This article was downloaded by: [UOV University of Oviedo]On: 04 November 2014, At: 10:32Publisher: Taylor & FrancisInforma Ltd Registered in England and Wales Registered Number: 1072954 Registeredoffice: Mortimer House, 37-41 Mortimer Street, London W1T 3JH, UK

New Zealand Journal of Crop andHorticultural SciencePublication details, including instructions for authors andsubscription information:http://www.tandfonline.com/loi/tnzc20

An expert system for diagnosis of pests,diseases, and disorders in apple cropsR. H. Kemp a , T. M. Stewart b & A. Boorman ca Department of Computer Science , Massey University , PalmerstonNorth , New Zealandb Department of Plant Health , Massey University , PalmerstonNorth , New Zealandc Department of Horticultural Science (ex) , Massey University ,Palmerston North , New ZealandPublished online: 01 Jun 2012.

To cite this article: R. H. Kemp , T. M. Stewart & A. Boorman (1989) An expert system for diagnosis ofpests, diseases, and disorders in apple crops, New Zealand Journal of Crop and Horticultural Science,17:1, 89-96, DOI: 10.1080/01140671.1989.10428014

To link to this article: http://dx.doi.org/10.1080/01140671.1989.10428014

PLEASE SCROLL DOWN FOR ARTICLE

Taylor & Francis makes every effort to ensure the accuracy of all the information (the“Content”) contained in the publications on our platform. However, Taylor & Francis,our agents, and our licensors make no representations or warranties whatsoever as tothe accuracy, completeness, or suitability for any purpose of the Content. Any opinionsand views expressed in this publication are the opinions and views of the authors,and are not the views of or endorsed by Taylor & Francis. The accuracy of the Contentshould not be relied upon and should be independently verified with primary sourcesof information. Taylor and Francis shall not be liable for any losses, actions, claims,proceedings, demands, costs, expenses, damages, and other liabilities whatsoever orhowsoever caused arising directly or indirectly in connection with, in relation to or arisingout of the use of the Content.

This article may be used for research, teaching, and private study purposes. Anysubstantial or systematic reproduction, redistribution, reselling, loan, sub-licensing,systematic supply, or distribution in any form to anyone is expressly forbidden. Terms &

Conditions of access and use can be found at http://www.tandfonline.com/page/terms-and-conditions

Dow

nloa

ded

by [

UO

V U

nive

rsity

of

Ovi

edo]

at 1

0:32

04

Nov

embe

r 20

14

New Zealand Journal ofCrop and Horticultural Science, 1989, Vol. 17 : 89-960114-0671/89/1701-0089$2.50/0 © Crown copyright 1989

An expert system for diagnosis of pests, diseases, and disordersin apple crops

89

R.H.KEMPDepartment of Computer ScienceMassey UniversityPalmerston North, New Zealand

T. M. STEWART

Department of Plant HealthMassey UniversityPalmerston North, New Zealand

A. BOORMAN

Department of Horticultural Science (ex)Massey UniversityPalmerston North, New Zealand

Abstract A prototype expert system is beingdeveloped to diagnose pests, diseases, and disordersin apple trees. The system has the dual aims ofsupplying diagnostic advice to apple growers andteaching students to diagnose problems effectively.Notable features of the system include: clearseparationbetween heuristicknowledge and domain!real world facts, a flexible certainty factor system,and a user-friendly environment.

Keywords knowledge-based systems; expertsystems; apples; pest; disease; nutrient disorder;diagnosis

INTRODUCTION

Since the 1970s, computers have been usedincreasingly in cropprotection throughout the world.Popular uses include data collection and analysis,predicting and managing pestand diseaseoutbreaks,and providing information from on-line databases.A recent development has been the growth of

Received 29 August 1988; accepted 12 October 1988

computer based 'expert systems' for the prediction,diagnosis, and/or treatment ofplantproblems (Sandset al. 1986; Latin et al. 1987).

True rule-based expert systems differ fromtraditional decision-making programs that follow asimple pattern (Buchanan & Shortliffe 1985). Thelatteruse strict algorithms working on precise data tocome to a definite conclusion. This poses problemsif the data are in any way incomplete or incorrect, asthe conditional branching may lead the user on aone-way trip down the wrongpath; a similar situationto a taxonomic key where one observation has beenwrongly interpreted. Also, these programs do nothandle new data easily or allow the user to ask whythey came to the conclusions they did. An expertsystem however, contains a knowledge base whichis analysed by a logic interpreter (interface engine),prompting for and 'weighing up' all the relevantinput from the user before offering likely possibilities.In this way, the software acts very much like ahuman expert, who must ask probing, relevantquestions, assess all the facts available (oftenincomplete) and come to the most likely conclusion.Like the human expert, the system can be asked tojustify its decision.

At Massey University, an experimental systemhas been designed to diagnose pests, diseases, andnutritional disorders in apple crops. The system hasan unusual two-tier structure that is particularlysuited to this type of application.

Developingan expertsystem is a lengthyprocess,and a fully completed, validated apple pest, disease,and disorder diagnostic system may be some yearsaway. It was the authors opinion however, that aninterim report describing the design of the systemwould be of interest to agricultural and horticulturalscientists.

OUTLINE OF THE SYSTEM

The system is completely interactive, gatheringinformation from the user (who may be an orchardist,advisor, or student) about their particularplant health

Dow

nloa

ded

by [

UO

V U

nive

rsity

of

Ovi

edo]

at 1

0:32

04

Nov

embe

r 20

14

90 New ZealandJournalof Crop and Horticultural Science, 1989, Vol. 17

• File Selection

Leaf Test

Input uelues for known minerals.(Ieaue blank any that are unknown).

range range

Nitrogen Q 0-3.5 Sulphur CJ 0-0.7

Phosphorus CJ 0-0.35 Manganese CJ 0-250ppm

Potassium c=J 0-2.5 Iron c=J 0-600ppm

Magnesium CJ 0-0.5 Zinc CJ 0-90ppm

cetctum CJ 0-3.0 Boron CJ 0-90ppm

OK [ancel

Fig. 1 Full screenwith leaftestdialoguebox.

problem, during a consultative session. From theinformationobtained,a list of possibleproblemsisset up and the likelihood of each being present isestimated. In future versions, a treatmentrecommendation may be provided by the systemitself.

Anapplepest,disease, anddisorderexpertsystemmaycovera largenumberof planthealthproblems.Some,suchas spraydamage, canbe socomplexthatdiagnosis not only requires an expert, but reliesheavilyon the growerhavingcloselymonitoredthemanagement of theproperty,andbeingable to giveprecisedetailson thehistoryof thecropoverthelast(and perhaps previous) season. It was decided todevelop the expert system stage by stage, startingwithsimple,well-documentedproblems andworkingup to the moredifficultones suchas the above.Thepackagecurrentlycoverspests,fungalandbacterialdiseases,and nutrientdeficiencies/toxicities.

The program is written in Prolog and wasdevelopedona Prime750. It wasthentransferred toan AppleMacintosh* usingLPAProlog in order toimprovethe interfaceby usingthe standardWIMPs(Windows, Icons, Mouse, Pull-down menus)facilities. Prologwaschosenbecauseavailable shellsdidnotprovidetheflexibility requiredtoexperimentwithappropriatestructuresandcontrolmechanisms.Prolog not only has low level facilities to enablespecialised structures and processes to be directlymodelled, butalsoincludes powerful meta-interpreter

*Apple Macintosh is a trade mark licensed to AppleComputer, Inc.

functions to allow economical implementation ofthese features.

A USER VIEW OF THE SYSTEM

Atthe startof a session,theuseris askedif he wantsmoredetailon how the systemworks.Experiencedoperators can by-pass this stage. Questions of ageneralnaturearethenaskedin ordertoeliminateasmanyproblemsaspossibleandalsotodetermine themostlikelyones.Theseareexaminedin moredetailin the next phase. At the end of the session anestimate of each of the possible problems andlikelihood of its occurrence is presentedto the user.

The initialquestions askedincludedetailsaboutlocation, types of apples affected,soil and leaf testresults (Fig. 1), and general questions concerningleafandfruitsymptoms. Infieldtests,theanswerstomany questions were found to be a straight 'yes','no', or 'unknown'. For this reason the answersavailableon many screens have these as the mainalternatives (Fig.2). Sometimes, however, the usermaywishto shadehisanswer. For example,he maynotbe surewhethertheleaveshavechlorosis. In thiscase he can select 'possibly' or 'possibly not' fromthe alternatives and is then presentedwith a followupscreenaskinghimtogivemoredetailonhowsureof his replyhe is (seeFig. 3). This two-level systemseemstobe quiteeffectiveinpractice.In the90%ofthe cases when a straight answer is applicable, theuser needs only select from a small number ofalternatives and does not have to answer twoquestions. In the few(butoftencrucial)caseswhere

Dow

nloa

ded

by [

UO

V U

nive

rsity

of

Ovi

edo]

at 1

0:32

04

Nov

embe

r 20

14

Kemp et al.-An expert system

Initial Questions

definitely ~Are the Ieauas particularly possiblysmall ? possibly not

definitely not

fOunlmown

( OK ~l ( Cancel 1( Why? 1( Clarify) 4f

Fig. 2 Standard question and answer format.

Fig.4 Presentingtheuserwitha list of possibilities.

91

I nitial Questions

unsure rHow sure are you about fairly confidentyour answer? quite confident

i I .

( OK ~J ( Cancel 1( Whl.f? 1( CI<1l"if9 1 9

Fig. 3 Graded scale for uncertain answers.

Possible disorders

On the right is a list of problemsthat appear to merit furtherinuestigation.Click OK If you accept these.I f there are any you do NOTwish to be inuestigated furtherthen highlight the appropriateproblem(s) and then click OK.

Calcium deficiency QNitrogen deficiencyMagnesium deficiency ;0

OK IJlHhlJ? 1

Cancel

"there is some doubt, he may express his degree ofuncertainty as a natural language word or phrase.

After the initial stage, the system will haveaccumulated enough information to arrive at sometentative conclusions. A set of 'hypotheses' is thenpresented to the user. These are the problems thatwill be investigated further (see Fig. 4). The user hasthe option of removing any of these from the list (athis own risk). This feature is provided since he mayhave information which would preclude a particularproblem. For example, chloride toxicity would onlynormally be considered if the orchard were near thesea or subject to sea breezes and so the user mightremove this from the list if it was not applicable.

Now, each of the hypotheses is investigated inturn in more detail. An affirmative response to somequestions provides positive evidence. For example,

the presence oftiny orange-white maggots in tightlyrolled young leaves gives strong evidence of leafcurling midge (Dasyneura mall). In other cases ananswer of 'yes' reduces the likelihood of a problem.Necrotic spotting on the leaves of apple trees, forinstance, indicates that magnesium deficiency isunlikely to be the major problem. It may be that theanswers to individual questions allow the system tomake positive or negative inferences regarding otherhypotheses, too, but this is performed automaticallyby the system.

At any stage during the questioning, the usermay request further information relating to what isbeing asked. Currently the system allows for threekinds of elaboration: 'clarify', 'why', and 'explain'.

Users will differ in their understanding of plantscience jargon. Terminology is, therefore, kept to a

Dow

nloa

ded

by [

UO

V U

nive

rsity

of

Ovi

edo]

at 1

0:32

04

Nov

embe

r 20

14

92 New Zealand Journal of Crop and Horticultural Science, 1989, Vol. 17

Current problem under inuestigation is copper deficiency

Clarification has been requested for the question:

'Is there any euidence of ..-osetting l'

It tnara are short Internodes on the shoots 0..- there Islittle eHtension at the tips then this is called rosetting

Fig. 5 Clarification displaybox.

OK ) ( Cancel) •Cur..-ent problem under inuestigation is copper deficiency

The 'Why' option has been selected fur the question:

'Is tnere a nrontem with the young shoots l'

Fig. 6 Example of 'why'explanation.

If the answer is 'yes' to the aboue question thenthere may be rosettlng.

rosetting is a symptom of copper deficiency

OK ) ( Cancel

( EHplain ) •minimum in the system, and where used can beexplained by way of the 'clarify' option (see Fig. 5).

It is often useful to know the reason behind aquestion to understand the point of it. The 'why'facility provides for this (see Fig. 6). As in manyexpert systems, the user can be given some explan­ation of why a question is being asked i.e., what thequestioner is leading up to.The question may pertaindirectly to a characteristic symptom or, in othercases, to a piece of evidence that would tend toreduce the likelihood of a condition (see Fig. 7).

The third feature ('explain') is mainly used foreducational purposes. This enables the user to seewhy a heuristic relationship holds. For example, ifhe is asked if there is a flour-like coating on theleaves he can select 'explain' and find out why thatphenomenon is associated with powdery mildew(Fig. 8).

After the second part of the session, the systemuses the accumulatedevidence toproduce likelihoods

ranging through from 'almost certain' to 'almostcertainly not' for each of the problems considered.This information may be accessed in graphical form(see Fig. 9).

KNOWLEDGE BASE STRUCTURE

The knowledge base has two parts-the heuristicknowledge and taxonomic/real world generalknowledge. Heuristic knowledge is the domainexpert's knowledge of the links between symptoms,causes, indicators etc. and the pests, diseases, ordisorders that may be associated with these. Theseparation is in line with Clancey's heuristicclassification model (Clancey 1985).

The heuristics that link symptoms to problemsare easy to pin-point for the expert. For example,alkaline soil may cause manganese deficiency in aplant. Thus if the soil is known to be alkaline then

Dow

nloa

ded

by [

UO

V U

nive

rsity

of

Ovi

edo]

at 1

0:32

04

Nov

embe

r 20

14

Kemp et al.-An expert system

Fig. 7 Example of 'why'explanation with negativeevidence.

93

Current problem under inuestigation is magnesiumdeficiency

The 'Why' option has been selected for the question:

'Is there euidence of necrotic spotting on the leaues l'

If plants haue necrotic spotting they are unlikely to besuffering from magnesium deficiency

OK ) [ Cancel

( [Hplain "Fig.8 Example of 'explain'display box. Current problem under inuestigation is powdery mildew

The '[Hplain' option has been selected for the question:

'Is there a flour-like coating on the leaues l'

Powdery mildew is a fungus disease that attacks youngplant tissue. Unlike most other fungi, this species growsmainly on the outside of the host feeding off epidermalcells. Because it deuelops eHternally, it is easily seen,giuing young leaues and twigs the appearance of beingdusted with flour.

,--_O_K_ttJ [ Cancel

manganese deficiency may be suspected. Conversely,if manganese deficiency is being investigated thenevidence may be obtained by checking whether thesoil is alkaline. A rule of this description may berepresented as:

alkaline soil JE manganese deficiency

In fact, of course, an unqualified relationship of thiskind in plant problem diagnosis would not be helpfulin isolation since, not only does alkaline soil causeother problems but also manganese deficiency maybe present even if the soil is not alkaline. What isneeded is some measure of uncertainty linking theindicator with the condition. For instance, the rule:

alkaline soil JE manganese deficiency

may have an associated degree of belief of 0.8 that,when there is alkaline soil, there is manganese

deficiency. However, if the soil is not alkaline, thenthe likelihood of the deficiency may only be decreasedby 0.3. The first figure represents sufficiency and thesecond one is a measure of necessity. These twofactors may be represented as parameters within aProlog statement:

manganese_deficiency (0.8 ,0.3) if alkaline_soil

By allowing the parameters to range between -1and +I, the corresponding measure of disbeliefgiven by positive evidence and measure of beliefcontributed by lack of evidence can be included. Inthis way the full range of necessity and sufficiencyfor the presence and absence of symptoms may beconcisely represented.

The system also needs to cope with disjunctions,conjunctions, and negations such as those in thefollowing rule:

Dow

nloa

ded

by [

UO

V U

nive

rsity

of

Ovi

edo]

at 1

0:32

04

Nov

embe

r 20

14

94 New Zealand Journal of Crop and Horticultural Science, 1989, Vol. 17

* File Selection

Results of nnolysisFig. 9 Presentation of resultsin graphical form.

Definilely Nol

I I

Unknown

Degrees of Belief

Defi nilel y

Choose Run from the Selection menu to try another query

Choose Quit from the File menu to finish

boron_deficiency (0.6, 0) if (rough_bark orsplitbark) and not bark_measles.

Typically there will be about a dozen heuristic ruleslinking indicators to problems for each pest, disease,or disorder.

The remainder of the knowledge base containshierarchical information about the problems. This isrepresented in tree structures such as that shown inFig. 10. There is no point in storing this hierarchy asproduction rules, although it can be done if a shell isbeing used to implement the system. A much moreconvenient approach is to use a standard treerepresentation method employing facts or lists. Suchstructures can easily be set up in Prolog (Clocksin1984).

THE INFERENCE ENGINE

The inference engine performs a standard initialforward chainingprocedure to gather enough data toproduce hypotheses and then switches to a goaldirected mode for the secondpartof the consultation.However, the goal directed mode works only on theheuristic partof theknowledge base. For the problemhierarchy, the engine works in a top down fashion.This is done to reduce the number of questions thatare likely to be needed.

To illustrate this: suppose we suspect that thetrees are suffering from apple black spot (Venturiainaequalis) for which one of the symptoms is corkyblack spots on the fruit. The system would obtain thelink between the disease and the symptom from a

production rule. The first stage is to ascertain whetherthe system already knows if this symptom is present.If so, then the user need not be asked and theevidence can be directly added to the system'sinformation store. If the answer is not known then,instead of inquiring directly, the system checks tosee if a question higher in the problem hierarchy hasbeen asked; in this case whether there are necroticspots on the fruit. Applying the same logic at thislevel, if no answer is known then the system ascendsthe tree again and determines whether the fruit isshowing any signs at all of necrosis. Eventuallyeither the top of the tree or a node where the answerto the question above is known will be reached. Iftheanswer to that question is 'no' or 'unknown' thenthis answer can be propagated downwards. If theanswer is 'yes' then the more specific questions areeach asked in turn until a non-positive answer isfound or the leaf of the hierarchy tree is reached.

By asking these more general questions first, thesystem can quickly eliminate a large number ofproblems and associated questions. For example, ifthere is no fruit problem then not only can it bededuced that there are no corky spots on the fruit butalso the answers to questions concerning fruitdiscoloration, distorted fruit, diseased fruit and so oncan be inferred without asking the user.

In the current version of the program, each of thehypotheses is tested in tum with detailed questionsto build up the evidence either for or against eachproblem occurring. This is a simplistic approach butis surprisingly effective since it is straightforwardfor the user to follow. It is particularly appropriatefor teaching purposes.

Dow

nloa

ded

by [

UO

V U

nive

rsity

of

Ovi

edo]

at 1

0:32

04

Nov

embe

r 20

14

Kemp et al.-An expert system

~fruit necrosis fruit distortion fruit chlorosis

/necrotic spotting

Icorky black spots

Fig. 10 Part of problem taxonomy in the expert system.

As a hypothesis is followed through, it maybecome obvious that a single piece of evidencemakes that possibility a certainty or that thecumulative evidence is incontrovertible. In eithercase the system will break off the consideration ofthat hypothesis. The same happens if a large amountof negative evidence has been collected.

EXPLANATION FACILITIES

It has already been noted that the system containsthree types of explanation. Firstly, the 'clarify' optionallows the user to obtain more information regardingtechnical or imprecise terms used in the questions.We have seen several examples of the former. Anexample of the latter is the question:

"Is there any premature defoliation?"

In this instance, the clarification might specify that,if significant leaf fall occurs before early or midautumn, then this can be considered prematuredefoliation. This kind ofinformation can be stored intabular form.

Explanations of the 'why' form are usuallyprovided in expert systems. Often they involve apresentation of the rules currently being appliedwhich can be useful for debugging but the utility ofsuch explanations for end-users is debatable. Theyalso often follow through the deductive chain onestep at a time which can be tedious and annoying.

In our system, if 'why' is selected, then the useris given the text associated with the correspondingheuristic node at the bottom of the hierarchy tree aswell as the link between the currentquestion and thisbottom node (see Fig. 6). The intermediate nodes inthe problem hierarchy are automatically skippedsince these involve real world knowledge which is

95

important for the system but normally self-evidentto the user. The user does not want to be given all theintermediate information that leaf discoloration is aleaf symptom, and that necrosis is a form of leafdiscoloration and that marginal necrosis is a form ofnecrosis.

Implementation ofthe 'why' feature is facilitatedby the way the knowledge is stored. The ultimategoal is the question in the terminal node of the treeand the intermediate branches between the questionbeing asked and this one can be easily omitted.

The heuristic indicators at the bottom of thehierarchy tree are crucial to the problem-solvingprocess but will not necessarily be important to theuser who may just want a solution and may not wishto know the fundamental underlying theory behindthe diagnosis. However, clear reasons for the linksbetween symptoms and disorders, for example,should be available in case the user does want toincrease his understanding of the processes involved.In an educational environment, an appreciation ofthese links may be vital.

Such reasoned explanations are provided in oursystem when the user selects 'explain' on a 'why'screen. Technical information is then displayed suchas that shown in Fig. 8. Again, because of theseparation of the types of knowledge, it is simple toprovide appropriate explanations for each of theheuristic rules.

By incorporatingextensiveexplanation facilities,students undergoing training in the diagnosis ofplant pest or disease problems can use the programto assist in learning the procedures involved in thistask. Growers using the system benefit by gainingmore knowledge of crop health and the associatedterminology.

INTERFACE

The interface is an often neglected part of expertsystems. It has been noted by Stock (pers. comm.)that, however sophisticated a system is, if it does nothave an easy to understand and attractive interface itwill not receive regular use. In a system such as ours,the interface is of paramount importance. The userswill often not be computer literate (unlike users ofapplications in the business and commercial world).Also much detailed information has to be entered,such as location of the orchard, types of appleaffected, results of mineral tests and likelihoods ofsymptoms being present. Loss of concentration bythe user can lead to unreliable answers. One of the

Dow

nloa

ded

by [

UO

V U

nive

rsity

of

Ovi

edo]

at 1

0:32

04

Nov

embe

r 20

14

96 New Zealand Journal of Crop and Horticultural Science, 1989, Vol. 17

main aims of this project is to make expert systemsmore stimulating to use and to make the informationtransfer as painless as possible, hence the interfaceis one part of the system that is receiving muchattention.

KNOWLEDGEREmNEMENT

Following the knowledge acquisition stage, theknowledge engineer and domain expert produce aframework for the system. The details of the questionphrasing, certainty factors, problem hierarchy etc.should then be determined by the expert during'knowledge refinement'. In particular, certaintyfactors, although simple in concept, need to beobserved in action for the domain expert to decideon appropriate values.

This process is facilitated by allowing him to runthe system in a mode in which the paths that theprogram is taking and the rules being used can bedisplayed. Additionally, the effect of each answerthe user may make can be shown on the screen innumerical form. During a session the expert canmodify numerical factors, re-word questions, adjustthe problem hierarchy and then re-run the programto observe the effects of the changes. Not only doesthis refinement by the expert ensure the system isbehaving as expected but it also gives him a betterappreciation of the methodology used.

SUMMARYA major goal of this project has been to develop asuitable design for a plant health expert system that

will not be a toy, but will be of assistance to bothgrowers and those who wish to know more aboutplant health. Our program stands midway betweenthose that use surface representations to modelempirical relationships, and those that contain deeprepresentations of causal processes.

The currentversionof thesystemcanbe comparedto a telephone consultation between a user and theexpert. In such a scenario, a definite diagnosis canoften be made but sometimes there will be uncer­tainty. Descriptions of symptoms and signs are of atextualnature,and muchdepends ontheobservationalprowess of the grower in order to arrive at a cleardiagnosis. It is planned in the future to incorporatepictures with the text and, in this way, help growersand students to better identify the pests and diseasesconcerned.

REFERENCES

Buchanan, B. G.; Shortliffe, E. H. 1985: Rule based expertsystems. Massachusetts, Addison-Wesley, 748p.

Clancey, W. 1. 1985: Heuristic classification. Artificialintelligence 27: 289-350.

Clocksin, W. F.; Mellish C. S. 1984: Programming inProlog, (Second edition). Berlin, Springer-Verlag,279p.

Latin, R. X.; Miles, G. E.; Rettinger, J. C. 1987: Expertsystems in plant pathology. Plant disease 71:866-872.

Sands, D. C.; Sharp, E. L.; Scharen, A. L.; Slater, L. S.1986:Anexpert system for predicting crop diseaseepidemics. Phytopathology 76: 1083 (abstractonly).

Dow

nloa

ded

by [

UO

V U

nive

rsity

of

Ovi

edo]

at 1

0:32

04

Nov

embe

r 20

14