Land-Cover-Change Trajectories in Southern Cameroondusk.geo.orst.edu/prosem/PDFs/bosley_landcover.pdf · 2005-01-11 · Land-Cover-Change Trajectories 469 of von Thünen and Ricardo

Annals of the Association of American Geographers

, 90(3), 2000, p. 467–494© 2000 by Association of American GeographersPublished by Blackwell Publishers, 350 Main Street, Malden, MA 02148, and 108 Cowley Road, Oxford, OX4 1JF, UK.

Land-Cover-Change Trajectoriesin Southern Cameroon

Benoît Mertens and Eric F. Lambin

Department of Geography, Université Catholique de Louvain

The objective of this study is to better understand the complexity of deforestation processes insouthern Cameroon by testing a multivariate, spatial model of land-cover change trajectories as-sociated with deforestation. The spatial model integrates a spectrum of independent variablesthat characterize land rent on a spatially explicit basis. The use of a time series of high-spatial-resolution remote sensing images (Landsat MSS and SPOT XS), spanning two decades, allows athorough validation of spatial projections of future deforestation. Remote sensing observationsreveal a continuous trend of forest clearing and forest degradation in southern regions of Cam-eroon, but with a highly fluctuating rate. A significant proportion of the areas subject to a land-cover conversion experienced other changes in the following years. The study also demonstratesthat modeling land-cover change trajectories over several observation years allows a better pro-jection of areas with a high probability of change in land-cover than projecting such areas on thebasis of observations from the previous time period alone. Statistical results suggest that, in oursouthern Cameroon study area, roads mostly increased the accessibility of the forest for migrantsrather than providing incentives for a transformation of local subsistence agriculture into market-oriented farming systems. The spatial model developed in this study allows simulations of likelyimpacts of human actions, leading to a transformation of the landscape (e.g., road projects) on keylandscape attributes (e.g., biodiversity). Currently, several road projects or major logging conces-sions exist in southern Cameroon.

Key Words: deforestation, land-use change, Africa, geographic in-formation systems, landscape model.

nderstanding processes of land-use andland-cover change has always been animportant research area in geography.

This research is receiving renewed interest thanksto a combination of new demands for under-standing anthropogenic surface processes (Meyerand Turner 1994) and the maturation of tech-nologies used to collect and analyze largeamounts of spatially explicit data on the landsurface (Lambin 1997). Enhanced monitoringand modeling abilities allow researchers to im-prove the detailed spatial modeling of land-useand land-cover changes while still achieving thelevel of generality that is required for a globalperspective on such changes.

Land-cover changes are often conceived assimple and irreversible conversions from onecover type to another. In this idealized represen-tation, deforestation would be a total and per-manent change from a dense forest landscape toan area with a low tree cover, and agriculturalexpansion would mean the transition to perma-

nent cultivation. While such neat schemes wouldmake the monitoring and modeling of land-cover changes a straightforward task, reality ismore complex. Land-cover changes are most of-ten noncontinuous in space, leading to complexlandscape mosaics and mixtures of cover types.They are also reversible, and often follow timesequences of successive cover types (or mixtureof cover types). These sequences can be linear(e.g., from dense forest to smallholder agricul-ture to large ranches to degraded land) or cycli-cal, as in the case of long-fallow agriculture. It isnecessary to integrate this complexity into theanalysis of land-cover-change processes so as toacquire a better understanding of the causes ofchange and to better predict likely evolutionsbased on such models.

Most previous statistical analyses of the causesof deforestation have thus far used, as a depen-dent variable, a simple binary measure such asdeforestation versus no deforestation, or forestversus nonforest (see reviews by Brown and

U

468 Mertens and Lambin

Pearce 1994; Lambin 1994). The assumption insuch studies is that deforestation is permanentand that the presence of nonforest at one timeimplies that forest will remain absent for a longperiod. A few authors, analyzing time series ofremote sensing data, have instead emphasizedthe reversibility of land-cover conversions, whichin theory should lead to complex trajectories ofchange (Lucas et al. 1993; Alves and Skole1996). This issue is particularly important in theanalysis of secondary forests. Such forests accu-mulate biomass more rapidly than primary onesand therefore act as a net sink for atmosphericcarbon (Brown and Lugo 1990). Another exam-ple of the reversibility of land-cover changes isfound in semiarid regions that are highly vul-nerable to droughts, but that also exhibit consid-erable resiliency (Tucker et al. 1991).

The concept of

trajectories of change

hasattracted some attention from a theoreticalviewpoint. These trajectories, defined as trendsover time among the relationships betweenthe factors that shape the changing nature ofhuman-environment relations and their effectswithin a particular region (Kasperson et al.1995), take widely different forms and dependon circumstances, regional contexts, and gov-ernment policies. Lambin (1997) suggested thatgeneric paths of change can be identified, suchas the typical sequences of land-use changesfound across tropical regions. Trajectories ofchange were also analyzed as part of thelong-run processes of agricultural intensificationdriven by demographic phenomena, as de-scribed by Boserup (1965). Such trajectorieshave been characterized as defined by the stockof environmental resources and human well-being (Karshenas 1994), or as a function of time,in terms of degree of sustainability of human-environment relations (Kasperson et al. 1995).In this study, trajectories of land-cover changerefer to successions of land-cover types for agiven sampling unit over more than two obser-vation years.

The objective of this study is to better under-stand the land-use change trajectories associ-ated with different deforestation processes insouthern Cameroon. We test a multivariate spa-tial model of land-cover-change trajectoriesassociated with deforestation. Previous studieshave ignored complex sequences of land-coverchanges, as they simply measured the conver-sions from one category to another (Sader andJoyce 1988; Gastellu-Etchegorry and Sinulingga

1988; Ludeke et al. 1990; Brown et al. 1993; Liuet al. 1993; Sader 1995; Mertens and Lambin1997). Chomitz and Gray (1996) developed amultinomial logit model, with three potentialoutcomes (i.e., natural vegetation, commercialagriculture, and semisubsistence agriculture).All these previous studies measured changesin and-cover between only two time periods.In this paper, we test whether modeling land-cover-change trajectories over a sequence ofobservation years allows a better projectionof areas with a high probability of change inland-cover than on the basis of observationsfrom only the previous time period. In otherwords, do we better represent deforestation pro-cesses if we consider that land-use changes arenot first-order processes—i.e., if the probabilityof a given land use at any time does not onlydepend upon the most recent use, but also uponeven earlier ones?

After defining the conceptual model onwhich our analyses are based, this work consistsin the following steps: (1) measurement of thedependent variable, which is defined as sequencesof land-cover changes, using a time series ofhigh-spatial-resolution remote sensing images,spanning two decades; (2) measurements of in-dependent variables that characterize as realisti-cally as possible the determinants of land-use ona spatially explicit basis; (3) development ofmultivariate spatial models of land-cover con-versions and land-cover change trajectories; and(4) validation of spatial projections of areasat risk of deforestation, using recent remote-sensing data. The major improvements, com-pared to our previous study on the same region(Mertens and Lambin 1997), consist in the na-ture of the dependent variable, the integrationof a broader spectrum of independent variables,the multivariate model, the validation of themodel predictions, and the proposed applicationof the model’s projections to assess potential im-pacts of land-cover changes.

Background

Economic Models of Deforestation

In econometric approaches to land-use changes,the supply and demand functions of the landmarket, which is assumed to be competitive, areestimated. Most economic modeling of land-usechanges originates from the land-rent theories

Land-Cover-Change Trajectories 469

of von Thünen and Ricardo. Models of urbanand periurban land allocation are much moredeveloped than their rural counterpart (Rieb-same et al. 1994). Any parcel of land, given itsattributes and location, is assumed to be allo-cated to the use that earns the highest rent. De-forestation, for example, is driven by choices byland managers among alternative rents. Landconservation depends on investment decisionsby land managers, based on comparisons betweendiscounted benefits and costs. Such models allowinvestigation of the influence of various policymeasures on land allocation choices (Lambin1997).

Microeconomic models usually assume thatthe agents whose behavior is described withinthe model have the capacity to make informedpredictions and plans, and that they are riskminimizers (Fischer et al. 1996). After exploringall options available to them, individuals makerational decisions based on available informa-tion, obligations, and expectations (social as wellas economic), so as to balance anticipated re-turns and risks (Vanclay 1995). Given the in-herent unpredictability of some of the socioeco-nomic forces driving land-use changes, economicmodels often adopt a normative approach.

Panayotou and Sungsuwan (1989) constructeda theoretically based model of tropical defores-tation by introducing three demand functions:the demand for logging, the demand for fuel-wood, and the demand for conversion to agri-cultural land. They then obtained an aggregatedeforestation function and empirically esti-mated a simplified model for northern Thailand.In addition, several partial or general equilib-rium models of land use have been developed todescribe the trade-off between land clearancefor agriculture and resource conservation ormaintenance (Walker 1987; Southgate 1990;Jones and O’Neill 1992). Jones and O’Neill(1992) examined the impact of profit maximiza-tion decisions modeled at the individual levelon region-wide environmental outcomes. VonThünen land-use models were also applied todescribe periurban deforestation at regionalscales (Jones and O’Neill 1993; Chomitz andGray 1996). The spatial model of Chomitzand Gray (1996) applied to Belize indicatedthat intensification of the road network aroundmarket areas provides a better trade-off betweenspurring development and minimizing defores-tation than the extensification of the network.In a review of economic models of deforestation,

Kaimowitz and Angelsen (1997) distinguishedbetween household and firm-level models,regional-level models, and national and mac-rolevel models. Within their typology, this studyexamines a regional-level spatial regression model.The main contribution of our work, in this sub-stantive and methodological context, is to im-prove the representation of the processes of de-forestation in these economic models—not assimple forest conversions between two time pe-riods but as complex trajectories of changes af-fected by reversibility and fluctuations over suc-cessive observation periods.

Deforestation and Land Rent

Our approach to modeling land-use and land-cover changes assumes that land is devoted tothe use that generates the highest potentialrent. This rent is a function of the returns andcosts of forest conversion.

Returns

to forest con-version depend on:

1.

farmgate price of outputs

(P

O

) as a functionof the price at the market and the trans-portation cost; it is computed as:

P

O

5

price at town

2

S

(distance on road

i

* cost/kg-km on road

i

) (1)

2.

farmgate price of inputs

(P

I

) as a function ofthe price at the market and the transpor-tation cost; it is computed as:

P

I

5

price at town

1

S

(distance on road

i

* cost/kg-km on road

i

) (2)

3.

agroclimatic

conditions

, which can be rep-resented by classes of land aptitude foragriculture, which describe the influenceof soil, climate, and topography on thepotential yield of the main crops of the lo-cal farming systems.

Costs

of forest conversion depend on:

1.

physical accessibility to the forest

, which inturn depends on the road network and thenumber of openings in the forest coverthat facilitate access to the forested areas;this can be measured by the forest-coverfragmentation and by the distance of anyforest location to the nearest forest/non-forest edge;


2.

forest-clearing cost

, which is related to thetechnology used for forest clearing and tothe density of the vegetation cover;

3.

social accessibility to forested land

, which isrelated to human pressure on the land anddepends on the population density relativeto nonfarm employment opportunities inthe region and, as a proxy variable, on theaverage income level of the local popula-tion (low average income level is assumedto generate greater demand for land);

4.

land tenure

and, in particular, the degree oftenure security of the forest occupants.

A forest plot is assumed to be converted tononforest if it has a positive rent (i.e., if thereturns to conversion are positive). The rent isunobserved (Chomitz and Gray 1996), but areduced-form model can be estimated empiri-cally, as a cross-section, under the assumption ofa spatial equilibrium between the supply and de-mand for agricultural and forest commodities(Panayotou and Sungsuwan 1989). The modelis greatly simplified if market prices are made ex-ogenous, and if it is assumed that they remainconstant. In this case, land-use decisions leadingto changes in total production or consumptionof agricultural products cannot affect marketprices through, for example, market saturation.It also simplifies the model since we can then as-sume that decisions on any site are independentof those of other sites (Vanclay 1995).

A problem may arise in the above modelwhen roads are endogenous to agricultural landuse. As Chomitz and Gray note: “If roads arepreferentially routed through agriculturally suit-able areas and if some aspects of suitability arenot observed, then the model may overestimatethe effect of distance from the road” (1996:493). This bias can be reduced by explicitlyintroducing a variable measuring the suitabilityof land for agriculture. Also, in southern Cam-eroon, the placement of the road network isgenerally exogenous to agricultural land use (asit is related to the need to connect distant citiesand towns), though it is often endogenous tologging activities.

The conceptual model described above guidedthe selection of independent variables to be en-tered in the model. Unfortunately, no spatiallyexplicit information was available on land tenureor land pressure. On the latter, the demographicdata collected by local administrators were eithertoo spatially aggregated or too sparse to be ex-

ploited in a spatial model. These two variableswere therefore ignored in the analyses that fol-low; the forest-clearing technique is assumed tobe primitive (fire and hand-cutting of trees),and to be identical throughout the study area.

Study Area

The Institutional Context in Cameroon

Evergreen and semideciduous forests coverapproximately 44 percent of the area of Cam-eroon (FAO 1995), mostly in the southern partof the country. In the past several decades, a for-estry code has regulated the management offorests in Cameroon. Recent degradation of theforests and international concerns about envi-ronmental issues have encouraged the countryto revise its forest policy so as to promote a sus-tainable use of forest resources. This has resultedin the recent creation of the Office National deDéveloppement des Forêts (ONADEF) in 1990and the Ministry of Environment and Forest(MINEF) in 1992. Regulations governing theentire forestry sector have been significantlymodified with the creation in 1995 of the Na-tional Forestry Action Program (NFAP). Its ob-jectives (MINEF 1996) are to: (1) ensure theprotection of the forest heritage and participatein safeguarding the environment and preservingbiodiversity; (2) increase participation of localpopulations in forest conservation and manage-ment, so as to raise living standards; (3) developforest resources with a view to increasing GDPwhile conserving production potential of fuel-wood, timber, nontimber products, and wildliferesources; (4) ensure resource renewal throughregeneration and reforestation, with a view tosustaining land potential; (5) revitalize the for-estry sector by setting up an efficient institu-tional system that involves all parties concernedwith the management of the sector.

A land-use plan, consisting of a zoning of theterritory into large land-management units, hasalso been adopted. According to this plan, thesouthern part of Cameroon, which containsmost of the nation’s forest resources, is dividedinto two domains: (1) the

permanent forest do-main

, composed of forest zones owned by theState [

forêts domaniales

] or public collectives[

forêts communales

]; and (2) the

nonpermanentforest domain

, composed of land likely to be di-verted to agro-sylvo-pastoral uses.


It is interesting to monitor whether thesenew institutional tools have alleviated some ofthe problems of forest degradation in Cam-eroon. In its latest Forest Resource Assessmentreport, the Food and Agriculture Organizationof the U.N. (FAO 1995) reported that Camer-oon loses a net 0.6 percent of its forest everyyear. It is difficult, however, to produce accuratedeforestation figures since, in Cameroon, mostforest-cover changes take place as degradationdue to fires, agricultural encroachment, and se-lective logging activities (Sayer et al. 1992).

The Bertoua Region

The study area is around the town of Bertoua,in the East Province of Cameroon. Located onthe northeastern part of the southern Cameroonplateau (Figure 1), it is dominated by semidecid-uous forests with forest-savanna transition zoneand savanna woodlands in the northern part ofthe province. The East province still has thelowest population density of Cameroon (4.1 in-habitants/km

2

in 1987) and the highest rate ofrural population (70.1 percent). Slash-and-burn

Figure 1. Cameroon with the location of the study area. Only the main towns of interest are shown. The approx-imate spatial extent of the actual forest area is shaded in light grey. The Dja reserve area is shaded in dark grey.


agriculture is widespread. The low populationdensity in southern Cameroon provinces has ledto high rates of inmigration, mostly of workersinto agricultural and forest logging sectors. Allprovinces in the forest zone of Cameroon have alarge positive migratory balance, except for theSouth Province (Pokam 1996). In-migrationinto the East Province is related to agriculturalland availability, the presence of logging compa-nies, and improvements in road and railway in-frastructures. The road network, which, in thisregion, the logging companies often built andmaintained for wood transport, facilitates access toforest areas by shifting cultivators in search of newagricultural land (Karsenty et al. 1993). Most ofthese landless migrants originate from the adjacentprovinces or from the Central African Republic.

The East Province is a sensitive zone in termsof deforestation and biodiversity. Large areas de-fined in the new zoning plan as belonging to theforest domains have already been cleared or de-graded. Numerous logging concessions have beengranted to citizens as well as to foreign compa-nies, in small, one-year-lease concessions referredto as ventes de coupes. A large faunal reserve, theDja reservation, lies just south of the study area(Figure 1). It is one of the largest conservationareas in Cameroon (5,260 km

2

) and is a majorbiodiversity reserve both for fauna and vegeta-tion.

1

Due to the growth of the rural populationin the region, it is threatened by poaching andrapid encroachment of cocoa, coffee, and subsis-tence plots, particularly on the northern and west-ern borders. Several roads in the province, in-cluding at the border of the Dja reservation, havebeen rehabilitated or built through funding by in-ternational agencies. This raises worries that theroad developments will lead to further deforesta-tion at the edge or within the reservation for woodextraction, commercial hunting, and agriculturalencroachment. For all these reasons, and in orderto better manage potential ecological impacts offuture road projects, it is useful to develop predic-tive models for areas likely to be affected by futureforest conversion. This can best be achieved bycalibrating a spatial model of deforestation on ad-jacent areas in the province, where the process offorest-cover change is more advanced.

Data

A time series of four high-spatial-resolutionsatellite images was acquired for the years 1973,

1986, 1991, and 1996. Two Landsat MSS images(at a nominal spatial resolution of 50

3

70 m)were obtained at thirteen-year intervals (1973and 1986), as were two SPOT XS (at a nominalspatial resolution of 20

3

20 m) at five-year in-tervals (1991 and 1996). The Landsat MSS andSPOT data from the four years were geometri-cally rectified and registered to a common UTMprojection based on 1:200,000-scale topo-graphic maps of Cameroon, and were further re-sampled, using the nearest-neighbor technique,to the spatial resolution of Landsat MSS. Thebest fit was provided using a second-order poly-nomial based on more than one hundred groundcontrol points (root mean square error belowthe MSS pixel size). The area used to calibratethe model corresponds to the common zone be-tween the two Landsat MSS and the 1991SPOT images. It covers approximately 110,000ha. The area used for model validation is thecommon zone between the area used for modelcalibration and the 1996 SPOT image. It covers50,000 ha. We applied radiometric corrections,taking into account sensor parameters and solarillumination conditions, to transform the digitalnumbers into reflectances (using the formulasprovided in Markham and Barker 1987). Verti-cal panchromatic aerial photographs from 1979(at a 1:20,000 scale) and 1993 (at a 1:10,000scale) were also assembled to support the inter-pretation of the satellite data.

Digital maps of road networks and townswere produced by manual digitization of the1:200,000-scale topographic maps of Cameroon(Centre Géographique National 1978) and themore recent aerial photographs and remotesensing data for an update. Based on field obser-vations, we characterized each road segment byits surface material. Population data from the1976 and 1987 censuses, as well as agriculturaldata from 1984 to 1991, were compiled fromagricultural census and surveys at the level of“arrondissements,” administrative units with anaverage size of 5,700 km

2

.During three successive field excursions in

1993, 1994, and 1997, we assembled detailedground observations of vegetation cover, agri-cultural practices, wood exploitation activities,surface material of roads, crop prices at differentmarkets in the region, distribution systems, andqualitative information on land-use conflicts.We collected landscape observations along themain and secondary roads to cover all the acces-sible parts of the study area. All field observa-


tions were georeferenced (with an accuracyaround 100 m) using a global positioning system(GPS). We used these field data, and the aerialphotographs for the most remote areas, for vali-dation of the remote sensing-based land-coverand land-cover change maps, and to support theinterpretation of the statistical results. Theseobservations were represented with a minimummapping unit of 50

3

50 m (i.e., the MSS pixel).The spatial resolution of the modeling exper-

iment is determined by the coarser spatial reso-lution of the data used to compute the depen-dent variable (the Landsat MSS at 50

3

50 m, inthis case). This resolution is therefore the basisfor measuring all independent variables. As forany geographical study, the spatial resolutionmay have an impact on results (Arbia 1989)—for example, a finer spatial resolution couldhave led to the detection of finer-scale forestclearings and could, therefore, have altered thestatistical results. The spatial resolution usedhere, however, is consistent with the spatial ex-tent of the study, and the mapping unit more orless corresponds to the size of fields in small-holder agriculture.

Variables and Procedures

Measurement of the Dependent Variables

Below, two models with two different depen-dent variables are tested and compared. First,the dependent variable is defined as the pres-ence or absence of deforestation between twoobservation years. Second, it is defined as a set ofland-cover change trajectories over three obser-vation years. In both cases, it is measured in thesteps described below, using the time series ofhigh-spatial-resolution remote sensing images.

Land-Cover Classifications.

Land-cover mapsof the study area for 1973, 1986, 1991, and 1996were derived for each individual image, using asupervised classification procedure, based on amaximum likelihood classifier. The trainingareas were derived from the field observations incombination with an interpretation of the aerialphotographs at the closest date. Training areaswere only selected within the areas that did notchange between the dates of acquisition of re-mote sensing data and the period of field obser-vations. We estimated classification accuracy forall years, using an independent sample of 262

observations from the field campaigns and thelow-altitude aerial photographs. Contingencytables between the reference data and remotesensing-based classification were produced, andthe overall accuracy and Kappa index were com-puted (Congalton 1991). The population ofpixels was first stratified by land-cover class, andan equal number of pixel observations was ran-domly selected in each class, with a minimum of50 observations per class (as recommended byCongalton 1991).

Land-Cover-Change Analysis.

Changes inland-cover between the successive dates weredetected by combining two techniques in a mul-tistage approach (after Pilon et al. 1988; Sader1995; Macleod and Congalton 1998): (1) post-classification comparison—that is, overlayingand comparing the two successive land-coverclassifications—and (2) band differencing—that is, subtracting reflectances (in this case,from the near-infrared Landsat MSS band 0.8–1.1

m

m and SPOT band 0.79–0.89

m

m) mea-sured at the successive dates. The postclassifica-tion comparison leads to a categorical map thatindicates, for every pixel, the land-cover classesat the two successive observation years (Jaku-bauskas 1990; Sader 1995). The ability to char-acterize “from” and “to” identifiers is essentialfor the definition of land-cover change trajecto-ries. The disadvantage of this method is its de-pendence upon the classification accuracy of themaps compared.

2

We reduced this problem herethanks to an aggregation of the land-coverclasses into a binary map: forest versus nonforestbefore change detection. This simple classifica-tion scheme results in a much-improved classifi-cation accuracy at the successive dates.

In addition, the second change-detectiontechnique used in combination with the post-classification comparison further increases thereliability of the final land-cover change map. Theband differencing method leads to a measure ofchange along a continuum of change intensity.It indicates the degree to which the change hasmodified the vegetation cover, using the surfacereflectance as a proxy. While a direct evaluationof change areas is not possible with this lattermethod, it allows for the detection of forest deg-radation and thinning as well as of forest-coverconversion. Several recent studies have demon-strated the utility of the band differencingmethod for land-cover change detection (Mu-choney and Haack 1994; Coppin and Bauer


1996; Milchener and Houhoulis 1997; Macleodand Congalton 1998).

We have combined the advantage of the twomethods using the following procedure. Theunchanged areas identified, based on the imagedifferencing map, were labeled based on the first-date land-cover map. The changed areas identi-fied based on image differencing were labeledusing the result of the postclassification compar-ison method. This allows us to reduce the mis-classifications on the final land-cover changemap by reducing the area for which the results ofthe postclassification method are retained. Inother words, the results of the postclassificationmethod are only applied to the areas that havebeen identified as being affected by changebased on the image differencing method, whichis more reliable for change detection but fails tocharacterize the “from” and “to” land-coverclasses. To avoid any overestimation of defores-tation based on the image differencing method,a threshold value of change intensity was de-fined to highlight only the areas that have expe-rienced the highest change intensity. Thisthreshold was defined on the basis of field obser-vations and corresponds to

6

1.5 standard devi-ations from the mean change intensity for thestudy area. Note that, as the Landsat MSS andSPOT data were resampled to the same spatialresolution (50

3

50 m) before they were classi-fied, the combination of data from different sen-sors does not affect the estimates of land-coverchange over time. Note also that great care wastaken to avoid any bias related to climate varia-tion, first by working with remote-sensing datafrom the same season, and second, by adaptingthe change thresholds to the phenological con-dition of semideciduous forest.

Trajectories of Land-Cover Change.

The tra-jectories of land-cover change are defined by thesuccessive transitions between land-cover cate-gories over the observation years. Only threeobservation years were retained for the defini-tion of these trajectories, since the fourth obser-vation year was reserved for validation purposes.All possible land-cover trajectories were re-duced to eight major sequences based upon theforest and nonforest categories. The latter classincludes agriculture, heavily degraded forest,and savanna. These sequences and their inter-pretation are shown in Table 1.

These sequences are defined from a limitedtemporal sample of observations. Therefore, ex-pressions such as “permanent forest regrowth” or“old forest clearing” only have meaning relativeto the timing of observations—for example, anew clearing could take place anytime after thelast observation of forest regrowth. The se-quences that involve successions of clearingsand regrowths (numbers 4 and 5) are likely to beassociated with fallow agriculture.

Measurement of the Independent Variables

Several spatially explicit variables, generatedusing standard GIS analysis tools,

3

were evalu-ated as determinants of the land rent and, there-fore, as influencing decisions on forest clearing.These are described below.

Weighted distance to the nearest road.

This vari-able was calculated as a series of 50-m-wide buff-ers expanding from each arc of the road network.The distance from any point to the nearest roadwas weighted by the average transportation coston each road segment. Transport cost was based

Table 1.

Land-Cover-Change Trajectories between 1973 and 1991 with Area, Proportion of Total Area Affected by Each Trajectory (%)

Change Trajectory Coverage

1973 1986 1991 Description (ha) (%)

1 Forest Forest Forest Stable primary or secondary forest 92,451.25 81.102 Forest Forest Nonforest Recent forest clearing 4,143.75 3.643 Forest Nonforest Nonforest Old and permanent forest clearing 4,535.00 3.974 Forest Nonforest Forest Old forest clearing with regrowth 264.50 0.235 Nonforest Forest Nonforest Forest regrowth with new clearing 430.50 0.376 Nonforest Forest Forest Old and permanent forest regrowth 650.00 0.577 Nonforest Nonforest Forest Recent forest regrowth 562.50 0.49

8

Nonforest

Nonforest

Nonforest

Stable savanna or permanent agriculture

10,954.00

9.60


on the surface material of the road and was mea-sured by the hourly-average traveling speed, fol-lowing Chomitz and Gray (1996). A parcel ofland located at one km from an asphalt roadpassable throughout the year has a higher esti-mated rent than a parcel of land at the same dis-tance from a narrow dirt road usable only in thedry season. Four categories of roads were consid-ered, based on the hourly-average travelingspeed (from 20 to 80 km/h). The accessibility ofany point was described as a function of theshortest distance to the road multiplied by aweight varying from 1 (for principal, permanentroads with an average speed of 80 km/h) to 4(for tracks, which can only be used seasonally,with an average speed of 20 km/h), with valuesof 2 and 3 for, respectively, secondary roads(with an average speed of 60 km/h) and perma-nent tracks (with an average speed of 40 km/h).Detailed studies in Central Africa (Finifter andVerna 1996) and in the Brazilian Amazon(Stone 1998) have estimated the transportationcost for timber (in monetary units) for differentroad types. These studies have demonstratedthat the actual transportation cost on principalroads, secondary roads, and permanent tracksvaries by a factor proportional to the averagetraveling speed (1–3). They did not considerseasonal tracks. Field surveys suggest that thetransportation costs for agricultural productsand timber vary in a similar way for differentroad types (Forni, personal communication).

Weighted distance to the nearest market townalong the road network.

This variable was calcu-lated as the sum of the distance from each loca-tion to the nearest road and the distance, alongthe road, to the nearest town. The distance onthe road to the nearest town was weighted ac-cording to: (1) the transportation cost, byweighting the distances traveled respectively ondifferent sections of the road by a factor varyingfrom 1 to 4 according to the road quality (seeabove), and (2) the type of town market, byweighting the total distance to the town by afactor inversely proportional to the price level ofagricultural products at the market. Actually,the attractivity of a town for a farmer settled inthe surrounding area depends not only on its ac-cessibility but also on the average retail price ofagricultural products at its market. High outputprices at a given market increase the land rent ofparcels of surrounding land. Three types of mar-kets were considered: urban, semirural, and rural(Atayi and Knipscheer 1980). At rural markets,

producers sell their goods directly. At semiruralmarkets, there is a mix of producers and retailersselling agricultural products. At urban markets,there is a higher proportion of retailers, and re-tail prices are higher. Producers benefit fromthese higher retail prices if they sell their prod-ucts directly at the market, which is the case forfarmers living at a reasonable distance from thetown or if the retailers do not capture most ofthe price difference between rural and urbanmarkets. The weights assigned to the three cate-gories of market types in the distance calcula-tion were based on the average retail price of theagricultural products at the market under con-sideration, as observed by Atayi and Knipscheer(1980). The values are 1 for rural markets, 1.08for semirural markets, and 1.59 for urban mar-kets. Thus:

Distance to town

j

5

[distance to the nearest road

1

(distance on road segment

i

3

surface material of road

i

)]/index of average retail price at town

j

(3)

Only towns with more than 5,000 inhabitantsand with a permanent market were taken intoaccount. The towns just outside the study areawere also included in this calculation.

Soil aptitude for agriculture.

This variable wasdetermined from a combination of data on the1:1,000,000-scale ORSTOM soil map (Martinand Segalen 1966), and the 1:200,000-scale soilmap included in the zoning plan of the EastProvince (Ministry of Forestry and Environment1995). Even though the scales of the maps differ,the level of generalization of both is roughlyequivalent. The latter map provides slightlymore accurate information for the southern partof the study area. Five classes of soil aptitude foragriculture were distinguished with, in decreas-ing order of quality: (1) brown-red, ferraliticsoils, (2) red soils with sesquioxydes, (3) hydro-morphic soils, (4) hardened soils, and (5) later-itic duricrusts. This information was introducedas a categorical variable in the statistical model.The cartographic information on soil aptitude isthe least detailed and reliable of all the variablesincluded in the model.

Shortest distance to the nearest forest/nonforestedge at the initial time period.

This variable was se-lected as a measure of the physical accessibilityto the forest cover for farmers and/or loggers,and is hypothesized to influence forest-clearingcosts. It was calculated as a series of 50-m-wide

Si


buffers expanding from all interfaces betweenpixels classified as forest and those classified asnonforest. To remove the influence of isolatedpixels on this distance calculation, the land-cover map was first smoothed using a five-by-fivepixel low pass filter. For this smoothing, themost frequently occurring class in the windowwas assigned to the central pixel of a movingwindow (i.e., a modal filter).

Spatial fragmentation of the forest cover in theimmediate surroundings of each location.

This vari-able was calculated in nine-by-nine pixels win-dows using the Matheron (1970) index:

M

5

Number of edges between forestand other cover types pixels

Ö

no. of forest pixels

3

Ö

total no. of pixels (4)

The numerator measures the number of pairs ofadjacent pixels classified as forest and othercover types (i.e., number of forest/nonforest in-terfaces included within the window), while thedenominator normalizes this count by the size ofthe forest and the total area. The index charac-terizes the length of the perimeter line of forestpixels exposed to nonforest cover types. Thisvariable, and the previous one, are taken as prox-ies for the accessibility of the forest cover. Theyshould indirectly influence forest-clearing costs.

Logistic Multiple Regression Models

Background.

Many of the previous spatial,statistical analyses of deforestation have con-centrated on univariate relationships (Saderand Joyce 1988; Gastellu-Etchegorry and Sinul-ingga 1988; Brown et al. 1993; Liu et al. 1993;Sader 1995; Mertens and Lambin 1997). Theseauthors attempted to identify predictors of areaswith the greatest propensity for deforestation.Developing a multivariate model of the proxi-mate causes of deforestation allows us to ac-count for the interactions between independentvariables and to rank explanatory variables ac-cording to their degree of importance in ex-plaining the spatial-variation deforestation.

The technique used in this study is logisticmultiple regression (LMR). It is designed to es-timate the parameters of a multivariate explana-tory model in situations in which the dependentvariable is categorical, and the independentvariables are continuous or categorical. LMR is

more appropriate than discriminant analysis ifsome independent variables are qualitative innature (Press and Wilson 1978). Our objectiveis to predict the probability of the presence orabsence of deforestation (or any other land-coverchange trajectory), based on a series of environ-mental and locational descriptor variables.The LMR technique yields coefficients for eachindependent variable based on a sample of data.These coefficients are interpreted as weightsin an algorithm that generates a map depictingthe probability of a specific category of land-cover change for all sampling units. LMR hasalready been successfully used in wildlife habi-tat studies (Pereira and Itami 1991; Narumalaniet al. 1997; Bian and West 1997) and deforesta-tion analyses (Ludeke et al. 1990; Chomitz andGray 1996).

Statistical Procedure.

LMR identifies therole and intensity of explanatory variables

X

n

inthe prediction of the probability of one state ofthe dependent variable, which is defined as acategorical variable

Y.

Suppose

X

is a vector ofexplanatory variables and

p

is the response prob-ability to be modeled with, in the case of a di-chotomous dependent variable,

p

5

Pr(Y 51|X), with Y 5 0 meaning the absence of defor-estation and Y 5 1 meaning the presence ofdeforestation. The linear logistic model has theform:

(5)

where a is the intercept and bn are slope param-eters. The probability values can thus be quanti-tatively expressed in terms of explanatory vari-ables by:

(6)

Odds ratios can also be used to facilitate modelinterpretation (Stokes et al. 1995; Menard 1995).The odds ratio is a measure of association thatapproximates how much more likely (or un-likely) it is for the outcome to be present for aset of values of independent variables (Hosmerand Lemeshow 1989). The probability, the odds,and the logit are three different ways of expressingthe same thing (Menard 1995). The estimatedodds values are computed as the exponential of

logit(p) log p1 p2------------- 55

a b1X1 b2X2 ¼ 1 bnXn111

pe

a( b1X1 b2X2 ¼ bnXn )1 111

1 ea( b1X1 b2X2 ¼ bnXn )1 111

1--------------------------------------------------------------------------5


the parameter estimate values (Hosmer andLemeshow 1989; Agresti 1990):

(7)

In this study, LMR was performed using theCATMOD functions in the SAS/STAT soft-ware. The predictive ability of a logistic re-gression model is evaluated from the table ofmaximum likelihood estimate (MLE), whichcontains the maximum likelihood estimate ofthe parameter, the estimated standard errorof the parameter estimate, the Wald chi-squarestatistic, and the significance probability for theparameter estimate. In the case of logisticmodels, the goodness-of-fit measure is defined asthe ratio of maximized log likelihood. Thispseudo R2 or r2 is defined as (Wrigley 1985):

(8)

or one minus the ratio of the maximized log like-lihood values of the fitted (log[b]) and constant-term-only (log[C]) models. Although r2 rangesin value from 0 to 1, its value tends to be consid-erably lower than the value of the coefficient ofdetermination R2 of conventional regressionanalysis. It should not be judged by the stan-dards of what is normally considered a “good fit”in conventional regression analysis (Wrigley1985). Values between 0.2 and 0.4 should betaken to represent a very good fit of the model(Domencich and McFadden 1975).

Implementation for Southern Cameroon.Based on our data, spatial LMR models werebuilt with specific categories of land-coverchange as the dependent variable and, as de-scribed above, independent variables for roadproximity, town proximity, forest-cover frag-mentation, proximity to a forest/nonforest edge,and soil aptitude for agriculture (a categoricalvariable). We tested the existence of collinear-ity between the independent variables and de-veloped and compared two types of models. Wefirst defined the dependent variable as the pres-ence or absence of deforestation between twoobservation years, referred to as the forest con-version model. Second, we defined a set of land-cover-change trajectories over three observa-tion years as the land-cover trajectories model.These changes in land-cover were derived fromthe change-detection analysis performed on thetime series of remote sensing data.

odds (p) e a( b1X1 b2X2 ¼ bn1 Xn )1115

r2 1

b[ ]logC[ ]log

-----------------25

For the land-cover trajectories model, the de-pendent variable is categorical. All land-covertrajectories that start with a forest cover in theinitial period (1973) need to be modeled to-gether, using a single multinomial logistic re-gression model. The results of this model aretherefore conditional to the initial land-cover.These trajectories include stable forest, perma-nent deforestation (at the early or later period),and deforestation followed by reforestation. Amultinomial model accounts for the fact thatthese trajectories are not independent fromeach other, as the sum of the areas affected by allthese trajectories needs to be equal to the totalarea under forest in 1973. The same is true for alltrajectories starting with nonforest in 1973.

The statistical procedure specifies that the re-sponse function consists of generalized logits ofthe marginal probabilities for the dependentvariable. The generalized logits are obtained bytaking the logarithms of the ratios of two prob-abilities. The denominator of each ratio is themarginal probability corresponding to the lastobserved level of the variable—the baseline cat-egory (Agresti 1990) or reference group (Hosmerand Lemeshow 1989), noted as Pb. The numer-ator is the marginal probability correspondingto the trajectory under consideration, notedas Pc:

(9)

For continuous explanatory variables, positivevalues of the parameters b indicate that largervalues of the variable increase the likelihood ofcategories other than the baseline category and,inversely, negative values of the parameters bindicate that larger values of the explanatoryvariables decrease the likelihood of categoriesother than the baseline category (Jobson 1992).For categorical explanatory variables (i.e., soilaptitude), positive values of the parameter esti-mates indicate higher probability of observingthe considered trajectory rather than the base-line trajectory for the considered level of thevariable. For categorical variable (n categories),parameter estimates are computed for the n 2 1levels of the variable (differential effect for then 2 1 levels of the variable for the consideredlogit). The parameter estimate (P) for the lastlevel (in the case of n 5 3) can be obtained by:

P3 5 2P1 2 P2 (10)

logit Pc

Pc

Pb-----log5


If r is the number of response levels (i.e., trajec-tories) of the dependent variable, the procedurecomputes r 21 equations giving the likelihood ofobserving a considered trajectory rather than thebaseline trajectory for the set of independent vari-ables. The logit for comparing two nonbaselinetrajectories is the difference between the logit ofthese two trajectories obtained with the baselinetrajectory model (Hosmer and Lemeshow 1989).

To perform the LMR, a random sampling pro-cedure was used to select N observations pointsdistributed in the study area. Sampling of obser-vations was used due to the presence of spatial au-tocorrelation in the data. Selecting random ob-servations decreases the probability of selectingadjacent, and therefore spatially autocorrelated,observations. For every sample observation, thevalues of the dependent and independent vari-ables were recorded. For the forest-conversionmodel, a random sample of around 10,000 obser-vations was selected. For the land-cover trajecto-ries model, the sample size was approximately30,000 observations. This ensures that the sam-ple size of each response level of the dependentvariable (r 5 4 as there are four trajectories forthe model) is at least 75, that is, at least 25 obser-vations for each response function (r 2 1).4 In re-ality, there are many more sample observationsfor most trajectories but the least representedland-cover trajectory determines the total samplesize, given the random nature of the sampling.

Mapping the Probabilityof Land-Cover Changes

Applying LMR results to updated values of theindependent variables allows us to generate animage representing the surface probability of fu-ture land-cover changes using equation (6). Aprobability surface for the trajectory associatedwith recent deforestation can be generated, as de-forestation in the near-future is more likely to becontrolled by the same factors that controlled themost recent deforestation. This image can thenbe overlaid on the land-cover map correspondingto the year 1991 (the probability of deforestationwas only retained for those pixels that were stillforest in 1991). Forest locations characterized bya high probability concentrate spatial attributesthat, in the recent past, have been associatedwith a high frequency of deforestation. This mapwill therefore predict the forest areas that aremost likely to be affected by deforestation. It will

contain no indication as to when these land-cover changes might take place, but it does sug-gest where future changes will occur if similar re-cent causal processes observed are maintained.The predicted probabilities of land-cover changesonly become projections of future land-coverchanges if a dynamic model predicts future ratesof land-cover conversion. Such a dynamic modelcould be based on the assumption that, in theshort-term, changes will continue to take place inthe same way they did in the recent past and thatthe processes of land-cover change, as estimatedby the model’s coefficients, are stationary. Longer-term projections should incorporate the actualdriving forces, as opposed to the proximatecauses, of the process. The deforestation proba-bility map (shown below) allows for a rigorousvalidation of the models since it can be comparedto the actual deforestation taking place betweenthe last observation year used in the model cali-bration, and any subsequent observation years.

Simulation of the Impactsof Land-Cover Changes

The simulation of the potential impacts ofland-cover changes can be performed in two ways:first, by assessing the impacts on specific landscapeattributes of the changes predicted to occur in thedependent variable and, second, by simulatinganticipated changes in one or several of the in-dependent variables and assessing their predictedimpacts on specific landscape attributes throughchanges in the dependent variable.

For the first case, the overlay of the land-cover-change probability maps on cartographiclayers allows us to evaluate what will be thelikely consequences of predicted changes on im-portant issues such as agricultural sustainability,measured by the sensitivity to soil erosion ofareas that have been cleared for agriculturalpurposes; biodiversity conservation, measuredthrough the loss of important habitats; and therole of tropical forests and secondary growths onthe carbon cycle. In this approach, we take intoaccount only the impacts ensuing from a contin-uation of current land-use practices.

For the second case, it is quite likely that beforeinvesting in infrastructure development projects(e.g., road construction, rehabilitation, or upgrad-ing, or the attribution of a logging concession), aninstitution will be willing to conduct a detailed as-sessment of the possible ecological impacts of theproject. This simply requires that, after calibrating


the spatial model on past observations, theplanned developments (e.g., a road project) are in-tegrated in the cartographic layer(s) representingthe independent variable(s) that will be altered bythe project (e.g., distance to road and distance tomarket towns). A new projection of deforestationrisk zones can then be produced with these modi-fied variables, based on the calibrated model. Theoutput map displaying the new probabilities ofland-cover changes can then be overlaid on mapsdisplaying specific landscape attributes (e.g., biodi-versity reserves) to assess the impact of the roadproject on the resource of interest (as in the firstcase). To demonstrate the feasibility and usefulnessof this approach, we conducted a simulation of theimpact of a fictitious road project on a biodiversityreserve in our study area.

Results

Land-Cover Classifications

Five main land-cover classes were discrimi-nated with a high level of accuracy using the re-

mote sensing data: dense forest, fragmented for-est, agriculture, savanna, and bare soils. Thedense forest class corresponds to evergreen ormoist deciduous forest zones, dominated by treesof at least 5 m high and with a forest-cover pro-portion of 70 percent or more (i.e., closed can-opy). The fragmented forest class is a landscapemosaic with between 30 to 70 percent of forestmixed with other classes. The agriculture classcorresponds to areas that are at least partiallycovered by fields or recent fallows (i.e., in theearly successional stages), with a forest-coverproportion lower than 30 percent. Only forestfallows of less than 10 years are spectrally sepa-rable from forests. Older fallows are thus in-cluded in the forest class. The savanna class cor-responds to areas dominated by an herbaceouscover and shrubs. The bare soils also includeroads and built-up areas. The accuracy of theclassifications, before merging of classes, are sat-isfactory (see Table 2 for the 1991 classification,as an example). After merging of the classes intoa binary forest/nonforest map, the accuracies aremuch improved: overall accuracy and Kappa co-efficients of, respectively, 97 percent and 0.95

Table 2. Contingency Table and Accuracy Indicators for the Original (a) and the Binary Recoded (b) Classification of the Spot XS Image of 1991 Aggregated at a 50 3 50 m Resolution

(a) Original Classification

Field Observations

BareSoils Savanna

FallowAgriculture

FragmentedForest

DenseForest Total Accuracy

Bare soils 44 4 1 1 0 50 0.88Savanna 2 44 3 1 0 50 0.88Fallow/agriculture 1 3 47 1 0 52 0.90Fragmented forest 0 1 2 51 2 56 0.91Dense forest 0 0 0 3 51 54 0.94Total 47 52 53 57 53 262Producer accuracy 0.94 0.85 0.89 0.89 0.96Overall accuracy 0.90Kappa 0.88

(b) Binary Recoded (Forest/ Nonforest) Classification

Field Observations

Nonforest Forest Total Accuracy

Nonforest 149 3 152 0.98Forest 3 107 110 0.97Total 152 110 262Producer accuracy 0.98 0.97Overall accuracy 0.98

0.95Kappa


for 1973, 97 percent and 0.93 for 1986, and 98percent and 0.95 for 1991.

Land-Cover-Change Analysis

Results of the change detection analysis indi-cate a net reduction in forest cover areathroughout the study period. The change matri-ces shown in Tables 3, 4, and 5 represent thetransfer of area (in hectares) among land-covercategories for pairs of successive observationyears. The transition probabilities are computedas:

(11)

where aij is the area under the land-cover i at theinitial period that was converted to land-cover jat the subsequent period. As the matrices werecomputed over a different number of years, thetransition probabilities have been recomputedon an annual basis, using:

(12)

pij aijaij

j^/5

T i j,1 expm

1n--- mT i j,

nlog5

where T1i,j is the transition probability matrix

on an annual basis, expm is the matrix exponen-tial, n is the number of years for the consideredperiod, logm is matrix logarithm, and Tn

i,j is thetransition probability matrix for the consideredperiod.

Tables 3, 4, and 5 show that the net deforesta-tion rate as measured by the ratio between, onone hand, the difference between deforestedareas and areas affected by forest regrowth and,on the other hand, the forest-cover area at theinitial period, is 7.36 percent between 1973 and1991. If we only take into account the rate of de-struction of forest areas at the initial period—ignoring, therefore, the regrowth of secondaryforests that takes place either on the deforestedarea, a few years later, or elsewhere—the accu-mulated rate of deforestation from 1973 to 1991is 9.24 percent. This latter rate is a better indica-tion of potential impacts of deforestation on hab-itats and, therefore, biodiversity, while the formerrate is more relevant to establish a carbon budgetor to monitor the availability of wood resources.These tables also show that most forest clearings

Table 3. Land-Cover Changes, Transition Probabilities, and Deforestation Rate, 1973–1986a

1986

Forest Nonforest

Forest Area (ha) 96,620.5 4,800.751973 Annual transition probability (%) 99.61 0.39

Nonforest Area (ha) 1,081.5 11,488.75Annual transition probability (%) 0.70 99.30

Gross deforestation (ha)Net deforestation (ha)Net annual deforestation rate (%)

4,800.753,719.25

0.28a The top number in row class is the area in hectares; the bottom number is the transition probability on an annual basis (%).


1991

Forest Nonforest

Forest Area (ha) 93,101.25 4,574.251986 Annual transition probability (%) 99.03 0.97

Nonforest Area (ha) 827 15,489Annual transition probability (%) 1.06 98.94

Gross deforestation (ha)Net deforestation (ha)Net annual deforestation rate (%)

4,574.253,747.25

0.77a The top number in row class is the area in hectares; the bottom number is the transition probability on an annual basis (%).


take place for agricultural purposes. There is alsosome proportion of the agricultural areas affectedby forest regrowth. This is part of the long fallowfarming system of the region.

The analysis of these tables also reveals twoother important elements. First, the rates of defor-estation fluctuates through time (from 0.28 per-cent to 0.77 percent per year between 1973 to1986, and 1986 to 1991). Note that the net defor-estation rate between 1986 to 1991, and 1991 to1996 is similar (0.77 and 0.75 percent), while thetransition probabilities of forest conversion andreforestation are double in 1991–1996 comparedto 1986–1991. Therefore, we cannot assume de-forestation to be a process that takes place at aconstant rate in time. It is strongly controlled byexogenous socioeconomic forces that vary overtime, as demonstrated through household surveysin Mertens et al. (2000). Second, the transitionprobabilities between pairs of land-cover classesvary also through time. Most notable is the proba-bility of a transition from forest to nonforest,which increases from 0.39 for the first period to0.97 for the second period. Thus, the transitionmatrix is nonstationary. This has important con-sequences since the stationarity of the transitionmatrix is a major assumption of Markov-basedmodels of land-cover change (Lambin 1994).When dealing with nonstationary processes, thesemodels lose any predictive ability unless one mod-ifies the transition probabilities through time ac-cording to some complex model design or basedon the knowledge that a different land-use changeprocess is taking place at a given time.

Land-Cover-Change Trajectories

The analysis of the percentage of the totalarea affected by each land-cover change trajec-

tory reveals that the stable trajectories—that is,continuously forest or nonforest for the threeobservation years—account for 90.70 percentof the study area (see Table 1). Thus, nearly 10percent of the landscape has changed at leastonce over the observation period. In the areasthat changed, 90 percent did so just once overthe period while 10 percent have changed twicein an eighteen-year period.

The analysis of the spatial pattern of the dif-ferent land-cover trajectories is also interesting.The forest clearings that appear to be perma-nent are mostly located around towns or nearroads. Trajectories characterized by successivetransitions between forest and nonforest corre-spond to short-fallow areas, mostly locatedalong new roads at a greater distance fromtowns. In general, the spatial pattern of defores-tation clearly follows a process of expansion dif-fusion, as illustrated for the area of Dimako,southeast of our study area (Figure 2). There is aclear centrifugal expansion of deforestationaround the town and along the road network inthe logged-over area west of Dimako.

Spatial Regression Models of Deforestation

Preliminary Statistical Analyses. Follow-ing the results of univariate relationships be-tween the dependent and independent variables(Mertens and Lambin 1997), the logarithm ofthe distance from roads, towns and forest/non-forest edges was used in the multivariate models.Also, low levels of collinearity were found be-tween the independent variables, with values ofthe coefficients of determination (R2) of themultivariate relationships between one inde-pendent variable against all the others ranging


1996

Forest Nonforest

Forest Area (ha) 34,308.5 3,1051991 Annual transition probability (%) 98.18 1.82

Nonforest Area (ha) 1,702.5 12,061Annual transition probability (%) 2.70 97.29

Gross deforestation (ha) 3,105Net deforestation (ha) 1,402.5Net annual deforestation rate (%) 0.75a The top number in row class is the area in hectares; the bottom number is the transition probability on an annual basis (%).


between 0.45 and 0.52. It is only with R2 largerthan 0.80 for at least one of the independentvariables that the collinearity is considered asbeing problematic for LMR (Menard 1995). So,even though there is some dependence betweenthe explanatory variables, it is not to a level thatshould pose a problem for the logistic regression.The R2 of the univariate relationships betweenvariables that were expected to be related, suchas distance to roads and distance to forest/non-forest edges, was in fact low (0.26 in that case).

Forest-Cover Conversion Model. The re-sults of the two separate LMR models of defores-tation for the periods from 1973 to 1986, andfrom 1986 to 1991, are given in Tables 6 and 7.These tables present the values of the parameterestimates with their corresponding standard er-ror, chi-square statistics, and significance proba-bility. Positive values of parameters indicatethat larger values of the explanatory variablesincrease the likelihood of deforestation, whilenegative values indicate the opposite. The re-sulting equation allows us to calculate the prob-

ability of deforestation. The chi-square statisticsindicate the relative weight of the explanatoryvariables in the model and allow us to assessthe role of each variable in the prediction ofdeforestation.

The overall explanatory power of the modelsfor both periods was satisfactory, with values ofr2 of 0.26 for the first period and 0.29 for the sec-ond period, which represents a good fit of themodel. The significance level and the relativeexplanatory power of the independent variablesvaried between the two periods. For both peri-ods, the distance to forest/nonforest edges is themost important explanatory variable. In bothcases, the probability associated with the pres-ence of deforestation increases as the distancesfrom forest edges decreases. Forest areas near aforest edge are more accessible and, therefore,more likely to be cleared. The estimated odds ofobserving deforestation is exp(21.0007) 5 0.37times higher for each additional unit of distanceto forest/nonforest edge. As it is the logarithm ofthe variable, distance to forest edges, which wasused in the model, and as the distance unit was

Figure 2. The land-cover-change trajectories around the town of Dimako, southeast of the study area, illustratingthe process of expansion diffusion of deforestation away from the town and along the road network.


50 m, areas at 0.5 km from forest edges are (1/0.37) 5 2.72 times more likely to be clearedthan areas at 5 km from forest edges.

For the first period, the role of proximity totowns is important in explaining the probabilityof deforestation. Areas at 0.5 km from towns are2.14 times more likely to be cleared than areasat 5 km. Accessibility to towns is thereforean important factor in deforestation, eitherthrough the commercialization of agriculturaloutputs, or through attempts to gain access toinputs. The proximity to a town may also be re-lated to the presence of services in small urbancenters or because towns act as centers of diffu-sion for migrants. Surprisingly, for the secondperiod, the explanatory power of the distance totowns is not statistically significant at the 0.05probability level. The inverse situation is ob-served for the variable “distance to roads,”which is insignificant for the first period, buthighly significant for the next period. The roadsin the second period are nearly as important fordeforestation as the towns are in the first period.

Areas at 0.5 km from roads are 1.62 times morelikely to be cleared than areas 5 km. These re-sults can be explained by the fact that between1973 and 1986, forest colonization was at an ini-tial stage and mostly affected the immediate sur-roundings of towns. It is only in the mid-1980sthat migrants have started to move farther alongthe road, at a greater distance from towns.

Deforestation is taking place mostly on goodsoils during the first period while, in the secondperiod, the explanatory effect of good soils be-comes negligible. This suggests that, after theareas around towns have been cleared, roadsidelocations are the dominant factor in determin-ing deforestation (for accessibility reasons), andmigrants do not make incursions into the forest,away from roads, to search for the best soils. Thisconclusion, however, is tempered by the factthat the soil variable is the one most prone toerror in our database. Note also that a highlevel of forest fragmentation is not significantin explaining the location of deforestation foreither period.

Table 6. Forest Conversion LMR Model for the First Period (1973–1986)

Parameter Estimates Std. Error Chi-Square

SignificanceProbability

Distance from roadsl 20.0664 0.0717 0.86 0.3544Distance from townsl 20.7631* 0.0830 84.43 0.0000Distance from forest/nonforest edgel 21.0007* 0.0746 179.70 0.0000Forest fragmentation 0.0001 0.0013 0.01 0.9056High soil aptitudec 0.6481* 0.1967 10.86 0.0010Medium soil aptitudec 0.5298 0.2904 3.33 0.0681Number of Obs: 9,905 r2: 0.26

*Denotes parameter estimates significant at the 0.01 level of confidence.c Categorical variable: differential effect of n 2 1 levels of the variable.l The logarithm of the variable is employed.

Table 7. Forest Conversion LMR Model for the Second Period (1986–1991)

Parameter Estimates Std. Error Chi-Square

SignificanceProbability

Distance from roadsl 20.4823* 0.0653 54.60 0.0000Distance from townsl 20.0379 0.0877 0.19 0.6654Distance from forest/nonforest edgel 21.2367* 0.0788 246.03 0.0000Forest fragmentation 0.0015 0.0012 1.53 0.2155High soil aptitudec 0.3220 0.1763 3.34 0.0677Medium soil aptitudec 20.4663 0.3128 2.22 0.1361Number of Obs: 9,917 r2: 0.29

* Denotes parameter estimates significant at the 0.01 level of confidence.c Categorical variable: differential effect are computed for n 2 1 levels of the variable.l The logarithm of the variable is employed.


Land-Cover Trajectories Model. We onlypresent the results of the multinomial model de-veloped for all the areas that were forest at theinitial time period (1973), as forest clearing is byfar the dominant land-cover change process inthe region. Table 8 presents the values of the pa-rameter estimates, with their correspondingstandard error, chi-square statistics, and signifi-cance probability. Positive values of parametersindicate that larger values of the explanatoryvariables increase the likelihood of observingthe land-cover trajectory under considerationrather than the baseline trajectory (stable forestthroughout the study period in this case). Theoverall explanatory power of the model washigh, with a r2 5 0.36. Thus, sequences of land-cover changes are indeed controlled by the spa-tial determinants of land rent that were intro-duced in this model. Table 8 shows also the oddsvalues between all trajectories. For all the defor-estation trajectories, parameter estimates are

negative for all three distance variables. Thus,the probabilities of trajectories associated withdeforestation decrease, by reference to the stabletrajectory, with increasing distances from roads,towns, and forest edges. By contrast, the proba-bility of occurrence of stable trajectories ishigher at greater distances from forest edges,roads, and towns, and for low levels of fragmen-tation of the forest-cover. The best protectionagainst forest clearing in conservation areas(such as the neighboring Dja faunal reserve) istherefore a buffer of closed forest. No such bufferzone currently exists around this reserve.

As observed above, the distance from townsis only important for predicting early deforesta-tion, while distance to roads has a greater effecton the recent deforestation. As the deforesta-tion in the region progresses, farmers have tomove further from towns, along the road net-work, to find new agricultural areas. More recentdeforestation is therefore less influenced by prox-

Table 8. Multinomial Logit Model for the Land-Cover-Change Trajectories Beginning with Forest Cover in 1973, Using the Stable Forest Trajectory (Forest-Forest-Forest) as the Baseline Category

ParameterEstimates Std. Error Chi-Square

SignificanceProbability Odds

Early deforestation trajectory(forest/nonforest/nonforest)

Distance from roadsl 20.1702* 0.0438 15.11 0.0001 0.843Distance from townsl 20.8477* 0.0555 232.98 0.0000 0.428Distance from forest/nonforest edgel 20.9586* 0.0462 430.71 0.0000 0.383Forest fragmentation 0.0029* 0.0007 13.49 0.0002 1.002High soil aptitudec 1.0825* 0.1535 49.76 0.0000 2.952Medium soil aptitudec 20.2213 0.2628 0.71 0.3996 0.801

Recent deforestation trajectory(forest/forest/nonforest)

Distance from roadsl 20.4160* 0.0413 101.28 0.0000 0.659Distance from townsl 20.0925 0.0578 2.65 0.1034 0.911Distance from forest/nonforest edgel 21.0610* 0.0485 477.84 0.0000 0.346Forest fragmentation 0.0042* 0.0008 30.50 0.0000 1.004High soil aptitudec 0.6032* 0.1360 19.67 0.0000 1.827Medium soil aptitudec 21.0065* 0.2521 15.93 0.0001 0.365

Fallow agriculture trajectory(forest/nonforest/forest)

Distance from roadsl 20.1747 0.1602 1.19 0.2756 0.839Distance from townsl 20.8464* 0.1862 20.67 0.0000 0.428Distance from forest/nonforest edgel 20.9548* 0.1779 28.82 0.0000 0.384Forest fragmentation 0.0054 0.0027 3.80 0.0514 1.005High soil aptitudec 0.5001 0.3925 1.62 0.2026 1.64Medium soil aptitudec 1.1006 0.4964 4.92 0.0266 3.005

Number of Obs: 29,757 r2: 0.36

* Denotes parameter estimates significant at the 0.01 level of confidence.c Categorical variable: differential effect are computed for n 2 1 levels of the variable.l The logarithm of the variable is employed.


imity to towns. The decreasing explanatory powerof the distance to towns, as the deforestationprogresses, suggests that the role of roads in thedeforestation process is mostly to increase the ac-

cessibility of the forest for migrants (i.e., a coloni-zation process) rather than to spur the transfor-mation of local subsistence agriculture intomarket-oriented farming systems (i.e., a process

Figure 3. The predicted probabilities of the land-cover-change trajectories associated with deforestation (“defor-estation risk zones”). The predicted values are based on the model calibrated for the period 1973–1986–1991(land-cover trajectories model). Each location still forest in 1991 was assigned a probability (between 0 and 1) ofbeing cleared, given its landscape attributes. The areas predicted as having the highest probability of deforestationare mainly located at short distances from forest/nonforest edges, towns and roads.


of development of local agriculture in response toincentives from distant markets). And if roads doindeed facilitate the commercialization of agri-cultural products to the markets, it seems to be

mostly the case for new migrants who have set-tled along these roads rather than for farmers in-volved in shifting cultivation around towns.

The land-cover-change trajectory that corre-

Figure 4. Simulation of the impact of a fictitious road-upgrading project on the deforestation in the Deng Dengin situ conservation area: (a) map of the road network of the area and limits of the Deng Deng forest, (b) land-cover classification of the area, (c) areas with a predicted probability of deforestation .0.8 based on the currentsituation of the road network, and (d) areas with a predicted probability of deforestation .0.8 based on the roadnetwork after upgrading of the road north of the forest. It is clearly visible that the road-upgrading project wouldlead to an increase of the area at risk of deforestation within the forest reserve. It would therefore have a quantifi-able impact on biodiversity.


sponds to fallow agriculture (forest, nonforest,forest) is mostly predicted by proximity to forestedges and to towns. Interestingly, fallow agricul-ture is positively associated with the presence ofmedium aptitude soils while the more perma-nent deforestation, which is likely to be asso-ciated with more intensive, market-orientedagriculture, is associated with the presence ofsoil with a high aptitude for agriculture. Theestimated odds of observing the fallow agri-culture trajectory rather than the early defor-estation trajectory on medium-aptitude soil isexp[1.01 2 (20.22)] 5 3.41 times higher. Theestimated odds of observing the early deforesta-tion trajectory rather than the fallow agriculturetrajectory on high-aptitude soil is exp(1.08 20.50) 5 1.79 times higher.

Distance to forest/nonforest edges is thedominant variable to explain all land-coverchange trajectories. This indicates the “spatialinertia” of land-cover change processes: changesrecur where they have already taken place in therecent past. The rationale behind this spatialprocess is related to the importance of the acces-sibility of the forest cover, which is increased bythe existence of openings in the forest cover.There is therefore a strong “spatial spread” effectof land-cover changes, whether these are perma-nent conversions or temporary changes.

Mapping the Probability of Land-Cover Changes: Validation and Model Comparison

The probability of deforestation was first pre-dicted based on the model calibrated on two obser-vation years (1986–1991). On the predicted map,each location that was still forest in 1991 was as-signed a certain probability (between 0 and 1) ofbeing cleared, given its landscape attributes. Dueto the high explanatory power of the variablesmeasuring distances to forest/nonforest edges,towns, and roads, the areas predicted as having thehighest probability of deforestation are primarilylocated at small distances from these features.

Second, the probability of deforestation waspredicted based on the land-cover trajectoriesmodel calibrated over the period 1973–1986–1991, for locations forested in 1973. Figure 3represents the predicted probability of defores-tation using the equation corresponding to thetrajectory of recent deforestation. The areaswith the highest risk of deforestation (in red-orange) are concentrated around the maintowns, on the forest edges and along roads. Ev-

ery small opening in the forest cover is also asso-ciated with a higher risk of deforestation.5 (For asimulation of the impact of a fictitious project ondeforestation, see Figure 4, discussed below.)

Figure 5. Actual deforestation between 1991 and1996 in areas with different ranges of predicted prob-abilities of deforestation using: (a) the forest-conversionmodel (1986–1973), and (b) the land-cover trajecto-ries model (1973–1986–1991). In both cases, highrates of deforestation have been observed in areas pre-dicted to have a very high probability of deforesta-tion. The observed deforestation decreases sharplywith the predicted probabilities of deforestation,since the validation period is only the five subsequentyears. The model based on land-cover-change trajec-tories yields better predictions.


To validate these predictions of land-coverchange risk-zones, the predicted probabilities ofchange, based on models calibrated on observa-tions up to 1991, were compared to the observedland-cover changes between 1991 and 1996(Figure 5). This validation—and the compari-son of the performances of the two-periodsmodel versus the land-cover change trajectorymodel—were performed in two ways: (1) byquantifying the proportion of areas with a highpredicted probability of deforestation that haveactually been deforested, as measured by:

area with a high probability of deforestationand deforested in 1991–96

total area with a highprobability of deforestation

and (2) by quantifying the proportion of the ob-served deforestation taking place in areas thatwere predicted to have a high probability of de-forestation, as measured by:

area with a high probability of deforestationand deforested in 1991–96

total area deforested in 1991–96

In general, high rates of deforestation havebeen observed in areas predicted to have a veryhigh probability of deforestation. The observeddeforestation decreases sharply with the pre-dicted probabilities of deforestation, since thevalidation period is short (just the five subse-quent years). Concerning the first measure(Table 9 and Figure 5), 89.08 percent of the forestareas predicted to have a probability of defores-tation larger than 0.98, with the model cali-

brated on land-cover change trajectories be-tween 1973, 1986 and 1991, have actually beencleared in the five following years. For the two-dates model (1986–1991), this percentage isonly 63.57 percent. Thus, analyzing successiveland-cover changes rather than only conversionsover a short time period does improve modelperformance. The difference between the twomodels is quickly attenuated as we examineareas with a slightly lower predicted probabilityof deforestation. Concerning the second mea-sure (Table 10), for all intervals of predictedprobabilities of deforestation, there is a higherpercentage of observed deforestation if the land-cover-change trajectories model is used, ratherthan the two-periods model. Again, the differ-ence between the two models is attenuated forareas with a lower predicted probability of defor-estation. We can therefore conclude that themodeling of land-cover-change trajectories overmore than two observation years leads to a bet-ter prediction of the probabilities of land-coverchange than the modeling of simple land-cover conversions over just two periods of time.

More fundamentally, Table 10 reveals thatthere is an important trade-off between the sizeof the area a policy-maker is willing to consideras being “at risk” and the amount of deforesta-tion that will be found in that area. Areas atvery high risk (e.g., predicted probability of de-forestation of 0.80 or more) based on the trajec-

Table 9. Proportion of Observed Deforested Areas (1991–1996) within Predicted

Probability of Deforestation Zones Using Forest-Conversion Model and Land-Cover

Trajectories Model

Proportion of ObservedDeforested Area (%)

Predicted Probability ofDeforestation

Based on the1986–1991

Forest-Conversion

Model

Based on the1973–1986–1991

Land-CoverTrajectories

Model

.0.98 63.57 89.08

.0.95 61.85 74.46

.0.90 51.87 62.91

Table 10. Proportion of Observed Deforested Areas (1991–1996) versus Predicted

Probability of Deforestation Using the Land-Cover Trajectories Modela

Observed Deforested and Forested Area (%)

PredictedProbability ofDeforestation

Based on the1986–1991

Forest-Conversion

Model

Based on the1973–1986–1991

Land-Cover Trajectories

Model

.0.90 20.42 24.571.74 1.86

.0.80 48.97 50.188.97 8.61

.0.50 80.34 81.1250.76 45.65

a The top number in row class is the proportion of observeddeforested areas (%); the bottom number is the proportionof forested area (%) affected by a given range of probabilityof deforestation in 1991.


tory model, cover only about 8 percent of the to-tal forest area and concentrate around 50percent of the total deforestation of the area. Onone hand, it is feasible to monitor this territory

at frequent time intervals and to plan interven-tions to mitigate undesirable patterns or effectsof forest conversion. On the other hand, thesevery high-risk zones only include half of the to-

Figure 6. Observed deforestation between 1991 and 1996 within and outside areas with a predicted probabilityof deforestation .0.5. Most deforestation is concentrated in the areas predicted to be at risk of deforestation, ex-cept the linear patterns of deforestation associated with recent logging activities and diffuse patterns associatedwith shifting cultivation (zone B). Similar linear patterns of deforestation are also observed in the southwesternpart on the study area. These patterns appear in areas with high values of predicted probability of deforestation asthey correspond to older logging concessions already taken over by shifting cultivators.


tal deforestation. To control a much larger shareof the total deforestation (e.g., 80 percent of thetotal deforestation actually taking place), onewould have to consider all areas with a predictedprobability of deforestation of 0.50 or more,which represents 45 percent of the total forestarea (Table 10)—a much larger territory tocover. The optimal trade-off between the size ofthe territory to be monitored and the amountof deforestation taking place in that territory de-pends on each application.

Figure 6 displays a map of actual deforestedareas during the validation period (1991–1996)in relation to the areas predicted to have a highand low probability of deforestation. It is partic-ularly interesting to analyze the location of areaswith a weak correspondence between observedand predicted deforestation. These areas arecharacterized by a linear or a diffuse pattern offorest clearing. The linear pattern is particularlyvisible in the northwestern corner of zone B(Figure 6). These areas correspond to deforesta-tion along the logging roads. The diffuse patterncorresponds to areas of slash-and-burn agricul-ture. In both cases, the processes of deforesta-tion are weakly linked to the market and are lessdetermined by accessibility constraints. Rather,the location of such activities are strongly con-trolled by the land-tenure system, either for thegranting of logging concessions by the govern-ment or in the selection by farmers of the areasto be cleared for shifting cultivation. Since landtenure was not represented in the model, it doesnot predict well the location of logging andshifting cultivation activities. As seen in Table11, the model performances on an area withoutany logging activity (zone A, Figure 6) is supe-rior to the model performances on an area withlogging activity (zone B, Figure 6).

Impacts of Deforestation Processes

For illustrative purposes, we conducted a sim-ulation of the potential impact of a fictitiousroad-upgrading project on a biodiversity reserve.The Deng Deng forest is located northwest ofBertoua. It is a primary forest that the FAO(1996) identified as an area of “in situ conserva-tion of forest genetic resources.” It is surroundedwest, east, and south by permanent roads thatare attracting an important migration flux andhave led to significant roadside deforestation overthe last decades (Figure 4a, b). North of the forest,

the road connecting Ndemba II and Koundi is stilla poor dirt road, barely passable during the rainyseason. For this simulation, we assumed that therewas a project to upgrade this dirt road into an as-phalt road that would be permanently passable.

Our impact assessment of this improved roadwas performed by changing the weight of theroad north of the forest from a value of 4 (sea-sonal road with an average speed of 20 km/h) toa value of 1 (permanent roads with an averagespeed of 80 km/h). A new map of deforestationrisk zones was then produced, using the modelcalibrated on the past observations. The resultof this spatial projection (Figure 4d) can becompared to the map of deforestation risk zonebased on the current situation of the road net-work (Figure 4c). The upgrading of the roadwould lead to an increase of the areas at risk ofdeforestation in the Deng Deng forest (from 1.2percent to 2.5 percent having a probability ofbeing cleared larger than 0.80, which means adifference of 406 ha of forest at high risk of be-ing cleared). Outside the forest reserve, the in-crease in the forest area at risk is much smaller(from 28.3 percent to 28.5 percent, having aprobability of being cleared larger than 0.80,which means a difference of 65 ha of forest athigh risk of being cleared) (compare Figures 4c,d). The potential impact of this road projectwithin the biodiversity reserve would thus belarge in the absence of measures to enforce thereserve boundaries.

Table 11. Proportion of Observed Deforested Areas (1996) versus Predicted Probability of

Deforestation Using the Land-Cover Trajectories Model, for Zones with and without

Logging Concessionsa

Observed Deforested and Forested Area (%)

PredictedProbability ofDeforestation

In Zonewithout

Logging (A)

In Zonewith

Logging (B)

.0.90 31.20 17.121.91 1.96

.0.80 63.54 38.298.95 9.12

.0.50 85.79 82.9844.51 70.12

a The top number in row class is the proportion of observeddeforested areas (%); the bottom number is the proportionof forested area (%) affected by a given range of probabilityof deforestation in 1991.


Conclusion

Remote-sensing observations revealed a con-tinuous trend of forest clearing and forest degra-dation in southern regions of Cameroon, butwith a highly fluctuating rate. The rates of defor-estation have sharply increased in the area ofBertoua in the mid-1980s, followed by a slightdecline in the early 1990s. This study also re-veals that the new institutional tools for forestmanagement and land-use planning in Cam-eroon have not yet provided a sustainable answerto the problems of degradation of the forests.

Successive observations highlight the dy-namic character of land-cover changes. A sig-nificant proportion of the areas subject to aland-cover conversion are subject to anotherchange in the following years. The matrix oftransition probabilities between different land-cover types is nonstationary. Also, there areclear trajectories of land-cover changes that dis-play a spatial pattern suggesting a centrifugal ex-pansion of deforestation around towns andalong the road network. The spatial distributionof land-cover-change trajectories can be mod-eled on the basis of cultural and natural land-scape attributes that describe the land rent offorest areas. These spatial models indicate the“spatial inertia” of land-cover-change processes:changes take place around locations of recentprevious changes. Openings in the forest coverincrease accessibility into the forest. There istherefore a strong “spatial spread” effect of land-cover changes. Taking into account land-covertrajectories (i.e., succession of land-covers overmore than two dates) allows for more reliablespatial projections of areas at risk of being defor-ested than taking into consideration only theprevious observation. Land-cover changes aretherefore best described as multiple-order pro-cesses, as there are generic and predictable tra-jectories of land-cover changes.

The spatial relationships analyzed in thisstudy allowed us to better understand the pro-cess of deforestation in this region and, in par-ticular, to formulate a hypothesis on the role ofroads with respect to deforestation. Statisticalresults suggested that roads mostly increased theaccessibility of the forest for migrants ratherthan provided incentives for a transformation oflocal subsistence agriculture into market-orientedfarming systems. The process is therefore more acolonization of forest areas by migrants than thedevelopment of local agriculture in response to

an expansion of distant urban markets. An al-ternative hypothesis could be that, once theareas surrounding towns had been cleared, theattraction of roads is simply a new manifestationof the attraction of towns. At the end of the ob-servation period, however, the rate of deforesta-tion far from towns (but along the roads) is ashigh as the rate in the vicinity of towns at the be-ginning of the observation period. The fact thatforest clearing for fallow agriculture takes placeon medium rather than high-aptitude soilssuggests that deforestation is more strongly in-fluenced by accessibility variables than by soilsuitability.

The main weaknesses of the modeling ap-proach adopted in this study are that: (1) it is dif-ficult to separate correlation from causality or todetermine the direction of causality (Kaimo-witz and Angelsen 1997); (2) only measurable,location-specific variables influencing the deci-sion to convert land-cover have been integratedinto the model, while underlying and more dis-tant driving forces are ignored; and (3) themodel assumes that cultural and sociopoliticalfactors have a minor influence on the selectionof sites to be cleared. This study therefore onlyprovides a representation of reality from one spe-cific angle, which is due to be complemented byother disciplinary vantage points before a com-prehensive understanding of the land-coverchange processes in the region can be acquired.This study, however, generates several researchquestions to be addressed by other disciplinesand approaches. For example, the patterns ofland-use change revealed by this study havebeen related to household survey data that illu-minate the macroeconomic factors driving someof these changes (Mertens et al. 2000).

The spatial model developed in this study al-lows us to conduct simulations of likely impactsof human actions leading to a transformation ofthe landscape (e.g., road projects) on key land-scape attributes (e.g., biodiversity). Currently,several road projects or major logging conces-sions exist in southern Cameroon. They lead toheated debates, notably from nature conservationnongovernmental organizations, yet at presentthere is no objective way to conduct regional-scale environmental impact assessments. Themodeling framework developed in this studyprovides a most suitable basis to perform suchsimulations. The spatial-model design devel-oped here can lead to a decision support systemthat can be used as a policy tool to conduct en-


vironmental-impact assessment studies duringthe approval process of any project affecting thelandscape. It generates spatially explicit projec-tions that allows us to better understand the in-teraction between development projects andnature-conservation imperatives.

This research showed that land-cover changesassociated with the granting of logging conces-sions and with shifting cultivation not orientedtoward the market were not well predicted bythe current model design. Other variables (e.g.,land tenure) need to be integrated. Also, it islikely that the modeling of these activities needsto adopt a broader perspective (i.e., not exclu-sively spatial) and, in the case of logging, to in-tegrate more remote driving forces such as na-tional or international demand for timber.

Acknowledgments

This work was performed in the research program onsatellite remote sensing for the Belgian State, Services ofthe Prime Minister, Office for Scientific, Technical andCultural Affairs (contract T4/DD/001). We are gratefulto D. Margot for programming work done for this study,to D. Peeters for advice on the statistical analyses, and tofour anonymous reviewers who greatly helped in im-proving an earlier version of the manuscript.

Notes

1. It was recognized as Biosphere Reserve underUNESCO’s Man and the Biosphere Program in1981 and was inscribed on the World Heritage listin 1984 (Sayer et al. 1992).

2. The change map derived from a comparison oftwo classifications exhibits accuracies similar tothe product of multiplying the accuracies of eachindividual classification (Stow et al. 1980).

3. The image processing and GIS analyses were per-formed with the Modular GIS Environment soft-ware package of Intergraph corporation.

4. Because the dependent variable of each modelcontains four trajectories, and the standard re-sponse function is used to compute three general-ized logits for each population (SAS 1989).

5. By contrast, a map of the predicted probabilities offallow agriculture (not displayed on the figure)shows that the higher probabilities of occurrenceare less related to roads and are more concentratedaround the small forest openings.

References

Agresti, A. 1990. Categorical Data Analysis. NewYork: John Wiley.

Alves, D.S., and Skole, D.L. 1996. CharacterizingLand-Cover Dynamics Using Multi TemporalImagery. International Journal of Remote Sensing17: 835–39.

Arbia, G. 1989. Spatial Data Configuration in StatisticalAnalysis of Regional Economic and Related Prob-lems. Dordrecht: Kluwer Academic Publishers.

Atayi, E.A., and Knipscheer, H.C. 1980. The Market-ing of Food Crops in the Eastern Province of Cam-eroon. Agricultural Economics Discussion Paper11. Ibadan, Nigeria: International Institute ofTropical Agriculture.

Bian, L., and West, E. 1997. GIS Modeling of ElkCalving Habitat in a Prairie Environment withStatistics. Photogrammetric Engineering and Re-mote Sensing 63: 161–67.

Boserup, E. 1965. The Conditions of AgriculturalGrowth. London: Allen and Unwin.

Brown, K., and Pearce, D.W. 1994. The Causes ofTropical Deforestation. London: UCL Press.

Brown, S., and Lugo, A. 1990. Tropical SecondaryForests. Journal of Tropical Ecology 6: 1–32.

———; Iverson, L.R.; and Lugo, A. 1993. Land Useand Biomass Changes of Forests in PeninsularMalaysia during 1972–1982: Use of GIS Analy-sis. In Effects of Land-Use Change on AtmosphericCO2 Concentrations: Southeast Asia as a CaseStudy, ed. V.H. Dale. New York: Springler Verlag.

Centre Géographique National. 1978. Cartes duCameroun à l’échelle 1:200 000. Yaoundé, Cam-eroon: Ministère de l’Economie et du Plan.

Chomitz, K.M., and Gray, D.A. 1996. Roads, LandUse, and Deforestation: A Spatial Model Ap-plied to Belize. World Bank Economic Review 10:487–512.

Congalton, R.G. 1991. A Review of Assessing the Ac-curacy of Classifications of Remotely Sensed Data.Remote Sensing of Environment 37 (1): 35–46.

Coppin, P.R., and Bauer, M.E. 1996. Digital ChangeDetection in Forest Ecosystems with Remote Sens-ing Imagery. Remote Sensing Reviews 13: 207–34.

Domencich, T.A., and McFadden, D. 1975. UrbanTravel Demand: Behavioural Analysis. Amster-dam: North-Holland.

Finifter, R., and Verna, G. 1996. Le prix et la valeur:application au négoce international des boistropicaux africains. Note de recherche du réseauthématique “Entreprenariat,” AUPELF-UREF96: 1–15.

Fischer, G.; Ermoliev, Y.; Keyzer, M.A.; and Rosen-zweig, C. 1996. Simulating the Socio-Economic andBiogeophysical Driving Forces of Land-Use and Land-Cover Change: The IIASA Land-Use ChangeModel. Working Paper WP-96-010. Laxenburg: In-ternational Institute for Applied Systems Analysis.

Gastellu-Etchegorry, J.P., and Sinulingga, A.B. 1988.Designing a GIS for the Study of Forest Evolu-tion in Central Java. Tijdscrift voor Economischeen Sociaale Geografie 79 (2): 93–103.


Hosmer, D.W., and Lemeshow, S. 1989. Applied re-gression analysis. New York: John Wiley.

Jakubauskas, M.E.; Lulla, K.P.; Mausel, P.W. 1990.Assessment of Vegetation Change in a Fire-Altered Forest Landscape. Photogrammetric Engi-neering and Remote Sensing 56: 371–77.

Jobson, J.D. 1992. Applied Multivariate Data Analysis,vol. 2: Categorical and Multivariate Methods. NewYork: Springer.

Jones, D.W., and O’Neill, R.V. 1992. Endogenous En-vironmental Degradation and Land Conserva-tion: Agricultural Land Use in Large Region.Ecological Economics 6: 79–101.

——— and ———. 1993. Land Use in the Presenceof an Atmosphere Externality, with and withoutCorrective Taxes. Journal of Regional Science 33:457–80.

Kaimowitz, D., and Angelsen, A. 1997. EconomicModels of Tropical Deforestation: A Review.Bogor, Indonesia: Center for International For-estry Research (CIFOR).

Karshenas, M. 1994. Environment, Technology, andEmployment: Towards a New Definition of Sus-tainable Development. Development and Change25: 723–56.

Karsenty, A., and Maître, H.F. 1993. Etude des modal-ités d’exploitation du bois en liaison avec une gestiondurable des forêts tropicales humides. Brussels:Commission des Communautés EuropéennesDG XI.

Kasperson, J.X.; Kasperson, R.E.; and Turner II, B.L.1995. Regions at Risk. Tokyo: United NationsUniversity Press.

Lambin, E. 1994. Modeling Deforestation Processes: AReview. TREES series B: Research Report n81,European Commission, EUR 15744 EN.

———. 1997. Modeling and Monitoring Land-Cover Change Processes in Tropical Regions.Progress in Physical Geography 21: 375–93.

Liu, D.S.; Iverson, L.R.; and Brown, S. 1993. Ratesand Patterns of Deforestation in the Philippines:Application of Geographic Information Sys-tems Analysis. Forest Ecology and Management57: 1–16.

Lucas, R.M.; Honzak, M.; Foody, G.M.; Curran, P.J.;Corves, C. 1993. Characterizing Tropical Sec-ondary Forests Using Multi-Temporal LandsatSensor Imagery. International Journal of RemoteSensing 14: 3061–67.

Ludeke, A.K.; Maggio, R.C.; and Reid, L.M. 1990.An Analysis of Anthropogenic DeforestationUsing Logistic Regression and GIS. Journal ofEnvironmental Management 31: 247–59.

Macleod, R.D., and Congalton, R.G. 1998. A Quan-titative Comparison of Change-Detection Al-gorithms for Monitoring Eelgrass from RemotelySensed Data. Photogrammetric Engineering andRemote Sensing 64 (3): 207–16

Markham, B.L., and Barker, J.L. 1987. Radiometric

Properties of U.S.-Processed Landsat MSS Data.Remote Sensing of Environment 22: 39–71.

Martin, D., and Segalen, P. 1966. Carte pédologiquedu Cameroun Oriental. Paris: Office de la Re-cherche Scientifique et Technique Outre-Mer(ORSTOM).

Matheron, G. 1970. La théorie des variables régional-isées et ses applications. Les cahiers du Centre deMorphologie Mathématique de Fontainebleau,Fasc. 5.

Menard, S. 1995. Applied logistic regression analysis.Quantitative Applications in the Social Sci-ences n8106. London: Sage Publications.

Mertens, B., and Lambin, E. 1997. Spatial Modelingof Deforestation in Southern Cameroon: SpatialDisaggregation of Diverse Deforestation Pro-cesses. Applied Geography 17: 143–62.

———; Sunderlin, W.; Ndoye, O.; and Lambin, E.2000. Impact of Macro Economic Transforma-tions on Deforestation in South Cameroon: In-tegration of Household Survey and RemotelySensed Data. World Development 28: 983–1000.

Meyer, W.B., and Turner II, B.L. 1994. Changes inLand Use and Land-Cover: A Global Perspective.Cambridge: Cambridge University Press.

Milchener, W.K., and Houhoulis, P.F. 1997. Detec-tion of Vegetation Changes Accuracy with Ex-tensive Flooding in a Forested Ecosystem. Photo-grammetric Engineering and Remote Sensing 63:1363–74.

Ministry of Forestry and Environment (MINEF).1995. National Forestry Action Programme ofCameroon. Yaoundé, Cameroon: MINEF.

———. MINEF. 1996. Plan national de gestion de líen-vironnement au Cameroun, vol. 1, rapport princi-pal. Yaoundé, Cameroon: Ministère de l’Envi-ronnement et des Forêts.

Muchoney, D.M., and Haack, B.N. 1994. ChangeDetection for Monitoring Forest Defoliation.Photogrammetric Engineering and Remote Sensing60 (10): 1243–51.

Narumalani, S.; Jensen, J.R.; Althausen, J.D.;Burkhalter, S.; and Mackey, H.E., Jr. 1997.Aquatic Macrophyte Modeling Using GIS andLogistic Multiple Regression. PhotogrammetricEngineering & Remote Sensing 63: 41–49.

Panayotou, T., and Sungsuwan, S. 1989. An Economet-ric Study of the Causes of Tropical Deforestation:The Case of Northeast Thailand. DevelopmentDiscussion Paper 284, Cambridge: Harvard Insti-tute for International Development.

Pereira, J.M.C., and Itami, R.M. 1991. GIS-BasedHabitat Modeling Using Logistic Multiple Re-gression: A Study of the Mt. Graham Red Squir-rel. Photogrammetric Engineering and Remote Sens-ing 57: 1475–86.

Pilon, P.G.; Howarth, P.J.; Bullock, R.A.; Adeniyi,P.O. 1988. An Enhanced Classification Ap-proach to Change Detection in Semiarid Envi-


ronments. Photogrammetric Engineering and Re-mote Sensing 54: 1709–16.

Pokam, J. 1996. Les migrations dans la zone de forêthumide du Cameroun. Draft version. Bogor, Indo-nesia: Center for International Forestry Research.

Press, S.J., and Wilson, S. 1978. Choosing betweenLogistic Regression and Discriminant Analysis.Journal of the American Statistical Association 73:699–705.

Riebsame, W.E.; Meyer, W.B.; and Turner, B.L. II.1994. Modeling Land Use and Cover as Part ofGlobal Environmental Change. Climatic Change28: 45–64.

Sader, S.A., and Joyce, A.T. 1988. Deforestation Ratesand Trends in Costa Rica. Biotropica 20: 11–19.

———. 1995. Spatial Characteristics of Forest Clearingand Vegetation Regrowth as Detected by LandsatThematic Mapper Imagery. Photogrammetric Engi-neering and Remote Sensing 61: 1145–51.

SAS Institute. 1989. SAS/STAT User’s Guide, Version6, 4th ed., vol. 1 and 2. Cary, NC: SAS Institute.

Sayer, A.S.; Harcourt, C.S.; and Collins, N.M., ed.1992. The Conservation Atlas of Tropical Forests:Africa. London: Macmillan.

Southgate, D. 1990. The Causes of Land Degradationalong “Spontaneously” Expanding AgriculturalFrontiers in the Third World. Land Economics66: 93–101.

Stokes, M.A.; Davis, C.S.; Koch, G.G. 1995 Categor-ical Data Analysis Using the SAS System. Cary,NC: SAS Institute.

Stone, S.W. 1998. Using a Geographic InformationSystem for Applied Policy Analysis: The Case ofLogging in the Eastern Amazon.

Stow, D.A.; Tinney, L.R.; and Estes, J.E. 1980. Deriv-ing Land Use/Land Cover Change Statisticsfrom Landsat: A Study in Prime AgriculturalLand. Proceedings of the 14th International Sympo-sium on Remote Sensing of Environment, pp.1227–37. Ann Arbor, MI.

Tucker, C.J.; Dregne, H.E.; and Newcomb, W.W. 1991.Expansion and Contraction of the Sahara Desertfrom 1980 to 1990. Science 253: 299–301.

United Nations Food and Agriculture Organization(FAO). 1995. Forest Resources Assessment 1990.Tropical Countries. Forestry Paper 112. Rome:FAO.

———. 1996. Development of a Step-by-Step Guidefor the in-Situ Conservation of Forest GeneticResources Forest Genetic Resources 24.

Vanclay, J.K. 1995. Modeling Land Use Patterns at theForest Edge: Feasibility of a Static Spatial Model.In Ecological Economics Conference, pp. 78–84.Coffs Harbour, NSW, 1995, Centre for Agricul-tural and Resource Economics, University ofNew England, Armidalen, NSW, Australia.

Walker, R.T. 1987. Land Use Transition and Defores-tation in Developing Countries. GeographicalAnalysis 19: 18–30.

Wrigley, N. 1985 Categorical Data Analysis for Geogra-phers and Environmental Scientists. New York:Longman.

Correspondence: Centre for International Forestry Research, P.O. Box 6596, JKPWB, Jakarta 10065, Indonesia,email [email protected] (Mertens); Department of Geography, Catholic University of Louvain, placeLouis Pasteur 3, B-1348 Louvain-la-Neuve, Belgium, email [email protected] (Lambin).

Documents

Land-Cover-Change Trajectories in Southern Cameroondusk.geo.orst.edu/prosem/PDFs/bosley_landcover.pdf · 2005-01-11 · Land-Cover-Change Trajectories 469 of von Thünen and Ricardo