (transformations) of predictor variables (3) itdeals more satisfactorily with missing data values and outliers (4) it is more adept atcapturing nonadditive and nonlinear behaviour (5) it doesnrsquot make any assumptionsabout data distribution and (6) it is easily updateable as more data are collected(Moore DM et al 1991) The DTA model framework is especially appealing becauseof its capability to integrate a wide range of data sets as explanatory variables

DTA offers a unique opportunity for interaction between soil experts and soilmodellers because the output of the model is a set of rules that can be pedologicallyinterpreted by the soil expert In this sense expert knowledge is used in an implicitmanner in DTA (somewhat effectively achieving Goal 3) While these rules can often beexceedingly complex at minimum the expert can decide whether initial splits makesense given their understanding of the landscape The success of DTA results oftenhinges on the ability of the modeller to make key decisions during the model buildingprocess there is unfortunately no definitive way to determine the most optimal treeThe aid of the soil expert can potentially elucidate this problem

4 Expert systems

A variety of expert system approaches to PSM have been developed to utilize expertknowledge The purpose of such methods is to exploit the information the soil surveyoraccumulates while working in the field by integrating such knowledge into thepredictive model (McCracken and Cate 1986) Unlike the majority of the researchreviewed thus far the dependent variable in many expert systems models is often soiltaxa or mapping unit This apparent disadvantage of expert systems (using classifica-tion to characterize soil continuity) does make them easier to integrate into traditional

soil survey In addition several authors have developed methods to developcontinuous soil property maps from the output of expert system models designed topredict soil unit occurrence

Expert systems are composed of data (information on spatial environmentalvariables eg topography climate etc) a knowledge base (rules and facts related tosoil variation supplied from the soil surveyor) and an inference engine (whichcombines data and the knowledge base to infer logically valid conclusions) (Skidmoreet al 1996) Expert systems differ from conventional models in two ways (1) they storeand manipulate qualitative information (allowing them access to information thatcannot normally be used in other modelling frameworks) and (2) they are structuredas meta-models (the knowledge is separated from the model) (Davis 1993) This allowsthe model to selectively choose which information is relevant at various stages of themodelling process and it allows for information to be easily updated Davis (1993)reviews the application of expert systems to environmental modelling research conclu-ding that the technique is becoming more widely accepted He further states that theapplication of expert systems is constrained by an absence of fundamental knowledgefor rule generation a problem that would appear less relevant to soil mapping giventhe amount of untapped expert knowledge accumulated by the soil surveyor

The first mention of the use of expert systems in pedology was in a paper presenta-tion at the Northeast Committee Soil Survey Conference 1984 (Flach 1985) In herpaper Flach (1985) hinted that recent developments in computer science especiallyexpert systems and artificial intelligence could make modelling a practical mappingtool for soil scientists in the near future A year later McCracken and Cate (1986) hopedto encourage soil scientists to explore expert systems and its application to soil sciencethrough an optimistic article they wrote on the potential use of expert systemsHowever little research was actually conducted in the late 1980s and expert systemshad not yet begun to fulfil the lofty goals set forth by early practitioners (Dale et al1989)

In the early 1990s expert systems approaches to predictive soil mapping began toappear in the literature Skidmore et al (1991) used a Bayesian expert system to mapforest soil into different classes and their results compared favourably with availablesoil maps and actual field-collected data Their methods successfully incorporatedsurveyor knowledge and remotely sensed and digital terrain attributes but failed tobetter characterize continuous soilndashlandscape variability because their final productwas a choropleth map Skidmore et al (1996) revisited their earlier research andassessed the mapping accuracy of their results They found that the soil map producedby the expert system achieved an overall accuracy of 698 (sample size = 53) while themap derived from conventional methods had an accuracy of 736 In addition toknowledge provided by a soil scientist Skidmore et alrsquos (1996) expert system usedvegetation derived for aerial photography and topographic variables derived from a 10-m DEM (wetness index topographic position and slope)

Using an expert system Cook et al (1996) successfully produced a continuous soilproperty map for organic matter content using wetness index aspect and slope asexplanatory variables Their methods were somewhat inefficient because they requireda separate expert system for each soil property of interest However their researchrepresents the only example of expert systems used to directly predict the spatial dis-tribution of a soil property All of the other examples of expert systems in the literature

P Scull et al 189

190 Predictive soil mapping a review

deal with soil type or class The use of expert systems to map soil properties needs tobe explored further

The expert systems discussed thus far have all used Boolean logic within theirtheoretical framework whereby an observation can belong to one and only one classand the soil properties of that class are assigned to the observation With thedevelopment of fuzzy logic and the semantic import model (the second fuzzy logicapproach to creating continuous classes) the opportunity exists to develop fuzzy logic-based expert systems Such systems can be used in conjunction with expert knowledgein situations where experts have a good qualitative idea of how to group data but havetrouble dealing with observations that are not well represented by rigid classificationschemes (Burrough and McDonnell 1998) This method can be particularly useful insituations when taxonomic schemes have been previously developed as is the case insoil taxonomy Several examples of this type of approach were published in the mid-1990s by A Zhu and colleagues (Zhu and Band 1994 Zhu et al 1996 1997 Zhu1997ab) Such systems proved useful for mapping soil at unvisited locations usingsurveyor knowledge and were also capable of producing continuous soil propertymaps The use of fuzzy logic within the theoretical framework of the expert systemallows the complex nature of soil to propagate through the modelling process neversubjected to classification schemes that filter out potentially useful lsquonoisersquo The use offuzzy logic also gives the soil surveyor more latitude during the interview processwhen the knowledge base is defined for the expert system Zhu et al (1997) developeda fuzzy logic-based expert system called SoLIM to determine the similarity of eachgrid cell in a study area to the various taxonomic mapping units delimited by the soilsurvey Continuous soil attribute maps were calculated using the similarity values andtheir relative soil survey determined attribute values (effectively achieving Goal 2) Theresulting data (which proved more accurate than soil survey data once field checked)consists of a raster grid whose resolution is determined by the resolution of the inputenvironmental and digital elevation data As noted previously this type of data modelis more applicable to environmental modelling than the choropleth map (Burrough andMcDonnell 1998)

Expert system approaches to PSM are capable of exploiting soil surveyor knowledgeby developing rule-based systems that imitate the surveyorrsquos conceptual model of soilvariability (the primary focus of Goal 3) The method would appear extremely usefulfor mapping projects (such as those conducted by the NRCS-NCSS) where fieldwork isinitially conducted to determine soilndashlandscape relations Expert system developmentcould be directly inserted into the traditional soil survey mapping approach as asubstitute for the step where the surveyor converts hisher conceptual model into achoropleth map Rather that knowledge could be incorporated into the expert systemwhich could be used to predictively map soil The resulting raster map would be morescientifically based and explicit than the hand-drawn choropleth maps of the past Itsscale would not be limited to that of the aerial photography but rather to the scale ofthe environmental data

Despite these advantages the expert system approach has some drawbacks Becausethe method is dependent upon expert knowledge it cannot be applied whereenvironmentndashsoil relations are poorly understood Of course this criticism can be madeof all PSM models as inductive or deductive knowledge of soilndashenvironment relationsis a prerequisite for PSM However expert systems are deductive models and as such

are not driven by any specific field-collected data (although presumably the soil experthas field experience in the mapping area) Expert systems donrsquot afford the opportunityto first statistically document landscapendashsoil relations and then extrapolate the resultsbecause expert systems do not directly use sample soil data (lsquohard datarsquo) to determinesoilndashlandscape relations Thus the utilization of the relationship between environmen-tal variables and soil properties (Goal 1) is only indirectly achieved Expert systemsapproaches have been demonstrated to be extremely effective in a small number of casestudies The possibility of satisfying all three goals discussed in this review makesexpert systems a predictive soil mapping method that needs to be further tested ndash bothgeographically and across different scales of analysis

V Conclusion

Most of the predictive soil mapping research outlined in this review was conducted atvery large map scales (over small areas) In fact the majority of the research wasconcerned with assessing the spatial variability of soil character within individual fieldsor across soil toposequences The primary driving force behind this type of research hasbeen the need to provide accurate soils information for agriculture and ecologicalmodels It is clear that terrain attributes are powerful predictors at the local scaleGeostatistical tools have been successful at using terrain attributes and the spatialdependence of soil properties to interpolate between existing data points withinindividual fields Across soil toposequences statistical approaches provide a usefulmeans of predicting soil character PSM research has been most successful at the fieldscale because many of the soil-forming factors are held constant For example the neteffect of four of Jennyrsquos five soil-forming factors (climate organism parent material andtime) was minimal within many of the studies reviewed While some research hastackled larger areas of study there exists a deficiency in our ability to predictively mapsoils at smaller map scales Since the distribution of soil is scale-dependent differentPSM methods and predictors are likely to work better at different scales Focus in thefuture must continue to move toward working over larger spatial extents of study inorder to produce landscape-scale soil information

Further a large proportion of the research was conducted in semi-arid gentlysloping agricultural landscapes Humid forestlands mountainous regions and desertshave received little attention As such PSM methods need to continue to be testedandor developed in a wider variety of landscapes where spatial soil distributions canbe more complex Different methods will likely be successful to different degrees indifferent environments Whereas terrain attributes are dominant predictors of soilcharacter across toposequences in prairie lands other predictors are likely necessary inareas where soils do not develop into clear toposequences For example remote sensingdata has been a useful predictor of soil chronosequences in desert landscapes wheresurface appearance is often related to soil character Surficial geology is often reflectiveof soil character in mountain regions where soils are thin and significant bedrock existsat the surface Focus in the future must continue to determine which methods andpredictors work best in which environments

Since the most useful PSM approach will vary across spatial scales and environmen-tal gradients the method used should be driven by the mapping objectives of the

P Scull et al 191

192 Predictive soil mapping a review

project Whereas the traditional soil survey of the past was expected to meet the needsof a diverse group of end users PSM methods will vary given the objectives of thesurvey As a result the end users of the soil data need to play a more active role in thesurvey process Because precision agriculture and large-scale mapping has been theprimary focus of PSM there exists a deficiency in our ability to predictively map soilsfor the purpose of general land use planning and management Even though lessdetailed soil information is needed for these purposes the development of PSMmethods is complicated by the fact that soils exhibit complex spatial variability atsmall map scales where soil-forming environments vary greatly from one location tothe next

A number of alternative methods of characterizing the continuous nature of the soillandscape have been developed Thus far most of the PSM research has provided soilinformation in a nonobject form (as opposed to defining soil types as independententities) Soil data have been generated organized and presented in the form of eitherisorithmic maps or fine-scale raster grids Both of these data models are field-viewmodels of geographic space which allow the soil to be perceived as a constantlyvarying surface Two distinct approaches have been employed (1) mapping individualsoil properties and (2) mapping continuously varying (fuzzy) soil classes Mappingindividual properties is the most common approach and will likely continue todominate PSM research The use of fuzzy soil classes in the literature is less commonand more difficult to be integrated into standard mapping procedures (such as use ofsoil taxa) because the concept radically differs from the traditional view of the soillandscape However fuzzy soil classes (generated using the fuzzy-k means approach) dohave the potential to help further advance Soil taxonomy by identifying taxonomicclasses that are more reflective of pedologic processes at work

Expert systems have been greatly underutilized in PSM research especiallyconsidering how effective a small number of case studies have been Expert systemshave the potential to satisfy successfully all three goals discussed throughout thereview They also have the potential to bridge the gap between traditional approachesand PSM methods because field soil scientists do not have to change their conceptualapproach to mapping They can still conduct field reconnaissance to determinesoilndashlandscape relationships Afterwards though they can use expert systems to betterexploit the knowledge they have garnered Such an approach could help ease thetransition from traditional soil survey to more scientifically explicit methods Expertsystems are also the most fruitful approach to utilizing a wealth of data that has alreadybeen collected in a nonexplicit manner the wealth of expert knowledge (in the form ofthe conceptual model) that senior soil mappers have accumulated

Focus in the future must also move in the direction of operability The utility of manyPSM approaches has been clearly demonstrated yet changes in how soils are mappedand perceived has been slow In the USA various members of the National CooperativeSoil Survey have called for a change in the philosophy of survey (from mappingdiscrete soil types to mapping continuous varying soil properties McSweeny et al1994) but little change has actually taken place To realize the potential of PSM in con-tributing to an overall change in standard soil mapping procedures practitionersworking in the field of PSM need to form working relationships with field soilscientists PSM methods need also to be presented in a manner that is comprehensibleto the soil science trained field mapper Predictive soil mapping is a relatively recent

phenomenon yet much progress has been made In the process of mapping soilknowledge will continue to advance

P Scull et al 193

References

Agbu PA Fehrenbacher DJ and Jansen IJ1990 Statistical comparison of SPOT spectralmaps with field soil maps Soil Science Society ofAmerica Journal 54 818ndash18

Barrett L 1999 Particulars in contextmaintaining a balance in soil geography Annalsof the Association of America Geographers 89707ndash13

Beckett PHT and Webster R 1971 Soilvariability a review Soils and Fertilizers 341ndash15

Bell JC Grigal DF and Bates PC 2000 A soil-terrain model for estimating spatial patterns ofsoil organic carbon In Wilson J and Gallant Jeditors Terrain analysis principles and applica-tions New York City NY John Wiley and Sons295ndash310

Birkeland PW 1999 Soil and geomorphologyThird edition New York NY Oxford UniversityPress

Breiman L Friedman JH Olshen RA andStone CJ 1984 Classification and regressiontrees Belmont CA Wadsworth

Brule FJ 1996 Fuzzy systems ndash a tutorialhttp newsgroup compai httpwwwquadralaycom (last accessed 30 August 2001)

Bunkin FV and Bunkin AF 2000 Lidarsounding of water soil and plants Atmosphericand Oceanic Optics 13 54ndash72

Buol SW Hole FD McCracken RJ andSouthard RJ 1997 Soil genesis and classifica-tion Ames IA Iowa State University Press

Burgess TM and Webster R 1980a Optimalinterpolation and isarithmic mapping of soilproperties the semi-variogram and punctualkriging Journal of Soil Science 31 315ndash31

ndashndashndashndash 1980b Optimal interpolation and isarithmicmapping of soil properties block kringingJournal of Soil Science 31 331ndash41

Burrough PA 1989 Fuzzy mathematicalmethods for soil survey and land evaluationJournal of Soil Science 40 477ndash92

Burrough PA and McDonnell RA 1998Principles of geographic information systems(Revised edition) Oxford Clarendon Press

Burrough PA Beckett PHT and Jarvis MG1971 The relation between cost and utility insoil survey Journal of Soil Science 22 368ndash81

Burrough PA Van Gaans PMF and

Hootsman R 1997 Continuous classificationin soil survey spatial correlation confusionand boundaries Geoderma 77 115ndash35

Cambell JB 1977 Variation of selectedproperties across a soil boundary Soil ScienceSociety of America Journal 41 578ndash82

Cambell JB and Edmonds WJ 1984 Themissing geographic dimension to soiltaxonomy Annals of the Association of AmericanGeographers 74 83ndash97

Castrignano A Giugliarini L Risaliti R andMartinelli N 2000 Study of spatial relation-ships among some soil physico-chemicalproperties of a field in central Italy using multi-variate geostatistics Geoderma 97 39ndash60

Cialella AT Dubayah R Lawrence W andLevine E 1997 Predicting soil drainage classusing remotely sensed and digital elevationdata Journal of Soil Science 62(2) 171ndash78

Clark RN and Swayze GA 1996 Evolution inimaging spectroscopy analysis and sensorsignal-to-noise an examination of how far wehave come Summaries of the sixth annual JPLairborne Earth science workshop 4ndash8 March 1996AVIRIS Workshop Vol 1 5

Clarke LA and Pregibon D 1992 Tree-basedmodels In Chambers J and Hastie J editorsStatistical models in S Pacific Grove Wadsworthand Brooks 377ndash419

Coleman TL Agbu PA and MontgomeryOL 1993 Spectral differentiation of surfacesoils and soil properties ndash is it possible fromspace platforms Soil Science 155 283ndash93

Cook SE Corner RJ Grealish G GesslerPE and Chartres CJ 1996 A rule-basedsystem to map soil properties Soil ScienceSociety of America Journal 60 1893ndash900

Cruickshank JG 1972 Soil geography New YorkNY John Wiley amp Sons

Csillag F Pasztor L Biehl LL 1993 Spectralband selection for the characterization ofsalinity status of soils Remote Sensing ofEnvironment 43 231ndash42

Dale MB McBratney AB and Russell JS1989 On the role of expert systems andnumerical taxonomy in soil classificationJournal of Soil Science 40 223ndash34

Davis JR 1993 Expert systems and environ-

P Scull et al 181pH

N CEC

extractab

lecatio

ns (K Ca M

g)Hartemink et

al 20

01ndash

ndashndash

rp

Heu

velin

k an

d W

ebster 20

01ndash

ndashndash

rp

Hew

itt 19

93ndash

ndashndash

rp

Indo

rante et

al 19

96ndash

ndashndash

rp

King et

al 19

99Lo

gistic regression

Presen

ceabsen

ceSlop

e aspec

t po

t solar

Ec

Non

calc c

lay-loam

energy

Kno

tters eta

l 199

5Kriging

co

-kriging

So

ft layer de

pth

Hillslop

e po

sitio

nCx

regression

kriging

Laga

cherie and

Holmes 19

97Classifica

tion tree

Map

ping

unit

Geo

logy variou

s topo

Ec

indices

Laslett e

tal 198

7Kriging

splin

es tren

dpH

Non

ec

surfac

e nea

rest neigh

bor

McB

ratney 1

992

ndashndash

ndashrp

McB

ratney

eta

l 199

1Block kriging

Clay co

nten

tNon

eC

McB

ratney

eta

l 200

0ndash

ndashndash

rp

McB

ratney

and

Ode

h 199

7Fu

zzy sets in

soil scienc

endash

ndashRp

McB

ratney

and

de Gruijter

Fuzz

y-k-mea

ns w

ithFu

zzy classes

Field co

llected

phy

sical

C19

92ex

tragrade

sch

emical and

morph

olog

ical soil

prop

ertie

s

McC

rack

en and

Cate 198

6Artificial intellig

ence

ndash

ndashrp

expe

rt systems

McK

ensie an

d Austin

19

93Gen

eralized

line

ar m

odels

Clay co

nten

t CEC

Slop

e relief land

form

ec

(logit)

pH EC

COLE

slop

e po

sitio

nbu

lk den

sity and

othe

rs

Moo

re eta

l 199

3Line

ar reg

ression

A horizon

dep

th O

MSlop

e w

etne

ss and

strea

mEC

and P co

nten

t pH

power in

dices aspect

curvature

182 Predictive soil mapping a reviewTa

ble 1

Con

tinu

ed

Stud

yMod

ellin

g metho

dDep

ende

nt variables

Environm

ental v

ariables

Goa

ls atta

ined

a

McK

ensie an

d Ryan 199

9Reg

ression tree and

linea

rSo

lum dep

th P an

dElevation slope

CE

regression

N con

tent

curvature CTI

contribu

ting area do

wn-

slop

e mea

ns for slope

clim

ate da

ta Presco

ttInde

x G

amma

Rad

iometry La

ndsat T

M

and Geo

logy unit

Ode

h et

al 19

92ab

Fuzz

y-c-mea

ns and

Fuzzy classes

Field-co

llected

phy

sica

lC

kriging

chem

ical and

morph

olog

ical soil

prop

ertie

s

Ode

h et

al 19

94

Reg

ression kriging

co

-So

lum dep

th de

pth

Slop

e aspect cu

rvature

eC

1995

kriging regression

to bed

rock gravel

kriging

and clay con

tent

Skidmore et

al 19

91Bayesian expe

rt system

Soil land

scap

e un

itVeg type

wetne

ss in

dex

EcX

1996

grad

ient terrain po

sitio

n

Voltz and

Web

ster 19

90Kriging

cu

bic splin

eClay co

nten

tNon

eC

Web

ster 1

994

Dev

elop

men

t of

ndashndash

rppe

dometrics

Zhu

199

7ab

Fuzz

y logic expe

rt system

Soil series A

horizon

Elev pm aspe

ct c

anop

yECX

Zhu

and

Ban

d 199

4(SoL

IM)

depth in

dividu

alco

verage

grad

ient

Zhu

eta

l 199

7series m

aps

curvature

Not

es

a Letters refer to the de

gree

to w

hich

the

goa

ls of PS

M defined

in the

introdu

ction are achiev

ed

Soilndash

e nvironm

ent relatio

ns utilized

(letter E) be

tter represen

tatio

n of soil c o

ntinuity (C

) an

d ex

pert kno

wledg

e utilize

d (X) c

orrespon

d to goa

ls 1 2 and

3respec

tively from

the

introdu

ction and

cap

ital lette

rs (E C X) ind

icate the metho

d is relatively more successful th

an th

ose metho

ds den

oted

by lower

case le

tters (e c x

) rp ind

icates rev

iew pap

ers

P Scull et al 183

predictive mapping because landscape position is never quantified (rather positionwas often qualitatively defined eg toe-slope) and thus the documented relationshipscannot be generalized using environmental data and digital elevation models to predictsoil character at unvisited sites

Table 1 documents modelling methods model variables and the extent to which thereferenced research satisfies the previously defined goals of PSM (lsquoGoals attainedrsquocolumn) Soil-environment relations utilized (letter E) better representation of soilcontinuity (C) and expert knowledge utilized (X) correspond to goals 1 2 and 3respectively from the introduction and capital letters (E C X) indicate the method isrelatively more successful than those methods denoted by lower case letters (e c x)Citations with no letters present within the lsquoGoals attainedrsquo column do not address theaforementioned goals For example the Cialella et al (1997) received a rating of lsquoEcrsquomeaning that the methods employed successfully utilized environmentalndashsoil characterrelations (E) and somewhat successfully presented a better method of representing soilcontinuity (c) The ratings are provided simply to help organize the literature that wasreviewed Review papers are included within the table denoted by lsquorprsquo Geostatisticalmethods are not included in the table because they have been comprehensivelysurveyed elsewhere (Odeh et al 1994 Burrough et al 1997 McBratney et al 2000Heuvelink and Webster 2001) and because the objectives and assumptions of geostatis-tical methods differ slightly from other PSM research We briefly outline thesedifferences in the following section

1 Geostatistical methods

Geostatistics are a subset of traditional statistics that deal primarily with spatial dataand account for spatial autocorrelation using kriging as the spatial interpolator Theconcept is based upon the theory of regionalized variables which was mainlydeveloped by Matheron (1963) and Krige (1963) Kriging is a form of weighted localaveraging that uses a measure of spatial dependence the variogram to determine theweights applied to the data when computing the averages Geostatistical methods havebeen used in predictive soil mapping research to spatially interpolate soil propertyvalues at unmeasured sites from field-collected data

Burgess and Webster (1980a b) were the first to introduce ordinary kriging to the soilcommunity and since that time an enormous amount of work has been published Forexample ordinary kriging has been used to interpolate many different soil propertiesincluding pollution trace element deficiencies salinity and fertility (Heuvelink andWebster 2001) Ordinary kriging has been criticized for a variety of reasons Forexample Laslett et al (1987) reported that several authors had criticized geostatisticsbecause kriging is a global rather than local technique failing to take into accountknowledge of soil materials and processes Other authors have criticized geostatisticsbecause they are excessively data dependent requiring a large number of closelyspaced data points (Zhu 1997a) As Webster and Oliver (1992) suggest in excess of ahundred samples may be needed to use geostatistics at the field scale because of highspatial variability of soil in some areas Geostatistics also assume spatial autocorrela-tion which sometimes may be a poor assumption in complex terrain where abruptchanges in soil-forming factors occur (McBratney et al 2000) Ordinary kriging by itself

184 Predictive soil mapping a review

does not satisfy two of the three goals of PSM presented in the introduction ndash it does notadequately incorporate expert knowledge and it does not exploit the relationshipbetween environmental variables and soil properties

Ordinary kriging has been modified in a variety of ways to better incorporateancillary data and known soilndashlandscape relationships Block kriging involvesdetermining estimates over meaningful areas rather than at specific points (Burgess andWebster 1980a McBratney et al 1991) Using this method a study area can be stratifiedinto different regions that are reflective of the pedogenetic processes at work In orderto accommodate a trend within a dependent soil variable universal kriging has beenused (Webster 1994) Kriging with external drift is similar to universal kriging but ituses an ancillary variable to represent the trend (McBratney et al 2000) Co-krigingtakes advantage of correlation that may exist between the variable of interest and othermore easily measured variables (Odeh et al 1995) Regression kriging involvesspatially interpolating the residuals from a non-spatial model by kriging and addingthe result to the prediction obtained from that model (Goovaerts 1997 Castrignano etal 2000) Factorial kriging is another method to integrate multivariate data into thestandard kriging routine to extrapolate soil data (Goovaerts 1992) Many authors havecompared these various methods (Laslett et al 1987 Voltz and Webster 1990 Odeh etal 1994 Knotters et al 1995)

Fuzzy logic has been used with geostatistics by various authors to produce new kindsof fuzzy soil maps with continuous classes (Burrough 1989 McBratney and DeGruijter1992 Odeh et al 1992a and reviewed by McBratney and Odeh 1997) The processentails kriging the matrix of membership values determined by the fuzzy k-meansclassifier resulting in a continuous soil surface where individual locations are allowedto belong to more than one class and no rigid boundaries are designated to separate thesoil into discrete units or entities The results of such analysis can be used to assess thepedologic process validity of soil taxonomy by determining whether soils grouptogether into classes that are similar to taxonomic types Fuzzy classes wouldpresumably reflect the main pedologic features within a given area In this sense theclassification is quantitative whereas soil taxonomy is rooted in qualitative discrimina-tion A comparison between the two could provide insightful and help assist soiltaxonomy in making a classification that is more indicative of underlying soil processes

Geostatistics in soil research were originally introduced to quantitatively assess soilvariability within soil mapping units (McBratney et al 1991) in response to criticisms inthe early 1970s that soil unit composition was not well quantified (Beckett and Webster1971) In this regard geostatistics have been very useful having served well the originalgoals set forth by Burgess and Webster (1980b) when they drew kriging to the attentionof soil scientists as a means of spatial prediction At the field scale soil variation islargely due to the effect that topography has on soil genesis Geostatistics have beensuccessfully applied in such environments by using terrain attributes as ancillary datawithin many of the kriging routines described above Such quantitative within-unitvariability of soil properties is very useful in the field of precision agriculture and othersituations (eg pollutants) where very detailed soil attribute information is needed atthe field scale (Heuvelink and Webster 2001)

However geostatistics have not been applied in a wide variety of environments or atlarger scales In order to be successfully applied in different environments geostatisticswill likely require a different suite of ancillary data For example remote sensing data

P Scull et al 185

could be used in arid regions where soil toposequences are less well expressed Atlarger scales of prediction selection of different sets of ancillary variables is requiredbecause different processes define soil character at different scales The most obviousexample is that of climate which may control soil distribution at large scales(continental) but has little explanatory power at the field level Regardless of whetheradequate ancillary data exist the amount of data required to use geostatistics forlandscape-level prediction would be extremely difficult and costly to collect given thestrict sampling protocol required to characterize spatial dependence It is also unclearat what landscape scale soils exhibit spatial autocorrelation

Geostatistical approaches do provide a means of creating continuous soil attributesurfaces to better represent soil continuity (Goal 2) and they can be used to exploit therelationship between environmental variables and soil properties in order to moreefficiently collect soil data (Goal 1) However they do not sufficiently utilize expertknowledge (Goal 3) as no attempt has been made in geostatistical approaches todirectly integrate expert knowledge Fundamentally kriging is a process of interpola-tion designed to predict attribute values in between locations of measured samples Inthis sense geostatistics represent a middle ground between pure interpolation (egnearest neighbour type classifier) in which only measured points for the variable ofinterest are used to determine unknown values and other predictive models thatprimarily use soilndashenvironment correlation to create predictive maps

2 Statistical methods

Statistical methods can be used to exploit the relationship between quantifiablelandscape indices and soil properties to create predictive soil maps For exampleMcKensie and Austin (1993) used a regression to account for a large percentage ofvariation for many soil characteristics (A horizon clay content CEC EC pH bulkdensity and COLE B horizon clay content CEC ESP EC pH bulk density and COLE)using a variety of predictor variables (slope presence or absence of impeding layerrelief landform topographic position) Their results confirm the hypothesis of MooreID et al (1991) that soil character is related to quantifiable landscape indices Howevertheir methods do not provide inference of soil properties at unmeasured sites frommapped environment data because the topographic variables were measured in thefield Linear regression has also been used with terrain variables derived from a 15-mDEM in northeastern Colorado to predict soil attributes (organic matter contentextractable phosphorous pH and texture) at unvisited sites (Moore et al 1993) In thatparticular study 50 of the variance of A-horizon thickness was explained by slope andthe wetness index Gessler et al (1995) also used regression to model A-horizonthickness from topographic variables in southeastern Australia (plan curvature andwetness index r2 = 063 P = 0001) They modeled solum depth and used logisticregression to model E horizon presenceabsence Elsewhere logistic regression hasbeen used to model the presenceabsence of noncalcareous clay loam horizon in centralFrance using terrain attributes from a 20-m DEM (King et al 1999) Exponentialregression has been used to model soil organic carbon using terrain variables (Bell et al2000) in glacial outwash soils in east-central Minnesota Generalized additive models(GAM) have been used less frequently in PSM research Gessler (1996) used a GAM

186 Predictive soil mapping a review

model to predict total soil carbon A horizon depth and solum depth using a variety ofenvironmental predictors (slope elevation wetness index mean annual temperatureprecipitation and radiation)

This small body of research opened the door to more complex methods by demon-strating the existence of quantifiable relationships These authors were able to producesoil attribute maps using raster data models whose scale was dependent upon the gridcell resolution of the environmental data They were successful at exploiting the rela-tionship between quantifiable topographic attributes and soil profile character (Goal 1)The continuous soil attribute surfaces they produce also better represent soil continuitythan the choropleth soil maps produced by traditional soil survey (Goal 2) Howeverthe bulk of these methods (excepting GAMs) are limited by their assumed linear rela-tionship between soil and topographic attributes their assumptions of normallydistributed data and their high data requirements Standard statistical procedures arealso not flexible enough to allow robust integration with a variety of potential datasources such as expert knowledge (Goal 1)

Statistical methods do demonstrate in a quantitative manner that terrain analysis canbe used to predict soil attributes in relatively small areas with homogeneous parentmaterial A large proportion of the research using statistical methods was conducted insemi-arid landscapes at small scales (the largest study area of the entire group was~2000 ha) Obviously for statistical approaches to be most effective they need to bemore universal As such they need to be tested andor developed at larger scales andin more diverse landscapes

3 Decision tree analysis (DTA)

The use of decision tree analysis is just beginning to be explored in predictive soilmapping research although it has been used successfully in the related field ofpredictive vegetation mapping since the early 1990s (Lees and Ritman 1991 MooreDM et al 1991 Franklin 1998) DTA is a form of divisive classification The process oftree modelling involves successively partitioning data (called recursive partitioning inthe tree modelling literature) into increasingly homogeneous subsets which once thepartitioning has ceased are called terminal nodes (Lees and Ritman 1991) Splits orrules defining how to partition the data are selected based on information statistics thatdefine how well the split decreases impurity within the data set (Clark and Pregibon1992) Splits are based on threshold values of an explanatory variable selected bycomparing the increase in resulting purity of node membership for all possiblethresholds and variables The process is iterative growing from the root node (thecomplete data set) to the terminal nodes in a dendritic fashion (Friedl and Brodley1997) Once the tree has been constructed (or grown) it encodes a set of decision rulesthat describe the data partitioning process These rules can be used to classify or predictother data sets (Moore DM et al 1991) Pruning the tree is often necessary to preventthe tree from being overfit to the sample data and to reduce tree complexity Pruningentails combining pairs of terminal nodes into single nodes and can be accomplishedusing cross-validation which yields an initial indication of how large a tree makesrobust predictions (Safavian and Norvig 1991) Cross-validation involves systematical-ly removing portions of the data set and running the remaining sample through the tree

in an iterative manner eventually yielding estimates of the misclassification rates foreach class each node and the whole model (Breiman et al 1984) In this mannerdifferent sized trees can be compared in terms of parsimony

The term DTA is used to collectively refer to all types of tree-based modelling (theword lsquodecisionrsquo is used because it is descriptive indicating that the analysis eventuallyleads to a set of decision rules defining data partitions) The term should be distin-guished from classification tree analysis because the latter refers specifically to DTAwhere the response variable is categorical The term CART (classification andregression trees ndash Breiman et al 1984) is sometimes used but strictly speaking refers tospecific software Friedl and Brodley (1997) provide a review of the decision treealgorithms They divide types of DTA into two classes (1) homogeneous decision treesfor which a single algorithm is used to estimate each split (eg CART) and (2) hybriddecision trees (HDT) for which different splitting methods can be used at differentpoints in the tree (eg Quinlan 1993) They further divide homogeneous decision treesinto univariate (UDT) where single features of the input data define splits and multi-variate decision trees (MDT) where multiple features of the input data can define splitsAccording to this naming convention no distinction is made between methods withdifferent types of response variables although all methods can be used with bothcategorical and continuous response variables In a comparison of these various typesof algorithms on a variety of data sets Friedl and Brodley (1997) found that HDT hadthe highest classification accuracy

The overall aim of DTA is to design a set of predictive rules (eg if geology type Athen soil type B) developed from training data which can then be applied to ageographic data base to predict the value of a response variable (Michaelsen et al 1994)Therefore DTA explicitly uses soilndashlandscape correlation in model development (Goal2) The technique appears promising in soil research but needs to be further tested asnot many of the above types of DTA have been tested in the PSM literature in fact onlyunivariate approaches have been employed For example Lagacherie and Holmes(1997) successfully used univariate DTA to model a categorical response variable soiltype within a training set and then assuming that the training set was representativeextrapolated the model to a much larger region Their work is interesting becausealthough not a single sample came from the area they eventually mapped they wereable to produce a soil map that was more accurate (74 versus 69) than the existingmap produced from traditional methods Cialella et al (1997) also used univariate DTAto predict soil drainage class from a variety of terrain attributes and remotely senseddata They predicted soil drainage class with an average of 78 accuracy ndash impressivegiven that the variation accounted for by a typical soil survey ranges from about halfthe total variance for physical attributes to less than one-tenth for some soil chemicalattributes (Gessler et al 1995)

DTA has been compared with other approaches by several authors In the applicationof erosion modelling DTA results were similar to Artificial Neural Networks (Ellis1996) Both methods achieved high training accuracy (as measured by the KappaStatistic) but in terms of prediction accuracy both methods performed poorly Gessleret al (1995) compared DTA with generalized linear models (GLM) and generalizedadditive models (GAM) to predict A-horizon thickness and concluded that GLM waspreferable to both DTA and GAM McKensie and Ryan (1999) compared regressiontrees and standard linear regression to predict soil properties (total solum depth soil

P Scull et al 187

188 Predictive soil mapping a review

phophorus and soil carbon) using a large variety of predictors (elevation slopecurvature wetness index climate variables geology unit Landsat TM data andGamma radiometric data) and found that different methods work best for differentproperties and overall success hinged on the strength of the relationship between soiland environmental variables Their study of a 50 000 ha forested landscape in southernNew South Wales Australia is one of the few studies at such a small map scale (largearea) or in forested landscape The most extensive comparison of techniques involvingDTA was conducted by McBratney et al (2000) They compared regression techniques(GLM GAM DTA-regression tree) geostatistical techniques (kriging and heterotopicco-kriging) and a hybrid technique (regression kriging) DTA was found to be thepoorest performing of the regression techniques because of the unrealistic predictionsurface generated by the DTA model DTA has also been criticized by other authorsbecause of the stepped prediction surface (Gessler 1996) This phenomenon can beespecially apparent in situations where predictor variables have different resolutionsThe extent to which DTA yields a better spatial representation of soil continuity is afunction of the scale and type of predictor variables used and therefore varies withindividual models

DTA is gaining widespread popularity as a means to develop prediction rules thatcan be rapidly and repeatedly evaluated (Cialella et al 1997 Franklin et al 2000) DTAprovides the following advantages over standard statistical techniques (1) it is easier tointerpret when explanatory variables are both nominal and continuous (2) it isinvariant to monotone re-expressions (transformations) of predictor variables (3) itdeals more satisfactorily with missing data values and outliers (4) it is more adept atcapturing nonadditive and nonlinear behaviour (5) it doesnrsquot make any assumptionsabout data distribution and (6) it is easily updateable as more data are collected(Moore DM et al 1991) The DTA model framework is especially appealing becauseof its capability to integrate a wide range of data sets as explanatory variables

DTA offers a unique opportunity for interaction between soil experts and soilmodellers because the output of the model is a set of rules that can be pedologicallyinterpreted by the soil expert In this sense expert knowledge is used in an implicitmanner in DTA (somewhat effectively achieving Goal 3) While these rules can often beexceedingly complex at minimum the expert can decide whether initial splits makesense given their understanding of the landscape The success of DTA results oftenhinges on the ability of the modeller to make key decisions during the model buildingprocess there is unfortunately no definitive way to determine the most optimal treeThe aid of the soil expert can potentially elucidate this problem

4 Expert systems

A variety of expert system approaches to PSM have been developed to utilize expertknowledge The purpose of such methods is to exploit the information the soil surveyoraccumulates while working in the field by integrating such knowledge into thepredictive model (McCracken and Cate 1986) Unlike the majority of the researchreviewed thus far the dependent variable in many expert systems models is often soiltaxa or mapping unit This apparent disadvantage of expert systems (using classifica-tion to characterize soil continuity) does make them easier to integrate into traditional

soil survey In addition several authors have developed methods to developcontinuous soil property maps from the output of expert system models designed topredict soil unit occurrence

Expert systems are composed of data (information on spatial environmentalvariables eg topography climate etc) a knowledge base (rules and facts related tosoil variation supplied from the soil surveyor) and an inference engine (whichcombines data and the knowledge base to infer logically valid conclusions) (Skidmoreet al 1996) Expert systems differ from conventional models in two ways (1) they storeand manipulate qualitative information (allowing them access to information thatcannot normally be used in other modelling frameworks) and (2) they are structuredas meta-models (the knowledge is separated from the model) (Davis 1993) This allowsthe model to selectively choose which information is relevant at various stages of themodelling process and it allows for information to be easily updated Davis (1993)reviews the application of expert systems to environmental modelling research conclu-ding that the technique is becoming more widely accepted He further states that theapplication of expert systems is constrained by an absence of fundamental knowledgefor rule generation a problem that would appear less relevant to soil mapping giventhe amount of untapped expert knowledge accumulated by the soil surveyor

The first mention of the use of expert systems in pedology was in a paper presenta-tion at the Northeast Committee Soil Survey Conference 1984 (Flach 1985) In herpaper Flach (1985) hinted that recent developments in computer science especiallyexpert systems and artificial intelligence could make modelling a practical mappingtool for soil scientists in the near future A year later McCracken and Cate (1986) hopedto encourage soil scientists to explore expert systems and its application to soil sciencethrough an optimistic article they wrote on the potential use of expert systemsHowever little research was actually conducted in the late 1980s and expert systemshad not yet begun to fulfil the lofty goals set forth by early practitioners (Dale et al1989)

In the early 1990s expert systems approaches to predictive soil mapping began toappear in the literature Skidmore et al (1991) used a Bayesian expert system to mapforest soil into different classes and their results compared favourably with availablesoil maps and actual field-collected data Their methods successfully incorporatedsurveyor knowledge and remotely sensed and digital terrain attributes but failed tobetter characterize continuous soilndashlandscape variability because their final productwas a choropleth map Skidmore et al (1996) revisited their earlier research andassessed the mapping accuracy of their results They found that the soil map producedby the expert system achieved an overall accuracy of 698 (sample size = 53) while themap derived from conventional methods had an accuracy of 736 In addition toknowledge provided by a soil scientist Skidmore et alrsquos (1996) expert system usedvegetation derived for aerial photography and topographic variables derived from a 10-m DEM (wetness index topographic position and slope)

Using an expert system Cook et al (1996) successfully produced a continuous soilproperty map for organic matter content using wetness index aspect and slope asexplanatory variables Their methods were somewhat inefficient because they requireda separate expert system for each soil property of interest However their researchrepresents the only example of expert systems used to directly predict the spatial dis-tribution of a soil property All of the other examples of expert systems in the literature

P Scull et al 189

190 Predictive soil mapping a review

deal with soil type or class The use of expert systems to map soil properties needs tobe explored further

The expert systems discussed thus far have all used Boolean logic within theirtheoretical framework whereby an observation can belong to one and only one classand the soil properties of that class are assigned to the observation With thedevelopment of fuzzy logic and the semantic import model (the second fuzzy logicapproach to creating continuous classes) the opportunity exists to develop fuzzy logic-based expert systems Such systems can be used in conjunction with expert knowledgein situations where experts have a good qualitative idea of how to group data but havetrouble dealing with observations that are not well represented by rigid classificationschemes (Burrough and McDonnell 1998) This method can be particularly useful insituations when taxonomic schemes have been previously developed as is the case insoil taxonomy Several examples of this type of approach were published in the mid-1990s by A Zhu and colleagues (Zhu and Band 1994 Zhu et al 1996 1997 Zhu1997ab) Such systems proved useful for mapping soil at unvisited locations usingsurveyor knowledge and were also capable of producing continuous soil propertymaps The use of fuzzy logic within the theoretical framework of the expert systemallows the complex nature of soil to propagate through the modelling process neversubjected to classification schemes that filter out potentially useful lsquonoisersquo The use offuzzy logic also gives the soil surveyor more latitude during the interview processwhen the knowledge base is defined for the expert system Zhu et al (1997) developeda fuzzy logic-based expert system called SoLIM to determine the similarity of eachgrid cell in a study area to the various taxonomic mapping units delimited by the soilsurvey Continuous soil attribute maps were calculated using the similarity values andtheir relative soil survey determined attribute values (effectively achieving Goal 2) Theresulting data (which proved more accurate than soil survey data once field checked)consists of a raster grid whose resolution is determined by the resolution of the inputenvironmental and digital elevation data As noted previously this type of data modelis more applicable to environmental modelling than the choropleth map (Burrough andMcDonnell 1998)

Expert system approaches to PSM are capable of exploiting soil surveyor knowledgeby developing rule-based systems that imitate the surveyorrsquos conceptual model of soilvariability (the primary focus of Goal 3) The method would appear extremely usefulfor mapping projects (such as those conducted by the NRCS-NCSS) where fieldwork isinitially conducted to determine soilndashlandscape relations Expert system developmentcould be directly inserted into the traditional soil survey mapping approach as asubstitute for the step where the surveyor converts hisher conceptual model into achoropleth map Rather that knowledge could be incorporated into the expert systemwhich could be used to predictively map soil The resulting raster map would be morescientifically based and explicit than the hand-drawn choropleth maps of the past Itsscale would not be limited to that of the aerial photography but rather to the scale ofthe environmental data

Despite these advantages the expert system approach has some drawbacks Becausethe method is dependent upon expert knowledge it cannot be applied whereenvironmentndashsoil relations are poorly understood Of course this criticism can be madeof all PSM models as inductive or deductive knowledge of soilndashenvironment relationsis a prerequisite for PSM However expert systems are deductive models and as such

are not driven by any specific field-collected data (although presumably the soil experthas field experience in the mapping area) Expert systems donrsquot afford the opportunityto first statistically document landscapendashsoil relations and then extrapolate the resultsbecause expert systems do not directly use sample soil data (lsquohard datarsquo) to determinesoilndashlandscape relations Thus the utilization of the relationship between environmen-tal variables and soil properties (Goal 1) is only indirectly achieved Expert systemsapproaches have been demonstrated to be extremely effective in a small number of casestudies The possibility of satisfying all three goals discussed in this review makesexpert systems a predictive soil mapping method that needs to be further tested ndash bothgeographically and across different scales of analysis

V Conclusion

Most of the predictive soil mapping research outlined in this review was conducted atvery large map scales (over small areas) In fact the majority of the research wasconcerned with assessing the spatial variability of soil character within individual fieldsor across soil toposequences The primary driving force behind this type of research hasbeen the need to provide accurate soils information for agriculture and ecologicalmodels It is clear that terrain attributes are powerful predictors at the local scaleGeostatistical tools have been successful at using terrain attributes and the spatialdependence of soil properties to interpolate between existing data points withinindividual fields Across soil toposequences statistical approaches provide a usefulmeans of predicting soil character PSM research has been most successful at the fieldscale because many of the soil-forming factors are held constant For example the neteffect of four of Jennyrsquos five soil-forming factors (climate organism parent material andtime) was minimal within many of the studies reviewed While some research hastackled larger areas of study there exists a deficiency in our ability to predictively mapsoils at smaller map scales Since the distribution of soil is scale-dependent differentPSM methods and predictors are likely to work better at different scales Focus in thefuture must continue to move toward working over larger spatial extents of study inorder to produce landscape-scale soil information

Further a large proportion of the research was conducted in semi-arid gentlysloping agricultural landscapes Humid forestlands mountainous regions and desertshave received little attention As such PSM methods need to continue to be testedandor developed in a wider variety of landscapes where spatial soil distributions canbe more complex Different methods will likely be successful to different degrees indifferent environments Whereas terrain attributes are dominant predictors of soilcharacter across toposequences in prairie lands other predictors are likely necessary inareas where soils do not develop into clear toposequences For example remote sensingdata has been a useful predictor of soil chronosequences in desert landscapes wheresurface appearance is often related to soil character Surficial geology is often reflectiveof soil character in mountain regions where soils are thin and significant bedrock existsat the surface Focus in the future must continue to determine which methods andpredictors work best in which environments

Since the most useful PSM approach will vary across spatial scales and environmen-tal gradients the method used should be driven by the mapping objectives of the

P Scull et al 191

192 Predictive soil mapping a review

project Whereas the traditional soil survey of the past was expected to meet the needsof a diverse group of end users PSM methods will vary given the objectives of thesurvey As a result the end users of the soil data need to play a more active role in thesurvey process Because precision agriculture and large-scale mapping has been theprimary focus of PSM there exists a deficiency in our ability to predictively map soilsfor the purpose of general land use planning and management Even though lessdetailed soil information is needed for these purposes the development of PSMmethods is complicated by the fact that soils exhibit complex spatial variability atsmall map scales where soil-forming environments vary greatly from one location tothe next

A number of alternative methods of characterizing the continuous nature of the soillandscape have been developed Thus far most of the PSM research has provided soilinformation in a nonobject form (as opposed to defining soil types as independententities) Soil data have been generated organized and presented in the form of eitherisorithmic maps or fine-scale raster grids Both of these data models are field-viewmodels of geographic space which allow the soil to be perceived as a constantlyvarying surface Two distinct approaches have been employed (1) mapping individualsoil properties and (2) mapping continuously varying (fuzzy) soil classes Mappingindividual properties is the most common approach and will likely continue todominate PSM research The use of fuzzy soil classes in the literature is less commonand more difficult to be integrated into standard mapping procedures (such as use ofsoil taxa) because the concept radically differs from the traditional view of the soillandscape However fuzzy soil classes (generated using the fuzzy-k means approach) dohave the potential to help further advance Soil taxonomy by identifying taxonomicclasses that are more reflective of pedologic processes at work

Expert systems have been greatly underutilized in PSM research especiallyconsidering how effective a small number of case studies have been Expert systemshave the potential to satisfy successfully all three goals discussed throughout thereview They also have the potential to bridge the gap between traditional approachesand PSM methods because field soil scientists do not have to change their conceptualapproach to mapping They can still conduct field reconnaissance to determinesoilndashlandscape relationships Afterwards though they can use expert systems to betterexploit the knowledge they have garnered Such an approach could help ease thetransition from traditional soil survey to more scientifically explicit methods Expertsystems are also the most fruitful approach to utilizing a wealth of data that has alreadybeen collected in a nonexplicit manner the wealth of expert knowledge (in the form ofthe conceptual model) that senior soil mappers have accumulated

Focus in the future must also move in the direction of operability The utility of manyPSM approaches has been clearly demonstrated yet changes in how soils are mappedand perceived has been slow In the USA various members of the National CooperativeSoil Survey have called for a change in the philosophy of survey (from mappingdiscrete soil types to mapping continuous varying soil properties McSweeny et al1994) but little change has actually taken place To realize the potential of PSM in con-tributing to an overall change in standard soil mapping procedures practitionersworking in the field of PSM need to form working relationships with field soilscientists PSM methods need also to be presented in a manner that is comprehensibleto the soil science trained field mapper Predictive soil mapping is a relatively recent

phenomenon yet much progress has been made In the process of mapping soilknowledge will continue to advance

P Scull et al 193

References

Agbu PA Fehrenbacher DJ and Jansen IJ1990 Statistical comparison of SPOT spectralmaps with field soil maps Soil Science Society ofAmerica Journal 54 818ndash18

Barrett L 1999 Particulars in contextmaintaining a balance in soil geography Annalsof the Association of America Geographers 89707ndash13

Beckett PHT and Webster R 1971 Soilvariability a review Soils and Fertilizers 341ndash15

Bell JC Grigal DF and Bates PC 2000 A soil-terrain model for estimating spatial patterns ofsoil organic carbon In Wilson J and Gallant Jeditors Terrain analysis principles and applica-tions New York City NY John Wiley and Sons295ndash310

Birkeland PW 1999 Soil and geomorphologyThird edition New York NY Oxford UniversityPress

Breiman L Friedman JH Olshen RA andStone CJ 1984 Classification and regressiontrees Belmont CA Wadsworth

Brule FJ 1996 Fuzzy systems ndash a tutorialhttp newsgroup compai httpwwwquadralaycom (last accessed 30 August 2001)

Bunkin FV and Bunkin AF 2000 Lidarsounding of water soil and plants Atmosphericand Oceanic Optics 13 54ndash72

Buol SW Hole FD McCracken RJ andSouthard RJ 1997 Soil genesis and classifica-tion Ames IA Iowa State University Press

Burgess TM and Webster R 1980a Optimalinterpolation and isarithmic mapping of soilproperties the semi-variogram and punctualkriging Journal of Soil Science 31 315ndash31

ndashndashndashndash 1980b Optimal interpolation and isarithmicmapping of soil properties block kringingJournal of Soil Science 31 331ndash41

Burrough PA 1989 Fuzzy mathematicalmethods for soil survey and land evaluationJournal of Soil Science 40 477ndash92

Burrough PA and McDonnell RA 1998Principles of geographic information systems(Revised edition) Oxford Clarendon Press

Burrough PA Beckett PHT and Jarvis MG1971 The relation between cost and utility insoil survey Journal of Soil Science 22 368ndash81

Burrough PA Van Gaans PMF and

Hootsman R 1997 Continuous classificationin soil survey spatial correlation confusionand boundaries Geoderma 77 115ndash35

Cambell JB 1977 Variation of selectedproperties across a soil boundary Soil ScienceSociety of America Journal 41 578ndash82

Cambell JB and Edmonds WJ 1984 Themissing geographic dimension to soiltaxonomy Annals of the Association of AmericanGeographers 74 83ndash97

Castrignano A Giugliarini L Risaliti R andMartinelli N 2000 Study of spatial relation-ships among some soil physico-chemicalproperties of a field in central Italy using multi-variate geostatistics Geoderma 97 39ndash60

Cialella AT Dubayah R Lawrence W andLevine E 1997 Predicting soil drainage classusing remotely sensed and digital elevationdata Journal of Soil Science 62(2) 171ndash78

Clark RN and Swayze GA 1996 Evolution inimaging spectroscopy analysis and sensorsignal-to-noise an examination of how far wehave come Summaries of the sixth annual JPLairborne Earth science workshop 4ndash8 March 1996AVIRIS Workshop Vol 1 5

Clarke LA and Pregibon D 1992 Tree-basedmodels In Chambers J and Hastie J editorsStatistical models in S Pacific Grove Wadsworthand Brooks 377ndash419

Coleman TL Agbu PA and MontgomeryOL 1993 Spectral differentiation of surfacesoils and soil properties ndash is it possible fromspace platforms Soil Science 155 283ndash93

Cook SE Corner RJ Grealish G GesslerPE and Chartres CJ 1996 A rule-basedsystem to map soil properties Soil ScienceSociety of America Journal 60 1893ndash900

Cruickshank JG 1972 Soil geography New YorkNY John Wiley amp Sons

Csillag F Pasztor L Biehl LL 1993 Spectralband selection for the characterization ofsalinity status of soils Remote Sensing ofEnvironment 43 231ndash42

Dale MB McBratney AB and Russell JS1989 On the role of expert systems andnumerical taxonomy in soil classificationJournal of Soil Science 40 223ndash34

Davis JR 1993 Expert systems and environ-

182 Predictive soil mapping a reviewTa

ble 1

Con

tinu

ed

Stud

yMod

ellin

g metho

dDep

ende

nt variables

Environm

ental v

ariables

Goa

ls atta

ined

a

McK

ensie an

d Ryan 199

9Reg

ression tree and

linea

rSo

lum dep

th P an

dElevation slope

CE

regression

N con

tent

curvature CTI

contribu

ting area do

wn-

slop

e mea

ns for slope

clim

ate da

ta Presco

ttInde

x G

amma

Rad

iometry La

ndsat T

M

and Geo

logy unit

Ode

h et

al 19

92ab

Fuzz

y-c-mea

ns and

Fuzzy classes

Field-co

llected

phy

sica

lC

kriging

chem

ical and

morph

olog

ical soil

prop

ertie

s

Ode

h et

al 19

94

Reg

ression kriging

co

-So

lum dep

th de

pth

Slop

e aspect cu

rvature

eC

1995

kriging regression

to bed

rock gravel

kriging

and clay con

tent

Skidmore et

al 19

91Bayesian expe

rt system

Soil land

scap

e un

itVeg type

wetne

ss in

dex

EcX

1996

grad

ient terrain po

sitio

n

Voltz and

Web

ster 19

90Kriging

cu

bic splin

eClay co

nten

tNon

eC

Web

ster 1

994

Dev

elop

men

t of

ndashndash

rppe

dometrics

Zhu

199

7ab

Fuzz

y logic expe

rt system

Soil series A

horizon

Elev pm aspe

ct c

anop

yECX

Zhu

and

Ban

d 199

4(SoL

IM)

depth in

dividu

alco

verage

grad

ient

Zhu

eta

l 199

7series m

aps

curvature

Not

es

a Letters refer to the de

gree

to w

hich

the

goa

ls of PS

M defined

in the

introdu

ction are achiev

ed

Soilndash

e nvironm

ent relatio

ns utilized

(letter E) be

tter represen

tatio

n of soil c o

ntinuity (C

) an

d ex

pert kno

wledg

e utilize

d (X) c

orrespon

d to goa

ls 1 2 and

3respec

tively from

the

introdu

ction and

cap

ital lette

rs (E C X) ind

icate the metho

d is relatively more successful th

an th

ose metho

ds den

oted

by lower

case le

tters (e c x

) rp ind

icates rev

iew pap

ers

P Scull et al 183

predictive mapping because landscape position is never quantified (rather positionwas often qualitatively defined eg toe-slope) and thus the documented relationshipscannot be generalized using environmental data and digital elevation models to predictsoil character at unvisited sites

Table 1 documents modelling methods model variables and the extent to which thereferenced research satisfies the previously defined goals of PSM (lsquoGoals attainedrsquocolumn) Soil-environment relations utilized (letter E) better representation of soilcontinuity (C) and expert knowledge utilized (X) correspond to goals 1 2 and 3respectively from the introduction and capital letters (E C X) indicate the method isrelatively more successful than those methods denoted by lower case letters (e c x)Citations with no letters present within the lsquoGoals attainedrsquo column do not address theaforementioned goals For example the Cialella et al (1997) received a rating of lsquoEcrsquomeaning that the methods employed successfully utilized environmentalndashsoil characterrelations (E) and somewhat successfully presented a better method of representing soilcontinuity (c) The ratings are provided simply to help organize the literature that wasreviewed Review papers are included within the table denoted by lsquorprsquo Geostatisticalmethods are not included in the table because they have been comprehensivelysurveyed elsewhere (Odeh et al 1994 Burrough et al 1997 McBratney et al 2000Heuvelink and Webster 2001) and because the objectives and assumptions of geostatis-tical methods differ slightly from other PSM research We briefly outline thesedifferences in the following section

1 Geostatistical methods

Geostatistics are a subset of traditional statistics that deal primarily with spatial dataand account for spatial autocorrelation using kriging as the spatial interpolator Theconcept is based upon the theory of regionalized variables which was mainlydeveloped by Matheron (1963) and Krige (1963) Kriging is a form of weighted localaveraging that uses a measure of spatial dependence the variogram to determine theweights applied to the data when computing the averages Geostatistical methods havebeen used in predictive soil mapping research to spatially interpolate soil propertyvalues at unmeasured sites from field-collected data

Burgess and Webster (1980a b) were the first to introduce ordinary kriging to the soilcommunity and since that time an enormous amount of work has been published Forexample ordinary kriging has been used to interpolate many different soil propertiesincluding pollution trace element deficiencies salinity and fertility (Heuvelink andWebster 2001) Ordinary kriging has been criticized for a variety of reasons Forexample Laslett et al (1987) reported that several authors had criticized geostatisticsbecause kriging is a global rather than local technique failing to take into accountknowledge of soil materials and processes Other authors have criticized geostatisticsbecause they are excessively data dependent requiring a large number of closelyspaced data points (Zhu 1997a) As Webster and Oliver (1992) suggest in excess of ahundred samples may be needed to use geostatistics at the field scale because of highspatial variability of soil in some areas Geostatistics also assume spatial autocorrela-tion which sometimes may be a poor assumption in complex terrain where abruptchanges in soil-forming factors occur (McBratney et al 2000) Ordinary kriging by itself

184 Predictive soil mapping a review

does not satisfy two of the three goals of PSM presented in the introduction ndash it does notadequately incorporate expert knowledge and it does not exploit the relationshipbetween environmental variables and soil properties

Ordinary kriging has been modified in a variety of ways to better incorporateancillary data and known soilndashlandscape relationships Block kriging involvesdetermining estimates over meaningful areas rather than at specific points (Burgess andWebster 1980a McBratney et al 1991) Using this method a study area can be stratifiedinto different regions that are reflective of the pedogenetic processes at work In orderto accommodate a trend within a dependent soil variable universal kriging has beenused (Webster 1994) Kriging with external drift is similar to universal kriging but ituses an ancillary variable to represent the trend (McBratney et al 2000) Co-krigingtakes advantage of correlation that may exist between the variable of interest and othermore easily measured variables (Odeh et al 1995) Regression kriging involvesspatially interpolating the residuals from a non-spatial model by kriging and addingthe result to the prediction obtained from that model (Goovaerts 1997 Castrignano etal 2000) Factorial kriging is another method to integrate multivariate data into thestandard kriging routine to extrapolate soil data (Goovaerts 1992) Many authors havecompared these various methods (Laslett et al 1987 Voltz and Webster 1990 Odeh etal 1994 Knotters et al 1995)

Fuzzy logic has been used with geostatistics by various authors to produce new kindsof fuzzy soil maps with continuous classes (Burrough 1989 McBratney and DeGruijter1992 Odeh et al 1992a and reviewed by McBratney and Odeh 1997) The processentails kriging the matrix of membership values determined by the fuzzy k-meansclassifier resulting in a continuous soil surface where individual locations are allowedto belong to more than one class and no rigid boundaries are designated to separate thesoil into discrete units or entities The results of such analysis can be used to assess thepedologic process validity of soil taxonomy by determining whether soils grouptogether into classes that are similar to taxonomic types Fuzzy classes wouldpresumably reflect the main pedologic features within a given area In this sense theclassification is quantitative whereas soil taxonomy is rooted in qualitative discrimina-tion A comparison between the two could provide insightful and help assist soiltaxonomy in making a classification that is more indicative of underlying soil processes

Geostatistics in soil research were originally introduced to quantitatively assess soilvariability within soil mapping units (McBratney et al 1991) in response to criticisms inthe early 1970s that soil unit composition was not well quantified (Beckett and Webster1971) In this regard geostatistics have been very useful having served well the originalgoals set forth by Burgess and Webster (1980b) when they drew kriging to the attentionof soil scientists as a means of spatial prediction At the field scale soil variation islargely due to the effect that topography has on soil genesis Geostatistics have beensuccessfully applied in such environments by using terrain attributes as ancillary datawithin many of the kriging routines described above Such quantitative within-unitvariability of soil properties is very useful in the field of precision agriculture and othersituations (eg pollutants) where very detailed soil attribute information is needed atthe field scale (Heuvelink and Webster 2001)

However geostatistics have not been applied in a wide variety of environments or atlarger scales In order to be successfully applied in different environments geostatisticswill likely require a different suite of ancillary data For example remote sensing data

P Scull et al 185

could be used in arid regions where soil toposequences are less well expressed Atlarger scales of prediction selection of different sets of ancillary variables is requiredbecause different processes define soil character at different scales The most obviousexample is that of climate which may control soil distribution at large scales(continental) but has little explanatory power at the field level Regardless of whetheradequate ancillary data exist the amount of data required to use geostatistics forlandscape-level prediction would be extremely difficult and costly to collect given thestrict sampling protocol required to characterize spatial dependence It is also unclearat what landscape scale soils exhibit spatial autocorrelation

Geostatistical approaches do provide a means of creating continuous soil attributesurfaces to better represent soil continuity (Goal 2) and they can be used to exploit therelationship between environmental variables and soil properties in order to moreefficiently collect soil data (Goal 1) However they do not sufficiently utilize expertknowledge (Goal 3) as no attempt has been made in geostatistical approaches todirectly integrate expert knowledge Fundamentally kriging is a process of interpola-tion designed to predict attribute values in between locations of measured samples Inthis sense geostatistics represent a middle ground between pure interpolation (egnearest neighbour type classifier) in which only measured points for the variable ofinterest are used to determine unknown values and other predictive models thatprimarily use soilndashenvironment correlation to create predictive maps

2 Statistical methods

Statistical methods can be used to exploit the relationship between quantifiablelandscape indices and soil properties to create predictive soil maps For exampleMcKensie and Austin (1993) used a regression to account for a large percentage ofvariation for many soil characteristics (A horizon clay content CEC EC pH bulkdensity and COLE B horizon clay content CEC ESP EC pH bulk density and COLE)using a variety of predictor variables (slope presence or absence of impeding layerrelief landform topographic position) Their results confirm the hypothesis of MooreID et al (1991) that soil character is related to quantifiable landscape indices Howevertheir methods do not provide inference of soil properties at unmeasured sites frommapped environment data because the topographic variables were measured in thefield Linear regression has also been used with terrain variables derived from a 15-mDEM in northeastern Colorado to predict soil attributes (organic matter contentextractable phosphorous pH and texture) at unvisited sites (Moore et al 1993) In thatparticular study 50 of the variance of A-horizon thickness was explained by slope andthe wetness index Gessler et al (1995) also used regression to model A-horizonthickness from topographic variables in southeastern Australia (plan curvature andwetness index r2 = 063 P = 0001) They modeled solum depth and used logisticregression to model E horizon presenceabsence Elsewhere logistic regression hasbeen used to model the presenceabsence of noncalcareous clay loam horizon in centralFrance using terrain attributes from a 20-m DEM (King et al 1999) Exponentialregression has been used to model soil organic carbon using terrain variables (Bell et al2000) in glacial outwash soils in east-central Minnesota Generalized additive models(GAM) have been used less frequently in PSM research Gessler (1996) used a GAM

186 Predictive soil mapping a review

model to predict total soil carbon A horizon depth and solum depth using a variety ofenvironmental predictors (slope elevation wetness index mean annual temperatureprecipitation and radiation)

This small body of research opened the door to more complex methods by demon-strating the existence of quantifiable relationships These authors were able to producesoil attribute maps using raster data models whose scale was dependent upon the gridcell resolution of the environmental data They were successful at exploiting the rela-tionship between quantifiable topographic attributes and soil profile character (Goal 1)The continuous soil attribute surfaces they produce also better represent soil continuitythan the choropleth soil maps produced by traditional soil survey (Goal 2) Howeverthe bulk of these methods (excepting GAMs) are limited by their assumed linear rela-tionship between soil and topographic attributes their assumptions of normallydistributed data and their high data requirements Standard statistical procedures arealso not flexible enough to allow robust integration with a variety of potential datasources such as expert knowledge (Goal 1)

Statistical methods do demonstrate in a quantitative manner that terrain analysis canbe used to predict soil attributes in relatively small areas with homogeneous parentmaterial A large proportion of the research using statistical methods was conducted insemi-arid landscapes at small scales (the largest study area of the entire group was~2000 ha) Obviously for statistical approaches to be most effective they need to bemore universal As such they need to be tested andor developed at larger scales andin more diverse landscapes

3 Decision tree analysis (DTA)

The use of decision tree analysis is just beginning to be explored in predictive soilmapping research although it has been used successfully in the related field ofpredictive vegetation mapping since the early 1990s (Lees and Ritman 1991 MooreDM et al 1991 Franklin 1998) DTA is a form of divisive classification The process oftree modelling involves successively partitioning data (called recursive partitioning inthe tree modelling literature) into increasingly homogeneous subsets which once thepartitioning has ceased are called terminal nodes (Lees and Ritman 1991) Splits orrules defining how to partition the data are selected based on information statistics thatdefine how well the split decreases impurity within the data set (Clark and Pregibon1992) Splits are based on threshold values of an explanatory variable selected bycomparing the increase in resulting purity of node membership for all possiblethresholds and variables The process is iterative growing from the root node (thecomplete data set) to the terminal nodes in a dendritic fashion (Friedl and Brodley1997) Once the tree has been constructed (or grown) it encodes a set of decision rulesthat describe the data partitioning process These rules can be used to classify or predictother data sets (Moore DM et al 1991) Pruning the tree is often necessary to preventthe tree from being overfit to the sample data and to reduce tree complexity Pruningentails combining pairs of terminal nodes into single nodes and can be accomplishedusing cross-validation which yields an initial indication of how large a tree makesrobust predictions (Safavian and Norvig 1991) Cross-validation involves systematical-ly removing portions of the data set and running the remaining sample through the tree

in an iterative manner eventually yielding estimates of the misclassification rates foreach class each node and the whole model (Breiman et al 1984) In this mannerdifferent sized trees can be compared in terms of parsimony

The term DTA is used to collectively refer to all types of tree-based modelling (theword lsquodecisionrsquo is used because it is descriptive indicating that the analysis eventuallyleads to a set of decision rules defining data partitions) The term should be distin-guished from classification tree analysis because the latter refers specifically to DTAwhere the response variable is categorical The term CART (classification andregression trees ndash Breiman et al 1984) is sometimes used but strictly speaking refers tospecific software Friedl and Brodley (1997) provide a review of the decision treealgorithms They divide types of DTA into two classes (1) homogeneous decision treesfor which a single algorithm is used to estimate each split (eg CART) and (2) hybriddecision trees (HDT) for which different splitting methods can be used at differentpoints in the tree (eg Quinlan 1993) They further divide homogeneous decision treesinto univariate (UDT) where single features of the input data define splits and multi-variate decision trees (MDT) where multiple features of the input data can define splitsAccording to this naming convention no distinction is made between methods withdifferent types of response variables although all methods can be used with bothcategorical and continuous response variables In a comparison of these various typesof algorithms on a variety of data sets Friedl and Brodley (1997) found that HDT hadthe highest classification accuracy

The overall aim of DTA is to design a set of predictive rules (eg if geology type Athen soil type B) developed from training data which can then be applied to ageographic data base to predict the value of a response variable (Michaelsen et al 1994)Therefore DTA explicitly uses soilndashlandscape correlation in model development (Goal2) The technique appears promising in soil research but needs to be further tested asnot many of the above types of DTA have been tested in the PSM literature in fact onlyunivariate approaches have been employed For example Lagacherie and Holmes(1997) successfully used univariate DTA to model a categorical response variable soiltype within a training set and then assuming that the training set was representativeextrapolated the model to a much larger region Their work is interesting becausealthough not a single sample came from the area they eventually mapped they wereable to produce a soil map that was more accurate (74 versus 69) than the existingmap produced from traditional methods Cialella et al (1997) also used univariate DTAto predict soil drainage class from a variety of terrain attributes and remotely senseddata They predicted soil drainage class with an average of 78 accuracy ndash impressivegiven that the variation accounted for by a typical soil survey ranges from about halfthe total variance for physical attributes to less than one-tenth for some soil chemicalattributes (Gessler et al 1995)

DTA has been compared with other approaches by several authors In the applicationof erosion modelling DTA results were similar to Artificial Neural Networks (Ellis1996) Both methods achieved high training accuracy (as measured by the KappaStatistic) but in terms of prediction accuracy both methods performed poorly Gessleret al (1995) compared DTA with generalized linear models (GLM) and generalizedadditive models (GAM) to predict A-horizon thickness and concluded that GLM waspreferable to both DTA and GAM McKensie and Ryan (1999) compared regressiontrees and standard linear regression to predict soil properties (total solum depth soil

P Scull et al 187

188 Predictive soil mapping a review

phophorus and soil carbon) using a large variety of predictors (elevation slopecurvature wetness index climate variables geology unit Landsat TM data andGamma radiometric data) and found that different methods work best for differentproperties and overall success hinged on the strength of the relationship between soiland environmental variables Their study of a 50 000 ha forested landscape in southernNew South Wales Australia is one of the few studies at such a small map scale (largearea) or in forested landscape The most extensive comparison of techniques involvingDTA was conducted by McBratney et al (2000) They compared regression techniques(GLM GAM DTA-regression tree) geostatistical techniques (kriging and heterotopicco-kriging) and a hybrid technique (regression kriging) DTA was found to be thepoorest performing of the regression techniques because of the unrealistic predictionsurface generated by the DTA model DTA has also been criticized by other authorsbecause of the stepped prediction surface (Gessler 1996) This phenomenon can beespecially apparent in situations where predictor variables have different resolutionsThe extent to which DTA yields a better spatial representation of soil continuity is afunction of the scale and type of predictor variables used and therefore varies withindividual models

DTA is gaining widespread popularity as a means to develop prediction rules thatcan be rapidly and repeatedly evaluated (Cialella et al 1997 Franklin et al 2000) DTAprovides the following advantages over standard statistical techniques (1) it is easier tointerpret when explanatory variables are both nominal and continuous (2) it isinvariant to monotone re-expressions (transformations) of predictor variables (3) itdeals more satisfactorily with missing data values and outliers (4) it is more adept atcapturing nonadditive and nonlinear behaviour (5) it doesnrsquot make any assumptionsabout data distribution and (6) it is easily updateable as more data are collected(Moore DM et al 1991) The DTA model framework is especially appealing becauseof its capability to integrate a wide range of data sets as explanatory variables

DTA offers a unique opportunity for interaction between soil experts and soilmodellers because the output of the model is a set of rules that can be pedologicallyinterpreted by the soil expert In this sense expert knowledge is used in an implicitmanner in DTA (somewhat effectively achieving Goal 3) While these rules can often beexceedingly complex at minimum the expert can decide whether initial splits makesense given their understanding of the landscape The success of DTA results oftenhinges on the ability of the modeller to make key decisions during the model buildingprocess there is unfortunately no definitive way to determine the most optimal treeThe aid of the soil expert can potentially elucidate this problem

4 Expert systems

A variety of expert system approaches to PSM have been developed to utilize expertknowledge The purpose of such methods is to exploit the information the soil surveyoraccumulates while working in the field by integrating such knowledge into thepredictive model (McCracken and Cate 1986) Unlike the majority of the researchreviewed thus far the dependent variable in many expert systems models is often soiltaxa or mapping unit This apparent disadvantage of expert systems (using classifica-tion to characterize soil continuity) does make them easier to integrate into traditional

soil survey In addition several authors have developed methods to developcontinuous soil property maps from the output of expert system models designed topredict soil unit occurrence

Expert systems are composed of data (information on spatial environmentalvariables eg topography climate etc) a knowledge base (rules and facts related tosoil variation supplied from the soil surveyor) and an inference engine (whichcombines data and the knowledge base to infer logically valid conclusions) (Skidmoreet al 1996) Expert systems differ from conventional models in two ways (1) they storeand manipulate qualitative information (allowing them access to information thatcannot normally be used in other modelling frameworks) and (2) they are structuredas meta-models (the knowledge is separated from the model) (Davis 1993) This allowsthe model to selectively choose which information is relevant at various stages of themodelling process and it allows for information to be easily updated Davis (1993)reviews the application of expert systems to environmental modelling research conclu-ding that the technique is becoming more widely accepted He further states that theapplication of expert systems is constrained by an absence of fundamental knowledgefor rule generation a problem that would appear less relevant to soil mapping giventhe amount of untapped expert knowledge accumulated by the soil surveyor

The first mention of the use of expert systems in pedology was in a paper presenta-tion at the Northeast Committee Soil Survey Conference 1984 (Flach 1985) In herpaper Flach (1985) hinted that recent developments in computer science especiallyexpert systems and artificial intelligence could make modelling a practical mappingtool for soil scientists in the near future A year later McCracken and Cate (1986) hopedto encourage soil scientists to explore expert systems and its application to soil sciencethrough an optimistic article they wrote on the potential use of expert systemsHowever little research was actually conducted in the late 1980s and expert systemshad not yet begun to fulfil the lofty goals set forth by early practitioners (Dale et al1989)

In the early 1990s expert systems approaches to predictive soil mapping began toappear in the literature Skidmore et al (1991) used a Bayesian expert system to mapforest soil into different classes and their results compared favourably with availablesoil maps and actual field-collected data Their methods successfully incorporatedsurveyor knowledge and remotely sensed and digital terrain attributes but failed tobetter characterize continuous soilndashlandscape variability because their final productwas a choropleth map Skidmore et al (1996) revisited their earlier research andassessed the mapping accuracy of their results They found that the soil map producedby the expert system achieved an overall accuracy of 698 (sample size = 53) while themap derived from conventional methods had an accuracy of 736 In addition toknowledge provided by a soil scientist Skidmore et alrsquos (1996) expert system usedvegetation derived for aerial photography and topographic variables derived from a 10-m DEM (wetness index topographic position and slope)

Using an expert system Cook et al (1996) successfully produced a continuous soilproperty map for organic matter content using wetness index aspect and slope asexplanatory variables Their methods were somewhat inefficient because they requireda separate expert system for each soil property of interest However their researchrepresents the only example of expert systems used to directly predict the spatial dis-tribution of a soil property All of the other examples of expert systems in the literature

P Scull et al 189

190 Predictive soil mapping a review

deal with soil type or class The use of expert systems to map soil properties needs tobe explored further

The expert systems discussed thus far have all used Boolean logic within theirtheoretical framework whereby an observation can belong to one and only one classand the soil properties of that class are assigned to the observation With thedevelopment of fuzzy logic and the semantic import model (the second fuzzy logicapproach to creating continuous classes) the opportunity exists to develop fuzzy logic-based expert systems Such systems can be used in conjunction with expert knowledgein situations where experts have a good qualitative idea of how to group data but havetrouble dealing with observations that are not well represented by rigid classificationschemes (Burrough and McDonnell 1998) This method can be particularly useful insituations when taxonomic schemes have been previously developed as is the case insoil taxonomy Several examples of this type of approach were published in the mid-1990s by A Zhu and colleagues (Zhu and Band 1994 Zhu et al 1996 1997 Zhu1997ab) Such systems proved useful for mapping soil at unvisited locations usingsurveyor knowledge and were also capable of producing continuous soil propertymaps The use of fuzzy logic within the theoretical framework of the expert systemallows the complex nature of soil to propagate through the modelling process neversubjected to classification schemes that filter out potentially useful lsquonoisersquo The use offuzzy logic also gives the soil surveyor more latitude during the interview processwhen the knowledge base is defined for the expert system Zhu et al (1997) developeda fuzzy logic-based expert system called SoLIM to determine the similarity of eachgrid cell in a study area to the various taxonomic mapping units delimited by the soilsurvey Continuous soil attribute maps were calculated using the similarity values andtheir relative soil survey determined attribute values (effectively achieving Goal 2) Theresulting data (which proved more accurate than soil survey data once field checked)consists of a raster grid whose resolution is determined by the resolution of the inputenvironmental and digital elevation data As noted previously this type of data modelis more applicable to environmental modelling than the choropleth map (Burrough andMcDonnell 1998)

Expert system approaches to PSM are capable of exploiting soil surveyor knowledgeby developing rule-based systems that imitate the surveyorrsquos conceptual model of soilvariability (the primary focus of Goal 3) The method would appear extremely usefulfor mapping projects (such as those conducted by the NRCS-NCSS) where fieldwork isinitially conducted to determine soilndashlandscape relations Expert system developmentcould be directly inserted into the traditional soil survey mapping approach as asubstitute for the step where the surveyor converts hisher conceptual model into achoropleth map Rather that knowledge could be incorporated into the expert systemwhich could be used to predictively map soil The resulting raster map would be morescientifically based and explicit than the hand-drawn choropleth maps of the past Itsscale would not be limited to that of the aerial photography but rather to the scale ofthe environmental data

Despite these advantages the expert system approach has some drawbacks Becausethe method is dependent upon expert knowledge it cannot be applied whereenvironmentndashsoil relations are poorly understood Of course this criticism can be madeof all PSM models as inductive or deductive knowledge of soilndashenvironment relationsis a prerequisite for PSM However expert systems are deductive models and as such

are not driven by any specific field-collected data (although presumably the soil experthas field experience in the mapping area) Expert systems donrsquot afford the opportunityto first statistically document landscapendashsoil relations and then extrapolate the resultsbecause expert systems do not directly use sample soil data (lsquohard datarsquo) to determinesoilndashlandscape relations Thus the utilization of the relationship between environmen-tal variables and soil properties (Goal 1) is only indirectly achieved Expert systemsapproaches have been demonstrated to be extremely effective in a small number of casestudies The possibility of satisfying all three goals discussed in this review makesexpert systems a predictive soil mapping method that needs to be further tested ndash bothgeographically and across different scales of analysis

V Conclusion

Most of the predictive soil mapping research outlined in this review was conducted atvery large map scales (over small areas) In fact the majority of the research wasconcerned with assessing the spatial variability of soil character within individual fieldsor across soil toposequences The primary driving force behind this type of research hasbeen the need to provide accurate soils information for agriculture and ecologicalmodels It is clear that terrain attributes are powerful predictors at the local scaleGeostatistical tools have been successful at using terrain attributes and the spatialdependence of soil properties to interpolate between existing data points withinindividual fields Across soil toposequences statistical approaches provide a usefulmeans of predicting soil character PSM research has been most successful at the fieldscale because many of the soil-forming factors are held constant For example the neteffect of four of Jennyrsquos five soil-forming factors (climate organism parent material andtime) was minimal within many of the studies reviewed While some research hastackled larger areas of study there exists a deficiency in our ability to predictively mapsoils at smaller map scales Since the distribution of soil is scale-dependent differentPSM methods and predictors are likely to work better at different scales Focus in thefuture must continue to move toward working over larger spatial extents of study inorder to produce landscape-scale soil information

Further a large proportion of the research was conducted in semi-arid gentlysloping agricultural landscapes Humid forestlands mountainous regions and desertshave received little attention As such PSM methods need to continue to be testedandor developed in a wider variety of landscapes where spatial soil distributions canbe more complex Different methods will likely be successful to different degrees indifferent environments Whereas terrain attributes are dominant predictors of soilcharacter across toposequences in prairie lands other predictors are likely necessary inareas where soils do not develop into clear toposequences For example remote sensingdata has been a useful predictor of soil chronosequences in desert landscapes wheresurface appearance is often related to soil character Surficial geology is often reflectiveof soil character in mountain regions where soils are thin and significant bedrock existsat the surface Focus in the future must continue to determine which methods andpredictors work best in which environments

Since the most useful PSM approach will vary across spatial scales and environmen-tal gradients the method used should be driven by the mapping objectives of the

P Scull et al 191

192 Predictive soil mapping a review

project Whereas the traditional soil survey of the past was expected to meet the needsof a diverse group of end users PSM methods will vary given the objectives of thesurvey As a result the end users of the soil data need to play a more active role in thesurvey process Because precision agriculture and large-scale mapping has been theprimary focus of PSM there exists a deficiency in our ability to predictively map soilsfor the purpose of general land use planning and management Even though lessdetailed soil information is needed for these purposes the development of PSMmethods is complicated by the fact that soils exhibit complex spatial variability atsmall map scales where soil-forming environments vary greatly from one location tothe next

A number of alternative methods of characterizing the continuous nature of the soillandscape have been developed Thus far most of the PSM research has provided soilinformation in a nonobject form (as opposed to defining soil types as independententities) Soil data have been generated organized and presented in the form of eitherisorithmic maps or fine-scale raster grids Both of these data models are field-viewmodels of geographic space which allow the soil to be perceived as a constantlyvarying surface Two distinct approaches have been employed (1) mapping individualsoil properties and (2) mapping continuously varying (fuzzy) soil classes Mappingindividual properties is the most common approach and will likely continue todominate PSM research The use of fuzzy soil classes in the literature is less commonand more difficult to be integrated into standard mapping procedures (such as use ofsoil taxa) because the concept radically differs from the traditional view of the soillandscape However fuzzy soil classes (generated using the fuzzy-k means approach) dohave the potential to help further advance Soil taxonomy by identifying taxonomicclasses that are more reflective of pedologic processes at work

Expert systems have been greatly underutilized in PSM research especiallyconsidering how effective a small number of case studies have been Expert systemshave the potential to satisfy successfully all three goals discussed throughout thereview They also have the potential to bridge the gap between traditional approachesand PSM methods because field soil scientists do not have to change their conceptualapproach to mapping They can still conduct field reconnaissance to determinesoilndashlandscape relationships Afterwards though they can use expert systems to betterexploit the knowledge they have garnered Such an approach could help ease thetransition from traditional soil survey to more scientifically explicit methods Expertsystems are also the most fruitful approach to utilizing a wealth of data that has alreadybeen collected in a nonexplicit manner the wealth of expert knowledge (in the form ofthe conceptual model) that senior soil mappers have accumulated

Focus in the future must also move in the direction of operability The utility of manyPSM approaches has been clearly demonstrated yet changes in how soils are mappedand perceived has been slow In the USA various members of the National CooperativeSoil Survey have called for a change in the philosophy of survey (from mappingdiscrete soil types to mapping continuous varying soil properties McSweeny et al1994) but little change has actually taken place To realize the potential of PSM in con-tributing to an overall change in standard soil mapping procedures practitionersworking in the field of PSM need to form working relationships with field soilscientists PSM methods need also to be presented in a manner that is comprehensibleto the soil science trained field mapper Predictive soil mapping is a relatively recent

phenomenon yet much progress has been made In the process of mapping soilknowledge will continue to advance

P Scull et al 193

References

Agbu PA Fehrenbacher DJ and Jansen IJ1990 Statistical comparison of SPOT spectralmaps with field soil maps Soil Science Society ofAmerica Journal 54 818ndash18

Barrett L 1999 Particulars in contextmaintaining a balance in soil geography Annalsof the Association of America Geographers 89707ndash13

Beckett PHT and Webster R 1971 Soilvariability a review Soils and Fertilizers 341ndash15

Bell JC Grigal DF and Bates PC 2000 A soil-terrain model for estimating spatial patterns ofsoil organic carbon In Wilson J and Gallant Jeditors Terrain analysis principles and applica-tions New York City NY John Wiley and Sons295ndash310

Birkeland PW 1999 Soil and geomorphologyThird edition New York NY Oxford UniversityPress

Breiman L Friedman JH Olshen RA andStone CJ 1984 Classification and regressiontrees Belmont CA Wadsworth

Brule FJ 1996 Fuzzy systems ndash a tutorialhttp newsgroup compai httpwwwquadralaycom (last accessed 30 August 2001)

Bunkin FV and Bunkin AF 2000 Lidarsounding of water soil and plants Atmosphericand Oceanic Optics 13 54ndash72

Buol SW Hole FD McCracken RJ andSouthard RJ 1997 Soil genesis and classifica-tion Ames IA Iowa State University Press

Burgess TM and Webster R 1980a Optimalinterpolation and isarithmic mapping of soilproperties the semi-variogram and punctualkriging Journal of Soil Science 31 315ndash31

ndashndashndashndash 1980b Optimal interpolation and isarithmicmapping of soil properties block kringingJournal of Soil Science 31 331ndash41

Burrough PA 1989 Fuzzy mathematicalmethods for soil survey and land evaluationJournal of Soil Science 40 477ndash92

Burrough PA and McDonnell RA 1998Principles of geographic information systems(Revised edition) Oxford Clarendon Press

Burrough PA Beckett PHT and Jarvis MG1971 The relation between cost and utility insoil survey Journal of Soil Science 22 368ndash81

Burrough PA Van Gaans PMF and

Hootsman R 1997 Continuous classificationin soil survey spatial correlation confusionand boundaries Geoderma 77 115ndash35

Cambell JB 1977 Variation of selectedproperties across a soil boundary Soil ScienceSociety of America Journal 41 578ndash82

Cambell JB and Edmonds WJ 1984 Themissing geographic dimension to soiltaxonomy Annals of the Association of AmericanGeographers 74 83ndash97

Castrignano A Giugliarini L Risaliti R andMartinelli N 2000 Study of spatial relation-ships among some soil physico-chemicalproperties of a field in central Italy using multi-variate geostatistics Geoderma 97 39ndash60

Cialella AT Dubayah R Lawrence W andLevine E 1997 Predicting soil drainage classusing remotely sensed and digital elevationdata Journal of Soil Science 62(2) 171ndash78

Clark RN and Swayze GA 1996 Evolution inimaging spectroscopy analysis and sensorsignal-to-noise an examination of how far wehave come Summaries of the sixth annual JPLairborne Earth science workshop 4ndash8 March 1996AVIRIS Workshop Vol 1 5

Clarke LA and Pregibon D 1992 Tree-basedmodels In Chambers J and Hastie J editorsStatistical models in S Pacific Grove Wadsworthand Brooks 377ndash419

Coleman TL Agbu PA and MontgomeryOL 1993 Spectral differentiation of surfacesoils and soil properties ndash is it possible fromspace platforms Soil Science 155 283ndash93

Cook SE Corner RJ Grealish G GesslerPE and Chartres CJ 1996 A rule-basedsystem to map soil properties Soil ScienceSociety of America Journal 60 1893ndash900

Cruickshank JG 1972 Soil geography New YorkNY John Wiley amp Sons

Csillag F Pasztor L Biehl LL 1993 Spectralband selection for the characterization ofsalinity status of soils Remote Sensing ofEnvironment 43 231ndash42

Dale MB McBratney AB and Russell JS1989 On the role of expert systems andnumerical taxonomy in soil classificationJournal of Soil Science 40 223ndash34

Davis JR 1993 Expert systems and environ-

P Scull et al 183

predictive mapping because landscape position is never quantified (rather positionwas often qualitatively defined eg toe-slope) and thus the documented relationshipscannot be generalized using environmental data and digital elevation models to predictsoil character at unvisited sites

Table 1 documents modelling methods model variables and the extent to which thereferenced research satisfies the previously defined goals of PSM (lsquoGoals attainedrsquocolumn) Soil-environment relations utilized (letter E) better representation of soilcontinuity (C) and expert knowledge utilized (X) correspond to goals 1 2 and 3respectively from the introduction and capital letters (E C X) indicate the method isrelatively more successful than those methods denoted by lower case letters (e c x)Citations with no letters present within the lsquoGoals attainedrsquo column do not address theaforementioned goals For example the Cialella et al (1997) received a rating of lsquoEcrsquomeaning that the methods employed successfully utilized environmentalndashsoil characterrelations (E) and somewhat successfully presented a better method of representing soilcontinuity (c) The ratings are provided simply to help organize the literature that wasreviewed Review papers are included within the table denoted by lsquorprsquo Geostatisticalmethods are not included in the table because they have been comprehensivelysurveyed elsewhere (Odeh et al 1994 Burrough et al 1997 McBratney et al 2000Heuvelink and Webster 2001) and because the objectives and assumptions of geostatis-tical methods differ slightly from other PSM research We briefly outline thesedifferences in the following section

1 Geostatistical methods

Geostatistics are a subset of traditional statistics that deal primarily with spatial dataand account for spatial autocorrelation using kriging as the spatial interpolator Theconcept is based upon the theory of regionalized variables which was mainlydeveloped by Matheron (1963) and Krige (1963) Kriging is a form of weighted localaveraging that uses a measure of spatial dependence the variogram to determine theweights applied to the data when computing the averages Geostatistical methods havebeen used in predictive soil mapping research to spatially interpolate soil propertyvalues at unmeasured sites from field-collected data

Burgess and Webster (1980a b) were the first to introduce ordinary kriging to the soilcommunity and since that time an enormous amount of work has been published Forexample ordinary kriging has been used to interpolate many different soil propertiesincluding pollution trace element deficiencies salinity and fertility (Heuvelink andWebster 2001) Ordinary kriging has been criticized for a variety of reasons Forexample Laslett et al (1987) reported that several authors had criticized geostatisticsbecause kriging is a global rather than local technique failing to take into accountknowledge of soil materials and processes Other authors have criticized geostatisticsbecause they are excessively data dependent requiring a large number of closelyspaced data points (Zhu 1997a) As Webster and Oliver (1992) suggest in excess of ahundred samples may be needed to use geostatistics at the field scale because of highspatial variability of soil in some areas Geostatistics also assume spatial autocorrela-tion which sometimes may be a poor assumption in complex terrain where abruptchanges in soil-forming factors occur (McBratney et al 2000) Ordinary kriging by itself

184 Predictive soil mapping a review

does not satisfy two of the three goals of PSM presented in the introduction ndash it does notadequately incorporate expert knowledge and it does not exploit the relationshipbetween environmental variables and soil properties

Ordinary kriging has been modified in a variety of ways to better incorporateancillary data and known soilndashlandscape relationships Block kriging involvesdetermining estimates over meaningful areas rather than at specific points (Burgess andWebster 1980a McBratney et al 1991) Using this method a study area can be stratifiedinto different regions that are reflective of the pedogenetic processes at work In orderto accommodate a trend within a dependent soil variable universal kriging has beenused (Webster 1994) Kriging with external drift is similar to universal kriging but ituses an ancillary variable to represent the trend (McBratney et al 2000) Co-krigingtakes advantage of correlation that may exist between the variable of interest and othermore easily measured variables (Odeh et al 1995) Regression kriging involvesspatially interpolating the residuals from a non-spatial model by kriging and addingthe result to the prediction obtained from that model (Goovaerts 1997 Castrignano etal 2000) Factorial kriging is another method to integrate multivariate data into thestandard kriging routine to extrapolate soil data (Goovaerts 1992) Many authors havecompared these various methods (Laslett et al 1987 Voltz and Webster 1990 Odeh etal 1994 Knotters et al 1995)

Fuzzy logic has been used with geostatistics by various authors to produce new kindsof fuzzy soil maps with continuous classes (Burrough 1989 McBratney and DeGruijter1992 Odeh et al 1992a and reviewed by McBratney and Odeh 1997) The processentails kriging the matrix of membership values determined by the fuzzy k-meansclassifier resulting in a continuous soil surface where individual locations are allowedto belong to more than one class and no rigid boundaries are designated to separate thesoil into discrete units or entities The results of such analysis can be used to assess thepedologic process validity of soil taxonomy by determining whether soils grouptogether into classes that are similar to taxonomic types Fuzzy classes wouldpresumably reflect the main pedologic features within a given area In this sense theclassification is quantitative whereas soil taxonomy is rooted in qualitative discrimina-tion A comparison between the two could provide insightful and help assist soiltaxonomy in making a classification that is more indicative of underlying soil processes

Geostatistics in soil research were originally introduced to quantitatively assess soilvariability within soil mapping units (McBratney et al 1991) in response to criticisms inthe early 1970s that soil unit composition was not well quantified (Beckett and Webster1971) In this regard geostatistics have been very useful having served well the originalgoals set forth by Burgess and Webster (1980b) when they drew kriging to the attentionof soil scientists as a means of spatial prediction At the field scale soil variation islargely due to the effect that topography has on soil genesis Geostatistics have beensuccessfully applied in such environments by using terrain attributes as ancillary datawithin many of the kriging routines described above Such quantitative within-unitvariability of soil properties is very useful in the field of precision agriculture and othersituations (eg pollutants) where very detailed soil attribute information is needed atthe field scale (Heuvelink and Webster 2001)

However geostatistics have not been applied in a wide variety of environments or atlarger scales In order to be successfully applied in different environments geostatisticswill likely require a different suite of ancillary data For example remote sensing data

P Scull et al 185

could be used in arid regions where soil toposequences are less well expressed Atlarger scales of prediction selection of different sets of ancillary variables is requiredbecause different processes define soil character at different scales The most obviousexample is that of climate which may control soil distribution at large scales(continental) but has little explanatory power at the field level Regardless of whetheradequate ancillary data exist the amount of data required to use geostatistics forlandscape-level prediction would be extremely difficult and costly to collect given thestrict sampling protocol required to characterize spatial dependence It is also unclearat what landscape scale soils exhibit spatial autocorrelation

Geostatistical approaches do provide a means of creating continuous soil attributesurfaces to better represent soil continuity (Goal 2) and they can be used to exploit therelationship between environmental variables and soil properties in order to moreefficiently collect soil data (Goal 1) However they do not sufficiently utilize expertknowledge (Goal 3) as no attempt has been made in geostatistical approaches todirectly integrate expert knowledge Fundamentally kriging is a process of interpola-tion designed to predict attribute values in between locations of measured samples Inthis sense geostatistics represent a middle ground between pure interpolation (egnearest neighbour type classifier) in which only measured points for the variable ofinterest are used to determine unknown values and other predictive models thatprimarily use soilndashenvironment correlation to create predictive maps

2 Statistical methods

Statistical methods can be used to exploit the relationship between quantifiablelandscape indices and soil properties to create predictive soil maps For exampleMcKensie and Austin (1993) used a regression to account for a large percentage ofvariation for many soil characteristics (A horizon clay content CEC EC pH bulkdensity and COLE B horizon clay content CEC ESP EC pH bulk density and COLE)using a variety of predictor variables (slope presence or absence of impeding layerrelief landform topographic position) Their results confirm the hypothesis of MooreID et al (1991) that soil character is related to quantifiable landscape indices Howevertheir methods do not provide inference of soil properties at unmeasured sites frommapped environment data because the topographic variables were measured in thefield Linear regression has also been used with terrain variables derived from a 15-mDEM in northeastern Colorado to predict soil attributes (organic matter contentextractable phosphorous pH and texture) at unvisited sites (Moore et al 1993) In thatparticular study 50 of the variance of A-horizon thickness was explained by slope andthe wetness index Gessler et al (1995) also used regression to model A-horizonthickness from topographic variables in southeastern Australia (plan curvature andwetness index r2 = 063 P = 0001) They modeled solum depth and used logisticregression to model E horizon presenceabsence Elsewhere logistic regression hasbeen used to model the presenceabsence of noncalcareous clay loam horizon in centralFrance using terrain attributes from a 20-m DEM (King et al 1999) Exponentialregression has been used to model soil organic carbon using terrain variables (Bell et al2000) in glacial outwash soils in east-central Minnesota Generalized additive models(GAM) have been used less frequently in PSM research Gessler (1996) used a GAM

186 Predictive soil mapping a review

model to predict total soil carbon A horizon depth and solum depth using a variety ofenvironmental predictors (slope elevation wetness index mean annual temperatureprecipitation and radiation)

This small body of research opened the door to more complex methods by demon-strating the existence of quantifiable relationships These authors were able to producesoil attribute maps using raster data models whose scale was dependent upon the gridcell resolution of the environmental data They were successful at exploiting the rela-tionship between quantifiable topographic attributes and soil profile character (Goal 1)The continuous soil attribute surfaces they produce also better represent soil continuitythan the choropleth soil maps produced by traditional soil survey (Goal 2) Howeverthe bulk of these methods (excepting GAMs) are limited by their assumed linear rela-tionship between soil and topographic attributes their assumptions of normallydistributed data and their high data requirements Standard statistical procedures arealso not flexible enough to allow robust integration with a variety of potential datasources such as expert knowledge (Goal 1)

Statistical methods do demonstrate in a quantitative manner that terrain analysis canbe used to predict soil attributes in relatively small areas with homogeneous parentmaterial A large proportion of the research using statistical methods was conducted insemi-arid landscapes at small scales (the largest study area of the entire group was~2000 ha) Obviously for statistical approaches to be most effective they need to bemore universal As such they need to be tested andor developed at larger scales andin more diverse landscapes

3 Decision tree analysis (DTA)

The use of decision tree analysis is just beginning to be explored in predictive soilmapping research although it has been used successfully in the related field ofpredictive vegetation mapping since the early 1990s (Lees and Ritman 1991 MooreDM et al 1991 Franklin 1998) DTA is a form of divisive classification The process oftree modelling involves successively partitioning data (called recursive partitioning inthe tree modelling literature) into increasingly homogeneous subsets which once thepartitioning has ceased are called terminal nodes (Lees and Ritman 1991) Splits orrules defining how to partition the data are selected based on information statistics thatdefine how well the split decreases impurity within the data set (Clark and Pregibon1992) Splits are based on threshold values of an explanatory variable selected bycomparing the increase in resulting purity of node membership for all possiblethresholds and variables The process is iterative growing from the root node (thecomplete data set) to the terminal nodes in a dendritic fashion (Friedl and Brodley1997) Once the tree has been constructed (or grown) it encodes a set of decision rulesthat describe the data partitioning process These rules can be used to classify or predictother data sets (Moore DM et al 1991) Pruning the tree is often necessary to preventthe tree from being overfit to the sample data and to reduce tree complexity Pruningentails combining pairs of terminal nodes into single nodes and can be accomplishedusing cross-validation which yields an initial indication of how large a tree makesrobust predictions (Safavian and Norvig 1991) Cross-validation involves systematical-ly removing portions of the data set and running the remaining sample through the tree

in an iterative manner eventually yielding estimates of the misclassification rates foreach class each node and the whole model (Breiman et al 1984) In this mannerdifferent sized trees can be compared in terms of parsimony

The term DTA is used to collectively refer to all types of tree-based modelling (theword lsquodecisionrsquo is used because it is descriptive indicating that the analysis eventuallyleads to a set of decision rules defining data partitions) The term should be distin-guished from classification tree analysis because the latter refers specifically to DTAwhere the response variable is categorical The term CART (classification andregression trees ndash Breiman et al 1984) is sometimes used but strictly speaking refers tospecific software Friedl and Brodley (1997) provide a review of the decision treealgorithms They divide types of DTA into two classes (1) homogeneous decision treesfor which a single algorithm is used to estimate each split (eg CART) and (2) hybriddecision trees (HDT) for which different splitting methods can be used at differentpoints in the tree (eg Quinlan 1993) They further divide homogeneous decision treesinto univariate (UDT) where single features of the input data define splits and multi-variate decision trees (MDT) where multiple features of the input data can define splitsAccording to this naming convention no distinction is made between methods withdifferent types of response variables although all methods can be used with bothcategorical and continuous response variables In a comparison of these various typesof algorithms on a variety of data sets Friedl and Brodley (1997) found that HDT hadthe highest classification accuracy

The overall aim of DTA is to design a set of predictive rules (eg if geology type Athen soil type B) developed from training data which can then be applied to ageographic data base to predict the value of a response variable (Michaelsen et al 1994)Therefore DTA explicitly uses soilndashlandscape correlation in model development (Goal2) The technique appears promising in soil research but needs to be further tested asnot many of the above types of DTA have been tested in the PSM literature in fact onlyunivariate approaches have been employed For example Lagacherie and Holmes(1997) successfully used univariate DTA to model a categorical response variable soiltype within a training set and then assuming that the training set was representativeextrapolated the model to a much larger region Their work is interesting becausealthough not a single sample came from the area they eventually mapped they wereable to produce a soil map that was more accurate (74 versus 69) than the existingmap produced from traditional methods Cialella et al (1997) also used univariate DTAto predict soil drainage class from a variety of terrain attributes and remotely senseddata They predicted soil drainage class with an average of 78 accuracy ndash impressivegiven that the variation accounted for by a typical soil survey ranges from about halfthe total variance for physical attributes to less than one-tenth for some soil chemicalattributes (Gessler et al 1995)

DTA has been compared with other approaches by several authors In the applicationof erosion modelling DTA results were similar to Artificial Neural Networks (Ellis1996) Both methods achieved high training accuracy (as measured by the KappaStatistic) but in terms of prediction accuracy both methods performed poorly Gessleret al (1995) compared DTA with generalized linear models (GLM) and generalizedadditive models (GAM) to predict A-horizon thickness and concluded that GLM waspreferable to both DTA and GAM McKensie and Ryan (1999) compared regressiontrees and standard linear regression to predict soil properties (total solum depth soil

P Scull et al 187

188 Predictive soil mapping a review

phophorus and soil carbon) using a large variety of predictors (elevation slopecurvature wetness index climate variables geology unit Landsat TM data andGamma radiometric data) and found that different methods work best for differentproperties and overall success hinged on the strength of the relationship between soiland environmental variables Their study of a 50 000 ha forested landscape in southernNew South Wales Australia is one of the few studies at such a small map scale (largearea) or in forested landscape The most extensive comparison of techniques involvingDTA was conducted by McBratney et al (2000) They compared regression techniques(GLM GAM DTA-regression tree) geostatistical techniques (kriging and heterotopicco-kriging) and a hybrid technique (regression kriging) DTA was found to be thepoorest performing of the regression techniques because of the unrealistic predictionsurface generated by the DTA model DTA has also been criticized by other authorsbecause of the stepped prediction surface (Gessler 1996) This phenomenon can beespecially apparent in situations where predictor variables have different resolutionsThe extent to which DTA yields a better spatial representation of soil continuity is afunction of the scale and type of predictor variables used and therefore varies withindividual models

DTA is gaining widespread popularity as a means to develop prediction rules thatcan be rapidly and repeatedly evaluated (Cialella et al 1997 Franklin et al 2000) DTAprovides the following advantages over standard statistical techniques (1) it is easier tointerpret when explanatory variables are both nominal and continuous (2) it isinvariant to monotone re-expressions (transformations) of predictor variables (3) itdeals more satisfactorily with missing data values and outliers (4) it is more adept atcapturing nonadditive and nonlinear behaviour (5) it doesnrsquot make any assumptionsabout data distribution and (6) it is easily updateable as more data are collected(Moore DM et al 1991) The DTA model framework is especially appealing becauseof its capability to integrate a wide range of data sets as explanatory variables

DTA offers a unique opportunity for interaction between soil experts and soilmodellers because the output of the model is a set of rules that can be pedologicallyinterpreted by the soil expert In this sense expert knowledge is used in an implicitmanner in DTA (somewhat effectively achieving Goal 3) While these rules can often beexceedingly complex at minimum the expert can decide whether initial splits makesense given their understanding of the landscape The success of DTA results oftenhinges on the ability of the modeller to make key decisions during the model buildingprocess there is unfortunately no definitive way to determine the most optimal treeThe aid of the soil expert can potentially elucidate this problem

4 Expert systems

A variety of expert system approaches to PSM have been developed to utilize expertknowledge The purpose of such methods is to exploit the information the soil surveyoraccumulates while working in the field by integrating such knowledge into thepredictive model (McCracken and Cate 1986) Unlike the majority of the researchreviewed thus far the dependent variable in many expert systems models is often soiltaxa or mapping unit This apparent disadvantage of expert systems (using classifica-tion to characterize soil continuity) does make them easier to integrate into traditional

soil survey In addition several authors have developed methods to developcontinuous soil property maps from the output of expert system models designed topredict soil unit occurrence

Expert systems are composed of data (information on spatial environmentalvariables eg topography climate etc) a knowledge base (rules and facts related tosoil variation supplied from the soil surveyor) and an inference engine (whichcombines data and the knowledge base to infer logically valid conclusions) (Skidmoreet al 1996) Expert systems differ from conventional models in two ways (1) they storeand manipulate qualitative information (allowing them access to information thatcannot normally be used in other modelling frameworks) and (2) they are structuredas meta-models (the knowledge is separated from the model) (Davis 1993) This allowsthe model to selectively choose which information is relevant at various stages of themodelling process and it allows for information to be easily updated Davis (1993)reviews the application of expert systems to environmental modelling research conclu-ding that the technique is becoming more widely accepted He further states that theapplication of expert systems is constrained by an absence of fundamental knowledgefor rule generation a problem that would appear less relevant to soil mapping giventhe amount of untapped expert knowledge accumulated by the soil surveyor

The first mention of the use of expert systems in pedology was in a paper presenta-tion at the Northeast Committee Soil Survey Conference 1984 (Flach 1985) In herpaper Flach (1985) hinted that recent developments in computer science especiallyexpert systems and artificial intelligence could make modelling a practical mappingtool for soil scientists in the near future A year later McCracken and Cate (1986) hopedto encourage soil scientists to explore expert systems and its application to soil sciencethrough an optimistic article they wrote on the potential use of expert systemsHowever little research was actually conducted in the late 1980s and expert systemshad not yet begun to fulfil the lofty goals set forth by early practitioners (Dale et al1989)

In the early 1990s expert systems approaches to predictive soil mapping began toappear in the literature Skidmore et al (1991) used a Bayesian expert system to mapforest soil into different classes and their results compared favourably with availablesoil maps and actual field-collected data Their methods successfully incorporatedsurveyor knowledge and remotely sensed and digital terrain attributes but failed tobetter characterize continuous soilndashlandscape variability because their final productwas a choropleth map Skidmore et al (1996) revisited their earlier research andassessed the mapping accuracy of their results They found that the soil map producedby the expert system achieved an overall accuracy of 698 (sample size = 53) while themap derived from conventional methods had an accuracy of 736 In addition toknowledge provided by a soil scientist Skidmore et alrsquos (1996) expert system usedvegetation derived for aerial photography and topographic variables derived from a 10-m DEM (wetness index topographic position and slope)

Using an expert system Cook et al (1996) successfully produced a continuous soilproperty map for organic matter content using wetness index aspect and slope asexplanatory variables Their methods were somewhat inefficient because they requireda separate expert system for each soil property of interest However their researchrepresents the only example of expert systems used to directly predict the spatial dis-tribution of a soil property All of the other examples of expert systems in the literature

P Scull et al 189

190 Predictive soil mapping a review

deal with soil type or class The use of expert systems to map soil properties needs tobe explored further

The expert systems discussed thus far have all used Boolean logic within theirtheoretical framework whereby an observation can belong to one and only one classand the soil properties of that class are assigned to the observation With thedevelopment of fuzzy logic and the semantic import model (the second fuzzy logicapproach to creating continuous classes) the opportunity exists to develop fuzzy logic-based expert systems Such systems can be used in conjunction with expert knowledgein situations where experts have a good qualitative idea of how to group data but havetrouble dealing with observations that are not well represented by rigid classificationschemes (Burrough and McDonnell 1998) This method can be particularly useful insituations when taxonomic schemes have been previously developed as is the case insoil taxonomy Several examples of this type of approach were published in the mid-1990s by A Zhu and colleagues (Zhu and Band 1994 Zhu et al 1996 1997 Zhu1997ab) Such systems proved useful for mapping soil at unvisited locations usingsurveyor knowledge and were also capable of producing continuous soil propertymaps The use of fuzzy logic within the theoretical framework of the expert systemallows the complex nature of soil to propagate through the modelling process neversubjected to classification schemes that filter out potentially useful lsquonoisersquo The use offuzzy logic also gives the soil surveyor more latitude during the interview processwhen the knowledge base is defined for the expert system Zhu et al (1997) developeda fuzzy logic-based expert system called SoLIM to determine the similarity of eachgrid cell in a study area to the various taxonomic mapping units delimited by the soilsurvey Continuous soil attribute maps were calculated using the similarity values andtheir relative soil survey determined attribute values (effectively achieving Goal 2) Theresulting data (which proved more accurate than soil survey data once field checked)consists of a raster grid whose resolution is determined by the resolution of the inputenvironmental and digital elevation data As noted previously this type of data modelis more applicable to environmental modelling than the choropleth map (Burrough andMcDonnell 1998)

Expert system approaches to PSM are capable of exploiting soil surveyor knowledgeby developing rule-based systems that imitate the surveyorrsquos conceptual model of soilvariability (the primary focus of Goal 3) The method would appear extremely usefulfor mapping projects (such as those conducted by the NRCS-NCSS) where fieldwork isinitially conducted to determine soilndashlandscape relations Expert system developmentcould be directly inserted into the traditional soil survey mapping approach as asubstitute for the step where the surveyor converts hisher conceptual model into achoropleth map Rather that knowledge could be incorporated into the expert systemwhich could be used to predictively map soil The resulting raster map would be morescientifically based and explicit than the hand-drawn choropleth maps of the past Itsscale would not be limited to that of the aerial photography but rather to the scale ofthe environmental data

Despite these advantages the expert system approach has some drawbacks Becausethe method is dependent upon expert knowledge it cannot be applied whereenvironmentndashsoil relations are poorly understood Of course this criticism can be madeof all PSM models as inductive or deductive knowledge of soilndashenvironment relationsis a prerequisite for PSM However expert systems are deductive models and as such

are not driven by any specific field-collected data (although presumably the soil experthas field experience in the mapping area) Expert systems donrsquot afford the opportunityto first statistically document landscapendashsoil relations and then extrapolate the resultsbecause expert systems do not directly use sample soil data (lsquohard datarsquo) to determinesoilndashlandscape relations Thus the utilization of the relationship between environmen-tal variables and soil properties (Goal 1) is only indirectly achieved Expert systemsapproaches have been demonstrated to be extremely effective in a small number of casestudies The possibility of satisfying all three goals discussed in this review makesexpert systems a predictive soil mapping method that needs to be further tested ndash bothgeographically and across different scales of analysis

V Conclusion

Most of the predictive soil mapping research outlined in this review was conducted atvery large map scales (over small areas) In fact the majority of the research wasconcerned with assessing the spatial variability of soil character within individual fieldsor across soil toposequences The primary driving force behind this type of research hasbeen the need to provide accurate soils information for agriculture and ecologicalmodels It is clear that terrain attributes are powerful predictors at the local scaleGeostatistical tools have been successful at using terrain attributes and the spatialdependence of soil properties to interpolate between existing data points withinindividual fields Across soil toposequences statistical approaches provide a usefulmeans of predicting soil character PSM research has been most successful at the fieldscale because many of the soil-forming factors are held constant For example the neteffect of four of Jennyrsquos five soil-forming factors (climate organism parent material andtime) was minimal within many of the studies reviewed While some research hastackled larger areas of study there exists a deficiency in our ability to predictively mapsoils at smaller map scales Since the distribution of soil is scale-dependent differentPSM methods and predictors are likely to work better at different scales Focus in thefuture must continue to move toward working over larger spatial extents of study inorder to produce landscape-scale soil information

Further a large proportion of the research was conducted in semi-arid gentlysloping agricultural landscapes Humid forestlands mountainous regions and desertshave received little attention As such PSM methods need to continue to be testedandor developed in a wider variety of landscapes where spatial soil distributions canbe more complex Different methods will likely be successful to different degrees indifferent environments Whereas terrain attributes are dominant predictors of soilcharacter across toposequences in prairie lands other predictors are likely necessary inareas where soils do not develop into clear toposequences For example remote sensingdata has been a useful predictor of soil chronosequences in desert landscapes wheresurface appearance is often related to soil character Surficial geology is often reflectiveof soil character in mountain regions where soils are thin and significant bedrock existsat the surface Focus in the future must continue to determine which methods andpredictors work best in which environments

Since the most useful PSM approach will vary across spatial scales and environmen-tal gradients the method used should be driven by the mapping objectives of the

P Scull et al 191

192 Predictive soil mapping a review

project Whereas the traditional soil survey of the past was expected to meet the needsof a diverse group of end users PSM methods will vary given the objectives of thesurvey As a result the end users of the soil data need to play a more active role in thesurvey process Because precision agriculture and large-scale mapping has been theprimary focus of PSM there exists a deficiency in our ability to predictively map soilsfor the purpose of general land use planning and management Even though lessdetailed soil information is needed for these purposes the development of PSMmethods is complicated by the fact that soils exhibit complex spatial variability atsmall map scales where soil-forming environments vary greatly from one location tothe next

A number of alternative methods of characterizing the continuous nature of the soillandscape have been developed Thus far most of the PSM research has provided soilinformation in a nonobject form (as opposed to defining soil types as independententities) Soil data have been generated organized and presented in the form of eitherisorithmic maps or fine-scale raster grids Both of these data models are field-viewmodels of geographic space which allow the soil to be perceived as a constantlyvarying surface Two distinct approaches have been employed (1) mapping individualsoil properties and (2) mapping continuously varying (fuzzy) soil classes Mappingindividual properties is the most common approach and will likely continue todominate PSM research The use of fuzzy soil classes in the literature is less commonand more difficult to be integrated into standard mapping procedures (such as use ofsoil taxa) because the concept radically differs from the traditional view of the soillandscape However fuzzy soil classes (generated using the fuzzy-k means approach) dohave the potential to help further advance Soil taxonomy by identifying taxonomicclasses that are more reflective of pedologic processes at work

Expert systems have been greatly underutilized in PSM research especiallyconsidering how effective a small number of case studies have been Expert systemshave the potential to satisfy successfully all three goals discussed throughout thereview They also have the potential to bridge the gap between traditional approachesand PSM methods because field soil scientists do not have to change their conceptualapproach to mapping They can still conduct field reconnaissance to determinesoilndashlandscape relationships Afterwards though they can use expert systems to betterexploit the knowledge they have garnered Such an approach could help ease thetransition from traditional soil survey to more scientifically explicit methods Expertsystems are also the most fruitful approach to utilizing a wealth of data that has alreadybeen collected in a nonexplicit manner the wealth of expert knowledge (in the form ofthe conceptual model) that senior soil mappers have accumulated

Focus in the future must also move in the direction of operability The utility of manyPSM approaches has been clearly demonstrated yet changes in how soils are mappedand perceived has been slow In the USA various members of the National CooperativeSoil Survey have called for a change in the philosophy of survey (from mappingdiscrete soil types to mapping continuous varying soil properties McSweeny et al1994) but little change has actually taken place To realize the potential of PSM in con-tributing to an overall change in standard soil mapping procedures practitionersworking in the field of PSM need to form working relationships with field soilscientists PSM methods need also to be presented in a manner that is comprehensibleto the soil science trained field mapper Predictive soil mapping is a relatively recent

phenomenon yet much progress has been made In the process of mapping soilknowledge will continue to advance

P Scull et al 193

References

Agbu PA Fehrenbacher DJ and Jansen IJ1990 Statistical comparison of SPOT spectralmaps with field soil maps Soil Science Society ofAmerica Journal 54 818ndash18

Barrett L 1999 Particulars in contextmaintaining a balance in soil geography Annalsof the Association of America Geographers 89707ndash13

Beckett PHT and Webster R 1971 Soilvariability a review Soils and Fertilizers 341ndash15

Bell JC Grigal DF and Bates PC 2000 A soil-terrain model for estimating spatial patterns ofsoil organic carbon In Wilson J and Gallant Jeditors Terrain analysis principles and applica-tions New York City NY John Wiley and Sons295ndash310

Birkeland PW 1999 Soil and geomorphologyThird edition New York NY Oxford UniversityPress

Breiman L Friedman JH Olshen RA andStone CJ 1984 Classification and regressiontrees Belmont CA Wadsworth

Brule FJ 1996 Fuzzy systems ndash a tutorialhttp newsgroup compai httpwwwquadralaycom (last accessed 30 August 2001)

Bunkin FV and Bunkin AF 2000 Lidarsounding of water soil and plants Atmosphericand Oceanic Optics 13 54ndash72

Buol SW Hole FD McCracken RJ andSouthard RJ 1997 Soil genesis and classifica-tion Ames IA Iowa State University Press

Burgess TM and Webster R 1980a Optimalinterpolation and isarithmic mapping of soilproperties the semi-variogram and punctualkriging Journal of Soil Science 31 315ndash31

ndashndashndashndash 1980b Optimal interpolation and isarithmicmapping of soil properties block kringingJournal of Soil Science 31 331ndash41

Burrough PA 1989 Fuzzy mathematicalmethods for soil survey and land evaluationJournal of Soil Science 40 477ndash92

Burrough PA and McDonnell RA 1998Principles of geographic information systems(Revised edition) Oxford Clarendon Press

Burrough PA Beckett PHT and Jarvis MG1971 The relation between cost and utility insoil survey Journal of Soil Science 22 368ndash81

Burrough PA Van Gaans PMF and

Hootsman R 1997 Continuous classificationin soil survey spatial correlation confusionand boundaries Geoderma 77 115ndash35

Cambell JB 1977 Variation of selectedproperties across a soil boundary Soil ScienceSociety of America Journal 41 578ndash82

Cambell JB and Edmonds WJ 1984 Themissing geographic dimension to soiltaxonomy Annals of the Association of AmericanGeographers 74 83ndash97

Castrignano A Giugliarini L Risaliti R andMartinelli N 2000 Study of spatial relation-ships among some soil physico-chemicalproperties of a field in central Italy using multi-variate geostatistics Geoderma 97 39ndash60

Cialella AT Dubayah R Lawrence W andLevine E 1997 Predicting soil drainage classusing remotely sensed and digital elevationdata Journal of Soil Science 62(2) 171ndash78

Clark RN and Swayze GA 1996 Evolution inimaging spectroscopy analysis and sensorsignal-to-noise an examination of how far wehave come Summaries of the sixth annual JPLairborne Earth science workshop 4ndash8 March 1996AVIRIS Workshop Vol 1 5

Clarke LA and Pregibon D 1992 Tree-basedmodels In Chambers J and Hastie J editorsStatistical models in S Pacific Grove Wadsworthand Brooks 377ndash419

Coleman TL Agbu PA and MontgomeryOL 1993 Spectral differentiation of surfacesoils and soil properties ndash is it possible fromspace platforms Soil Science 155 283ndash93

Cook SE Corner RJ Grealish G GesslerPE and Chartres CJ 1996 A rule-basedsystem to map soil properties Soil ScienceSociety of America Journal 60 1893ndash900

Cruickshank JG 1972 Soil geography New YorkNY John Wiley amp Sons

Csillag F Pasztor L Biehl LL 1993 Spectralband selection for the characterization ofsalinity status of soils Remote Sensing ofEnvironment 43 231ndash42

Dale MB McBratney AB and Russell JS1989 On the role of expert systems andnumerical taxonomy in soil classificationJournal of Soil Science 40 223ndash34

Davis JR 1993 Expert systems and environ-

184 Predictive soil mapping a review

does not satisfy two of the three goals of PSM presented in the introduction ndash it does notadequately incorporate expert knowledge and it does not exploit the relationshipbetween environmental variables and soil properties

Ordinary kriging has been modified in a variety of ways to better incorporateancillary data and known soilndashlandscape relationships Block kriging involvesdetermining estimates over meaningful areas rather than at specific points (Burgess andWebster 1980a McBratney et al 1991) Using this method a study area can be stratifiedinto different regions that are reflective of the pedogenetic processes at work In orderto accommodate a trend within a dependent soil variable universal kriging has beenused (Webster 1994) Kriging with external drift is similar to universal kriging but ituses an ancillary variable to represent the trend (McBratney et al 2000) Co-krigingtakes advantage of correlation that may exist between the variable of interest and othermore easily measured variables (Odeh et al 1995) Regression kriging involvesspatially interpolating the residuals from a non-spatial model by kriging and addingthe result to the prediction obtained from that model (Goovaerts 1997 Castrignano etal 2000) Factorial kriging is another method to integrate multivariate data into thestandard kriging routine to extrapolate soil data (Goovaerts 1992) Many authors havecompared these various methods (Laslett et al 1987 Voltz and Webster 1990 Odeh etal 1994 Knotters et al 1995)

Fuzzy logic has been used with geostatistics by various authors to produce new kindsof fuzzy soil maps with continuous classes (Burrough 1989 McBratney and DeGruijter1992 Odeh et al 1992a and reviewed by McBratney and Odeh 1997) The processentails kriging the matrix of membership values determined by the fuzzy k-meansclassifier resulting in a continuous soil surface where individual locations are allowedto belong to more than one class and no rigid boundaries are designated to separate thesoil into discrete units or entities The results of such analysis can be used to assess thepedologic process validity of soil taxonomy by determining whether soils grouptogether into classes that are similar to taxonomic types Fuzzy classes wouldpresumably reflect the main pedologic features within a given area In this sense theclassification is quantitative whereas soil taxonomy is rooted in qualitative discrimina-tion A comparison between the two could provide insightful and help assist soiltaxonomy in making a classification that is more indicative of underlying soil processes

Geostatistics in soil research were originally introduced to quantitatively assess soilvariability within soil mapping units (McBratney et al 1991) in response to criticisms inthe early 1970s that soil unit composition was not well quantified (Beckett and Webster1971) In this regard geostatistics have been very useful having served well the originalgoals set forth by Burgess and Webster (1980b) when they drew kriging to the attentionof soil scientists as a means of spatial prediction At the field scale soil variation islargely due to the effect that topography has on soil genesis Geostatistics have beensuccessfully applied in such environments by using terrain attributes as ancillary datawithin many of the kriging routines described above Such quantitative within-unitvariability of soil properties is very useful in the field of precision agriculture and othersituations (eg pollutants) where very detailed soil attribute information is needed atthe field scale (Heuvelink and Webster 2001)

However geostatistics have not been applied in a wide variety of environments or atlarger scales In order to be successfully applied in different environments geostatisticswill likely require a different suite of ancillary data For example remote sensing data

P Scull et al 185

could be used in arid regions where soil toposequences are less well expressed Atlarger scales of prediction selection of different sets of ancillary variables is requiredbecause different processes define soil character at different scales The most obviousexample is that of climate which may control soil distribution at large scales(continental) but has little explanatory power at the field level Regardless of whetheradequate ancillary data exist the amount of data required to use geostatistics forlandscape-level prediction would be extremely difficult and costly to collect given thestrict sampling protocol required to characterize spatial dependence It is also unclearat what landscape scale soils exhibit spatial autocorrelation

Geostatistical approaches do provide a means of creating continuous soil attributesurfaces to better represent soil continuity (Goal 2) and they can be used to exploit therelationship between environmental variables and soil properties in order to moreefficiently collect soil data (Goal 1) However they do not sufficiently utilize expertknowledge (Goal 3) as no attempt has been made in geostatistical approaches todirectly integrate expert knowledge Fundamentally kriging is a process of interpola-tion designed to predict attribute values in between locations of measured samples Inthis sense geostatistics represent a middle ground between pure interpolation (egnearest neighbour type classifier) in which only measured points for the variable ofinterest are used to determine unknown values and other predictive models thatprimarily use soilndashenvironment correlation to create predictive maps

2 Statistical methods

Statistical methods can be used to exploit the relationship between quantifiablelandscape indices and soil properties to create predictive soil maps For exampleMcKensie and Austin (1993) used a regression to account for a large percentage ofvariation for many soil characteristics (A horizon clay content CEC EC pH bulkdensity and COLE B horizon clay content CEC ESP EC pH bulk density and COLE)using a variety of predictor variables (slope presence or absence of impeding layerrelief landform topographic position) Their results confirm the hypothesis of MooreID et al (1991) that soil character is related to quantifiable landscape indices Howevertheir methods do not provide inference of soil properties at unmeasured sites frommapped environment data because the topographic variables were measured in thefield Linear regression has also been used with terrain variables derived from a 15-mDEM in northeastern Colorado to predict soil attributes (organic matter contentextractable phosphorous pH and texture) at unvisited sites (Moore et al 1993) In thatparticular study 50 of the variance of A-horizon thickness was explained by slope andthe wetness index Gessler et al (1995) also used regression to model A-horizonthickness from topographic variables in southeastern Australia (plan curvature andwetness index r2 = 063 P = 0001) They modeled solum depth and used logisticregression to model E horizon presenceabsence Elsewhere logistic regression hasbeen used to model the presenceabsence of noncalcareous clay loam horizon in centralFrance using terrain attributes from a 20-m DEM (King et al 1999) Exponentialregression has been used to model soil organic carbon using terrain variables (Bell et al2000) in glacial outwash soils in east-central Minnesota Generalized additive models(GAM) have been used less frequently in PSM research Gessler (1996) used a GAM

186 Predictive soil mapping a review

model to predict total soil carbon A horizon depth and solum depth using a variety ofenvironmental predictors (slope elevation wetness index mean annual temperatureprecipitation and radiation)

This small body of research opened the door to more complex methods by demon-strating the existence of quantifiable relationships These authors were able to producesoil attribute maps using raster data models whose scale was dependent upon the gridcell resolution of the environmental data They were successful at exploiting the rela-tionship between quantifiable topographic attributes and soil profile character (Goal 1)The continuous soil attribute surfaces they produce also better represent soil continuitythan the choropleth soil maps produced by traditional soil survey (Goal 2) Howeverthe bulk of these methods (excepting GAMs) are limited by their assumed linear rela-tionship between soil and topographic attributes their assumptions of normallydistributed data and their high data requirements Standard statistical procedures arealso not flexible enough to allow robust integration with a variety of potential datasources such as expert knowledge (Goal 1)

Statistical methods do demonstrate in a quantitative manner that terrain analysis canbe used to predict soil attributes in relatively small areas with homogeneous parentmaterial A large proportion of the research using statistical methods was conducted insemi-arid landscapes at small scales (the largest study area of the entire group was~2000 ha) Obviously for statistical approaches to be most effective they need to bemore universal As such they need to be tested andor developed at larger scales andin more diverse landscapes

3 Decision tree analysis (DTA)

The use of decision tree analysis is just beginning to be explored in predictive soilmapping research although it has been used successfully in the related field ofpredictive vegetation mapping since the early 1990s (Lees and Ritman 1991 MooreDM et al 1991 Franklin 1998) DTA is a form of divisive classification The process oftree modelling involves successively partitioning data (called recursive partitioning inthe tree modelling literature) into increasingly homogeneous subsets which once thepartitioning has ceased are called terminal nodes (Lees and Ritman 1991) Splits orrules defining how to partition the data are selected based on information statistics thatdefine how well the split decreases impurity within the data set (Clark and Pregibon1992) Splits are based on threshold values of an explanatory variable selected bycomparing the increase in resulting purity of node membership for all possiblethresholds and variables The process is iterative growing from the root node (thecomplete data set) to the terminal nodes in a dendritic fashion (Friedl and Brodley1997) Once the tree has been constructed (or grown) it encodes a set of decision rulesthat describe the data partitioning process These rules can be used to classify or predictother data sets (Moore DM et al 1991) Pruning the tree is often necessary to preventthe tree from being overfit to the sample data and to reduce tree complexity Pruningentails combining pairs of terminal nodes into single nodes and can be accomplishedusing cross-validation which yields an initial indication of how large a tree makesrobust predictions (Safavian and Norvig 1991) Cross-validation involves systematical-ly removing portions of the data set and running the remaining sample through the tree

in an iterative manner eventually yielding estimates of the misclassification rates foreach class each node and the whole model (Breiman et al 1984) In this mannerdifferent sized trees can be compared in terms of parsimony

The term DTA is used to collectively refer to all types of tree-based modelling (theword lsquodecisionrsquo is used because it is descriptive indicating that the analysis eventuallyleads to a set of decision rules defining data partitions) The term should be distin-guished from classification tree analysis because the latter refers specifically to DTAwhere the response variable is categorical The term CART (classification andregression trees ndash Breiman et al 1984) is sometimes used but strictly speaking refers tospecific software Friedl and Brodley (1997) provide a review of the decision treealgorithms They divide types of DTA into two classes (1) homogeneous decision treesfor which a single algorithm is used to estimate each split (eg CART) and (2) hybriddecision trees (HDT) for which different splitting methods can be used at differentpoints in the tree (eg Quinlan 1993) They further divide homogeneous decision treesinto univariate (UDT) where single features of the input data define splits and multi-variate decision trees (MDT) where multiple features of the input data can define splitsAccording to this naming convention no distinction is made between methods withdifferent types of response variables although all methods can be used with bothcategorical and continuous response variables In a comparison of these various typesof algorithms on a variety of data sets Friedl and Brodley (1997) found that HDT hadthe highest classification accuracy

The overall aim of DTA is to design a set of predictive rules (eg if geology type Athen soil type B) developed from training data which can then be applied to ageographic data base to predict the value of a response variable (Michaelsen et al 1994)Therefore DTA explicitly uses soilndashlandscape correlation in model development (Goal2) The technique appears promising in soil research but needs to be further tested asnot many of the above types of DTA have been tested in the PSM literature in fact onlyunivariate approaches have been employed For example Lagacherie and Holmes(1997) successfully used univariate DTA to model a categorical response variable soiltype within a training set and then assuming that the training set was representativeextrapolated the model to a much larger region Their work is interesting becausealthough not a single sample came from the area they eventually mapped they wereable to produce a soil map that was more accurate (74 versus 69) than the existingmap produced from traditional methods Cialella et al (1997) also used univariate DTAto predict soil drainage class from a variety of terrain attributes and remotely senseddata They predicted soil drainage class with an average of 78 accuracy ndash impressivegiven that the variation accounted for by a typical soil survey ranges from about halfthe total variance for physical attributes to less than one-tenth for some soil chemicalattributes (Gessler et al 1995)

DTA has been compared with other approaches by several authors In the applicationof erosion modelling DTA results were similar to Artificial Neural Networks (Ellis1996) Both methods achieved high training accuracy (as measured by the KappaStatistic) but in terms of prediction accuracy both methods performed poorly Gessleret al (1995) compared DTA with generalized linear models (GLM) and generalizedadditive models (GAM) to predict A-horizon thickness and concluded that GLM waspreferable to both DTA and GAM McKensie and Ryan (1999) compared regressiontrees and standard linear regression to predict soil properties (total solum depth soil

P Scull et al 187

188 Predictive soil mapping a review

phophorus and soil carbon) using a large variety of predictors (elevation slopecurvature wetness index climate variables geology unit Landsat TM data andGamma radiometric data) and found that different methods work best for differentproperties and overall success hinged on the strength of the relationship between soiland environmental variables Their study of a 50 000 ha forested landscape in southernNew South Wales Australia is one of the few studies at such a small map scale (largearea) or in forested landscape The most extensive comparison of techniques involvingDTA was conducted by McBratney et al (2000) They compared regression techniques(GLM GAM DTA-regression tree) geostatistical techniques (kriging and heterotopicco-kriging) and a hybrid technique (regression kriging) DTA was found to be thepoorest performing of the regression techniques because of the unrealistic predictionsurface generated by the DTA model DTA has also been criticized by other authorsbecause of the stepped prediction surface (Gessler 1996) This phenomenon can beespecially apparent in situations where predictor variables have different resolutionsThe extent to which DTA yields a better spatial representation of soil continuity is afunction of the scale and type of predictor variables used and therefore varies withindividual models

DTA is gaining widespread popularity as a means to develop prediction rules thatcan be rapidly and repeatedly evaluated (Cialella et al 1997 Franklin et al 2000) DTAprovides the following advantages over standard statistical techniques (1) it is easier tointerpret when explanatory variables are both nominal and continuous (2) it isinvariant to monotone re-expressions (transformations) of predictor variables (3) itdeals more satisfactorily with missing data values and outliers (4) it is more adept atcapturing nonadditive and nonlinear behaviour (5) it doesnrsquot make any assumptionsabout data distribution and (6) it is easily updateable as more data are collected(Moore DM et al 1991) The DTA model framework is especially appealing becauseof its capability to integrate a wide range of data sets as explanatory variables

DTA offers a unique opportunity for interaction between soil experts and soilmodellers because the output of the model is a set of rules that can be pedologicallyinterpreted by the soil expert In this sense expert knowledge is used in an implicitmanner in DTA (somewhat effectively achieving Goal 3) While these rules can often beexceedingly complex at minimum the expert can decide whether initial splits makesense given their understanding of the landscape The success of DTA results oftenhinges on the ability of the modeller to make key decisions during the model buildingprocess there is unfortunately no definitive way to determine the most optimal treeThe aid of the soil expert can potentially elucidate this problem

4 Expert systems

A variety of expert system approaches to PSM have been developed to utilize expertknowledge The purpose of such methods is to exploit the information the soil surveyoraccumulates while working in the field by integrating such knowledge into thepredictive model (McCracken and Cate 1986) Unlike the majority of the researchreviewed thus far the dependent variable in many expert systems models is often soiltaxa or mapping unit This apparent disadvantage of expert systems (using classifica-tion to characterize soil continuity) does make them easier to integrate into traditional

soil survey In addition several authors have developed methods to developcontinuous soil property maps from the output of expert system models designed topredict soil unit occurrence

Expert systems are composed of data (information on spatial environmentalvariables eg topography climate etc) a knowledge base (rules and facts related tosoil variation supplied from the soil surveyor) and an inference engine (whichcombines data and the knowledge base to infer logically valid conclusions) (Skidmoreet al 1996) Expert systems differ from conventional models in two ways (1) they storeand manipulate qualitative information (allowing them access to information thatcannot normally be used in other modelling frameworks) and (2) they are structuredas meta-models (the knowledge is separated from the model) (Davis 1993) This allowsthe model to selectively choose which information is relevant at various stages of themodelling process and it allows for information to be easily updated Davis (1993)reviews the application of expert systems to environmental modelling research conclu-ding that the technique is becoming more widely accepted He further states that theapplication of expert systems is constrained by an absence of fundamental knowledgefor rule generation a problem that would appear less relevant to soil mapping giventhe amount of untapped expert knowledge accumulated by the soil surveyor

The first mention of the use of expert systems in pedology was in a paper presenta-tion at the Northeast Committee Soil Survey Conference 1984 (Flach 1985) In herpaper Flach (1985) hinted that recent developments in computer science especiallyexpert systems and artificial intelligence could make modelling a practical mappingtool for soil scientists in the near future A year later McCracken and Cate (1986) hopedto encourage soil scientists to explore expert systems and its application to soil sciencethrough an optimistic article they wrote on the potential use of expert systemsHowever little research was actually conducted in the late 1980s and expert systemshad not yet begun to fulfil the lofty goals set forth by early practitioners (Dale et al1989)

In the early 1990s expert systems approaches to predictive soil mapping began toappear in the literature Skidmore et al (1991) used a Bayesian expert system to mapforest soil into different classes and their results compared favourably with availablesoil maps and actual field-collected data Their methods successfully incorporatedsurveyor knowledge and remotely sensed and digital terrain attributes but failed tobetter characterize continuous soilndashlandscape variability because their final productwas a choropleth map Skidmore et al (1996) revisited their earlier research andassessed the mapping accuracy of their results They found that the soil map producedby the expert system achieved an overall accuracy of 698 (sample size = 53) while themap derived from conventional methods had an accuracy of 736 In addition toknowledge provided by a soil scientist Skidmore et alrsquos (1996) expert system usedvegetation derived for aerial photography and topographic variables derived from a 10-m DEM (wetness index topographic position and slope)

Using an expert system Cook et al (1996) successfully produced a continuous soilproperty map for organic matter content using wetness index aspect and slope asexplanatory variables Their methods were somewhat inefficient because they requireda separate expert system for each soil property of interest However their researchrepresents the only example of expert systems used to directly predict the spatial dis-tribution of a soil property All of the other examples of expert systems in the literature

P Scull et al 189

190 Predictive soil mapping a review

deal with soil type or class The use of expert systems to map soil properties needs tobe explored further

The expert systems discussed thus far have all used Boolean logic within theirtheoretical framework whereby an observation can belong to one and only one classand the soil properties of that class are assigned to the observation With thedevelopment of fuzzy logic and the semantic import model (the second fuzzy logicapproach to creating continuous classes) the opportunity exists to develop fuzzy logic-based expert systems Such systems can be used in conjunction with expert knowledgein situations where experts have a good qualitative idea of how to group data but havetrouble dealing with observations that are not well represented by rigid classificationschemes (Burrough and McDonnell 1998) This method can be particularly useful insituations when taxonomic schemes have been previously developed as is the case insoil taxonomy Several examples of this type of approach were published in the mid-1990s by A Zhu and colleagues (Zhu and Band 1994 Zhu et al 1996 1997 Zhu1997ab) Such systems proved useful for mapping soil at unvisited locations usingsurveyor knowledge and were also capable of producing continuous soil propertymaps The use of fuzzy logic within the theoretical framework of the expert systemallows the complex nature of soil to propagate through the modelling process neversubjected to classification schemes that filter out potentially useful lsquonoisersquo The use offuzzy logic also gives the soil surveyor more latitude during the interview processwhen the knowledge base is defined for the expert system Zhu et al (1997) developeda fuzzy logic-based expert system called SoLIM to determine the similarity of eachgrid cell in a study area to the various taxonomic mapping units delimited by the soilsurvey Continuous soil attribute maps were calculated using the similarity values andtheir relative soil survey determined attribute values (effectively achieving Goal 2) Theresulting data (which proved more accurate than soil survey data once field checked)consists of a raster grid whose resolution is determined by the resolution of the inputenvironmental and digital elevation data As noted previously this type of data modelis more applicable to environmental modelling than the choropleth map (Burrough andMcDonnell 1998)

Expert system approaches to PSM are capable of exploiting soil surveyor knowledgeby developing rule-based systems that imitate the surveyorrsquos conceptual model of soilvariability (the primary focus of Goal 3) The method would appear extremely usefulfor mapping projects (such as those conducted by the NRCS-NCSS) where fieldwork isinitially conducted to determine soilndashlandscape relations Expert system developmentcould be directly inserted into the traditional soil survey mapping approach as asubstitute for the step where the surveyor converts hisher conceptual model into achoropleth map Rather that knowledge could be incorporated into the expert systemwhich could be used to predictively map soil The resulting raster map would be morescientifically based and explicit than the hand-drawn choropleth maps of the past Itsscale would not be limited to that of the aerial photography but rather to the scale ofthe environmental data

Despite these advantages the expert system approach has some drawbacks Becausethe method is dependent upon expert knowledge it cannot be applied whereenvironmentndashsoil relations are poorly understood Of course this criticism can be madeof all PSM models as inductive or deductive knowledge of soilndashenvironment relationsis a prerequisite for PSM However expert systems are deductive models and as such

are not driven by any specific field-collected data (although presumably the soil experthas field experience in the mapping area) Expert systems donrsquot afford the opportunityto first statistically document landscapendashsoil relations and then extrapolate the resultsbecause expert systems do not directly use sample soil data (lsquohard datarsquo) to determinesoilndashlandscape relations Thus the utilization of the relationship between environmen-tal variables and soil properties (Goal 1) is only indirectly achieved Expert systemsapproaches have been demonstrated to be extremely effective in a small number of casestudies The possibility of satisfying all three goals discussed in this review makesexpert systems a predictive soil mapping method that needs to be further tested ndash bothgeographically and across different scales of analysis

V Conclusion

Most of the predictive soil mapping research outlined in this review was conducted atvery large map scales (over small areas) In fact the majority of the research wasconcerned with assessing the spatial variability of soil character within individual fieldsor across soil toposequences The primary driving force behind this type of research hasbeen the need to provide accurate soils information for agriculture and ecologicalmodels It is clear that terrain attributes are powerful predictors at the local scaleGeostatistical tools have been successful at using terrain attributes and the spatialdependence of soil properties to interpolate between existing data points withinindividual fields Across soil toposequences statistical approaches provide a usefulmeans of predicting soil character PSM research has been most successful at the fieldscale because many of the soil-forming factors are held constant For example the neteffect of four of Jennyrsquos five soil-forming factors (climate organism parent material andtime) was minimal within many of the studies reviewed While some research hastackled larger areas of study there exists a deficiency in our ability to predictively mapsoils at smaller map scales Since the distribution of soil is scale-dependent differentPSM methods and predictors are likely to work better at different scales Focus in thefuture must continue to move toward working over larger spatial extents of study inorder to produce landscape-scale soil information

Further a large proportion of the research was conducted in semi-arid gentlysloping agricultural landscapes Humid forestlands mountainous regions and desertshave received little attention As such PSM methods need to continue to be testedandor developed in a wider variety of landscapes where spatial soil distributions canbe more complex Different methods will likely be successful to different degrees indifferent environments Whereas terrain attributes are dominant predictors of soilcharacter across toposequences in prairie lands other predictors are likely necessary inareas where soils do not develop into clear toposequences For example remote sensingdata has been a useful predictor of soil chronosequences in desert landscapes wheresurface appearance is often related to soil character Surficial geology is often reflectiveof soil character in mountain regions where soils are thin and significant bedrock existsat the surface Focus in the future must continue to determine which methods andpredictors work best in which environments

Since the most useful PSM approach will vary across spatial scales and environmen-tal gradients the method used should be driven by the mapping objectives of the

P Scull et al 191

192 Predictive soil mapping a review

project Whereas the traditional soil survey of the past was expected to meet the needsof a diverse group of end users PSM methods will vary given the objectives of thesurvey As a result the end users of the soil data need to play a more active role in thesurvey process Because precision agriculture and large-scale mapping has been theprimary focus of PSM there exists a deficiency in our ability to predictively map soilsfor the purpose of general land use planning and management Even though lessdetailed soil information is needed for these purposes the development of PSMmethods is complicated by the fact that soils exhibit complex spatial variability atsmall map scales where soil-forming environments vary greatly from one location tothe next

A number of alternative methods of characterizing the continuous nature of the soillandscape have been developed Thus far most of the PSM research has provided soilinformation in a nonobject form (as opposed to defining soil types as independententities) Soil data have been generated organized and presented in the form of eitherisorithmic maps or fine-scale raster grids Both of these data models are field-viewmodels of geographic space which allow the soil to be perceived as a constantlyvarying surface Two distinct approaches have been employed (1) mapping individualsoil properties and (2) mapping continuously varying (fuzzy) soil classes Mappingindividual properties is the most common approach and will likely continue todominate PSM research The use of fuzzy soil classes in the literature is less commonand more difficult to be integrated into standard mapping procedures (such as use ofsoil taxa) because the concept radically differs from the traditional view of the soillandscape However fuzzy soil classes (generated using the fuzzy-k means approach) dohave the potential to help further advance Soil taxonomy by identifying taxonomicclasses that are more reflective of pedologic processes at work

Expert systems have been greatly underutilized in PSM research especiallyconsidering how effective a small number of case studies have been Expert systemshave the potential to satisfy successfully all three goals discussed throughout thereview They also have the potential to bridge the gap between traditional approachesand PSM methods because field soil scientists do not have to change their conceptualapproach to mapping They can still conduct field reconnaissance to determinesoilndashlandscape relationships Afterwards though they can use expert systems to betterexploit the knowledge they have garnered Such an approach could help ease thetransition from traditional soil survey to more scientifically explicit methods Expertsystems are also the most fruitful approach to utilizing a wealth of data that has alreadybeen collected in a nonexplicit manner the wealth of expert knowledge (in the form ofthe conceptual model) that senior soil mappers have accumulated

Focus in the future must also move in the direction of operability The utility of manyPSM approaches has been clearly demonstrated yet changes in how soils are mappedand perceived has been slow In the USA various members of the National CooperativeSoil Survey have called for a change in the philosophy of survey (from mappingdiscrete soil types to mapping continuous varying soil properties McSweeny et al1994) but little change has actually taken place To realize the potential of PSM in con-tributing to an overall change in standard soil mapping procedures practitionersworking in the field of PSM need to form working relationships with field soilscientists PSM methods need also to be presented in a manner that is comprehensibleto the soil science trained field mapper Predictive soil mapping is a relatively recent

phenomenon yet much progress has been made In the process of mapping soilknowledge will continue to advance

P Scull et al 193

References

Agbu PA Fehrenbacher DJ and Jansen IJ1990 Statistical comparison of SPOT spectralmaps with field soil maps Soil Science Society ofAmerica Journal 54 818ndash18

Barrett L 1999 Particulars in contextmaintaining a balance in soil geography Annalsof the Association of America Geographers 89707ndash13

Beckett PHT and Webster R 1971 Soilvariability a review Soils and Fertilizers 341ndash15

Bell JC Grigal DF and Bates PC 2000 A soil-terrain model for estimating spatial patterns ofsoil organic carbon In Wilson J and Gallant Jeditors Terrain analysis principles and applica-tions New York City NY John Wiley and Sons295ndash310

Birkeland PW 1999 Soil and geomorphologyThird edition New York NY Oxford UniversityPress

Breiman L Friedman JH Olshen RA andStone CJ 1984 Classification and regressiontrees Belmont CA Wadsworth

Brule FJ 1996 Fuzzy systems ndash a tutorialhttp newsgroup compai httpwwwquadralaycom (last accessed 30 August 2001)

Bunkin FV and Bunkin AF 2000 Lidarsounding of water soil and plants Atmosphericand Oceanic Optics 13 54ndash72

Buol SW Hole FD McCracken RJ andSouthard RJ 1997 Soil genesis and classifica-tion Ames IA Iowa State University Press

Burgess TM and Webster R 1980a Optimalinterpolation and isarithmic mapping of soilproperties the semi-variogram and punctualkriging Journal of Soil Science 31 315ndash31

ndashndashndashndash 1980b Optimal interpolation and isarithmicmapping of soil properties block kringingJournal of Soil Science 31 331ndash41

Burrough PA 1989 Fuzzy mathematicalmethods for soil survey and land evaluationJournal of Soil Science 40 477ndash92

Burrough PA and McDonnell RA 1998Principles of geographic information systems(Revised edition) Oxford Clarendon Press

Burrough PA Beckett PHT and Jarvis MG1971 The relation between cost and utility insoil survey Journal of Soil Science 22 368ndash81

Burrough PA Van Gaans PMF and

Hootsman R 1997 Continuous classificationin soil survey spatial correlation confusionand boundaries Geoderma 77 115ndash35

Cambell JB 1977 Variation of selectedproperties across a soil boundary Soil ScienceSociety of America Journal 41 578ndash82

Cambell JB and Edmonds WJ 1984 Themissing geographic dimension to soiltaxonomy Annals of the Association of AmericanGeographers 74 83ndash97

Castrignano A Giugliarini L Risaliti R andMartinelli N 2000 Study of spatial relation-ships among some soil physico-chemicalproperties of a field in central Italy using multi-variate geostatistics Geoderma 97 39ndash60

Cialella AT Dubayah R Lawrence W andLevine E 1997 Predicting soil drainage classusing remotely sensed and digital elevationdata Journal of Soil Science 62(2) 171ndash78

Clark RN and Swayze GA 1996 Evolution inimaging spectroscopy analysis and sensorsignal-to-noise an examination of how far wehave come Summaries of the sixth annual JPLairborne Earth science workshop 4ndash8 March 1996AVIRIS Workshop Vol 1 5

Clarke LA and Pregibon D 1992 Tree-basedmodels In Chambers J and Hastie J editorsStatistical models in S Pacific Grove Wadsworthand Brooks 377ndash419

Coleman TL Agbu PA and MontgomeryOL 1993 Spectral differentiation of surfacesoils and soil properties ndash is it possible fromspace platforms Soil Science 155 283ndash93

Cook SE Corner RJ Grealish G GesslerPE and Chartres CJ 1996 A rule-basedsystem to map soil properties Soil ScienceSociety of America Journal 60 1893ndash900

Cruickshank JG 1972 Soil geography New YorkNY John Wiley amp Sons

Csillag F Pasztor L Biehl LL 1993 Spectralband selection for the characterization ofsalinity status of soils Remote Sensing ofEnvironment 43 231ndash42

Dale MB McBratney AB and Russell JS1989 On the role of expert systems andnumerical taxonomy in soil classificationJournal of Soil Science 40 223ndash34

Davis JR 1993 Expert systems and environ-

P Scull et al 185

could be used in arid regions where soil toposequences are less well expressed Atlarger scales of prediction selection of different sets of ancillary variables is requiredbecause different processes define soil character at different scales The most obviousexample is that of climate which may control soil distribution at large scales(continental) but has little explanatory power at the field level Regardless of whetheradequate ancillary data exist the amount of data required to use geostatistics forlandscape-level prediction would be extremely difficult and costly to collect given thestrict sampling protocol required to characterize spatial dependence It is also unclearat what landscape scale soils exhibit spatial autocorrelation

Geostatistical approaches do provide a means of creating continuous soil attributesurfaces to better represent soil continuity (Goal 2) and they can be used to exploit therelationship between environmental variables and soil properties in order to moreefficiently collect soil data (Goal 1) However they do not sufficiently utilize expertknowledge (Goal 3) as no attempt has been made in geostatistical approaches todirectly integrate expert knowledge Fundamentally kriging is a process of interpola-tion designed to predict attribute values in between locations of measured samples Inthis sense geostatistics represent a middle ground between pure interpolation (egnearest neighbour type classifier) in which only measured points for the variable ofinterest are used to determine unknown values and other predictive models thatprimarily use soilndashenvironment correlation to create predictive maps

2 Statistical methods

Statistical methods can be used to exploit the relationship between quantifiablelandscape indices and soil properties to create predictive soil maps For exampleMcKensie and Austin (1993) used a regression to account for a large percentage ofvariation for many soil characteristics (A horizon clay content CEC EC pH bulkdensity and COLE B horizon clay content CEC ESP EC pH bulk density and COLE)using a variety of predictor variables (slope presence or absence of impeding layerrelief landform topographic position) Their results confirm the hypothesis of MooreID et al (1991) that soil character is related to quantifiable landscape indices Howevertheir methods do not provide inference of soil properties at unmeasured sites frommapped environment data because the topographic variables were measured in thefield Linear regression has also been used with terrain variables derived from a 15-mDEM in northeastern Colorado to predict soil attributes (organic matter contentextractable phosphorous pH and texture) at unvisited sites (Moore et al 1993) In thatparticular study 50 of the variance of A-horizon thickness was explained by slope andthe wetness index Gessler et al (1995) also used regression to model A-horizonthickness from topographic variables in southeastern Australia (plan curvature andwetness index r2 = 063 P = 0001) They modeled solum depth and used logisticregression to model E horizon presenceabsence Elsewhere logistic regression hasbeen used to model the presenceabsence of noncalcareous clay loam horizon in centralFrance using terrain attributes from a 20-m DEM (King et al 1999) Exponentialregression has been used to model soil organic carbon using terrain variables (Bell et al2000) in glacial outwash soils in east-central Minnesota Generalized additive models(GAM) have been used less frequently in PSM research Gessler (1996) used a GAM

186 Predictive soil mapping a review

model to predict total soil carbon A horizon depth and solum depth using a variety ofenvironmental predictors (slope elevation wetness index mean annual temperatureprecipitation and radiation)

This small body of research opened the door to more complex methods by demon-strating the existence of quantifiable relationships These authors were able to producesoil attribute maps using raster data models whose scale was dependent upon the gridcell resolution of the environmental data They were successful at exploiting the rela-tionship between quantifiable topographic attributes and soil profile character (Goal 1)The continuous soil attribute surfaces they produce also better represent soil continuitythan the choropleth soil maps produced by traditional soil survey (Goal 2) Howeverthe bulk of these methods (excepting GAMs) are limited by their assumed linear rela-tionship between soil and topographic attributes their assumptions of normallydistributed data and their high data requirements Standard statistical procedures arealso not flexible enough to allow robust integration with a variety of potential datasources such as expert knowledge (Goal 1)

Statistical methods do demonstrate in a quantitative manner that terrain analysis canbe used to predict soil attributes in relatively small areas with homogeneous parentmaterial A large proportion of the research using statistical methods was conducted insemi-arid landscapes at small scales (the largest study area of the entire group was~2000 ha) Obviously for statistical approaches to be most effective they need to bemore universal As such they need to be tested andor developed at larger scales andin more diverse landscapes

3 Decision tree analysis (DTA)

The use of decision tree analysis is just beginning to be explored in predictive soilmapping research although it has been used successfully in the related field ofpredictive vegetation mapping since the early 1990s (Lees and Ritman 1991 MooreDM et al 1991 Franklin 1998) DTA is a form of divisive classification The process oftree modelling involves successively partitioning data (called recursive partitioning inthe tree modelling literature) into increasingly homogeneous subsets which once thepartitioning has ceased are called terminal nodes (Lees and Ritman 1991) Splits orrules defining how to partition the data are selected based on information statistics thatdefine how well the split decreases impurity within the data set (Clark and Pregibon1992) Splits are based on threshold values of an explanatory variable selected bycomparing the increase in resulting purity of node membership for all possiblethresholds and variables The process is iterative growing from the root node (thecomplete data set) to the terminal nodes in a dendritic fashion (Friedl and Brodley1997) Once the tree has been constructed (or grown) it encodes a set of decision rulesthat describe the data partitioning process These rules can be used to classify or predictother data sets (Moore DM et al 1991) Pruning the tree is often necessary to preventthe tree from being overfit to the sample data and to reduce tree complexity Pruningentails combining pairs of terminal nodes into single nodes and can be accomplishedusing cross-validation which yields an initial indication of how large a tree makesrobust predictions (Safavian and Norvig 1991) Cross-validation involves systematical-ly removing portions of the data set and running the remaining sample through the tree

in an iterative manner eventually yielding estimates of the misclassification rates foreach class each node and the whole model (Breiman et al 1984) In this mannerdifferent sized trees can be compared in terms of parsimony

The term DTA is used to collectively refer to all types of tree-based modelling (theword lsquodecisionrsquo is used because it is descriptive indicating that the analysis eventuallyleads to a set of decision rules defining data partitions) The term should be distin-guished from classification tree analysis because the latter refers specifically to DTAwhere the response variable is categorical The term CART (classification andregression trees ndash Breiman et al 1984) is sometimes used but strictly speaking refers tospecific software Friedl and Brodley (1997) provide a review of the decision treealgorithms They divide types of DTA into two classes (1) homogeneous decision treesfor which a single algorithm is used to estimate each split (eg CART) and (2) hybriddecision trees (HDT) for which different splitting methods can be used at differentpoints in the tree (eg Quinlan 1993) They further divide homogeneous decision treesinto univariate (UDT) where single features of the input data define splits and multi-variate decision trees (MDT) where multiple features of the input data can define splitsAccording to this naming convention no distinction is made between methods withdifferent types of response variables although all methods can be used with bothcategorical and continuous response variables In a comparison of these various typesof algorithms on a variety of data sets Friedl and Brodley (1997) found that HDT hadthe highest classification accuracy

The overall aim of DTA is to design a set of predictive rules (eg if geology type Athen soil type B) developed from training data which can then be applied to ageographic data base to predict the value of a response variable (Michaelsen et al 1994)Therefore DTA explicitly uses soilndashlandscape correlation in model development (Goal2) The technique appears promising in soil research but needs to be further tested asnot many of the above types of DTA have been tested in the PSM literature in fact onlyunivariate approaches have been employed For example Lagacherie and Holmes(1997) successfully used univariate DTA to model a categorical response variable soiltype within a training set and then assuming that the training set was representativeextrapolated the model to a much larger region Their work is interesting becausealthough not a single sample came from the area they eventually mapped they wereable to produce a soil map that was more accurate (74 versus 69) than the existingmap produced from traditional methods Cialella et al (1997) also used univariate DTAto predict soil drainage class from a variety of terrain attributes and remotely senseddata They predicted soil drainage class with an average of 78 accuracy ndash impressivegiven that the variation accounted for by a typical soil survey ranges from about halfthe total variance for physical attributes to less than one-tenth for some soil chemicalattributes (Gessler et al 1995)

DTA has been compared with other approaches by several authors In the applicationof erosion modelling DTA results were similar to Artificial Neural Networks (Ellis1996) Both methods achieved high training accuracy (as measured by the KappaStatistic) but in terms of prediction accuracy both methods performed poorly Gessleret al (1995) compared DTA with generalized linear models (GLM) and generalizedadditive models (GAM) to predict A-horizon thickness and concluded that GLM waspreferable to both DTA and GAM McKensie and Ryan (1999) compared regressiontrees and standard linear regression to predict soil properties (total solum depth soil

P Scull et al 187

188 Predictive soil mapping a review

phophorus and soil carbon) using a large variety of predictors (elevation slopecurvature wetness index climate variables geology unit Landsat TM data andGamma radiometric data) and found that different methods work best for differentproperties and overall success hinged on the strength of the relationship between soiland environmental variables Their study of a 50 000 ha forested landscape in southernNew South Wales Australia is one of the few studies at such a small map scale (largearea) or in forested landscape The most extensive comparison of techniques involvingDTA was conducted by McBratney et al (2000) They compared regression techniques(GLM GAM DTA-regression tree) geostatistical techniques (kriging and heterotopicco-kriging) and a hybrid technique (regression kriging) DTA was found to be thepoorest performing of the regression techniques because of the unrealistic predictionsurface generated by the DTA model DTA has also been criticized by other authorsbecause of the stepped prediction surface (Gessler 1996) This phenomenon can beespecially apparent in situations where predictor variables have different resolutionsThe extent to which DTA yields a better spatial representation of soil continuity is afunction of the scale and type of predictor variables used and therefore varies withindividual models

DTA is gaining widespread popularity as a means to develop prediction rules thatcan be rapidly and repeatedly evaluated (Cialella et al 1997 Franklin et al 2000) DTAprovides the following advantages over standard statistical techniques (1) it is easier tointerpret when explanatory variables are both nominal and continuous (2) it isinvariant to monotone re-expressions (transformations) of predictor variables (3) itdeals more satisfactorily with missing data values and outliers (4) it is more adept atcapturing nonadditive and nonlinear behaviour (5) it doesnrsquot make any assumptionsabout data distribution and (6) it is easily updateable as more data are collected(Moore DM et al 1991) The DTA model framework is especially appealing becauseof its capability to integrate a wide range of data sets as explanatory variables

DTA offers a unique opportunity for interaction between soil experts and soilmodellers because the output of the model is a set of rules that can be pedologicallyinterpreted by the soil expert In this sense expert knowledge is used in an implicitmanner in DTA (somewhat effectively achieving Goal 3) While these rules can often beexceedingly complex at minimum the expert can decide whether initial splits makesense given their understanding of the landscape The success of DTA results oftenhinges on the ability of the modeller to make key decisions during the model buildingprocess there is unfortunately no definitive way to determine the most optimal treeThe aid of the soil expert can potentially elucidate this problem

4 Expert systems

A variety of expert system approaches to PSM have been developed to utilize expertknowledge The purpose of such methods is to exploit the information the soil surveyoraccumulates while working in the field by integrating such knowledge into thepredictive model (McCracken and Cate 1986) Unlike the majority of the researchreviewed thus far the dependent variable in many expert systems models is often soiltaxa or mapping unit This apparent disadvantage of expert systems (using classifica-tion to characterize soil continuity) does make them easier to integrate into traditional

soil survey In addition several authors have developed methods to developcontinuous soil property maps from the output of expert system models designed topredict soil unit occurrence

Expert systems are composed of data (information on spatial environmentalvariables eg topography climate etc) a knowledge base (rules and facts related tosoil variation supplied from the soil surveyor) and an inference engine (whichcombines data and the knowledge base to infer logically valid conclusions) (Skidmoreet al 1996) Expert systems differ from conventional models in two ways (1) they storeand manipulate qualitative information (allowing them access to information thatcannot normally be used in other modelling frameworks) and (2) they are structuredas meta-models (the knowledge is separated from the model) (Davis 1993) This allowsthe model to selectively choose which information is relevant at various stages of themodelling process and it allows for information to be easily updated Davis (1993)reviews the application of expert systems to environmental modelling research conclu-ding that the technique is becoming more widely accepted He further states that theapplication of expert systems is constrained by an absence of fundamental knowledgefor rule generation a problem that would appear less relevant to soil mapping giventhe amount of untapped expert knowledge accumulated by the soil surveyor

The first mention of the use of expert systems in pedology was in a paper presenta-tion at the Northeast Committee Soil Survey Conference 1984 (Flach 1985) In herpaper Flach (1985) hinted that recent developments in computer science especiallyexpert systems and artificial intelligence could make modelling a practical mappingtool for soil scientists in the near future A year later McCracken and Cate (1986) hopedto encourage soil scientists to explore expert systems and its application to soil sciencethrough an optimistic article they wrote on the potential use of expert systemsHowever little research was actually conducted in the late 1980s and expert systemshad not yet begun to fulfil the lofty goals set forth by early practitioners (Dale et al1989)

In the early 1990s expert systems approaches to predictive soil mapping began toappear in the literature Skidmore et al (1991) used a Bayesian expert system to mapforest soil into different classes and their results compared favourably with availablesoil maps and actual field-collected data Their methods successfully incorporatedsurveyor knowledge and remotely sensed and digital terrain attributes but failed tobetter characterize continuous soilndashlandscape variability because their final productwas a choropleth map Skidmore et al (1996) revisited their earlier research andassessed the mapping accuracy of their results They found that the soil map producedby the expert system achieved an overall accuracy of 698 (sample size = 53) while themap derived from conventional methods had an accuracy of 736 In addition toknowledge provided by a soil scientist Skidmore et alrsquos (1996) expert system usedvegetation derived for aerial photography and topographic variables derived from a 10-m DEM (wetness index topographic position and slope)

Using an expert system Cook et al (1996) successfully produced a continuous soilproperty map for organic matter content using wetness index aspect and slope asexplanatory variables Their methods were somewhat inefficient because they requireda separate expert system for each soil property of interest However their researchrepresents the only example of expert systems used to directly predict the spatial dis-tribution of a soil property All of the other examples of expert systems in the literature

P Scull et al 189

190 Predictive soil mapping a review

deal with soil type or class The use of expert systems to map soil properties needs tobe explored further

The expert systems discussed thus far have all used Boolean logic within theirtheoretical framework whereby an observation can belong to one and only one classand the soil properties of that class are assigned to the observation With thedevelopment of fuzzy logic and the semantic import model (the second fuzzy logicapproach to creating continuous classes) the opportunity exists to develop fuzzy logic-based expert systems Such systems can be used in conjunction with expert knowledgein situations where experts have a good qualitative idea of how to group data but havetrouble dealing with observations that are not well represented by rigid classificationschemes (Burrough and McDonnell 1998) This method can be particularly useful insituations when taxonomic schemes have been previously developed as is the case insoil taxonomy Several examples of this type of approach were published in the mid-1990s by A Zhu and colleagues (Zhu and Band 1994 Zhu et al 1996 1997 Zhu1997ab) Such systems proved useful for mapping soil at unvisited locations usingsurveyor knowledge and were also capable of producing continuous soil propertymaps The use of fuzzy logic within the theoretical framework of the expert systemallows the complex nature of soil to propagate through the modelling process neversubjected to classification schemes that filter out potentially useful lsquonoisersquo The use offuzzy logic also gives the soil surveyor more latitude during the interview processwhen the knowledge base is defined for the expert system Zhu et al (1997) developeda fuzzy logic-based expert system called SoLIM to determine the similarity of eachgrid cell in a study area to the various taxonomic mapping units delimited by the soilsurvey Continuous soil attribute maps were calculated using the similarity values andtheir relative soil survey determined attribute values (effectively achieving Goal 2) Theresulting data (which proved more accurate than soil survey data once field checked)consists of a raster grid whose resolution is determined by the resolution of the inputenvironmental and digital elevation data As noted previously this type of data modelis more applicable to environmental modelling than the choropleth map (Burrough andMcDonnell 1998)

Expert system approaches to PSM are capable of exploiting soil surveyor knowledgeby developing rule-based systems that imitate the surveyorrsquos conceptual model of soilvariability (the primary focus of Goal 3) The method would appear extremely usefulfor mapping projects (such as those conducted by the NRCS-NCSS) where fieldwork isinitially conducted to determine soilndashlandscape relations Expert system developmentcould be directly inserted into the traditional soil survey mapping approach as asubstitute for the step where the surveyor converts hisher conceptual model into achoropleth map Rather that knowledge could be incorporated into the expert systemwhich could be used to predictively map soil The resulting raster map would be morescientifically based and explicit than the hand-drawn choropleth maps of the past Itsscale would not be limited to that of the aerial photography but rather to the scale ofthe environmental data

Despite these advantages the expert system approach has some drawbacks Becausethe method is dependent upon expert knowledge it cannot be applied whereenvironmentndashsoil relations are poorly understood Of course this criticism can be madeof all PSM models as inductive or deductive knowledge of soilndashenvironment relationsis a prerequisite for PSM However expert systems are deductive models and as such

are not driven by any specific field-collected data (although presumably the soil experthas field experience in the mapping area) Expert systems donrsquot afford the opportunityto first statistically document landscapendashsoil relations and then extrapolate the resultsbecause expert systems do not directly use sample soil data (lsquohard datarsquo) to determinesoilndashlandscape relations Thus the utilization of the relationship between environmen-tal variables and soil properties (Goal 1) is only indirectly achieved Expert systemsapproaches have been demonstrated to be extremely effective in a small number of casestudies The possibility of satisfying all three goals discussed in this review makesexpert systems a predictive soil mapping method that needs to be further tested ndash bothgeographically and across different scales of analysis

V Conclusion

Most of the predictive soil mapping research outlined in this review was conducted atvery large map scales (over small areas) In fact the majority of the research wasconcerned with assessing the spatial variability of soil character within individual fieldsor across soil toposequences The primary driving force behind this type of research hasbeen the need to provide accurate soils information for agriculture and ecologicalmodels It is clear that terrain attributes are powerful predictors at the local scaleGeostatistical tools have been successful at using terrain attributes and the spatialdependence of soil properties to interpolate between existing data points withinindividual fields Across soil toposequences statistical approaches provide a usefulmeans of predicting soil character PSM research has been most successful at the fieldscale because many of the soil-forming factors are held constant For example the neteffect of four of Jennyrsquos five soil-forming factors (climate organism parent material andtime) was minimal within many of the studies reviewed While some research hastackled larger areas of study there exists a deficiency in our ability to predictively mapsoils at smaller map scales Since the distribution of soil is scale-dependent differentPSM methods and predictors are likely to work better at different scales Focus in thefuture must continue to move toward working over larger spatial extents of study inorder to produce landscape-scale soil information

Further a large proportion of the research was conducted in semi-arid gentlysloping agricultural landscapes Humid forestlands mountainous regions and desertshave received little attention As such PSM methods need to continue to be testedandor developed in a wider variety of landscapes where spatial soil distributions canbe more complex Different methods will likely be successful to different degrees indifferent environments Whereas terrain attributes are dominant predictors of soilcharacter across toposequences in prairie lands other predictors are likely necessary inareas where soils do not develop into clear toposequences For example remote sensingdata has been a useful predictor of soil chronosequences in desert landscapes wheresurface appearance is often related to soil character Surficial geology is often reflectiveof soil character in mountain regions where soils are thin and significant bedrock existsat the surface Focus in the future must continue to determine which methods andpredictors work best in which environments

Since the most useful PSM approach will vary across spatial scales and environmen-tal gradients the method used should be driven by the mapping objectives of the

P Scull et al 191

192 Predictive soil mapping a review

project Whereas the traditional soil survey of the past was expected to meet the needsof a diverse group of end users PSM methods will vary given the objectives of thesurvey As a result the end users of the soil data need to play a more active role in thesurvey process Because precision agriculture and large-scale mapping has been theprimary focus of PSM there exists a deficiency in our ability to predictively map soilsfor the purpose of general land use planning and management Even though lessdetailed soil information is needed for these purposes the development of PSMmethods is complicated by the fact that soils exhibit complex spatial variability atsmall map scales where soil-forming environments vary greatly from one location tothe next

A number of alternative methods of characterizing the continuous nature of the soillandscape have been developed Thus far most of the PSM research has provided soilinformation in a nonobject form (as opposed to defining soil types as independententities) Soil data have been generated organized and presented in the form of eitherisorithmic maps or fine-scale raster grids Both of these data models are field-viewmodels of geographic space which allow the soil to be perceived as a constantlyvarying surface Two distinct approaches have been employed (1) mapping individualsoil properties and (2) mapping continuously varying (fuzzy) soil classes Mappingindividual properties is the most common approach and will likely continue todominate PSM research The use of fuzzy soil classes in the literature is less commonand more difficult to be integrated into standard mapping procedures (such as use ofsoil taxa) because the concept radically differs from the traditional view of the soillandscape However fuzzy soil classes (generated using the fuzzy-k means approach) dohave the potential to help further advance Soil taxonomy by identifying taxonomicclasses that are more reflective of pedologic processes at work

Expert systems have been greatly underutilized in PSM research especiallyconsidering how effective a small number of case studies have been Expert systemshave the potential to satisfy successfully all three goals discussed throughout thereview They also have the potential to bridge the gap between traditional approachesand PSM methods because field soil scientists do not have to change their conceptualapproach to mapping They can still conduct field reconnaissance to determinesoilndashlandscape relationships Afterwards though they can use expert systems to betterexploit the knowledge they have garnered Such an approach could help ease thetransition from traditional soil survey to more scientifically explicit methods Expertsystems are also the most fruitful approach to utilizing a wealth of data that has alreadybeen collected in a nonexplicit manner the wealth of expert knowledge (in the form ofthe conceptual model) that senior soil mappers have accumulated

Focus in the future must also move in the direction of operability The utility of manyPSM approaches has been clearly demonstrated yet changes in how soils are mappedand perceived has been slow In the USA various members of the National CooperativeSoil Survey have called for a change in the philosophy of survey (from mappingdiscrete soil types to mapping continuous varying soil properties McSweeny et al1994) but little change has actually taken place To realize the potential of PSM in con-tributing to an overall change in standard soil mapping procedures practitionersworking in the field of PSM need to form working relationships with field soilscientists PSM methods need also to be presented in a manner that is comprehensibleto the soil science trained field mapper Predictive soil mapping is a relatively recent

phenomenon yet much progress has been made In the process of mapping soilknowledge will continue to advance

P Scull et al 193

References

Agbu PA Fehrenbacher DJ and Jansen IJ1990 Statistical comparison of SPOT spectralmaps with field soil maps Soil Science Society ofAmerica Journal 54 818ndash18

Barrett L 1999 Particulars in contextmaintaining a balance in soil geography Annalsof the Association of America Geographers 89707ndash13

Beckett PHT and Webster R 1971 Soilvariability a review Soils and Fertilizers 341ndash15

Bell JC Grigal DF and Bates PC 2000 A soil-terrain model for estimating spatial patterns ofsoil organic carbon In Wilson J and Gallant Jeditors Terrain analysis principles and applica-tions New York City NY John Wiley and Sons295ndash310

Birkeland PW 1999 Soil and geomorphologyThird edition New York NY Oxford UniversityPress

Breiman L Friedman JH Olshen RA andStone CJ 1984 Classification and regressiontrees Belmont CA Wadsworth

Brule FJ 1996 Fuzzy systems ndash a tutorialhttp newsgroup compai httpwwwquadralaycom (last accessed 30 August 2001)

Bunkin FV and Bunkin AF 2000 Lidarsounding of water soil and plants Atmosphericand Oceanic Optics 13 54ndash72

Buol SW Hole FD McCracken RJ andSouthard RJ 1997 Soil genesis and classifica-tion Ames IA Iowa State University Press

Burgess TM and Webster R 1980a Optimalinterpolation and isarithmic mapping of soilproperties the semi-variogram and punctualkriging Journal of Soil Science 31 315ndash31

ndashndashndashndash 1980b Optimal interpolation and isarithmicmapping of soil properties block kringingJournal of Soil Science 31 331ndash41

Burrough PA 1989 Fuzzy mathematicalmethods for soil survey and land evaluationJournal of Soil Science 40 477ndash92

Burrough PA and McDonnell RA 1998Principles of geographic information systems(Revised edition) Oxford Clarendon Press

Burrough PA Beckett PHT and Jarvis MG1971 The relation between cost and utility insoil survey Journal of Soil Science 22 368ndash81

Burrough PA Van Gaans PMF and

Hootsman R 1997 Continuous classificationin soil survey spatial correlation confusionand boundaries Geoderma 77 115ndash35

Cambell JB 1977 Variation of selectedproperties across a soil boundary Soil ScienceSociety of America Journal 41 578ndash82

Cambell JB and Edmonds WJ 1984 Themissing geographic dimension to soiltaxonomy Annals of the Association of AmericanGeographers 74 83ndash97

Castrignano A Giugliarini L Risaliti R andMartinelli N 2000 Study of spatial relation-ships among some soil physico-chemicalproperties of a field in central Italy using multi-variate geostatistics Geoderma 97 39ndash60

Cialella AT Dubayah R Lawrence W andLevine E 1997 Predicting soil drainage classusing remotely sensed and digital elevationdata Journal of Soil Science 62(2) 171ndash78

Clark RN and Swayze GA 1996 Evolution inimaging spectroscopy analysis and sensorsignal-to-noise an examination of how far wehave come Summaries of the sixth annual JPLairborne Earth science workshop 4ndash8 March 1996AVIRIS Workshop Vol 1 5

Clarke LA and Pregibon D 1992 Tree-basedmodels In Chambers J and Hastie J editorsStatistical models in S Pacific Grove Wadsworthand Brooks 377ndash419

Coleman TL Agbu PA and MontgomeryOL 1993 Spectral differentiation of surfacesoils and soil properties ndash is it possible fromspace platforms Soil Science 155 283ndash93

Cook SE Corner RJ Grealish G GesslerPE and Chartres CJ 1996 A rule-basedsystem to map soil properties Soil ScienceSociety of America Journal 60 1893ndash900

Cruickshank JG 1972 Soil geography New YorkNY John Wiley amp Sons

Csillag F Pasztor L Biehl LL 1993 Spectralband selection for the characterization ofsalinity status of soils Remote Sensing ofEnvironment 43 231ndash42

Dale MB McBratney AB and Russell JS1989 On the role of expert systems andnumerical taxonomy in soil classificationJournal of Soil Science 40 223ndash34

Davis JR 1993 Expert systems and environ-

186 Predictive soil mapping a review

model to predict total soil carbon A horizon depth and solum depth using a variety ofenvironmental predictors (slope elevation wetness index mean annual temperatureprecipitation and radiation)

This small body of research opened the door to more complex methods by demon-strating the existence of quantifiable relationships These authors were able to producesoil attribute maps using raster data models whose scale was dependent upon the gridcell resolution of the environmental data They were successful at exploiting the rela-tionship between quantifiable topographic attributes and soil profile character (Goal 1)The continuous soil attribute surfaces they produce also better represent soil continuitythan the choropleth soil maps produced by traditional soil survey (Goal 2) Howeverthe bulk of these methods (excepting GAMs) are limited by their assumed linear rela-tionship between soil and topographic attributes their assumptions of normallydistributed data and their high data requirements Standard statistical procedures arealso not flexible enough to allow robust integration with a variety of potential datasources such as expert knowledge (Goal 1)

Statistical methods do demonstrate in a quantitative manner that terrain analysis canbe used to predict soil attributes in relatively small areas with homogeneous parentmaterial A large proportion of the research using statistical methods was conducted insemi-arid landscapes at small scales (the largest study area of the entire group was~2000 ha) Obviously for statistical approaches to be most effective they need to bemore universal As such they need to be tested andor developed at larger scales andin more diverse landscapes

3 Decision tree analysis (DTA)

The use of decision tree analysis is just beginning to be explored in predictive soilmapping research although it has been used successfully in the related field ofpredictive vegetation mapping since the early 1990s (Lees and Ritman 1991 MooreDM et al 1991 Franklin 1998) DTA is a form of divisive classification The process oftree modelling involves successively partitioning data (called recursive partitioning inthe tree modelling literature) into increasingly homogeneous subsets which once thepartitioning has ceased are called terminal nodes (Lees and Ritman 1991) Splits orrules defining how to partition the data are selected based on information statistics thatdefine how well the split decreases impurity within the data set (Clark and Pregibon1992) Splits are based on threshold values of an explanatory variable selected bycomparing the increase in resulting purity of node membership for all possiblethresholds and variables The process is iterative growing from the root node (thecomplete data set) to the terminal nodes in a dendritic fashion (Friedl and Brodley1997) Once the tree has been constructed (or grown) it encodes a set of decision rulesthat describe the data partitioning process These rules can be used to classify or predictother data sets (Moore DM et al 1991) Pruning the tree is often necessary to preventthe tree from being overfit to the sample data and to reduce tree complexity Pruningentails combining pairs of terminal nodes into single nodes and can be accomplishedusing cross-validation which yields an initial indication of how large a tree makesrobust predictions (Safavian and Norvig 1991) Cross-validation involves systematical-ly removing portions of the data set and running the remaining sample through the tree

in an iterative manner eventually yielding estimates of the misclassification rates foreach class each node and the whole model (Breiman et al 1984) In this mannerdifferent sized trees can be compared in terms of parsimony

The term DTA is used to collectively refer to all types of tree-based modelling (theword lsquodecisionrsquo is used because it is descriptive indicating that the analysis eventuallyleads to a set of decision rules defining data partitions) The term should be distin-guished from classification tree analysis because the latter refers specifically to DTAwhere the response variable is categorical The term CART (classification andregression trees ndash Breiman et al 1984) is sometimes used but strictly speaking refers tospecific software Friedl and Brodley (1997) provide a review of the decision treealgorithms They divide types of DTA into two classes (1) homogeneous decision treesfor which a single algorithm is used to estimate each split (eg CART) and (2) hybriddecision trees (HDT) for which different splitting methods can be used at differentpoints in the tree (eg Quinlan 1993) They further divide homogeneous decision treesinto univariate (UDT) where single features of the input data define splits and multi-variate decision trees (MDT) where multiple features of the input data can define splitsAccording to this naming convention no distinction is made between methods withdifferent types of response variables although all methods can be used with bothcategorical and continuous response variables In a comparison of these various typesof algorithms on a variety of data sets Friedl and Brodley (1997) found that HDT hadthe highest classification accuracy

The overall aim of DTA is to design a set of predictive rules (eg if geology type Athen soil type B) developed from training data which can then be applied to ageographic data base to predict the value of a response variable (Michaelsen et al 1994)Therefore DTA explicitly uses soilndashlandscape correlation in model development (Goal2) The technique appears promising in soil research but needs to be further tested asnot many of the above types of DTA have been tested in the PSM literature in fact onlyunivariate approaches have been employed For example Lagacherie and Holmes(1997) successfully used univariate DTA to model a categorical response variable soiltype within a training set and then assuming that the training set was representativeextrapolated the model to a much larger region Their work is interesting becausealthough not a single sample came from the area they eventually mapped they wereable to produce a soil map that was more accurate (74 versus 69) than the existingmap produced from traditional methods Cialella et al (1997) also used univariate DTAto predict soil drainage class from a variety of terrain attributes and remotely senseddata They predicted soil drainage class with an average of 78 accuracy ndash impressivegiven that the variation accounted for by a typical soil survey ranges from about halfthe total variance for physical attributes to less than one-tenth for some soil chemicalattributes (Gessler et al 1995)

DTA has been compared with other approaches by several authors In the applicationof erosion modelling DTA results were similar to Artificial Neural Networks (Ellis1996) Both methods achieved high training accuracy (as measured by the KappaStatistic) but in terms of prediction accuracy both methods performed poorly Gessleret al (1995) compared DTA with generalized linear models (GLM) and generalizedadditive models (GAM) to predict A-horizon thickness and concluded that GLM waspreferable to both DTA and GAM McKensie and Ryan (1999) compared regressiontrees and standard linear regression to predict soil properties (total solum depth soil

P Scull et al 187

188 Predictive soil mapping a review

phophorus and soil carbon) using a large variety of predictors (elevation slopecurvature wetness index climate variables geology unit Landsat TM data andGamma radiometric data) and found that different methods work best for differentproperties and overall success hinged on the strength of the relationship between soiland environmental variables Their study of a 50 000 ha forested landscape in southernNew South Wales Australia is one of the few studies at such a small map scale (largearea) or in forested landscape The most extensive comparison of techniques involvingDTA was conducted by McBratney et al (2000) They compared regression techniques(GLM GAM DTA-regression tree) geostatistical techniques (kriging and heterotopicco-kriging) and a hybrid technique (regression kriging) DTA was found to be thepoorest performing of the regression techniques because of the unrealistic predictionsurface generated by the DTA model DTA has also been criticized by other authorsbecause of the stepped prediction surface (Gessler 1996) This phenomenon can beespecially apparent in situations where predictor variables have different resolutionsThe extent to which DTA yields a better spatial representation of soil continuity is afunction of the scale and type of predictor variables used and therefore varies withindividual models

DTA is gaining widespread popularity as a means to develop prediction rules thatcan be rapidly and repeatedly evaluated (Cialella et al 1997 Franklin et al 2000) DTAprovides the following advantages over standard statistical techniques (1) it is easier tointerpret when explanatory variables are both nominal and continuous (2) it isinvariant to monotone re-expressions (transformations) of predictor variables (3) itdeals more satisfactorily with missing data values and outliers (4) it is more adept atcapturing nonadditive and nonlinear behaviour (5) it doesnrsquot make any assumptionsabout data distribution and (6) it is easily updateable as more data are collected(Moore DM et al 1991) The DTA model framework is especially appealing becauseof its capability to integrate a wide range of data sets as explanatory variables

DTA offers a unique opportunity for interaction between soil experts and soilmodellers because the output of the model is a set of rules that can be pedologicallyinterpreted by the soil expert In this sense expert knowledge is used in an implicitmanner in DTA (somewhat effectively achieving Goal 3) While these rules can often beexceedingly complex at minimum the expert can decide whether initial splits makesense given their understanding of the landscape The success of DTA results oftenhinges on the ability of the modeller to make key decisions during the model buildingprocess there is unfortunately no definitive way to determine the most optimal treeThe aid of the soil expert can potentially elucidate this problem

4 Expert systems

A variety of expert system approaches to PSM have been developed to utilize expertknowledge The purpose of such methods is to exploit the information the soil surveyoraccumulates while working in the field by integrating such knowledge into thepredictive model (McCracken and Cate 1986) Unlike the majority of the researchreviewed thus far the dependent variable in many expert systems models is often soiltaxa or mapping unit This apparent disadvantage of expert systems (using classifica-tion to characterize soil continuity) does make them easier to integrate into traditional

soil survey In addition several authors have developed methods to developcontinuous soil property maps from the output of expert system models designed topredict soil unit occurrence

Expert systems are composed of data (information on spatial environmentalvariables eg topography climate etc) a knowledge base (rules and facts related tosoil variation supplied from the soil surveyor) and an inference engine (whichcombines data and the knowledge base to infer logically valid conclusions) (Skidmoreet al 1996) Expert systems differ from conventional models in two ways (1) they storeand manipulate qualitative information (allowing them access to information thatcannot normally be used in other modelling frameworks) and (2) they are structuredas meta-models (the knowledge is separated from the model) (Davis 1993) This allowsthe model to selectively choose which information is relevant at various stages of themodelling process and it allows for information to be easily updated Davis (1993)reviews the application of expert systems to environmental modelling research conclu-ding that the technique is becoming more widely accepted He further states that theapplication of expert systems is constrained by an absence of fundamental knowledgefor rule generation a problem that would appear less relevant to soil mapping giventhe amount of untapped expert knowledge accumulated by the soil surveyor

The first mention of the use of expert systems in pedology was in a paper presenta-tion at the Northeast Committee Soil Survey Conference 1984 (Flach 1985) In herpaper Flach (1985) hinted that recent developments in computer science especiallyexpert systems and artificial intelligence could make modelling a practical mappingtool for soil scientists in the near future A year later McCracken and Cate (1986) hopedto encourage soil scientists to explore expert systems and its application to soil sciencethrough an optimistic article they wrote on the potential use of expert systemsHowever little research was actually conducted in the late 1980s and expert systemshad not yet begun to fulfil the lofty goals set forth by early practitioners (Dale et al1989)

In the early 1990s expert systems approaches to predictive soil mapping began toappear in the literature Skidmore et al (1991) used a Bayesian expert system to mapforest soil into different classes and their results compared favourably with availablesoil maps and actual field-collected data Their methods successfully incorporatedsurveyor knowledge and remotely sensed and digital terrain attributes but failed tobetter characterize continuous soilndashlandscape variability because their final productwas a choropleth map Skidmore et al (1996) revisited their earlier research andassessed the mapping accuracy of their results They found that the soil map producedby the expert system achieved an overall accuracy of 698 (sample size = 53) while themap derived from conventional methods had an accuracy of 736 In addition toknowledge provided by a soil scientist Skidmore et alrsquos (1996) expert system usedvegetation derived for aerial photography and topographic variables derived from a 10-m DEM (wetness index topographic position and slope)

Using an expert system Cook et al (1996) successfully produced a continuous soilproperty map for organic matter content using wetness index aspect and slope asexplanatory variables Their methods were somewhat inefficient because they requireda separate expert system for each soil property of interest However their researchrepresents the only example of expert systems used to directly predict the spatial dis-tribution of a soil property All of the other examples of expert systems in the literature

P Scull et al 189

190 Predictive soil mapping a review

deal with soil type or class The use of expert systems to map soil properties needs tobe explored further

The expert systems discussed thus far have all used Boolean logic within theirtheoretical framework whereby an observation can belong to one and only one classand the soil properties of that class are assigned to the observation With thedevelopment of fuzzy logic and the semantic import model (the second fuzzy logicapproach to creating continuous classes) the opportunity exists to develop fuzzy logic-based expert systems Such systems can be used in conjunction with expert knowledgein situations where experts have a good qualitative idea of how to group data but havetrouble dealing with observations that are not well represented by rigid classificationschemes (Burrough and McDonnell 1998) This method can be particularly useful insituations when taxonomic schemes have been previously developed as is the case insoil taxonomy Several examples of this type of approach were published in the mid-1990s by A Zhu and colleagues (Zhu and Band 1994 Zhu et al 1996 1997 Zhu1997ab) Such systems proved useful for mapping soil at unvisited locations usingsurveyor knowledge and were also capable of producing continuous soil propertymaps The use of fuzzy logic within the theoretical framework of the expert systemallows the complex nature of soil to propagate through the modelling process neversubjected to classification schemes that filter out potentially useful lsquonoisersquo The use offuzzy logic also gives the soil surveyor more latitude during the interview processwhen the knowledge base is defined for the expert system Zhu et al (1997) developeda fuzzy logic-based expert system called SoLIM to determine the similarity of eachgrid cell in a study area to the various taxonomic mapping units delimited by the soilsurvey Continuous soil attribute maps were calculated using the similarity values andtheir relative soil survey determined attribute values (effectively achieving Goal 2) Theresulting data (which proved more accurate than soil survey data once field checked)consists of a raster grid whose resolution is determined by the resolution of the inputenvironmental and digital elevation data As noted previously this type of data modelis more applicable to environmental modelling than the choropleth map (Burrough andMcDonnell 1998)

Expert system approaches to PSM are capable of exploiting soil surveyor knowledgeby developing rule-based systems that imitate the surveyorrsquos conceptual model of soilvariability (the primary focus of Goal 3) The method would appear extremely usefulfor mapping projects (such as those conducted by the NRCS-NCSS) where fieldwork isinitially conducted to determine soilndashlandscape relations Expert system developmentcould be directly inserted into the traditional soil survey mapping approach as asubstitute for the step where the surveyor converts hisher conceptual model into achoropleth map Rather that knowledge could be incorporated into the expert systemwhich could be used to predictively map soil The resulting raster map would be morescientifically based and explicit than the hand-drawn choropleth maps of the past Itsscale would not be limited to that of the aerial photography but rather to the scale ofthe environmental data

Despite these advantages the expert system approach has some drawbacks Becausethe method is dependent upon expert knowledge it cannot be applied whereenvironmentndashsoil relations are poorly understood Of course this criticism can be madeof all PSM models as inductive or deductive knowledge of soilndashenvironment relationsis a prerequisite for PSM However expert systems are deductive models and as such

are not driven by any specific field-collected data (although presumably the soil experthas field experience in the mapping area) Expert systems donrsquot afford the opportunityto first statistically document landscapendashsoil relations and then extrapolate the resultsbecause expert systems do not directly use sample soil data (lsquohard datarsquo) to determinesoilndashlandscape relations Thus the utilization of the relationship between environmen-tal variables and soil properties (Goal 1) is only indirectly achieved Expert systemsapproaches have been demonstrated to be extremely effective in a small number of casestudies The possibility of satisfying all three goals discussed in this review makesexpert systems a predictive soil mapping method that needs to be further tested ndash bothgeographically and across different scales of analysis

V Conclusion

Most of the predictive soil mapping research outlined in this review was conducted atvery large map scales (over small areas) In fact the majority of the research wasconcerned with assessing the spatial variability of soil character within individual fieldsor across soil toposequences The primary driving force behind this type of research hasbeen the need to provide accurate soils information for agriculture and ecologicalmodels It is clear that terrain attributes are powerful predictors at the local scaleGeostatistical tools have been successful at using terrain attributes and the spatialdependence of soil properties to interpolate between existing data points withinindividual fields Across soil toposequences statistical approaches provide a usefulmeans of predicting soil character PSM research has been most successful at the fieldscale because many of the soil-forming factors are held constant For example the neteffect of four of Jennyrsquos five soil-forming factors (climate organism parent material andtime) was minimal within many of the studies reviewed While some research hastackled larger areas of study there exists a deficiency in our ability to predictively mapsoils at smaller map scales Since the distribution of soil is scale-dependent differentPSM methods and predictors are likely to work better at different scales Focus in thefuture must continue to move toward working over larger spatial extents of study inorder to produce landscape-scale soil information

Further a large proportion of the research was conducted in semi-arid gentlysloping agricultural landscapes Humid forestlands mountainous regions and desertshave received little attention As such PSM methods need to continue to be testedandor developed in a wider variety of landscapes where spatial soil distributions canbe more complex Different methods will likely be successful to different degrees indifferent environments Whereas terrain attributes are dominant predictors of soilcharacter across toposequences in prairie lands other predictors are likely necessary inareas where soils do not develop into clear toposequences For example remote sensingdata has been a useful predictor of soil chronosequences in desert landscapes wheresurface appearance is often related to soil character Surficial geology is often reflectiveof soil character in mountain regions where soils are thin and significant bedrock existsat the surface Focus in the future must continue to determine which methods andpredictors work best in which environments

Since the most useful PSM approach will vary across spatial scales and environmen-tal gradients the method used should be driven by the mapping objectives of the

P Scull et al 191

192 Predictive soil mapping a review

project Whereas the traditional soil survey of the past was expected to meet the needsof a diverse group of end users PSM methods will vary given the objectives of thesurvey As a result the end users of the soil data need to play a more active role in thesurvey process Because precision agriculture and large-scale mapping has been theprimary focus of PSM there exists a deficiency in our ability to predictively map soilsfor the purpose of general land use planning and management Even though lessdetailed soil information is needed for these purposes the development of PSMmethods is complicated by the fact that soils exhibit complex spatial variability atsmall map scales where soil-forming environments vary greatly from one location tothe next

A number of alternative methods of characterizing the continuous nature of the soillandscape have been developed Thus far most of the PSM research has provided soilinformation in a nonobject form (as opposed to defining soil types as independententities) Soil data have been generated organized and presented in the form of eitherisorithmic maps or fine-scale raster grids Both of these data models are field-viewmodels of geographic space which allow the soil to be perceived as a constantlyvarying surface Two distinct approaches have been employed (1) mapping individualsoil properties and (2) mapping continuously varying (fuzzy) soil classes Mappingindividual properties is the most common approach and will likely continue todominate PSM research The use of fuzzy soil classes in the literature is less commonand more difficult to be integrated into standard mapping procedures (such as use ofsoil taxa) because the concept radically differs from the traditional view of the soillandscape However fuzzy soil classes (generated using the fuzzy-k means approach) dohave the potential to help further advance Soil taxonomy by identifying taxonomicclasses that are more reflective of pedologic processes at work

Expert systems have been greatly underutilized in PSM research especiallyconsidering how effective a small number of case studies have been Expert systemshave the potential to satisfy successfully all three goals discussed throughout thereview They also have the potential to bridge the gap between traditional approachesand PSM methods because field soil scientists do not have to change their conceptualapproach to mapping They can still conduct field reconnaissance to determinesoilndashlandscape relationships Afterwards though they can use expert systems to betterexploit the knowledge they have garnered Such an approach could help ease thetransition from traditional soil survey to more scientifically explicit methods Expertsystems are also the most fruitful approach to utilizing a wealth of data that has alreadybeen collected in a nonexplicit manner the wealth of expert knowledge (in the form ofthe conceptual model) that senior soil mappers have accumulated

Focus in the future must also move in the direction of operability The utility of manyPSM approaches has been clearly demonstrated yet changes in how soils are mappedand perceived has been slow In the USA various members of the National CooperativeSoil Survey have called for a change in the philosophy of survey (from mappingdiscrete soil types to mapping continuous varying soil properties McSweeny et al1994) but little change has actually taken place To realize the potential of PSM in con-tributing to an overall change in standard soil mapping procedures practitionersworking in the field of PSM need to form working relationships with field soilscientists PSM methods need also to be presented in a manner that is comprehensibleto the soil science trained field mapper Predictive soil mapping is a relatively recent

phenomenon yet much progress has been made In the process of mapping soilknowledge will continue to advance

P Scull et al 193

References

Agbu PA Fehrenbacher DJ and Jansen IJ1990 Statistical comparison of SPOT spectralmaps with field soil maps Soil Science Society ofAmerica Journal 54 818ndash18

Barrett L 1999 Particulars in contextmaintaining a balance in soil geography Annalsof the Association of America Geographers 89707ndash13

Beckett PHT and Webster R 1971 Soilvariability a review Soils and Fertilizers 341ndash15

Bell JC Grigal DF and Bates PC 2000 A soil-terrain model for estimating spatial patterns ofsoil organic carbon In Wilson J and Gallant Jeditors Terrain analysis principles and applica-tions New York City NY John Wiley and Sons295ndash310

Birkeland PW 1999 Soil and geomorphologyThird edition New York NY Oxford UniversityPress

Breiman L Friedman JH Olshen RA andStone CJ 1984 Classification and regressiontrees Belmont CA Wadsworth

Brule FJ 1996 Fuzzy systems ndash a tutorialhttp newsgroup compai httpwwwquadralaycom (last accessed 30 August 2001)

Bunkin FV and Bunkin AF 2000 Lidarsounding of water soil and plants Atmosphericand Oceanic Optics 13 54ndash72

Buol SW Hole FD McCracken RJ andSouthard RJ 1997 Soil genesis and classifica-tion Ames IA Iowa State University Press

Burgess TM and Webster R 1980a Optimalinterpolation and isarithmic mapping of soilproperties the semi-variogram and punctualkriging Journal of Soil Science 31 315ndash31

ndashndashndashndash 1980b Optimal interpolation and isarithmicmapping of soil properties block kringingJournal of Soil Science 31 331ndash41

Burrough PA 1989 Fuzzy mathematicalmethods for soil survey and land evaluationJournal of Soil Science 40 477ndash92

Burrough PA and McDonnell RA 1998Principles of geographic information systems(Revised edition) Oxford Clarendon Press

Burrough PA Beckett PHT and Jarvis MG1971 The relation between cost and utility insoil survey Journal of Soil Science 22 368ndash81

Burrough PA Van Gaans PMF and

Hootsman R 1997 Continuous classificationin soil survey spatial correlation confusionand boundaries Geoderma 77 115ndash35

Cambell JB 1977 Variation of selectedproperties across a soil boundary Soil ScienceSociety of America Journal 41 578ndash82

Cambell JB and Edmonds WJ 1984 Themissing geographic dimension to soiltaxonomy Annals of the Association of AmericanGeographers 74 83ndash97

Castrignano A Giugliarini L Risaliti R andMartinelli N 2000 Study of spatial relation-ships among some soil physico-chemicalproperties of a field in central Italy using multi-variate geostatistics Geoderma 97 39ndash60

Cialella AT Dubayah R Lawrence W andLevine E 1997 Predicting soil drainage classusing remotely sensed and digital elevationdata Journal of Soil Science 62(2) 171ndash78

Clark RN and Swayze GA 1996 Evolution inimaging spectroscopy analysis and sensorsignal-to-noise an examination of how far wehave come Summaries of the sixth annual JPLairborne Earth science workshop 4ndash8 March 1996AVIRIS Workshop Vol 1 5

Clarke LA and Pregibon D 1992 Tree-basedmodels In Chambers J and Hastie J editorsStatistical models in S Pacific Grove Wadsworthand Brooks 377ndash419

Coleman TL Agbu PA and MontgomeryOL 1993 Spectral differentiation of surfacesoils and soil properties ndash is it possible fromspace platforms Soil Science 155 283ndash93

Cook SE Corner RJ Grealish G GesslerPE and Chartres CJ 1996 A rule-basedsystem to map soil properties Soil ScienceSociety of America Journal 60 1893ndash900

Cruickshank JG 1972 Soil geography New YorkNY John Wiley amp Sons

Csillag F Pasztor L Biehl LL 1993 Spectralband selection for the characterization ofsalinity status of soils Remote Sensing ofEnvironment 43 231ndash42

Dale MB McBratney AB and Russell JS1989 On the role of expert systems andnumerical taxonomy in soil classificationJournal of Soil Science 40 223ndash34

Davis JR 1993 Expert systems and environ-

in an iterative manner eventually yielding estimates of the misclassification rates foreach class each node and the whole model (Breiman et al 1984) In this mannerdifferent sized trees can be compared in terms of parsimony

The term DTA is used to collectively refer to all types of tree-based modelling (theword lsquodecisionrsquo is used because it is descriptive indicating that the analysis eventuallyleads to a set of decision rules defining data partitions) The term should be distin-guished from classification tree analysis because the latter refers specifically to DTAwhere the response variable is categorical The term CART (classification andregression trees ndash Breiman et al 1984) is sometimes used but strictly speaking refers tospecific software Friedl and Brodley (1997) provide a review of the decision treealgorithms They divide types of DTA into two classes (1) homogeneous decision treesfor which a single algorithm is used to estimate each split (eg CART) and (2) hybriddecision trees (HDT) for which different splitting methods can be used at differentpoints in the tree (eg Quinlan 1993) They further divide homogeneous decision treesinto univariate (UDT) where single features of the input data define splits and multi-variate decision trees (MDT) where multiple features of the input data can define splitsAccording to this naming convention no distinction is made between methods withdifferent types of response variables although all methods can be used with bothcategorical and continuous response variables In a comparison of these various typesof algorithms on a variety of data sets Friedl and Brodley (1997) found that HDT hadthe highest classification accuracy

The overall aim of DTA is to design a set of predictive rules (eg if geology type Athen soil type B) developed from training data which can then be applied to ageographic data base to predict the value of a response variable (Michaelsen et al 1994)Therefore DTA explicitly uses soilndashlandscape correlation in model development (Goal2) The technique appears promising in soil research but needs to be further tested asnot many of the above types of DTA have been tested in the PSM literature in fact onlyunivariate approaches have been employed For example Lagacherie and Holmes(1997) successfully used univariate DTA to model a categorical response variable soiltype within a training set and then assuming that the training set was representativeextrapolated the model to a much larger region Their work is interesting becausealthough not a single sample came from the area they eventually mapped they wereable to produce a soil map that was more accurate (74 versus 69) than the existingmap produced from traditional methods Cialella et al (1997) also used univariate DTAto predict soil drainage class from a variety of terrain attributes and remotely senseddata They predicted soil drainage class with an average of 78 accuracy ndash impressivegiven that the variation accounted for by a typical soil survey ranges from about halfthe total variance for physical attributes to less than one-tenth for some soil chemicalattributes (Gessler et al 1995)

DTA has been compared with other approaches by several authors In the applicationof erosion modelling DTA results were similar to Artificial Neural Networks (Ellis1996) Both methods achieved high training accuracy (as measured by the KappaStatistic) but in terms of prediction accuracy both methods performed poorly Gessleret al (1995) compared DTA with generalized linear models (GLM) and generalizedadditive models (GAM) to predict A-horizon thickness and concluded that GLM waspreferable to both DTA and GAM McKensie and Ryan (1999) compared regressiontrees and standard linear regression to predict soil properties (total solum depth soil

P Scull et al 187

188 Predictive soil mapping a review

phophorus and soil carbon) using a large variety of predictors (elevation slopecurvature wetness index climate variables geology unit Landsat TM data andGamma radiometric data) and found that different methods work best for differentproperties and overall success hinged on the strength of the relationship between soiland environmental variables Their study of a 50 000 ha forested landscape in southernNew South Wales Australia is one of the few studies at such a small map scale (largearea) or in forested landscape The most extensive comparison of techniques involvingDTA was conducted by McBratney et al (2000) They compared regression techniques(GLM GAM DTA-regression tree) geostatistical techniques (kriging and heterotopicco-kriging) and a hybrid technique (regression kriging) DTA was found to be thepoorest performing of the regression techniques because of the unrealistic predictionsurface generated by the DTA model DTA has also been criticized by other authorsbecause of the stepped prediction surface (Gessler 1996) This phenomenon can beespecially apparent in situations where predictor variables have different resolutionsThe extent to which DTA yields a better spatial representation of soil continuity is afunction of the scale and type of predictor variables used and therefore varies withindividual models

DTA is gaining widespread popularity as a means to develop prediction rules thatcan be rapidly and repeatedly evaluated (Cialella et al 1997 Franklin et al 2000) DTAprovides the following advantages over standard statistical techniques (1) it is easier tointerpret when explanatory variables are both nominal and continuous (2) it isinvariant to monotone re-expressions (transformations) of predictor variables (3) itdeals more satisfactorily with missing data values and outliers (4) it is more adept atcapturing nonadditive and nonlinear behaviour (5) it doesnrsquot make any assumptionsabout data distribution and (6) it is easily updateable as more data are collected(Moore DM et al 1991) The DTA model framework is especially appealing becauseof its capability to integrate a wide range of data sets as explanatory variables

DTA offers a unique opportunity for interaction between soil experts and soilmodellers because the output of the model is a set of rules that can be pedologicallyinterpreted by the soil expert In this sense expert knowledge is used in an implicitmanner in DTA (somewhat effectively achieving Goal 3) While these rules can often beexceedingly complex at minimum the expert can decide whether initial splits makesense given their understanding of the landscape The success of DTA results oftenhinges on the ability of the modeller to make key decisions during the model buildingprocess there is unfortunately no definitive way to determine the most optimal treeThe aid of the soil expert can potentially elucidate this problem

4 Expert systems

A variety of expert system approaches to PSM have been developed to utilize expertknowledge The purpose of such methods is to exploit the information the soil surveyoraccumulates while working in the field by integrating such knowledge into thepredictive model (McCracken and Cate 1986) Unlike the majority of the researchreviewed thus far the dependent variable in many expert systems models is often soiltaxa or mapping unit This apparent disadvantage of expert systems (using classifica-tion to characterize soil continuity) does make them easier to integrate into traditional

soil survey In addition several authors have developed methods to developcontinuous soil property maps from the output of expert system models designed topredict soil unit occurrence

Expert systems are composed of data (information on spatial environmentalvariables eg topography climate etc) a knowledge base (rules and facts related tosoil variation supplied from the soil surveyor) and an inference engine (whichcombines data and the knowledge base to infer logically valid conclusions) (Skidmoreet al 1996) Expert systems differ from conventional models in two ways (1) they storeand manipulate qualitative information (allowing them access to information thatcannot normally be used in other modelling frameworks) and (2) they are structuredas meta-models (the knowledge is separated from the model) (Davis 1993) This allowsthe model to selectively choose which information is relevant at various stages of themodelling process and it allows for information to be easily updated Davis (1993)reviews the application of expert systems to environmental modelling research conclu-ding that the technique is becoming more widely accepted He further states that theapplication of expert systems is constrained by an absence of fundamental knowledgefor rule generation a problem that would appear less relevant to soil mapping giventhe amount of untapped expert knowledge accumulated by the soil surveyor

The first mention of the use of expert systems in pedology was in a paper presenta-tion at the Northeast Committee Soil Survey Conference 1984 (Flach 1985) In herpaper Flach (1985) hinted that recent developments in computer science especiallyexpert systems and artificial intelligence could make modelling a practical mappingtool for soil scientists in the near future A year later McCracken and Cate (1986) hopedto encourage soil scientists to explore expert systems and its application to soil sciencethrough an optimistic article they wrote on the potential use of expert systemsHowever little research was actually conducted in the late 1980s and expert systemshad not yet begun to fulfil the lofty goals set forth by early practitioners (Dale et al1989)

In the early 1990s expert systems approaches to predictive soil mapping began toappear in the literature Skidmore et al (1991) used a Bayesian expert system to mapforest soil into different classes and their results compared favourably with availablesoil maps and actual field-collected data Their methods successfully incorporatedsurveyor knowledge and remotely sensed and digital terrain attributes but failed tobetter characterize continuous soilndashlandscape variability because their final productwas a choropleth map Skidmore et al (1996) revisited their earlier research andassessed the mapping accuracy of their results They found that the soil map producedby the expert system achieved an overall accuracy of 698 (sample size = 53) while themap derived from conventional methods had an accuracy of 736 In addition toknowledge provided by a soil scientist Skidmore et alrsquos (1996) expert system usedvegetation derived for aerial photography and topographic variables derived from a 10-m DEM (wetness index topographic position and slope)

Using an expert system Cook et al (1996) successfully produced a continuous soilproperty map for organic matter content using wetness index aspect and slope asexplanatory variables Their methods were somewhat inefficient because they requireda separate expert system for each soil property of interest However their researchrepresents the only example of expert systems used to directly predict the spatial dis-tribution of a soil property All of the other examples of expert systems in the literature

P Scull et al 189

190 Predictive soil mapping a review

deal with soil type or class The use of expert systems to map soil properties needs tobe explored further

The expert systems discussed thus far have all used Boolean logic within theirtheoretical framework whereby an observation can belong to one and only one classand the soil properties of that class are assigned to the observation With thedevelopment of fuzzy logic and the semantic import model (the second fuzzy logicapproach to creating continuous classes) the opportunity exists to develop fuzzy logic-based expert systems Such systems can be used in conjunction with expert knowledgein situations where experts have a good qualitative idea of how to group data but havetrouble dealing with observations that are not well represented by rigid classificationschemes (Burrough and McDonnell 1998) This method can be particularly useful insituations when taxonomic schemes have been previously developed as is the case insoil taxonomy Several examples of this type of approach were published in the mid-1990s by A Zhu and colleagues (Zhu and Band 1994 Zhu et al 1996 1997 Zhu1997ab) Such systems proved useful for mapping soil at unvisited locations usingsurveyor knowledge and were also capable of producing continuous soil propertymaps The use of fuzzy logic within the theoretical framework of the expert systemallows the complex nature of soil to propagate through the modelling process neversubjected to classification schemes that filter out potentially useful lsquonoisersquo The use offuzzy logic also gives the soil surveyor more latitude during the interview processwhen the knowledge base is defined for the expert system Zhu et al (1997) developeda fuzzy logic-based expert system called SoLIM to determine the similarity of eachgrid cell in a study area to the various taxonomic mapping units delimited by the soilsurvey Continuous soil attribute maps were calculated using the similarity values andtheir relative soil survey determined attribute values (effectively achieving Goal 2) Theresulting data (which proved more accurate than soil survey data once field checked)consists of a raster grid whose resolution is determined by the resolution of the inputenvironmental and digital elevation data As noted previously this type of data modelis more applicable to environmental modelling than the choropleth map (Burrough andMcDonnell 1998)

Expert system approaches to PSM are capable of exploiting soil surveyor knowledgeby developing rule-based systems that imitate the surveyorrsquos conceptual model of soilvariability (the primary focus of Goal 3) The method would appear extremely usefulfor mapping projects (such as those conducted by the NRCS-NCSS) where fieldwork isinitially conducted to determine soilndashlandscape relations Expert system developmentcould be directly inserted into the traditional soil survey mapping approach as asubstitute for the step where the surveyor converts hisher conceptual model into achoropleth map Rather that knowledge could be incorporated into the expert systemwhich could be used to predictively map soil The resulting raster map would be morescientifically based and explicit than the hand-drawn choropleth maps of the past Itsscale would not be limited to that of the aerial photography but rather to the scale ofthe environmental data

Despite these advantages the expert system approach has some drawbacks Becausethe method is dependent upon expert knowledge it cannot be applied whereenvironmentndashsoil relations are poorly understood Of course this criticism can be madeof all PSM models as inductive or deductive knowledge of soilndashenvironment relationsis a prerequisite for PSM However expert systems are deductive models and as such

are not driven by any specific field-collected data (although presumably the soil experthas field experience in the mapping area) Expert systems donrsquot afford the opportunityto first statistically document landscapendashsoil relations and then extrapolate the resultsbecause expert systems do not directly use sample soil data (lsquohard datarsquo) to determinesoilndashlandscape relations Thus the utilization of the relationship between environmen-tal variables and soil properties (Goal 1) is only indirectly achieved Expert systemsapproaches have been demonstrated to be extremely effective in a small number of casestudies The possibility of satisfying all three goals discussed in this review makesexpert systems a predictive soil mapping method that needs to be further tested ndash bothgeographically and across different scales of analysis

V Conclusion

Most of the predictive soil mapping research outlined in this review was conducted atvery large map scales (over small areas) In fact the majority of the research wasconcerned with assessing the spatial variability of soil character within individual fieldsor across soil toposequences The primary driving force behind this type of research hasbeen the need to provide accurate soils information for agriculture and ecologicalmodels It is clear that terrain attributes are powerful predictors at the local scaleGeostatistical tools have been successful at using terrain attributes and the spatialdependence of soil properties to interpolate between existing data points withinindividual fields Across soil toposequences statistical approaches provide a usefulmeans of predicting soil character PSM research has been most successful at the fieldscale because many of the soil-forming factors are held constant For example the neteffect of four of Jennyrsquos five soil-forming factors (climate organism parent material andtime) was minimal within many of the studies reviewed While some research hastackled larger areas of study there exists a deficiency in our ability to predictively mapsoils at smaller map scales Since the distribution of soil is scale-dependent differentPSM methods and predictors are likely to work better at different scales Focus in thefuture must continue to move toward working over larger spatial extents of study inorder to produce landscape-scale soil information

Further a large proportion of the research was conducted in semi-arid gentlysloping agricultural landscapes Humid forestlands mountainous regions and desertshave received little attention As such PSM methods need to continue to be testedandor developed in a wider variety of landscapes where spatial soil distributions canbe more complex Different methods will likely be successful to different degrees indifferent environments Whereas terrain attributes are dominant predictors of soilcharacter across toposequences in prairie lands other predictors are likely necessary inareas where soils do not develop into clear toposequences For example remote sensingdata has been a useful predictor of soil chronosequences in desert landscapes wheresurface appearance is often related to soil character Surficial geology is often reflectiveof soil character in mountain regions where soils are thin and significant bedrock existsat the surface Focus in the future must continue to determine which methods andpredictors work best in which environments

Since the most useful PSM approach will vary across spatial scales and environmen-tal gradients the method used should be driven by the mapping objectives of the

P Scull et al 191

192 Predictive soil mapping a review

project Whereas the traditional soil survey of the past was expected to meet the needsof a diverse group of end users PSM methods will vary given the objectives of thesurvey As a result the end users of the soil data need to play a more active role in thesurvey process Because precision agriculture and large-scale mapping has been theprimary focus of PSM there exists a deficiency in our ability to predictively map soilsfor the purpose of general land use planning and management Even though lessdetailed soil information is needed for these purposes the development of PSMmethods is complicated by the fact that soils exhibit complex spatial variability atsmall map scales where soil-forming environments vary greatly from one location tothe next

A number of alternative methods of characterizing the continuous nature of the soillandscape have been developed Thus far most of the PSM research has provided soilinformation in a nonobject form (as opposed to defining soil types as independententities) Soil data have been generated organized and presented in the form of eitherisorithmic maps or fine-scale raster grids Both of these data models are field-viewmodels of geographic space which allow the soil to be perceived as a constantlyvarying surface Two distinct approaches have been employed (1) mapping individualsoil properties and (2) mapping continuously varying (fuzzy) soil classes Mappingindividual properties is the most common approach and will likely continue todominate PSM research The use of fuzzy soil classes in the literature is less commonand more difficult to be integrated into standard mapping procedures (such as use ofsoil taxa) because the concept radically differs from the traditional view of the soillandscape However fuzzy soil classes (generated using the fuzzy-k means approach) dohave the potential to help further advance Soil taxonomy by identifying taxonomicclasses that are more reflective of pedologic processes at work

Expert systems have been greatly underutilized in PSM research especiallyconsidering how effective a small number of case studies have been Expert systemshave the potential to satisfy successfully all three goals discussed throughout thereview They also have the potential to bridge the gap between traditional approachesand PSM methods because field soil scientists do not have to change their conceptualapproach to mapping They can still conduct field reconnaissance to determinesoilndashlandscape relationships Afterwards though they can use expert systems to betterexploit the knowledge they have garnered Such an approach could help ease thetransition from traditional soil survey to more scientifically explicit methods Expertsystems are also the most fruitful approach to utilizing a wealth of data that has alreadybeen collected in a nonexplicit manner the wealth of expert knowledge (in the form ofthe conceptual model) that senior soil mappers have accumulated

Focus in the future must also move in the direction of operability The utility of manyPSM approaches has been clearly demonstrated yet changes in how soils are mappedand perceived has been slow In the USA various members of the National CooperativeSoil Survey have called for a change in the philosophy of survey (from mappingdiscrete soil types to mapping continuous varying soil properties McSweeny et al1994) but little change has actually taken place To realize the potential of PSM in con-tributing to an overall change in standard soil mapping procedures practitionersworking in the field of PSM need to form working relationships with field soilscientists PSM methods need also to be presented in a manner that is comprehensibleto the soil science trained field mapper Predictive soil mapping is a relatively recent

phenomenon yet much progress has been made In the process of mapping soilknowledge will continue to advance

P Scull et al 193

References

Agbu PA Fehrenbacher DJ and Jansen IJ1990 Statistical comparison of SPOT spectralmaps with field soil maps Soil Science Society ofAmerica Journal 54 818ndash18

Barrett L 1999 Particulars in contextmaintaining a balance in soil geography Annalsof the Association of America Geographers 89707ndash13

Beckett PHT and Webster R 1971 Soilvariability a review Soils and Fertilizers 341ndash15

Bell JC Grigal DF and Bates PC 2000 A soil-terrain model for estimating spatial patterns ofsoil organic carbon In Wilson J and Gallant Jeditors Terrain analysis principles and applica-tions New York City NY John Wiley and Sons295ndash310

Birkeland PW 1999 Soil and geomorphologyThird edition New York NY Oxford UniversityPress

Breiman L Friedman JH Olshen RA andStone CJ 1984 Classification and regressiontrees Belmont CA Wadsworth

Brule FJ 1996 Fuzzy systems ndash a tutorialhttp newsgroup compai httpwwwquadralaycom (last accessed 30 August 2001)

Bunkin FV and Bunkin AF 2000 Lidarsounding of water soil and plants Atmosphericand Oceanic Optics 13 54ndash72

Buol SW Hole FD McCracken RJ andSouthard RJ 1997 Soil genesis and classifica-tion Ames IA Iowa State University Press

Burgess TM and Webster R 1980a Optimalinterpolation and isarithmic mapping of soilproperties the semi-variogram and punctualkriging Journal of Soil Science 31 315ndash31

ndashndashndashndash 1980b Optimal interpolation and isarithmicmapping of soil properties block kringingJournal of Soil Science 31 331ndash41

Burrough PA 1989 Fuzzy mathematicalmethods for soil survey and land evaluationJournal of Soil Science 40 477ndash92

Burrough PA and McDonnell RA 1998Principles of geographic information systems(Revised edition) Oxford Clarendon Press

Burrough PA Beckett PHT and Jarvis MG1971 The relation between cost and utility insoil survey Journal of Soil Science 22 368ndash81

Burrough PA Van Gaans PMF and

Hootsman R 1997 Continuous classificationin soil survey spatial correlation confusionand boundaries Geoderma 77 115ndash35

Cambell JB 1977 Variation of selectedproperties across a soil boundary Soil ScienceSociety of America Journal 41 578ndash82

Cambell JB and Edmonds WJ 1984 Themissing geographic dimension to soiltaxonomy Annals of the Association of AmericanGeographers 74 83ndash97

Castrignano A Giugliarini L Risaliti R andMartinelli N 2000 Study of spatial relation-ships among some soil physico-chemicalproperties of a field in central Italy using multi-variate geostatistics Geoderma 97 39ndash60

Cialella AT Dubayah R Lawrence W andLevine E 1997 Predicting soil drainage classusing remotely sensed and digital elevationdata Journal of Soil Science 62(2) 171ndash78

Clark RN and Swayze GA 1996 Evolution inimaging spectroscopy analysis and sensorsignal-to-noise an examination of how far wehave come Summaries of the sixth annual JPLairborne Earth science workshop 4ndash8 March 1996AVIRIS Workshop Vol 1 5

Clarke LA and Pregibon D 1992 Tree-basedmodels In Chambers J and Hastie J editorsStatistical models in S Pacific Grove Wadsworthand Brooks 377ndash419

Coleman TL Agbu PA and MontgomeryOL 1993 Spectral differentiation of surfacesoils and soil properties ndash is it possible fromspace platforms Soil Science 155 283ndash93

Cook SE Corner RJ Grealish G GesslerPE and Chartres CJ 1996 A rule-basedsystem to map soil properties Soil ScienceSociety of America Journal 60 1893ndash900

Cruickshank JG 1972 Soil geography New YorkNY John Wiley amp Sons

Csillag F Pasztor L Biehl LL 1993 Spectralband selection for the characterization ofsalinity status of soils Remote Sensing ofEnvironment 43 231ndash42

Dale MB McBratney AB and Russell JS1989 On the role of expert systems andnumerical taxonomy in soil classificationJournal of Soil Science 40 223ndash34

Davis JR 1993 Expert systems and environ-

188 Predictive soil mapping a review

phophorus and soil carbon) using a large variety of predictors (elevation slopecurvature wetness index climate variables geology unit Landsat TM data andGamma radiometric data) and found that different methods work best for differentproperties and overall success hinged on the strength of the relationship between soiland environmental variables Their study of a 50 000 ha forested landscape in southernNew South Wales Australia is one of the few studies at such a small map scale (largearea) or in forested landscape The most extensive comparison of techniques involvingDTA was conducted by McBratney et al (2000) They compared regression techniques(GLM GAM DTA-regression tree) geostatistical techniques (kriging and heterotopicco-kriging) and a hybrid technique (regression kriging) DTA was found to be thepoorest performing of the regression techniques because of the unrealistic predictionsurface generated by the DTA model DTA has also been criticized by other authorsbecause of the stepped prediction surface (Gessler 1996) This phenomenon can beespecially apparent in situations where predictor variables have different resolutionsThe extent to which DTA yields a better spatial representation of soil continuity is afunction of the scale and type of predictor variables used and therefore varies withindividual models

DTA is gaining widespread popularity as a means to develop prediction rules thatcan be rapidly and repeatedly evaluated (Cialella et al 1997 Franklin et al 2000) DTAprovides the following advantages over standard statistical techniques (1) it is easier tointerpret when explanatory variables are both nominal and continuous (2) it isinvariant to monotone re-expressions (transformations) of predictor variables (3) itdeals more satisfactorily with missing data values and outliers (4) it is more adept atcapturing nonadditive and nonlinear behaviour (5) it doesnrsquot make any assumptionsabout data distribution and (6) it is easily updateable as more data are collected(Moore DM et al 1991) The DTA model framework is especially appealing becauseof its capability to integrate a wide range of data sets as explanatory variables

DTA offers a unique opportunity for interaction between soil experts and soilmodellers because the output of the model is a set of rules that can be pedologicallyinterpreted by the soil expert In this sense expert knowledge is used in an implicitmanner in DTA (somewhat effectively achieving Goal 3) While these rules can often beexceedingly complex at minimum the expert can decide whether initial splits makesense given their understanding of the landscape The success of DTA results oftenhinges on the ability of the modeller to make key decisions during the model buildingprocess there is unfortunately no definitive way to determine the most optimal treeThe aid of the soil expert can potentially elucidate this problem

4 Expert systems

A variety of expert system approaches to PSM have been developed to utilize expertknowledge The purpose of such methods is to exploit the information the soil surveyoraccumulates while working in the field by integrating such knowledge into thepredictive model (McCracken and Cate 1986) Unlike the majority of the researchreviewed thus far the dependent variable in many expert systems models is often soiltaxa or mapping unit This apparent disadvantage of expert systems (using classifica-tion to characterize soil continuity) does make them easier to integrate into traditional

soil survey In addition several authors have developed methods to developcontinuous soil property maps from the output of expert system models designed topredict soil unit occurrence

Expert systems are composed of data (information on spatial environmentalvariables eg topography climate etc) a knowledge base (rules and facts related tosoil variation supplied from the soil surveyor) and an inference engine (whichcombines data and the knowledge base to infer logically valid conclusions) (Skidmoreet al 1996) Expert systems differ from conventional models in two ways (1) they storeand manipulate qualitative information (allowing them access to information thatcannot normally be used in other modelling frameworks) and (2) they are structuredas meta-models (the knowledge is separated from the model) (Davis 1993) This allowsthe model to selectively choose which information is relevant at various stages of themodelling process and it allows for information to be easily updated Davis (1993)reviews the application of expert systems to environmental modelling research conclu-ding that the technique is becoming more widely accepted He further states that theapplication of expert systems is constrained by an absence of fundamental knowledgefor rule generation a problem that would appear less relevant to soil mapping giventhe amount of untapped expert knowledge accumulated by the soil surveyor

The first mention of the use of expert systems in pedology was in a paper presenta-tion at the Northeast Committee Soil Survey Conference 1984 (Flach 1985) In herpaper Flach (1985) hinted that recent developments in computer science especiallyexpert systems and artificial intelligence could make modelling a practical mappingtool for soil scientists in the near future A year later McCracken and Cate (1986) hopedto encourage soil scientists to explore expert systems and its application to soil sciencethrough an optimistic article they wrote on the potential use of expert systemsHowever little research was actually conducted in the late 1980s and expert systemshad not yet begun to fulfil the lofty goals set forth by early practitioners (Dale et al1989)

In the early 1990s expert systems approaches to predictive soil mapping began toappear in the literature Skidmore et al (1991) used a Bayesian expert system to mapforest soil into different classes and their results compared favourably with availablesoil maps and actual field-collected data Their methods successfully incorporatedsurveyor knowledge and remotely sensed and digital terrain attributes but failed tobetter characterize continuous soilndashlandscape variability because their final productwas a choropleth map Skidmore et al (1996) revisited their earlier research andassessed the mapping accuracy of their results They found that the soil map producedby the expert system achieved an overall accuracy of 698 (sample size = 53) while themap derived from conventional methods had an accuracy of 736 In addition toknowledge provided by a soil scientist Skidmore et alrsquos (1996) expert system usedvegetation derived for aerial photography and topographic variables derived from a 10-m DEM (wetness index topographic position and slope)

Using an expert system Cook et al (1996) successfully produced a continuous soilproperty map for organic matter content using wetness index aspect and slope asexplanatory variables Their methods were somewhat inefficient because they requireda separate expert system for each soil property of interest However their researchrepresents the only example of expert systems used to directly predict the spatial dis-tribution of a soil property All of the other examples of expert systems in the literature

P Scull et al 189

190 Predictive soil mapping a review

deal with soil type or class The use of expert systems to map soil properties needs tobe explored further

The expert systems discussed thus far have all used Boolean logic within theirtheoretical framework whereby an observation can belong to one and only one classand the soil properties of that class are assigned to the observation With thedevelopment of fuzzy logic and the semantic import model (the second fuzzy logicapproach to creating continuous classes) the opportunity exists to develop fuzzy logic-based expert systems Such systems can be used in conjunction with expert knowledgein situations where experts have a good qualitative idea of how to group data but havetrouble dealing with observations that are not well represented by rigid classificationschemes (Burrough and McDonnell 1998) This method can be particularly useful insituations when taxonomic schemes have been previously developed as is the case insoil taxonomy Several examples of this type of approach were published in the mid-1990s by A Zhu and colleagues (Zhu and Band 1994 Zhu et al 1996 1997 Zhu1997ab) Such systems proved useful for mapping soil at unvisited locations usingsurveyor knowledge and were also capable of producing continuous soil propertymaps The use of fuzzy logic within the theoretical framework of the expert systemallows the complex nature of soil to propagate through the modelling process neversubjected to classification schemes that filter out potentially useful lsquonoisersquo The use offuzzy logic also gives the soil surveyor more latitude during the interview processwhen the knowledge base is defined for the expert system Zhu et al (1997) developeda fuzzy logic-based expert system called SoLIM to determine the similarity of eachgrid cell in a study area to the various taxonomic mapping units delimited by the soilsurvey Continuous soil attribute maps were calculated using the similarity values andtheir relative soil survey determined attribute values (effectively achieving Goal 2) Theresulting data (which proved more accurate than soil survey data once field checked)consists of a raster grid whose resolution is determined by the resolution of the inputenvironmental and digital elevation data As noted previously this type of data modelis more applicable to environmental modelling than the choropleth map (Burrough andMcDonnell 1998)

Expert system approaches to PSM are capable of exploiting soil surveyor knowledgeby developing rule-based systems that imitate the surveyorrsquos conceptual model of soilvariability (the primary focus of Goal 3) The method would appear extremely usefulfor mapping projects (such as those conducted by the NRCS-NCSS) where fieldwork isinitially conducted to determine soilndashlandscape relations Expert system developmentcould be directly inserted into the traditional soil survey mapping approach as asubstitute for the step where the surveyor converts hisher conceptual model into achoropleth map Rather that knowledge could be incorporated into the expert systemwhich could be used to predictively map soil The resulting raster map would be morescientifically based and explicit than the hand-drawn choropleth maps of the past Itsscale would not be limited to that of the aerial photography but rather to the scale ofthe environmental data

Despite these advantages the expert system approach has some drawbacks Becausethe method is dependent upon expert knowledge it cannot be applied whereenvironmentndashsoil relations are poorly understood Of course this criticism can be madeof all PSM models as inductive or deductive knowledge of soilndashenvironment relationsis a prerequisite for PSM However expert systems are deductive models and as such

are not driven by any specific field-collected data (although presumably the soil experthas field experience in the mapping area) Expert systems donrsquot afford the opportunityto first statistically document landscapendashsoil relations and then extrapolate the resultsbecause expert systems do not directly use sample soil data (lsquohard datarsquo) to determinesoilndashlandscape relations Thus the utilization of the relationship between environmen-tal variables and soil properties (Goal 1) is only indirectly achieved Expert systemsapproaches have been demonstrated to be extremely effective in a small number of casestudies The possibility of satisfying all three goals discussed in this review makesexpert systems a predictive soil mapping method that needs to be further tested ndash bothgeographically and across different scales of analysis

V Conclusion

Most of the predictive soil mapping research outlined in this review was conducted atvery large map scales (over small areas) In fact the majority of the research wasconcerned with assessing the spatial variability of soil character within individual fieldsor across soil toposequences The primary driving force behind this type of research hasbeen the need to provide accurate soils information for agriculture and ecologicalmodels It is clear that terrain attributes are powerful predictors at the local scaleGeostatistical tools have been successful at using terrain attributes and the spatialdependence of soil properties to interpolate between existing data points withinindividual fields Across soil toposequences statistical approaches provide a usefulmeans of predicting soil character PSM research has been most successful at the fieldscale because many of the soil-forming factors are held constant For example the neteffect of four of Jennyrsquos five soil-forming factors (climate organism parent material andtime) was minimal within many of the studies reviewed While some research hastackled larger areas of study there exists a deficiency in our ability to predictively mapsoils at smaller map scales Since the distribution of soil is scale-dependent differentPSM methods and predictors are likely to work better at different scales Focus in thefuture must continue to move toward working over larger spatial extents of study inorder to produce landscape-scale soil information

Further a large proportion of the research was conducted in semi-arid gentlysloping agricultural landscapes Humid forestlands mountainous regions and desertshave received little attention As such PSM methods need to continue to be testedandor developed in a wider variety of landscapes where spatial soil distributions canbe more complex Different methods will likely be successful to different degrees indifferent environments Whereas terrain attributes are dominant predictors of soilcharacter across toposequences in prairie lands other predictors are likely necessary inareas where soils do not develop into clear toposequences For example remote sensingdata has been a useful predictor of soil chronosequences in desert landscapes wheresurface appearance is often related to soil character Surficial geology is often reflectiveof soil character in mountain regions where soils are thin and significant bedrock existsat the surface Focus in the future must continue to determine which methods andpredictors work best in which environments

Since the most useful PSM approach will vary across spatial scales and environmen-tal gradients the method used should be driven by the mapping objectives of the

P Scull et al 191

192 Predictive soil mapping a review

project Whereas the traditional soil survey of the past was expected to meet the needsof a diverse group of end users PSM methods will vary given the objectives of thesurvey As a result the end users of the soil data need to play a more active role in thesurvey process Because precision agriculture and large-scale mapping has been theprimary focus of PSM there exists a deficiency in our ability to predictively map soilsfor the purpose of general land use planning and management Even though lessdetailed soil information is needed for these purposes the development of PSMmethods is complicated by the fact that soils exhibit complex spatial variability atsmall map scales where soil-forming environments vary greatly from one location tothe next

A number of alternative methods of characterizing the continuous nature of the soillandscape have been developed Thus far most of the PSM research has provided soilinformation in a nonobject form (as opposed to defining soil types as independententities) Soil data have been generated organized and presented in the form of eitherisorithmic maps or fine-scale raster grids Both of these data models are field-viewmodels of geographic space which allow the soil to be perceived as a constantlyvarying surface Two distinct approaches have been employed (1) mapping individualsoil properties and (2) mapping continuously varying (fuzzy) soil classes Mappingindividual properties is the most common approach and will likely continue todominate PSM research The use of fuzzy soil classes in the literature is less commonand more difficult to be integrated into standard mapping procedures (such as use ofsoil taxa) because the concept radically differs from the traditional view of the soillandscape However fuzzy soil classes (generated using the fuzzy-k means approach) dohave the potential to help further advance Soil taxonomy by identifying taxonomicclasses that are more reflective of pedologic processes at work

Expert systems have been greatly underutilized in PSM research especiallyconsidering how effective a small number of case studies have been Expert systemshave the potential to satisfy successfully all three goals discussed throughout thereview They also have the potential to bridge the gap between traditional approachesand PSM methods because field soil scientists do not have to change their conceptualapproach to mapping They can still conduct field reconnaissance to determinesoilndashlandscape relationships Afterwards though they can use expert systems to betterexploit the knowledge they have garnered Such an approach could help ease thetransition from traditional soil survey to more scientifically explicit methods Expertsystems are also the most fruitful approach to utilizing a wealth of data that has alreadybeen collected in a nonexplicit manner the wealth of expert knowledge (in the form ofthe conceptual model) that senior soil mappers have accumulated

Focus in the future must also move in the direction of operability The utility of manyPSM approaches has been clearly demonstrated yet changes in how soils are mappedand perceived has been slow In the USA various members of the National CooperativeSoil Survey have called for a change in the philosophy of survey (from mappingdiscrete soil types to mapping continuous varying soil properties McSweeny et al1994) but little change has actually taken place To realize the potential of PSM in con-tributing to an overall change in standard soil mapping procedures practitionersworking in the field of PSM need to form working relationships with field soilscientists PSM methods need also to be presented in a manner that is comprehensibleto the soil science trained field mapper Predictive soil mapping is a relatively recent

phenomenon yet much progress has been made In the process of mapping soilknowledge will continue to advance

P Scull et al 193

References

Agbu PA Fehrenbacher DJ and Jansen IJ1990 Statistical comparison of SPOT spectralmaps with field soil maps Soil Science Society ofAmerica Journal 54 818ndash18

Barrett L 1999 Particulars in contextmaintaining a balance in soil geography Annalsof the Association of America Geographers 89707ndash13

Beckett PHT and Webster R 1971 Soilvariability a review Soils and Fertilizers 341ndash15

Bell JC Grigal DF and Bates PC 2000 A soil-terrain model for estimating spatial patterns ofsoil organic carbon In Wilson J and Gallant Jeditors Terrain analysis principles and applica-tions New York City NY John Wiley and Sons295ndash310

Birkeland PW 1999 Soil and geomorphologyThird edition New York NY Oxford UniversityPress

Breiman L Friedman JH Olshen RA andStone CJ 1984 Classification and regressiontrees Belmont CA Wadsworth

Brule FJ 1996 Fuzzy systems ndash a tutorialhttp newsgroup compai httpwwwquadralaycom (last accessed 30 August 2001)

Bunkin FV and Bunkin AF 2000 Lidarsounding of water soil and plants Atmosphericand Oceanic Optics 13 54ndash72

Buol SW Hole FD McCracken RJ andSouthard RJ 1997 Soil genesis and classifica-tion Ames IA Iowa State University Press

Burgess TM and Webster R 1980a Optimalinterpolation and isarithmic mapping of soilproperties the semi-variogram and punctualkriging Journal of Soil Science 31 315ndash31

ndashndashndashndash 1980b Optimal interpolation and isarithmicmapping of soil properties block kringingJournal of Soil Science 31 331ndash41

Burrough PA 1989 Fuzzy mathematicalmethods for soil survey and land evaluationJournal of Soil Science 40 477ndash92

Burrough PA and McDonnell RA 1998Principles of geographic information systems(Revised edition) Oxford Clarendon Press

Burrough PA Beckett PHT and Jarvis MG1971 The relation between cost and utility insoil survey Journal of Soil Science 22 368ndash81

Burrough PA Van Gaans PMF and

Hootsman R 1997 Continuous classificationin soil survey spatial correlation confusionand boundaries Geoderma 77 115ndash35

Cambell JB 1977 Variation of selectedproperties across a soil boundary Soil ScienceSociety of America Journal 41 578ndash82

Cambell JB and Edmonds WJ 1984 Themissing geographic dimension to soiltaxonomy Annals of the Association of AmericanGeographers 74 83ndash97

Castrignano A Giugliarini L Risaliti R andMartinelli N 2000 Study of spatial relation-ships among some soil physico-chemicalproperties of a field in central Italy using multi-variate geostatistics Geoderma 97 39ndash60

Cialella AT Dubayah R Lawrence W andLevine E 1997 Predicting soil drainage classusing remotely sensed and digital elevationdata Journal of Soil Science 62(2) 171ndash78

Clark RN and Swayze GA 1996 Evolution inimaging spectroscopy analysis and sensorsignal-to-noise an examination of how far wehave come Summaries of the sixth annual JPLairborne Earth science workshop 4ndash8 March 1996AVIRIS Workshop Vol 1 5

Clarke LA and Pregibon D 1992 Tree-basedmodels In Chambers J and Hastie J editorsStatistical models in S Pacific Grove Wadsworthand Brooks 377ndash419

Coleman TL Agbu PA and MontgomeryOL 1993 Spectral differentiation of surfacesoils and soil properties ndash is it possible fromspace platforms Soil Science 155 283ndash93

Cook SE Corner RJ Grealish G GesslerPE and Chartres CJ 1996 A rule-basedsystem to map soil properties Soil ScienceSociety of America Journal 60 1893ndash900

Cruickshank JG 1972 Soil geography New YorkNY John Wiley amp Sons

Csillag F Pasztor L Biehl LL 1993 Spectralband selection for the characterization ofsalinity status of soils Remote Sensing ofEnvironment 43 231ndash42

Dale MB McBratney AB and Russell JS1989 On the role of expert systems andnumerical taxonomy in soil classificationJournal of Soil Science 40 223ndash34

Davis JR 1993 Expert systems and environ-

soil survey In addition several authors have developed methods to developcontinuous soil property maps from the output of expert system models designed topredict soil unit occurrence

Expert systems are composed of data (information on spatial environmentalvariables eg topography climate etc) a knowledge base (rules and facts related tosoil variation supplied from the soil surveyor) and an inference engine (whichcombines data and the knowledge base to infer logically valid conclusions) (Skidmoreet al 1996) Expert systems differ from conventional models in two ways (1) they storeand manipulate qualitative information (allowing them access to information thatcannot normally be used in other modelling frameworks) and (2) they are structuredas meta-models (the knowledge is separated from the model) (Davis 1993) This allowsthe model to selectively choose which information is relevant at various stages of themodelling process and it allows for information to be easily updated Davis (1993)reviews the application of expert systems to environmental modelling research conclu-ding that the technique is becoming more widely accepted He further states that theapplication of expert systems is constrained by an absence of fundamental knowledgefor rule generation a problem that would appear less relevant to soil mapping giventhe amount of untapped expert knowledge accumulated by the soil surveyor

The first mention of the use of expert systems in pedology was in a paper presenta-tion at the Northeast Committee Soil Survey Conference 1984 (Flach 1985) In herpaper Flach (1985) hinted that recent developments in computer science especiallyexpert systems and artificial intelligence could make modelling a practical mappingtool for soil scientists in the near future A year later McCracken and Cate (1986) hopedto encourage soil scientists to explore expert systems and its application to soil sciencethrough an optimistic article they wrote on the potential use of expert systemsHowever little research was actually conducted in the late 1980s and expert systemshad not yet begun to fulfil the lofty goals set forth by early practitioners (Dale et al1989)

In the early 1990s expert systems approaches to predictive soil mapping began toappear in the literature Skidmore et al (1991) used a Bayesian expert system to mapforest soil into different classes and their results compared favourably with availablesoil maps and actual field-collected data Their methods successfully incorporatedsurveyor knowledge and remotely sensed and digital terrain attributes but failed tobetter characterize continuous soilndashlandscape variability because their final productwas a choropleth map Skidmore et al (1996) revisited their earlier research andassessed the mapping accuracy of their results They found that the soil map producedby the expert system achieved an overall accuracy of 698 (sample size = 53) while themap derived from conventional methods had an accuracy of 736 In addition toknowledge provided by a soil scientist Skidmore et alrsquos (1996) expert system usedvegetation derived for aerial photography and topographic variables derived from a 10-m DEM (wetness index topographic position and slope)

Using an expert system Cook et al (1996) successfully produced a continuous soilproperty map for organic matter content using wetness index aspect and slope asexplanatory variables Their methods were somewhat inefficient because they requireda separate expert system for each soil property of interest However their researchrepresents the only example of expert systems used to directly predict the spatial dis-tribution of a soil property All of the other examples of expert systems in the literature

P Scull et al 189

190 Predictive soil mapping a review

deal with soil type or class The use of expert systems to map soil properties needs tobe explored further

The expert systems discussed thus far have all used Boolean logic within theirtheoretical framework whereby an observation can belong to one and only one classand the soil properties of that class are assigned to the observation With thedevelopment of fuzzy logic and the semantic import model (the second fuzzy logicapproach to creating continuous classes) the opportunity exists to develop fuzzy logic-based expert systems Such systems can be used in conjunction with expert knowledgein situations where experts have a good qualitative idea of how to group data but havetrouble dealing with observations that are not well represented by rigid classificationschemes (Burrough and McDonnell 1998) This method can be particularly useful insituations when taxonomic schemes have been previously developed as is the case insoil taxonomy Several examples of this type of approach were published in the mid-1990s by A Zhu and colleagues (Zhu and Band 1994 Zhu et al 1996 1997 Zhu1997ab) Such systems proved useful for mapping soil at unvisited locations usingsurveyor knowledge and were also capable of producing continuous soil propertymaps The use of fuzzy logic within the theoretical framework of the expert systemallows the complex nature of soil to propagate through the modelling process neversubjected to classification schemes that filter out potentially useful lsquonoisersquo The use offuzzy logic also gives the soil surveyor more latitude during the interview processwhen the knowledge base is defined for the expert system Zhu et al (1997) developeda fuzzy logic-based expert system called SoLIM to determine the similarity of eachgrid cell in a study area to the various taxonomic mapping units delimited by the soilsurvey Continuous soil attribute maps were calculated using the similarity values andtheir relative soil survey determined attribute values (effectively achieving Goal 2) Theresulting data (which proved more accurate than soil survey data once field checked)consists of a raster grid whose resolution is determined by the resolution of the inputenvironmental and digital elevation data As noted previously this type of data modelis more applicable to environmental modelling than the choropleth map (Burrough andMcDonnell 1998)

Expert system approaches to PSM are capable of exploiting soil surveyor knowledgeby developing rule-based systems that imitate the surveyorrsquos conceptual model of soilvariability (the primary focus of Goal 3) The method would appear extremely usefulfor mapping projects (such as those conducted by the NRCS-NCSS) where fieldwork isinitially conducted to determine soilndashlandscape relations Expert system developmentcould be directly inserted into the traditional soil survey mapping approach as asubstitute for the step where the surveyor converts hisher conceptual model into achoropleth map Rather that knowledge could be incorporated into the expert systemwhich could be used to predictively map soil The resulting raster map would be morescientifically based and explicit than the hand-drawn choropleth maps of the past Itsscale would not be limited to that of the aerial photography but rather to the scale ofthe environmental data

Despite these advantages the expert system approach has some drawbacks Becausethe method is dependent upon expert knowledge it cannot be applied whereenvironmentndashsoil relations are poorly understood Of course this criticism can be madeof all PSM models as inductive or deductive knowledge of soilndashenvironment relationsis a prerequisite for PSM However expert systems are deductive models and as such

are not driven by any specific field-collected data (although presumably the soil experthas field experience in the mapping area) Expert systems donrsquot afford the opportunityto first statistically document landscapendashsoil relations and then extrapolate the resultsbecause expert systems do not directly use sample soil data (lsquohard datarsquo) to determinesoilndashlandscape relations Thus the utilization of the relationship between environmen-tal variables and soil properties (Goal 1) is only indirectly achieved Expert systemsapproaches have been demonstrated to be extremely effective in a small number of casestudies The possibility of satisfying all three goals discussed in this review makesexpert systems a predictive soil mapping method that needs to be further tested ndash bothgeographically and across different scales of analysis

V Conclusion

Most of the predictive soil mapping research outlined in this review was conducted atvery large map scales (over small areas) In fact the majority of the research wasconcerned with assessing the spatial variability of soil character within individual fieldsor across soil toposequences The primary driving force behind this type of research hasbeen the need to provide accurate soils information for agriculture and ecologicalmodels It is clear that terrain attributes are powerful predictors at the local scaleGeostatistical tools have been successful at using terrain attributes and the spatialdependence of soil properties to interpolate between existing data points withinindividual fields Across soil toposequences statistical approaches provide a usefulmeans of predicting soil character PSM research has been most successful at the fieldscale because many of the soil-forming factors are held constant For example the neteffect of four of Jennyrsquos five soil-forming factors (climate organism parent material andtime) was minimal within many of the studies reviewed While some research hastackled larger areas of study there exists a deficiency in our ability to predictively mapsoils at smaller map scales Since the distribution of soil is scale-dependent differentPSM methods and predictors are likely to work better at different scales Focus in thefuture must continue to move toward working over larger spatial extents of study inorder to produce landscape-scale soil information

Further a large proportion of the research was conducted in semi-arid gentlysloping agricultural landscapes Humid forestlands mountainous regions and desertshave received little attention As such PSM methods need to continue to be testedandor developed in a wider variety of landscapes where spatial soil distributions canbe more complex Different methods will likely be successful to different degrees indifferent environments Whereas terrain attributes are dominant predictors of soilcharacter across toposequences in prairie lands other predictors are likely necessary inareas where soils do not develop into clear toposequences For example remote sensingdata has been a useful predictor of soil chronosequences in desert landscapes wheresurface appearance is often related to soil character Surficial geology is often reflectiveof soil character in mountain regions where soils are thin and significant bedrock existsat the surface Focus in the future must continue to determine which methods andpredictors work best in which environments

Since the most useful PSM approach will vary across spatial scales and environmen-tal gradients the method used should be driven by the mapping objectives of the

P Scull et al 191

192 Predictive soil mapping a review

project Whereas the traditional soil survey of the past was expected to meet the needsof a diverse group of end users PSM methods will vary given the objectives of thesurvey As a result the end users of the soil data need to play a more active role in thesurvey process Because precision agriculture and large-scale mapping has been theprimary focus of PSM there exists a deficiency in our ability to predictively map soilsfor the purpose of general land use planning and management Even though lessdetailed soil information is needed for these purposes the development of PSMmethods is complicated by the fact that soils exhibit complex spatial variability atsmall map scales where soil-forming environments vary greatly from one location tothe next

A number of alternative methods of characterizing the continuous nature of the soillandscape have been developed Thus far most of the PSM research has provided soilinformation in a nonobject form (as opposed to defining soil types as independententities) Soil data have been generated organized and presented in the form of eitherisorithmic maps or fine-scale raster grids Both of these data models are field-viewmodels of geographic space which allow the soil to be perceived as a constantlyvarying surface Two distinct approaches have been employed (1) mapping individualsoil properties and (2) mapping continuously varying (fuzzy) soil classes Mappingindividual properties is the most common approach and will likely continue todominate PSM research The use of fuzzy soil classes in the literature is less commonand more difficult to be integrated into standard mapping procedures (such as use ofsoil taxa) because the concept radically differs from the traditional view of the soillandscape However fuzzy soil classes (generated using the fuzzy-k means approach) dohave the potential to help further advance Soil taxonomy by identifying taxonomicclasses that are more reflective of pedologic processes at work

Expert systems have been greatly underutilized in PSM research especiallyconsidering how effective a small number of case studies have been Expert systemshave the potential to satisfy successfully all three goals discussed throughout thereview They also have the potential to bridge the gap between traditional approachesand PSM methods because field soil scientists do not have to change their conceptualapproach to mapping They can still conduct field reconnaissance to determinesoilndashlandscape relationships Afterwards though they can use expert systems to betterexploit the knowledge they have garnered Such an approach could help ease thetransition from traditional soil survey to more scientifically explicit methods Expertsystems are also the most fruitful approach to utilizing a wealth of data that has alreadybeen collected in a nonexplicit manner the wealth of expert knowledge (in the form ofthe conceptual model) that senior soil mappers have accumulated

Focus in the future must also move in the direction of operability The utility of manyPSM approaches has been clearly demonstrated yet changes in how soils are mappedand perceived has been slow In the USA various members of the National CooperativeSoil Survey have called for a change in the philosophy of survey (from mappingdiscrete soil types to mapping continuous varying soil properties McSweeny et al1994) but little change has actually taken place To realize the potential of PSM in con-tributing to an overall change in standard soil mapping procedures practitionersworking in the field of PSM need to form working relationships with field soilscientists PSM methods need also to be presented in a manner that is comprehensibleto the soil science trained field mapper Predictive soil mapping is a relatively recent

phenomenon yet much progress has been made In the process of mapping soilknowledge will continue to advance

P Scull et al 193

References

Agbu PA Fehrenbacher DJ and Jansen IJ1990 Statistical comparison of SPOT spectralmaps with field soil maps Soil Science Society ofAmerica Journal 54 818ndash18

Barrett L 1999 Particulars in contextmaintaining a balance in soil geography Annalsof the Association of America Geographers 89707ndash13

Beckett PHT and Webster R 1971 Soilvariability a review Soils and Fertilizers 341ndash15

Bell JC Grigal DF and Bates PC 2000 A soil-terrain model for estimating spatial patterns ofsoil organic carbon In Wilson J and Gallant Jeditors Terrain analysis principles and applica-tions New York City NY John Wiley and Sons295ndash310

Birkeland PW 1999 Soil and geomorphologyThird edition New York NY Oxford UniversityPress

Breiman L Friedman JH Olshen RA andStone CJ 1984 Classification and regressiontrees Belmont CA Wadsworth

Brule FJ 1996 Fuzzy systems ndash a tutorialhttp newsgroup compai httpwwwquadralaycom (last accessed 30 August 2001)

Bunkin FV and Bunkin AF 2000 Lidarsounding of water soil and plants Atmosphericand Oceanic Optics 13 54ndash72

Buol SW Hole FD McCracken RJ andSouthard RJ 1997 Soil genesis and classifica-tion Ames IA Iowa State University Press

Burgess TM and Webster R 1980a Optimalinterpolation and isarithmic mapping of soilproperties the semi-variogram and punctualkriging Journal of Soil Science 31 315ndash31

ndashndashndashndash 1980b Optimal interpolation and isarithmicmapping of soil properties block kringingJournal of Soil Science 31 331ndash41

Burrough PA 1989 Fuzzy mathematicalmethods for soil survey and land evaluationJournal of Soil Science 40 477ndash92

Burrough PA and McDonnell RA 1998Principles of geographic information systems(Revised edition) Oxford Clarendon Press

Burrough PA Beckett PHT and Jarvis MG1971 The relation between cost and utility insoil survey Journal of Soil Science 22 368ndash81

Burrough PA Van Gaans PMF and

Hootsman R 1997 Continuous classificationin soil survey spatial correlation confusionand boundaries Geoderma 77 115ndash35

Cambell JB 1977 Variation of selectedproperties across a soil boundary Soil ScienceSociety of America Journal 41 578ndash82

Cambell JB and Edmonds WJ 1984 Themissing geographic dimension to soiltaxonomy Annals of the Association of AmericanGeographers 74 83ndash97

Castrignano A Giugliarini L Risaliti R andMartinelli N 2000 Study of spatial relation-ships among some soil physico-chemicalproperties of a field in central Italy using multi-variate geostatistics Geoderma 97 39ndash60

Cialella AT Dubayah R Lawrence W andLevine E 1997 Predicting soil drainage classusing remotely sensed and digital elevationdata Journal of Soil Science 62(2) 171ndash78

Clark RN and Swayze GA 1996 Evolution inimaging spectroscopy analysis and sensorsignal-to-noise an examination of how far wehave come Summaries of the sixth annual JPLairborne Earth science workshop 4ndash8 March 1996AVIRIS Workshop Vol 1 5

Clarke LA and Pregibon D 1992 Tree-basedmodels In Chambers J and Hastie J editorsStatistical models in S Pacific Grove Wadsworthand Brooks 377ndash419

Coleman TL Agbu PA and MontgomeryOL 1993 Spectral differentiation of surfacesoils and soil properties ndash is it possible fromspace platforms Soil Science 155 283ndash93

Cook SE Corner RJ Grealish G GesslerPE and Chartres CJ 1996 A rule-basedsystem to map soil properties Soil ScienceSociety of America Journal 60 1893ndash900

Cruickshank JG 1972 Soil geography New YorkNY John Wiley amp Sons

Csillag F Pasztor L Biehl LL 1993 Spectralband selection for the characterization ofsalinity status of soils Remote Sensing ofEnvironment 43 231ndash42

Dale MB McBratney AB and Russell JS1989 On the role of expert systems andnumerical taxonomy in soil classificationJournal of Soil Science 40 223ndash34

Davis JR 1993 Expert systems and environ-

190 Predictive soil mapping a review

deal with soil type or class The use of expert systems to map soil properties needs tobe explored further

The expert systems discussed thus far have all used Boolean logic within theirtheoretical framework whereby an observation can belong to one and only one classand the soil properties of that class are assigned to the observation With thedevelopment of fuzzy logic and the semantic import model (the second fuzzy logicapproach to creating continuous classes) the opportunity exists to develop fuzzy logic-based expert systems Such systems can be used in conjunction with expert knowledgein situations where experts have a good qualitative idea of how to group data but havetrouble dealing with observations that are not well represented by rigid classificationschemes (Burrough and McDonnell 1998) This method can be particularly useful insituations when taxonomic schemes have been previously developed as is the case insoil taxonomy Several examples of this type of approach were published in the mid-1990s by A Zhu and colleagues (Zhu and Band 1994 Zhu et al 1996 1997 Zhu1997ab) Such systems proved useful for mapping soil at unvisited locations usingsurveyor knowledge and were also capable of producing continuous soil propertymaps The use of fuzzy logic within the theoretical framework of the expert systemallows the complex nature of soil to propagate through the modelling process neversubjected to classification schemes that filter out potentially useful lsquonoisersquo The use offuzzy logic also gives the soil surveyor more latitude during the interview processwhen the knowledge base is defined for the expert system Zhu et al (1997) developeda fuzzy logic-based expert system called SoLIM to determine the similarity of eachgrid cell in a study area to the various taxonomic mapping units delimited by the soilsurvey Continuous soil attribute maps were calculated using the similarity values andtheir relative soil survey determined attribute values (effectively achieving Goal 2) Theresulting data (which proved more accurate than soil survey data once field checked)consists of a raster grid whose resolution is determined by the resolution of the inputenvironmental and digital elevation data As noted previously this type of data modelis more applicable to environmental modelling than the choropleth map (Burrough andMcDonnell 1998)

Expert system approaches to PSM are capable of exploiting soil surveyor knowledgeby developing rule-based systems that imitate the surveyorrsquos conceptual model of soilvariability (the primary focus of Goal 3) The method would appear extremely usefulfor mapping projects (such as those conducted by the NRCS-NCSS) where fieldwork isinitially conducted to determine soilndashlandscape relations Expert system developmentcould be directly inserted into the traditional soil survey mapping approach as asubstitute for the step where the surveyor converts hisher conceptual model into achoropleth map Rather that knowledge could be incorporated into the expert systemwhich could be used to predictively map soil The resulting raster map would be morescientifically based and explicit than the hand-drawn choropleth maps of the past Itsscale would not be limited to that of the aerial photography but rather to the scale ofthe environmental data

Despite these advantages the expert system approach has some drawbacks Becausethe method is dependent upon expert knowledge it cannot be applied whereenvironmentndashsoil relations are poorly understood Of course this criticism can be madeof all PSM models as inductive or deductive knowledge of soilndashenvironment relationsis a prerequisite for PSM However expert systems are deductive models and as such

are not driven by any specific field-collected data (although presumably the soil experthas field experience in the mapping area) Expert systems donrsquot afford the opportunityto first statistically document landscapendashsoil relations and then extrapolate the resultsbecause expert systems do not directly use sample soil data (lsquohard datarsquo) to determinesoilndashlandscape relations Thus the utilization of the relationship between environmen-tal variables and soil properties (Goal 1) is only indirectly achieved Expert systemsapproaches have been demonstrated to be extremely effective in a small number of casestudies The possibility of satisfying all three goals discussed in this review makesexpert systems a predictive soil mapping method that needs to be further tested ndash bothgeographically and across different scales of analysis

V Conclusion

Most of the predictive soil mapping research outlined in this review was conducted atvery large map scales (over small areas) In fact the majority of the research wasconcerned with assessing the spatial variability of soil character within individual fieldsor across soil toposequences The primary driving force behind this type of research hasbeen the need to provide accurate soils information for agriculture and ecologicalmodels It is clear that terrain attributes are powerful predictors at the local scaleGeostatistical tools have been successful at using terrain attributes and the spatialdependence of soil properties to interpolate between existing data points withinindividual fields Across soil toposequences statistical approaches provide a usefulmeans of predicting soil character PSM research has been most successful at the fieldscale because many of the soil-forming factors are held constant For example the neteffect of four of Jennyrsquos five soil-forming factors (climate organism parent material andtime) was minimal within many of the studies reviewed While some research hastackled larger areas of study there exists a deficiency in our ability to predictively mapsoils at smaller map scales Since the distribution of soil is scale-dependent differentPSM methods and predictors are likely to work better at different scales Focus in thefuture must continue to move toward working over larger spatial extents of study inorder to produce landscape-scale soil information

Further a large proportion of the research was conducted in semi-arid gentlysloping agricultural landscapes Humid forestlands mountainous regions and desertshave received little attention As such PSM methods need to continue to be testedandor developed in a wider variety of landscapes where spatial soil distributions canbe more complex Different methods will likely be successful to different degrees indifferent environments Whereas terrain attributes are dominant predictors of soilcharacter across toposequences in prairie lands other predictors are likely necessary inareas where soils do not develop into clear toposequences For example remote sensingdata has been a useful predictor of soil chronosequences in desert landscapes wheresurface appearance is often related to soil character Surficial geology is often reflectiveof soil character in mountain regions where soils are thin and significant bedrock existsat the surface Focus in the future must continue to determine which methods andpredictors work best in which environments

Since the most useful PSM approach will vary across spatial scales and environmen-tal gradients the method used should be driven by the mapping objectives of the

P Scull et al 191

192 Predictive soil mapping a review

project Whereas the traditional soil survey of the past was expected to meet the needsof a diverse group of end users PSM methods will vary given the objectives of thesurvey As a result the end users of the soil data need to play a more active role in thesurvey process Because precision agriculture and large-scale mapping has been theprimary focus of PSM there exists a deficiency in our ability to predictively map soilsfor the purpose of general land use planning and management Even though lessdetailed soil information is needed for these purposes the development of PSMmethods is complicated by the fact that soils exhibit complex spatial variability atsmall map scales where soil-forming environments vary greatly from one location tothe next

A number of alternative methods of characterizing the continuous nature of the soillandscape have been developed Thus far most of the PSM research has provided soilinformation in a nonobject form (as opposed to defining soil types as independententities) Soil data have been generated organized and presented in the form of eitherisorithmic maps or fine-scale raster grids Both of these data models are field-viewmodels of geographic space which allow the soil to be perceived as a constantlyvarying surface Two distinct approaches have been employed (1) mapping individualsoil properties and (2) mapping continuously varying (fuzzy) soil classes Mappingindividual properties is the most common approach and will likely continue todominate PSM research The use of fuzzy soil classes in the literature is less commonand more difficult to be integrated into standard mapping procedures (such as use ofsoil taxa) because the concept radically differs from the traditional view of the soillandscape However fuzzy soil classes (generated using the fuzzy-k means approach) dohave the potential to help further advance Soil taxonomy by identifying taxonomicclasses that are more reflective of pedologic processes at work

Expert systems have been greatly underutilized in PSM research especiallyconsidering how effective a small number of case studies have been Expert systemshave the potential to satisfy successfully all three goals discussed throughout thereview They also have the potential to bridge the gap between traditional approachesand PSM methods because field soil scientists do not have to change their conceptualapproach to mapping They can still conduct field reconnaissance to determinesoilndashlandscape relationships Afterwards though they can use expert systems to betterexploit the knowledge they have garnered Such an approach could help ease thetransition from traditional soil survey to more scientifically explicit methods Expertsystems are also the most fruitful approach to utilizing a wealth of data that has alreadybeen collected in a nonexplicit manner the wealth of expert knowledge (in the form ofthe conceptual model) that senior soil mappers have accumulated

Focus in the future must also move in the direction of operability The utility of manyPSM approaches has been clearly demonstrated yet changes in how soils are mappedand perceived has been slow In the USA various members of the National CooperativeSoil Survey have called for a change in the philosophy of survey (from mappingdiscrete soil types to mapping continuous varying soil properties McSweeny et al1994) but little change has actually taken place To realize the potential of PSM in con-tributing to an overall change in standard soil mapping procedures practitionersworking in the field of PSM need to form working relationships with field soilscientists PSM methods need also to be presented in a manner that is comprehensibleto the soil science trained field mapper Predictive soil mapping is a relatively recent

phenomenon yet much progress has been made In the process of mapping soilknowledge will continue to advance

P Scull et al 193

References

Agbu PA Fehrenbacher DJ and Jansen IJ1990 Statistical comparison of SPOT spectralmaps with field soil maps Soil Science Society ofAmerica Journal 54 818ndash18

Barrett L 1999 Particulars in contextmaintaining a balance in soil geography Annalsof the Association of America Geographers 89707ndash13

Beckett PHT and Webster R 1971 Soilvariability a review Soils and Fertilizers 341ndash15

Bell JC Grigal DF and Bates PC 2000 A soil-terrain model for estimating spatial patterns ofsoil organic carbon In Wilson J and Gallant Jeditors Terrain analysis principles and applica-tions New York City NY John Wiley and Sons295ndash310

Birkeland PW 1999 Soil and geomorphologyThird edition New York NY Oxford UniversityPress

Breiman L Friedman JH Olshen RA andStone CJ 1984 Classification and regressiontrees Belmont CA Wadsworth

Brule FJ 1996 Fuzzy systems ndash a tutorialhttp newsgroup compai httpwwwquadralaycom (last accessed 30 August 2001)

Bunkin FV and Bunkin AF 2000 Lidarsounding of water soil and plants Atmosphericand Oceanic Optics 13 54ndash72

Buol SW Hole FD McCracken RJ andSouthard RJ 1997 Soil genesis and classifica-tion Ames IA Iowa State University Press

Burgess TM and Webster R 1980a Optimalinterpolation and isarithmic mapping of soilproperties the semi-variogram and punctualkriging Journal of Soil Science 31 315ndash31

ndashndashndashndash 1980b Optimal interpolation and isarithmicmapping of soil properties block kringingJournal of Soil Science 31 331ndash41

Burrough PA 1989 Fuzzy mathematicalmethods for soil survey and land evaluationJournal of Soil Science 40 477ndash92

Burrough PA and McDonnell RA 1998Principles of geographic information systems(Revised edition) Oxford Clarendon Press

Burrough PA Beckett PHT and Jarvis MG1971 The relation between cost and utility insoil survey Journal of Soil Science 22 368ndash81

Burrough PA Van Gaans PMF and

Hootsman R 1997 Continuous classificationin soil survey spatial correlation confusionand boundaries Geoderma 77 115ndash35

Cambell JB 1977 Variation of selectedproperties across a soil boundary Soil ScienceSociety of America Journal 41 578ndash82

Cambell JB and Edmonds WJ 1984 Themissing geographic dimension to soiltaxonomy Annals of the Association of AmericanGeographers 74 83ndash97

Castrignano A Giugliarini L Risaliti R andMartinelli N 2000 Study of spatial relation-ships among some soil physico-chemicalproperties of a field in central Italy using multi-variate geostatistics Geoderma 97 39ndash60

Cialella AT Dubayah R Lawrence W andLevine E 1997 Predicting soil drainage classusing remotely sensed and digital elevationdata Journal of Soil Science 62(2) 171ndash78

Clark RN and Swayze GA 1996 Evolution inimaging spectroscopy analysis and sensorsignal-to-noise an examination of how far wehave come Summaries of the sixth annual JPLairborne Earth science workshop 4ndash8 March 1996AVIRIS Workshop Vol 1 5

Clarke LA and Pregibon D 1992 Tree-basedmodels In Chambers J and Hastie J editorsStatistical models in S Pacific Grove Wadsworthand Brooks 377ndash419

Coleman TL Agbu PA and MontgomeryOL 1993 Spectral differentiation of surfacesoils and soil properties ndash is it possible fromspace platforms Soil Science 155 283ndash93

Cook SE Corner RJ Grealish G GesslerPE and Chartres CJ 1996 A rule-basedsystem to map soil properties Soil ScienceSociety of America Journal 60 1893ndash900

Cruickshank JG 1972 Soil geography New YorkNY John Wiley amp Sons

Csillag F Pasztor L Biehl LL 1993 Spectralband selection for the characterization ofsalinity status of soils Remote Sensing ofEnvironment 43 231ndash42

Dale MB McBratney AB and Russell JS1989 On the role of expert systems andnumerical taxonomy in soil classificationJournal of Soil Science 40 223ndash34

Davis JR 1993 Expert systems and environ-

are not driven by any specific field-collected data (although presumably the soil experthas field experience in the mapping area) Expert systems donrsquot afford the opportunityto first statistically document landscapendashsoil relations and then extrapolate the resultsbecause expert systems do not directly use sample soil data (lsquohard datarsquo) to determinesoilndashlandscape relations Thus the utilization of the relationship between environmen-tal variables and soil properties (Goal 1) is only indirectly achieved Expert systemsapproaches have been demonstrated to be extremely effective in a small number of casestudies The possibility of satisfying all three goals discussed in this review makesexpert systems a predictive soil mapping method that needs to be further tested ndash bothgeographically and across different scales of analysis

V Conclusion

Most of the predictive soil mapping research outlined in this review was conducted atvery large map scales (over small areas) In fact the majority of the research wasconcerned with assessing the spatial variability of soil character within individual fieldsor across soil toposequences The primary driving force behind this type of research hasbeen the need to provide accurate soils information for agriculture and ecologicalmodels It is clear that terrain attributes are powerful predictors at the local scaleGeostatistical tools have been successful at using terrain attributes and the spatialdependence of soil properties to interpolate between existing data points withinindividual fields Across soil toposequences statistical approaches provide a usefulmeans of predicting soil character PSM research has been most successful at the fieldscale because many of the soil-forming factors are held constant For example the neteffect of four of Jennyrsquos five soil-forming factors (climate organism parent material andtime) was minimal within many of the studies reviewed While some research hastackled larger areas of study there exists a deficiency in our ability to predictively mapsoils at smaller map scales Since the distribution of soil is scale-dependent differentPSM methods and predictors are likely to work better at different scales Focus in thefuture must continue to move toward working over larger spatial extents of study inorder to produce landscape-scale soil information

Further a large proportion of the research was conducted in semi-arid gentlysloping agricultural landscapes Humid forestlands mountainous regions and desertshave received little attention As such PSM methods need to continue to be testedandor developed in a wider variety of landscapes where spatial soil distributions canbe more complex Different methods will likely be successful to different degrees indifferent environments Whereas terrain attributes are dominant predictors of soilcharacter across toposequences in prairie lands other predictors are likely necessary inareas where soils do not develop into clear toposequences For example remote sensingdata has been a useful predictor of soil chronosequences in desert landscapes wheresurface appearance is often related to soil character Surficial geology is often reflectiveof soil character in mountain regions where soils are thin and significant bedrock existsat the surface Focus in the future must continue to determine which methods andpredictors work best in which environments

Since the most useful PSM approach will vary across spatial scales and environmen-tal gradients the method used should be driven by the mapping objectives of the

P Scull et al 191

192 Predictive soil mapping a review

project Whereas the traditional soil survey of the past was expected to meet the needsof a diverse group of end users PSM methods will vary given the objectives of thesurvey As a result the end users of the soil data need to play a more active role in thesurvey process Because precision agriculture and large-scale mapping has been theprimary focus of PSM there exists a deficiency in our ability to predictively map soilsfor the purpose of general land use planning and management Even though lessdetailed soil information is needed for these purposes the development of PSMmethods is complicated by the fact that soils exhibit complex spatial variability atsmall map scales where soil-forming environments vary greatly from one location tothe next

A number of alternative methods of characterizing the continuous nature of the soillandscape have been developed Thus far most of the PSM research has provided soilinformation in a nonobject form (as opposed to defining soil types as independententities) Soil data have been generated organized and presented in the form of eitherisorithmic maps or fine-scale raster grids Both of these data models are field-viewmodels of geographic space which allow the soil to be perceived as a constantlyvarying surface Two distinct approaches have been employed (1) mapping individualsoil properties and (2) mapping continuously varying (fuzzy) soil classes Mappingindividual properties is the most common approach and will likely continue todominate PSM research The use of fuzzy soil classes in the literature is less commonand more difficult to be integrated into standard mapping procedures (such as use ofsoil taxa) because the concept radically differs from the traditional view of the soillandscape However fuzzy soil classes (generated using the fuzzy-k means approach) dohave the potential to help further advance Soil taxonomy by identifying taxonomicclasses that are more reflective of pedologic processes at work

Expert systems have been greatly underutilized in PSM research especiallyconsidering how effective a small number of case studies have been Expert systemshave the potential to satisfy successfully all three goals discussed throughout thereview They also have the potential to bridge the gap between traditional approachesand PSM methods because field soil scientists do not have to change their conceptualapproach to mapping They can still conduct field reconnaissance to determinesoilndashlandscape relationships Afterwards though they can use expert systems to betterexploit the knowledge they have garnered Such an approach could help ease thetransition from traditional soil survey to more scientifically explicit methods Expertsystems are also the most fruitful approach to utilizing a wealth of data that has alreadybeen collected in a nonexplicit manner the wealth of expert knowledge (in the form ofthe conceptual model) that senior soil mappers have accumulated

Focus in the future must also move in the direction of operability The utility of manyPSM approaches has been clearly demonstrated yet changes in how soils are mappedand perceived has been slow In the USA various members of the National CooperativeSoil Survey have called for a change in the philosophy of survey (from mappingdiscrete soil types to mapping continuous varying soil properties McSweeny et al1994) but little change has actually taken place To realize the potential of PSM in con-tributing to an overall change in standard soil mapping procedures practitionersworking in the field of PSM need to form working relationships with field soilscientists PSM methods need also to be presented in a manner that is comprehensibleto the soil science trained field mapper Predictive soil mapping is a relatively recent

phenomenon yet much progress has been made In the process of mapping soilknowledge will continue to advance

P Scull et al 193

References

Agbu PA Fehrenbacher DJ and Jansen IJ1990 Statistical comparison of SPOT spectralmaps with field soil maps Soil Science Society ofAmerica Journal 54 818ndash18

Barrett L 1999 Particulars in contextmaintaining a balance in soil geography Annalsof the Association of America Geographers 89707ndash13

Beckett PHT and Webster R 1971 Soilvariability a review Soils and Fertilizers 341ndash15

Bell JC Grigal DF and Bates PC 2000 A soil-terrain model for estimating spatial patterns ofsoil organic carbon In Wilson J and Gallant Jeditors Terrain analysis principles and applica-tions New York City NY John Wiley and Sons295ndash310

Birkeland PW 1999 Soil and geomorphologyThird edition New York NY Oxford UniversityPress

Breiman L Friedman JH Olshen RA andStone CJ 1984 Classification and regressiontrees Belmont CA Wadsworth

Brule FJ 1996 Fuzzy systems ndash a tutorialhttp newsgroup compai httpwwwquadralaycom (last accessed 30 August 2001)

Bunkin FV and Bunkin AF 2000 Lidarsounding of water soil and plants Atmosphericand Oceanic Optics 13 54ndash72

Buol SW Hole FD McCracken RJ andSouthard RJ 1997 Soil genesis and classifica-tion Ames IA Iowa State University Press

Burgess TM and Webster R 1980a Optimalinterpolation and isarithmic mapping of soilproperties the semi-variogram and punctualkriging Journal of Soil Science 31 315ndash31

ndashndashndashndash 1980b Optimal interpolation and isarithmicmapping of soil properties block kringingJournal of Soil Science 31 331ndash41

Burrough PA 1989 Fuzzy mathematicalmethods for soil survey and land evaluationJournal of Soil Science 40 477ndash92

Burrough PA and McDonnell RA 1998Principles of geographic information systems(Revised edition) Oxford Clarendon Press

Burrough PA Beckett PHT and Jarvis MG1971 The relation between cost and utility insoil survey Journal of Soil Science 22 368ndash81

Burrough PA Van Gaans PMF and

Hootsman R 1997 Continuous classificationin soil survey spatial correlation confusionand boundaries Geoderma 77 115ndash35

Cambell JB 1977 Variation of selectedproperties across a soil boundary Soil ScienceSociety of America Journal 41 578ndash82

Cambell JB and Edmonds WJ 1984 Themissing geographic dimension to soiltaxonomy Annals of the Association of AmericanGeographers 74 83ndash97

Castrignano A Giugliarini L Risaliti R andMartinelli N 2000 Study of spatial relation-ships among some soil physico-chemicalproperties of a field in central Italy using multi-variate geostatistics Geoderma 97 39ndash60

Cialella AT Dubayah R Lawrence W andLevine E 1997 Predicting soil drainage classusing remotely sensed and digital elevationdata Journal of Soil Science 62(2) 171ndash78

Clark RN and Swayze GA 1996 Evolution inimaging spectroscopy analysis and sensorsignal-to-noise an examination of how far wehave come Summaries of the sixth annual JPLairborne Earth science workshop 4ndash8 March 1996AVIRIS Workshop Vol 1 5

Clarke LA and Pregibon D 1992 Tree-basedmodels In Chambers J and Hastie J editorsStatistical models in S Pacific Grove Wadsworthand Brooks 377ndash419

Coleman TL Agbu PA and MontgomeryOL 1993 Spectral differentiation of surfacesoils and soil properties ndash is it possible fromspace platforms Soil Science 155 283ndash93

Cook SE Corner RJ Grealish G GesslerPE and Chartres CJ 1996 A rule-basedsystem to map soil properties Soil ScienceSociety of America Journal 60 1893ndash900

Cruickshank JG 1972 Soil geography New YorkNY John Wiley amp Sons

Csillag F Pasztor L Biehl LL 1993 Spectralband selection for the characterization ofsalinity status of soils Remote Sensing ofEnvironment 43 231ndash42

Dale MB McBratney AB and Russell JS1989 On the role of expert systems andnumerical taxonomy in soil classificationJournal of Soil Science 40 223ndash34

Davis JR 1993 Expert systems and environ-

192 Predictive soil mapping a review

project Whereas the traditional soil survey of the past was expected to meet the needsof a diverse group of end users PSM methods will vary given the objectives of thesurvey As a result the end users of the soil data need to play a more active role in thesurvey process Because precision agriculture and large-scale mapping has been theprimary focus of PSM there exists a deficiency in our ability to predictively map soilsfor the purpose of general land use planning and management Even though lessdetailed soil information is needed for these purposes the development of PSMmethods is complicated by the fact that soils exhibit complex spatial variability atsmall map scales where soil-forming environments vary greatly from one location tothe next

A number of alternative methods of characterizing the continuous nature of the soillandscape have been developed Thus far most of the PSM research has provided soilinformation in a nonobject form (as opposed to defining soil types as independententities) Soil data have been generated organized and presented in the form of eitherisorithmic maps or fine-scale raster grids Both of these data models are field-viewmodels of geographic space which allow the soil to be perceived as a constantlyvarying surface Two distinct approaches have been employed (1) mapping individualsoil properties and (2) mapping continuously varying (fuzzy) soil classes Mappingindividual properties is the most common approach and will likely continue todominate PSM research The use of fuzzy soil classes in the literature is less commonand more difficult to be integrated into standard mapping procedures (such as use ofsoil taxa) because the concept radically differs from the traditional view of the soillandscape However fuzzy soil classes (generated using the fuzzy-k means approach) dohave the potential to help further advance Soil taxonomy by identifying taxonomicclasses that are more reflective of pedologic processes at work

Expert systems have been greatly underutilized in PSM research especiallyconsidering how effective a small number of case studies have been Expert systemshave the potential to satisfy successfully all three goals discussed throughout thereview They also have the potential to bridge the gap between traditional approachesand PSM methods because field soil scientists do not have to change their conceptualapproach to mapping They can still conduct field reconnaissance to determinesoilndashlandscape relationships Afterwards though they can use expert systems to betterexploit the knowledge they have garnered Such an approach could help ease thetransition from traditional soil survey to more scientifically explicit methods Expertsystems are also the most fruitful approach to utilizing a wealth of data that has alreadybeen collected in a nonexplicit manner the wealth of expert knowledge (in the form ofthe conceptual model) that senior soil mappers have accumulated

Focus in the future must also move in the direction of operability The utility of manyPSM approaches has been clearly demonstrated yet changes in how soils are mappedand perceived has been slow In the USA various members of the National CooperativeSoil Survey have called for a change in the philosophy of survey (from mappingdiscrete soil types to mapping continuous varying soil properties McSweeny et al1994) but little change has actually taken place To realize the potential of PSM in con-tributing to an overall change in standard soil mapping procedures practitionersworking in the field of PSM need to form working relationships with field soilscientists PSM methods need also to be presented in a manner that is comprehensibleto the soil science trained field mapper Predictive soil mapping is a relatively recent

phenomenon yet much progress has been made In the process of mapping soilknowledge will continue to advance

P Scull et al 193

References

Agbu PA Fehrenbacher DJ and Jansen IJ1990 Statistical comparison of SPOT spectralmaps with field soil maps Soil Science Society ofAmerica Journal 54 818ndash18

Barrett L 1999 Particulars in contextmaintaining a balance in soil geography Annalsof the Association of America Geographers 89707ndash13

Beckett PHT and Webster R 1971 Soilvariability a review Soils and Fertilizers 341ndash15

Bell JC Grigal DF and Bates PC 2000 A soil-terrain model for estimating spatial patterns ofsoil organic carbon In Wilson J and Gallant Jeditors Terrain analysis principles and applica-tions New York City NY John Wiley and Sons295ndash310

Birkeland PW 1999 Soil and geomorphologyThird edition New York NY Oxford UniversityPress

Breiman L Friedman JH Olshen RA andStone CJ 1984 Classification and regressiontrees Belmont CA Wadsworth

Brule FJ 1996 Fuzzy systems ndash a tutorialhttp newsgroup compai httpwwwquadralaycom (last accessed 30 August 2001)

Bunkin FV and Bunkin AF 2000 Lidarsounding of water soil and plants Atmosphericand Oceanic Optics 13 54ndash72

Buol SW Hole FD McCracken RJ andSouthard RJ 1997 Soil genesis and classifica-tion Ames IA Iowa State University Press

Burgess TM and Webster R 1980a Optimalinterpolation and isarithmic mapping of soilproperties the semi-variogram and punctualkriging Journal of Soil Science 31 315ndash31

ndashndashndashndash 1980b Optimal interpolation and isarithmicmapping of soil properties block kringingJournal of Soil Science 31 331ndash41

Burrough PA 1989 Fuzzy mathematicalmethods for soil survey and land evaluationJournal of Soil Science 40 477ndash92

Burrough PA and McDonnell RA 1998Principles of geographic information systems(Revised edition) Oxford Clarendon Press

Burrough PA Beckett PHT and Jarvis MG1971 The relation between cost and utility insoil survey Journal of Soil Science 22 368ndash81

Burrough PA Van Gaans PMF and

Hootsman R 1997 Continuous classificationin soil survey spatial correlation confusionand boundaries Geoderma 77 115ndash35

Cambell JB 1977 Variation of selectedproperties across a soil boundary Soil ScienceSociety of America Journal 41 578ndash82

Cambell JB and Edmonds WJ 1984 Themissing geographic dimension to soiltaxonomy Annals of the Association of AmericanGeographers 74 83ndash97

Castrignano A Giugliarini L Risaliti R andMartinelli N 2000 Study of spatial relation-ships among some soil physico-chemicalproperties of a field in central Italy using multi-variate geostatistics Geoderma 97 39ndash60

Cialella AT Dubayah R Lawrence W andLevine E 1997 Predicting soil drainage classusing remotely sensed and digital elevationdata Journal of Soil Science 62(2) 171ndash78

Clark RN and Swayze GA 1996 Evolution inimaging spectroscopy analysis and sensorsignal-to-noise an examination of how far wehave come Summaries of the sixth annual JPLairborne Earth science workshop 4ndash8 March 1996AVIRIS Workshop Vol 1 5

Clarke LA and Pregibon D 1992 Tree-basedmodels In Chambers J and Hastie J editorsStatistical models in S Pacific Grove Wadsworthand Brooks 377ndash419

Coleman TL Agbu PA and MontgomeryOL 1993 Spectral differentiation of surfacesoils and soil properties ndash is it possible fromspace platforms Soil Science 155 283ndash93

Cook SE Corner RJ Grealish G GesslerPE and Chartres CJ 1996 A rule-basedsystem to map soil properties Soil ScienceSociety of America Journal 60 1893ndash900

Cruickshank JG 1972 Soil geography New YorkNY John Wiley amp Sons

Csillag F Pasztor L Biehl LL 1993 Spectralband selection for the characterization ofsalinity status of soils Remote Sensing ofEnvironment 43 231ndash42

Dale MB McBratney AB and Russell JS1989 On the role of expert systems andnumerical taxonomy in soil classificationJournal of Soil Science 40 223ndash34

Davis JR 1993 Expert systems and environ-

phenomenon yet much progress has been made In the process of mapping soilknowledge will continue to advance

P Scull et al 193

References

Agbu PA Fehrenbacher DJ and Jansen IJ1990 Statistical comparison of SPOT spectralmaps with field soil maps Soil Science Society ofAmerica Journal 54 818ndash18

Barrett L 1999 Particulars in contextmaintaining a balance in soil geography Annalsof the Association of America Geographers 89707ndash13

Beckett PHT and Webster R 1971 Soilvariability a review Soils and Fertilizers 341ndash15

Bell JC Grigal DF and Bates PC 2000 A soil-terrain model for estimating spatial patterns ofsoil organic carbon In Wilson J and Gallant Jeditors Terrain analysis principles and applica-tions New York City NY John Wiley and Sons295ndash310

Birkeland PW 1999 Soil and geomorphologyThird edition New York NY Oxford UniversityPress

Breiman L Friedman JH Olshen RA andStone CJ 1984 Classification and regressiontrees Belmont CA Wadsworth

Brule FJ 1996 Fuzzy systems ndash a tutorialhttp newsgroup compai httpwwwquadralaycom (last accessed 30 August 2001)

Bunkin FV and Bunkin AF 2000 Lidarsounding of water soil and plants Atmosphericand Oceanic Optics 13 54ndash72

Buol SW Hole FD McCracken RJ andSouthard RJ 1997 Soil genesis and classifica-tion Ames IA Iowa State University Press

Burgess TM and Webster R 1980a Optimalinterpolation and isarithmic mapping of soilproperties the semi-variogram and punctualkriging Journal of Soil Science 31 315ndash31

ndashndashndashndash 1980b Optimal interpolation and isarithmicmapping of soil properties block kringingJournal of Soil Science 31 331ndash41

Burrough PA 1989 Fuzzy mathematicalmethods for soil survey and land evaluationJournal of Soil Science 40 477ndash92

Burrough PA and McDonnell RA 1998Principles of geographic information systems(Revised edition) Oxford Clarendon Press

Burrough PA Beckett PHT and Jarvis MG1971 The relation between cost and utility insoil survey Journal of Soil Science 22 368ndash81

Burrough PA Van Gaans PMF and

Hootsman R 1997 Continuous classificationin soil survey spatial correlation confusionand boundaries Geoderma 77 115ndash35

Cambell JB 1977 Variation of selectedproperties across a soil boundary Soil ScienceSociety of America Journal 41 578ndash82

Cambell JB and Edmonds WJ 1984 Themissing geographic dimension to soiltaxonomy Annals of the Association of AmericanGeographers 74 83ndash97

Castrignano A Giugliarini L Risaliti R andMartinelli N 2000 Study of spatial relation-ships among some soil physico-chemicalproperties of a field in central Italy using multi-variate geostatistics Geoderma 97 39ndash60

Cialella AT Dubayah R Lawrence W andLevine E 1997 Predicting soil drainage classusing remotely sensed and digital elevationdata Journal of Soil Science 62(2) 171ndash78

Clark RN and Swayze GA 1996 Evolution inimaging spectroscopy analysis and sensorsignal-to-noise an examination of how far wehave come Summaries of the sixth annual JPLairborne Earth science workshop 4ndash8 March 1996AVIRIS Workshop Vol 1 5

Clarke LA and Pregibon D 1992 Tree-basedmodels In Chambers J and Hastie J editorsStatistical models in S Pacific Grove Wadsworthand Brooks 377ndash419

Coleman TL Agbu PA and MontgomeryOL 1993 Spectral differentiation of surfacesoils and soil properties ndash is it possible fromspace platforms Soil Science 155 283ndash93

Cook SE Corner RJ Grealish G GesslerPE and Chartres CJ 1996 A rule-basedsystem to map soil properties Soil ScienceSociety of America Journal 60 1893ndash900

Cruickshank JG 1972 Soil geography New YorkNY John Wiley amp Sons

Csillag F Pasztor L Biehl LL 1993 Spectralband selection for the characterization ofsalinity status of soils Remote Sensing ofEnvironment 43 231ndash42

Dale MB McBratney AB and Russell JS1989 On the role of expert systems andnumerical taxonomy in soil classificationJournal of Soil Science 40 223ndash34

Davis JR 1993 Expert systems and environ-

194 Predictive soil mapping a review

mental modelling In Jakeman AJ Beck MBand McAleer MJ editors Modelling change inenvironmental systems New York NY JohnWiley and Sons Ltd 3ndash35

De Gruijter JJ Walvoort DJJ and Van GaansPFM 1997 Continuous soil maps ndash a fuzzy setapproach to bridge the gap betweenaggregation levels of process and distributionmodels Geoderma 77 169ndash95

Dijkerman JC 1974 Pedology as a science therole of data models and theories in the studyof natural soil systems Geoderma 11 73ndash93

Dmitriev EA 1983 Continuity of soils and theproblem of soild classification MoscowUniversity Soil Science Bulletin 38 1ndash10

Ellis F 1996 The application of machine learningtechniques to erosion modelling InProceedings third international conference onintegrating GIS and environmental modelingSanta Fe NM 16ndash21 January 1996 httpwwwncgiaucsbeduconfSANTA_FE_CD-ROMmainhtml (last accessed 21 January2003) National Center for GeographicInformation and Analysis Santa Barbara CA

Engman ET and Chauhan N 1995 Status ofmicrowave soil moisture measurements withremote sensing Remote Sensing of Environment51 189ndash98

Fang Y 2000 DEM generation from multi-sensorSAR images International Archives ofPhotogrammetry and Remote Sensing 33 686ndash93

Flach KW 1985 Modeling and soil survey SoilSurvey Horizons 26 15ndash20

Franklin J 1995 Predictive vegetation mappinggeographic modeling of biospatial patterns inrelation to environmental gradients Progress inPhysical Geography 19 474ndash90

ndashndashndashndash 1998 Predicting the distributions of shrubspecies in California chaparral and coastal sagecommunities from climate and terrain-derivedvariables Journal of Vegetation Science 9 733ndash48

Franklin J McCullough P and Gray C 2000Terrain variables for predictive mapping ofvegetation communities in Southern CaliforniaIn Wilson J and Gallant J editors Terrainanalysis principles and applications New YorkCity NY John Wiley and Sons 331ndash53

Friedl MA and Brodley CE 1997 Decision treeclassification of land cover from remotelysensed data Remote Sensing of Environment 61399ndash409

Gessler PE 1996 Statistical soilndashlandscapemodelling for environmental managementDoctoral Dissertation The Australian National

University Canberra AustraliaGessler PE Moore ID McKensie NJ andRyan PJ 1995 Soil-landscape modelling andspatial prediction of soil attributes InternationalJournal Geographical Information Science 9421ndash32

Goetz AFH 1989 Spectral remote sensing ingeology In Asrar G editor Theory and applica-tions of optical remote sensing New York NYJohn Wiley and Sons 491ndash526

Goetz AFH Vane G Solomon JE and RockBN 1985 Imaging spectrometry for earthremote sensing Science 228 1147ndash53

Goodchild MF 1992a Geographical datamodeling Computers and Geosciences 18401ndash408

ndashndashndashndash 1992b Geographical information scienceInternational Journal Geographical InformationSystems 6 31ndash45

ndashndashndashndash 1994 Intergrating GIS and remote sensingfor vegetation analysis and modeling method-ological issues Journal of Vegetation Science 5615ndash26

Goovaerts P 1992 Factorial kriging analysis auseful tool for exploring the structure of multi-variate spatial soil information Journal of SoilScience 43 597ndash619

ndashndashndashndash 1997 Geostatistics for natural resourceevaluation New York City NY OxfordUniversity Press

Hall CAS and Olsen CG 1991 Predictingvariability of soil from landscape models InSpatial variability of soil and landforms SoilScience Society of America Special Publication28 9ndash24

Hartemink AE McBratney AB and CattleJA 2001 Developments and trends in soilscience 100 volumes of Geoderma 1967ndash2001Geoderma 100 217ndash68

Henderson TL Baumgardner MFFranzmeier DP Stott DE and Coster DC1992 High dimensional reflectance analysis ofsoil organic matter Soil Science Society ofAmerica Journal 56 865ndash72

Heuvelink GBM and Webster R 2001Modelling soil variation past present andfuture Geoderma 100 269ndash301

Hewitt AE 1993 Predictive modelling in soilsurvey Soil and Fertilizers 56 305ndash14

Horvath EH Post DF and Kelsey JB 1984The relationships of Landsat digital data to theproperties of Arizona rangelands Soil ScienceSociety of America Journal 48 1331ndash34

Hudson BD 1992 The soil survey as paradigmbased science Soil Science Society of AmericaJournal 56 836ndash41

P Scull et al 195

Huggett RJ 1975 Soil landscape systems amodel of soil genesis Geoderma 13 1ndash22

Indorante SJ McLeese RL Hammer RDThompson BW and Alexander DL 1996Positioning soil survey for the 21st centuryJournal of Soil and Water Conservation JanndashFeb21ndash28

Irons JR Weismiller RA and Petersen GW1989 Soil reflectance In Asrar G editor Theoryand applications of optical remote sensing NewYork NY John Wiley and Sons 66ndash106

Irvin BJ Ventura SJ and Slater BK 1996Landform classification for soilndashlandscapestudies In Proceedings third internationalconference on integrating GIS and environmentalmodeling Santa Fe NM 16ndash21 January 1996http wwwncgiaucsbeduconfSANTA_FE_CD-ROMmainhtml (last accessed 21January 2003) National Center for GeographicInformation and Analysis Santa Barbara CA

Jenny H 1941 Factors of soil formation New YorkNY McGraw-Hill

Johnson DL and Watson-Stegner D 1987Evolution model of pedogenesis Soil Science143 349ndash66

Johnson PE Smith MO Taylor-George Sand Adams JB 1983 A semiempirical methodfor analysis of the reflectance spectra of binarymineral mixtures Journal of Geophysical Research88 3557ndash61

Kemp KK 1992 Fields as a framework forintegrating GIS and environmental processmodels Part one representing spatialcontinuity Transactions in GIS 13 219ndash34

King D Bourennane H Isambert M andMacaire JJ 1999 Relationship of the presenceof a noncalcareous clay-loam horizon to DEMattributes in a gently sloping area Geoderma 8995ndash111

Kleshchenko VN Komarov SA MironovVL and Romanov AN 2000 Microwaveremote sensing of soil cover Proceedings ndash SPIEthe International Society for Optical Engineering4341 351ndash57

Knotters M Brus DJ and Oude Voshaar JH1995 A comparison of kriging co-kriging andkriging combined with regression for spatialinterpolation of horizon depth with censoredobservations Geoderma 67 227ndash46

Krige DG 1963 Two dimensional weightedmoving average trend surfaces for ore-evaluation Journal of the South AfricanInstitution of Mining and Metallurgy 66 13ndash38

Lagacherie P and Holmes S 1997 Addressinggeographical data errors in a classification treefor soil unit prediction International JournalGeographical Information Science 11 183ndash98

Laslett GM McBratney AB Pahl PJ andHutchinson MF 1987 Comparison of severalspatial prediction methods for soil pH Journalof Soil Science 38 325ndash41

Laymon CA Crosson WL Jackson TJManu A and Tsegaye TD 2001 Ground-based passive microwave remote sensingobservations of soil moisture at s-band and l-band with insight into measurement accuracyIEEE Transactions of Geoscience and RemoteSensing 39 1844ndash58

Lees BG and Ritman AK 1991 Decision-treeand rule induction approach to integration ofremotely sensed and GIS data in mappingvegetation in disturbed or hilly environmentsEnvironmental Management 15 823ndash31

Lillesand TM and Ralph Kiefer R 1994 Remotesensing and image processing New York NY JohnWiley and Sons

Mackay DS and Band LE 1998 Extraction andrepresentation of nested catchment areas fromdigital elevation models in lake-dominatedtopography Water Resources Research 34897ndash904

Matheron G 1963 Principals of geostatisticsEconomic Geology 58 1246ndash66

McBratney AB 1992 On variation uncertaintyand informatics in environmental soilmanagement Australian Journal of Soil Research30 913ndash35

McBratney AB and De Gruijter JJ 1992 Acontinuum approach to soil classification bymodified fuzzy k-means with extragradesJournal of Soil Science 43 159ndash75

McBratney AB and Odeh IOA 1997Application of fuzzy sets in soil science fuzzylogic fuzzy measurement and fuzzy decisionsGeoderma 77 85ndash113

McBratney AB Hart GA and McGarry D1991 The use of region partitioning to improvethe representation of geostatistically mappedsoil attributes Journal of Soil Science 42 513ndash32

McBratney AB Odeh IOA Bishop TFADunbar MS and Shatar TM 2000 Anoverview of pedometric techniques for use insoil survey Geoderma 97 293ndash327

McCracken RJ and Cate RB 1986 Artificialintelligence cognitive science andmeasurement theory applied in soil classifica-tion Soil Society of America Journal 50 557ndash61

McKensie NJ and Austin MP 1993 A quanti-

196 Predictive soil mapping a review

tative Australian approach to medium andsmall scale surveys based on soil stratigraphyand environmental correlations Geoderma 57329ndash55

McKensie MJ and Ryan PJ 1999 Spatialprediction fo soil properties using environmen-tal correlation Geoderma 89 67ndash94

McKensie NJ Gessler PE Ryan PJ andOrsquoConnell D 2000 The role of terrain analysisin soil mapping In Wilson J and Gallant Jeditors Terrain analysis principles and applica-tions New York City NY John Wiley and Sons245ndash65

McKnight TL 1993 Physical geography alandscape appreciation Fourth editionEnglewood Cliffs NJ Prentice Hall

McSweeney K Gessler PE Hammer J Bell Jand Peterson GW 1994 Towards a newframework for modeling the soil landscapecontinuum In Bryant RB and Arnold RWeditors Factors of soil formation a fiftiethanniversary retrospective Soil Science Society ofAmerica Special Publication 39 Madison WISoil Science Society of America 127ndash45

Metternicht GI 1998 Fuzzy classification ofJERS-1 SAR data an evaluation of its per-formance for soil salinity mapping EcologicalModelling 111 61ndash74

Michaelsen J Schimel DS Friedl MADavis FW and Dubayah RC 1994Regression tree analysis of satellite and terraindata to guide vegetation sampling and surveysJournal of Vegetation Science 5 673ndash86

Moore DM Lees BG and Davey SM 1991 Anew method for predicting vegetation distribu-tions using decision tree analysis in ageographic information system EnvironmentalManagement 15 59ndash71

Moore ID Grayson RB and Ladson AR1991 Digital terrain modelling a review ofhydrological geomorphological and biologicalapplications Hydrological Processes 5 3ndash30

Moore ID Gessler PE Nielsen GA andPeterson GA 1993 Soil attribute predictionusing terrain analysis Soil Science Society ofAmerica Journal 57 443ndash520

Narayanan RM and Hirsave PP 2001 Soilmoisture estimation models using SIR-C SARdata a case study in New Hampshire USARemote Sensing of Environment 75 385ndash96

Odeh IOA Chittleborough DJ andMcBratney AB 1991 Elucidation of soil-landform interrelationships by canonicalordination analysis Geoderma 49 1ndash32

Odeh IOA McBratney AB and

Chittleborough DJ 1992a Fuzzy-c-meansand kriging for mapping soil as a continuoussystem Soil Science Society America Journal 561848ndash54

ndashndashndashndash 1992b Soil pattern recognition with fuzzy-c-means Application to classification and soil-landform interrelationships Soil Science SocietyAmerica Journal 56 505ndash16

ndashndashndashndash 1994 Spatial prediction of soil propertiesfrom landform attributes derived from a digitalelevation model Geoderma 63 197ndash214

ndashndashndashndash 1995 Further results on prediction of soilproperties from terrain attributes heterotopiccokriging and regression-kriging Geoderma 67215ndash26

Okin GS Okin WJ Roberts DA andMurray BC 1998 Multiple endmemberspectral mixture analysis applications to anaridsemi-arid landscape Proceedings 7thAVIRIS Earth science workshop Pasadena CAJPL 8

Oliveira H 2000 Segmentation and classifica-tion of Landsat-TM images to monitor the soiluse International Archives of Photogrammetry andRemote Sensing 33 1065ndash72

Palacios-Orueta A and Ustin SL 1996Multivariate classification of soil spectraRemote Sensing of Environment 57 108ndash18

Palacios-Orueta A Pinzon JE Ustin SL andRoberts DA 1998 Remote sensing of soils inthe Santa Monica Mountains II Hierarchicalforeground and background analysis RemoteSensing of Environment 68 138ndash51

Quinlan JR 1993 C45 Algorithms for machinelearning San Mateo Morgan Kaufmann

Roberts DA Gardner M Church R UstinS Scheer G and Green RO 1998 Mappingchaparral in the Santa Monica Mountains usingmultiple endmember spectral mixture modelsRemote Sensing of Environment 65 267ndash79

Runge ECA 1973 Soil development sequencesand energy models Soil Science 115 183ndash93

Safavian SJ and Norvig P 1991 A survey oftree classifier methodology IEEE TransactionsSystems Man Cybernetics 21 660ndash74

Schaetzl RJ 1991 Factors affecting theformation of dark thick epipedons beneathforest vegetation Michigan USA Journal of SoilScience 42 501ndash12

Seyler F Bernoux M and Cerri CC 1998Landsat TM image texture and moisturevariations of the soil surface under the rain-forest of Rondonia state Brazil International

P Scull et al 197

Journal of Remote Sensing 19 1299ndash318Shipman H and Adams JB 1987 Detectabilityof minerals in desert alluvial fans usingreflectance spectra Journal Geophysical Research92 10391ndash492

Simonson RW 1959 Outline of a generalizedtheory of soil genesis Soil Science Society ofAmerica Proceedings 23 152ndash56

ndashndashndashndash 1997 History of soil science Early teachingin USA of Dokuchaiev factors of soil formationSoil Science Society of America Journal 61 11ndash16

Skidmore AK Ryan PJ Dawes W Short Dand Orsquoloughlin E 1991 Use of an expertsystem to map forest soil from a geographicalinformation system International JournalGeographical Information System 5 431ndash45

Skidmore AK Watford F Luckananurug Pand Ryan PJ 1996 An operational GIS expertsystem for mapping forest soil PhotogrammetricEngineering and Remote Sensing 62 501ndash11

Soil Survey Staff 1975 Soil taxonomy a basicsystem of soil classification for making and inter-preting soil surveys Agriculture handbook 436Washington DC Soil Conservation ServiceUS Department of Agriculture

Stoner ER Baumgardner MF WeismillerRA Biehl LL and Robinson BF 1980Extension of laboratory-measured soil spectrato field conditions Soil Science Society of AmericaJournal 44 572ndash74

Ventura SJ and Irvin BJ 1996 Automatedlandform classification methods for soillandscape studies In Proceedings third interna-tional conference on integrating GIS and environ-mental modeling Santa Fe NM 16ndash21 January1996 http wwwncgiaucsbeduconfSANTA_FE_CD-ROMmainhtml (lastaccessed 21 January 2003) National Center forGeographic Information and Analysis SantaBarbara CA

Voltz M and Webster R 1990 A comparison ofkriging cubic splines and classification forpredicting soil porperties from sampleinformation Journal of Soil Science 41 473ndash90

Wanchang Z Yamaguchi Y and Ogawa K2000 Evaluation of the effect of pre-processing

of the remotely sensed data on the actualevapotranspiration surface soil moisturemapping by an approach using Landsat DEMmeteorological data Geocarto International 1557ndash68

Webster R 1994 The development ofpedometrics Geoderma 62 1ndash15

Webster R and Beckett PHT 1968 Quality anusefulness of soil maps Nature 219 680ndash82

Webster R and Oliver MA 1992 Sampleadequately to estimate variograms of soilproperties Journal of Soil Science 43 177ndash92

Weibel R and Heller M 1991 Digital terrainmodeling In Goodchild MF and Rhind Deditors Geographic information systemsprinciples and applications New York NY Taylorand Francis 269ndash97

Wilcox CH Frazier BE and Ball ST 1994Relationship between soil organic carbon andLandsat TM data in eastern WashingtonPhotogrammetric Engineering and Remote Sensing60 777ndash81

Wilson J and Gallant J editors 2000 Terrainanalysis principles and applications New YorkNY John Wiley and Sons

Wright RL and Wilson SR 1979 On theanalysis of soil variability with an examplefrom Spain Geoderma 22 297ndash313

Zadeh LA 1965 Fuzzy sets Information andControl 8 338ndash53

Zhu A 1997a Measuring uncertainty in classassignment for natural resource maps underfuzzy logic Photogrammetric Engineering andRemote Sensing 63 1195ndash202

ndashndashndashndash 1997b A similarity model for representingsoil spatial information Geoderma 77 217ndash42

Zhu A and Band L 1994 A knowledge-basedapproach to data integration for soil mappingCanadian Journal of Remote Sensing 20 408ndash18

Zhu A Band L Dutton B and Nimlos TJ1996 Automated soil inference under fuzzylogic Ecological Modelling 90 123ndash45

Zhu A Band L Vertessy R and Dutton B1997 Derivation of soil properties using a soilland inference model SoLIM Soil ScienceSociety America Journal 61 523ndash33