您的当前位置：首页 English article

English article

来源：九壹网

ComputersandElectronicsinAgriculture71(2010)107–127

ContentslistsavailableatScienceDirect

ComputersandElectronicsinAgriculture

journalhomepage:www.elsevier.com/locate/compag

Review

Developmentofsoftcomputingandapplicationsinagriculturalandbiologicalengineering

YanboHuanga,∗,YubinLanb,StevenJ.Thomsona,AlexFangc,WesleyC.Hoffmannb,RonaldE.Laceyd

USDA-ARS,CPSRU,141ExperimentStationRoad,Stoneville,MS38776,UnitedStatesUSDA-ARS,APMRU,2771F&BRoad,CollegeStation,TX77845,UnitedStatesc

EngineeringTechnologyandIndustrialDistribution,TexasA&MUniversity,CollegeStation,TX77843,UnitedStatesd

BiologicalandAgriculturalEngineering,TexasA&MUniversity,CollegeStation,TX77843,UnitedStates

articleinfoabstract

Softcomputingisasetof“inexact”computingtechniques,whichareabletomodelandanalyzeverycom-plexproblems.Forthesecomplexproblems,moreconventionalmethodshavenotbeenabletoproducecost-effective,analytical,orcompletesolutions.Softcomputinghasbeenextensivelystudiedandappliedinthelastthreedecadesforscientiﬁcresearchandengineeringcomputing.Inagriculturalandbiologicalengineering,researchersandengineershavedevelopedmethodsoffuzzylogic,artiﬁcialneuralnetworks,geneticalgorithms,decisiontrees,andsupportvectormachinestostudysoilandwaterregimesrelatedtocropgrowth,analyzetheoperationoffoodprocessing,andsupportdecision-makinginprecisionfarming.Thispaperreviewsthedevelopmentofsoftcomputingtechniques.Withtheconceptsandmethods,appli-cationsofsoftcomputingintheﬁeldofagriculturalandbiologicalengineeringarepresented,especiallyinthesoilandwatercontextforcropmanagementanddecisionsupportinprecisionagriculture.Thefutureofdevelopmentandapplicationofsoftcomputinginagriculturalandbiologicalengineeringisdiscussed.

PublishedbyElsevierB.V.

Articlehistory:

Received1August2009

Receivedinrevisedform9January2010Accepted25January2010Keywords:

SoftcomputingFuzzylogic

ArtiﬁcialneuralnetworksGeneticalgorithmsCropmanagementPrecisionagriculture

Contents1.2.

Introduction.........................................................................................................................................Methodsofsoftcomputing.........................................................................................................................2.1.Fuzzylogic...................................................................................................................................2.2.Artiﬁcialneuralnetworks...................................................................................................................2.3.Geneticalgorithms..........................................................................................................................2.4.Bayesianinference...........................................................................................................................2.5.Decisiontree.................................................................................................................................Applicationsofsoftcomputinginagriculturalandbiologicalengineering........................................................................3.1.Fuzzylogicapplications.....................................................................................................................

3.1.1.Overview...........................................................................................................................3.1.2.Cropmanagement.................................................................................................................3.1.3.IrrigationandETcalculation.......................................................................................................3.1.4.Soilanalysis........................................................................................................................3.1.5.Precisionagriculture...............................................................................................................3.1.6.Chemicalapplication..............................................................................................................

3.2.Artiﬁcialneuralnetworkapplications.......................................................................................................

3.2.1.Overview...........................................................................................................................3.2.2.Cropmanagement.................................................................................................................3.2.3.IrrigationandETcalculation.......................................................................................................3.2.4.Soilanalysis........................................................................................................................3.2.5.Precisionagriculture...............................................................................................................3.2.6.Chemicalapplication..............................................................................................................

108

109109110110111111112112112112112112112113113113113115115116116

∗Correspondingauthor.Tel.:+16626865354;fax:+16626865372.E-mailaddress:yanbo.huang@ars.usda.gov(Y.Huang).0168-1699/$–seefrontmatter.PublishedbyElsevierB.V.doi:10.1016/j.compag.2010.01.001

108Y.Huangetal./ComputersandElectronicsinAgriculture71(2010)107–127

5.6.7.

Geneticalgorithmapplications..............................................................................................................3.3.1.Overview...........................................................................................................................3.3.2.Cropmanagement.................................................................................................................3.3.3.Irrigation...........................................................................................................................3.3.4.Soilanalysis........................................................................................................................3.3.5.Precisionagriculture...............................................................................................................3.3.6.Chemicalapplication..............................................................................................................

3.4.Bayesianinferenceapplications.............................................................................................................

3.4.1.Overview...........................................................................................................................3.4.2.Cropmanagement.................................................................................................................3.4.3.IrrigationandETcalculation.......................................................................................................3.4.4.Precisionagriculture...............................................................................................................3.4.5.Chemicalapplication..............................................................................................................

3.5.Decisiontreeapplications...................................................................................................................

3.5.1.Overview...........................................................................................................................3.5.2.Cropmanagement.................................................................................................................3.5.3.Precisionagriculture...............................................................................................................3.5.4.Chemicalapplication..............................................................................................................

Fusionofsoftcomputingmethodsinagriculturalandbiologicalengineering....................................................................4.1.Cropmanagement...........................................................................................................................4.2.ETcalculation................................................................................................................................4.3.Soilanalysis..................................................................................................................................4.4.Precisionagriculture.........................................................................................................................Supportvectormachines............................................................................................................................Comparisonandlimitationsofsoftcomputingtechniques........................................................................................Thefuture...........................................................................................................................................References...........................................................................................................................................

3.3.

116116116117117117117117117117117118118118118118118118118118118119119119121122123

1.Introduction

Softcomputingisasetofcomputingtechniques,suchasFuzzyLogic(FL),ArtiﬁcialNeuralNetworks(ANNs),andGeneticAlgorithms(GAs).Thesecomputingtechniques,unlikehardcom-puting,whichreferstoahugesetofconventionaltechniquessuchasstochasticandstatisticalmethods,offersomewhat“inexact”solutionsofverycomplexproblemsthroughmodelingandanal-ysiswithatoleranceofimprecision,uncertainty,partialtruth,andapproximation.Ineffect,softcomputingisanintegrationofbiologicalstructuresandcomputingtechniques.FLdevel-opesmulti-valued,non-numericlinguisticvariablesformodelinghumanreasoninginanimpreciseenvironment.ANNsprovidesconﬁgurationsmadeupofinterconnectingartiﬁcialneuronsthatmimicthepropertiesofbiologicalneurons.GAsareawayofsolv-ingproblemsbymimickingthesameprocessesnatureusesthroughselection,recombinationandmutation.

Softcomputingisusedtoachievetractability,robustness,andprovidealowcostsolutionwithatoleranceofimprecision,uncer-tainty,partialtruth,andapproximation.Thismakessoftcomputingcapableofsolvingproblemsthatmoreconventionalmethodshavenotyetbeenabletoprovideinacost-effective,analytical,orcom-pletemanner.Amongsoftcomputingtechniques,FLappearstobetheﬁrstonethathasestablishedfundamentalideasofsoftcomputing(Zadeh,1965,1973,1981).Theestablishedbasicideashaveinﬂuencedothertechniquesthatarrivedlater.In1986,theParallelDistributedProcessing(PDP)researchgrouppublishedaseriesofresultsandalgorithms(RumelhartandMcClelland,1986).ThisworkgaveastrongimpetustothestudyofmechanismsandstructureofthebrainandprovidedthecatalystformuchofthesubsequentresearchandapplicationofANNs.ThiscanbeviewedasthepointatwhichANNsbecameoneofthesoftcomputingtechniques.GAsweredevelopedbyJohnHollandin1975andpopularizedbyhisstudent,DavidGoldberg(Goldberg,19).Cur-rently,FL,ANNs,andGAsareconsideredascoretechniquesofsoftcomputing.Thecurrentlistofsoftcomputingtechniquesalsoincludesmachinelearning,probabilisticreasoning,andchaosthe-ory.Inthelastthreedecades,softcomputinghasbeenextensivelystudiedandappliedforscientiﬁcresearchandengineeringcom-puting.Althoughapplicationsofsoftcomputingtechniquesweresuccessfulinsolvingproblems,themethodologystillhasbeenadvancingtoprovidenewapproachesformoreefﬁcient,robust,andreliablesolutions.SupportVectorMachines(SVMs)(Burges,1998;CristianiniandTaylor,2000)emergedasasetofsupervisedgeneralizedlinearclassiﬁersandoftenprovidehigherclassiﬁcationaccuraciesthanmultilayerperceptronANNs.SVMshaveattractedgreaterinterestinrecentyears.Methodfusionisanotheradvance-mentinsoftcomputingdevelopment,whichcombinesorcascadesdifferentsoftcomputingtechniquestoimprovesystemperfor-manceoveranyindividualtechnique.Neuro-fuzzysystemsareatypicalexampleofsuchafusion(TakagiandHayashi,1991;Horikawaetal.,1992;NieandLinkens,1992;SimpsonandJahns,1993;MitraandPal,1994;JangandSun,1995).Thebehaviorandstabilityofhardcomputingaremorepredictable.Thecomputa-tionalburdenofalgorithmsistypicallyeitherlowormoderate.Thus,itisnaturaltoviewsoftcomputingandhardcomputingascomplementary.Thefusionofthemhasgreatpotentialfordevel-opinghigh-performance,cost-effective,andreliablecomputingschemesthatprovideinnovativesolutionstoproblems(Ovaskaetal.,2002).

Inagriculturalandbiologicalengineering,therehasbeensomeearlyresearchandapplicationsofsoftcomputing(Whittakeretal.,1991;ZhangandLitchﬁeld,1992;Eerikäinenetal.,1993)andinterestinsoftcomputinghasgrownsteadilyinthelastdecade.Asummaryofpapersandreportswerecollectedfromvar-ioussources,particularlythroughsearchesofthetechnicallibraryoftheASABE(AmericanSocietyofAgriculturalandBiologicalEngineers)(http://asae.frymulti.com/)andtheNationalAgricul-turalLibraryofUSDA(UnitedStatesDepartmentofAgriculture)(http://www.nal.usda.gov/).Itwasfoundthat,fromtheearly-1990sto2008,therehavebeen165reportsandpapers(65peerreviewed)onFL,348(193peerreviewed)onANNs,and83(36peerreviewed)onGAs.Itisinterestingtonotethat20reportsandpapers(13peerreviewed)werewrittenonSVMsfrom2003topresent,7(2peerreviewed)ofwhichwerepublishedin2008.Thismaysignify

Y.Huangetal./ComputersandElectronicsinAgriculture71(2010)107–127109

moreinterestinsoftcomputinginthenextdecadeinagriculturalandbiologicalengineering.Soilandwaterstudiesfocusedoncropmanagementanddecisionsupportsystemsforprecisionagricul-tureincludespecialtiessuchasthosehighlightedinGeoderma,aglobaljournalofsoilscience,SSSAJ(SoilScienceSocietyofAmer-icaJournal),JournalofSoil&WaterConservation,andJournalofPrecisionAgriculture.

Inagriculturalandbiologicalengineering,researchersandengineershavedevelopedmethodsforFL,ANNs,GAs,BayesianInference(BI),DecisionTree(DT),andSVMstostudysoilandwaterregimesrelatedtocropgrowth,analyzetheoperationoffoodpro-cessing,andsupportdecision-makinginprecisionfarming.TheyhavealsousedfusiontechniquesthatincludeFL,ANNandGAinsolvingproblems.However,wehavenotfoundapplicationswherefusionwasappliedtosoftcomputingandhardcomputing.Thiscouldbearesearchtopicwithgreatpotential.

Thispaperbrieﬂyreviewsthedevelopmentofsoftcomputingtechniques.Withtheconceptsandmethods,applicationsofsoftcomputingintheﬁeldofagriculturalandbiologicalengineeringarepresented.Thefutureofthedevelopmentandapplicationofsoftcomputinginagriculturalandbiologicalengineeringisdiscussed,especiallyinthesoilandwatercontextforcropmanagementandindecisionsupportinprecisionagriculture.2.Methodsofsoftcomputing

Ingeneral,softcomputingincludesthemethodsofFL,neuro-computing,evolutionarycomputing,probabilisticcomput-ing,beliefnetworks,chaoticsystems,andpartsoflearningtheory(http://www.cs.berkeley.edu/∼zadeh/acprco.html).Forresearchanddevelopmentinagriculturalandbiologicalengineering,pri-marymethodsofsigniﬁcantutilityincludeFL,ANNs,GAs,BIandDT.

ANNsaremodelarchitecturesandlearningalgorithmsofneuro-computing.GAsareaparticularclassofevolutionarycomputation.BIisamethodtorealizeprobabilisticcomputing.DTisoneoftheinterestingandcommonlyusedarchitecturesusedforlearning,reasoningandorganizationofdatasetsinsoftcomputing.FL,ANNs,GAs,BIandDThavebeenwidelyappliedforresearchanddevelop-mentinagriculturalandbiologicalengineering,especiallywithinthelastdecade.2.1.Fuzzylogic

FLisaformofmulti-valuedlogicderivedfromfuzzysetthe-orytodealwithreasoningthatisapproximate,ratherthanprecise.Incontrasttoyes/noor0/1binarylogic(crisp),FLprovidesasetofmembershipvaluesinclusivelybetween0and1toindicatethedegreeoftruth(fuzzy).Fig.1showsthedifferencebetweenacrispsetandafuzzysetbycomparingthecharacteristicfunctionsofthetwosets.Forthecrispset(Fig.1(a))thecharacteristicfunctionofAisassignedavalueof1or0toeachvalueinX,avaluesetofaphysicalproperty.1forAindicatesthatcorrespondingvaluesinXbelongtothesetA.0forAindicatesthatcorrespondingvaluesinXdonotbelongtothesetA.Theconceptofthecrispsetissufﬁ-cientformanyapplicationsbutisnotforsomeapplicationsthatrequireﬂexibility.InthefuzzysetthecharacteristicfunctionofAisassignedavaluebetween0and1,including0and1,toeachvalueinX.1or0forAstillindicatethatcorrespondingvaluesinXbelongordonotbelongtothesetA.Thevaluesbetween0and1forAindicatethatthecorrespondingvaluesinXbelongtothesetAinacertaindegreefromlow,medium,tohighwiththeincreaseoftheAvalue.Fig.1(b)showsatypicalcharacteristicfunctionofthefuzzyset.Withthecharacteristicfunctionthemembershipfunctionforafuzzysetcanbeconstructedtoquantifythemagnitudeofbelong-

Fig.1.Characteristicfunctionsofcrispandfuzzysets(b>d>a>c>0).Characteristicfunctionofacrispset(a)andfuzzyset(b).

ingofeachinput.Themembershipfunctionsmaybedifferentinshapeandinterval.ThelogicoperationssuchasANDandORcanimplementedwiththedifferentmembershipfunctionstogeneratearesultantmembershipfunction.

Inprocessmodelingandcontrol,systemsthatareill-deﬁnedandwithuncertaintiescanbemodeledwithafuzzyinferencesys-tememployingfuzzy‘If–Then’rulestoquantifyhumanknowledgeandreasoningprocesseswithoutemployingprecisequantitativeanalyses.Thefuzzyinferencesystemshouldincludethefollowingfunctionalblocks(Jang,1993):

•afuzziﬁcationinterfacethattransformsthecrispinputsintodegreesofmatchwithlinguisticvalues;•aknowledgebasethatincludes

◦arulebasecontaininganumberoffuzzy‘If–Then’rules;

◦adatabasethatdeﬁnesthemembershipfunctionsofthefuzzysetsusedinthefuzzyrules;

•adecision-makingunitthatperformstheinferenceoperationsontherules;and

•adefuzziﬁcationinterfacethattransformsthefuzzyresultsoftheinferenceintoacrispoutput.

Althoughsomeprecedingworkhasbeenindicated(Wilkinson,1963),Dr.LotﬁZadehhasbeengenerallyconsideredastheﬁrstper-sonwhointroducedFLtotheworld.In1965,Dr.Zadehaxiomatizedfuzzysettheory(Zadeh,1965).FLisaninterfacebetweenlogicandhumanreasoning.Humanshavearemarkablecapabilitytoreasonandmakedecisionsinanenvironmentofuncertainty,imprecision,incompletenessofinformation,andpartialityofknowledge,truthandclassmembership.FLisforformalization/mechanizationofthiscapability.FLmimicshumancontrollogic.Itcanbebuiltintoanythingfromsmall,hand-heldproductstolargecomputerizedprocesscontrolsystems.Itusesanimprecisebutverydescriptivelanguagetodealwithinputdataasahumanoperatorwould.Itisveryrobustandtolerantofoperatoranddatainputandoftenworkswhenﬁrstimplementedwithlittleornotuning.

FLhasbeenappliedtodiverseﬁeldssincethe1970swiththesufﬁcientdevelopmentofsmallcomputingcapability.Ithasbeenappliedinsuchvariedﬁeldsascontrolsystemsandartiﬁcialintel-ligence.In1974,theﬁrstsuccessfulapplicationoffuzzylogictothecontrolofalaboratory-scaleprocesswasreported(MamdaniandAssilian,1975).Controlofcementkilnswasanearlyindus-trialapplication(HolmbladandOstergaard,1982).Sincetheﬁrst

110Y.Huangetal./ComputersandElectronicsinAgriculture71(2010)107–127

consumerproductusingfuzzylogicwasmarketedin1987,theuseoffuzzycontrolhasincreasedsubstantially.AnumberofCAD(Computer-AidDesign)environmentsforfuzzycontroldesignhaveemergedtogetherwithVLSI(Very-Large-ScaleIntegration)hard-wareforfastexecution.Fuzzycontrolisbeingappliedtovarioussystemsintheprocessindustry(SanthanamandLangari,1994;Tanietal.,1994),consumerelectronics(Hirota,1993;Bonissone,1994),automatictrainoperation(YasunobuandMiyamoto,1985),trafﬁcsystemsingeneral(Hellendoorn,1993),andinotherﬁelds(Hirota,1993;Teranoetal.,1994).

AlthoughFLwasproposedbyProfessorZadehofUniversityofCaliforniaatBerkeley,FLhasbeengainingpopularityonlygradu-allyintheUnitedStates.Thismaybeduetodisappointmentwiththeunfulﬁlledpromisesof‘artiﬁcialintelligence’computingtech-niquesinthe1980s(KhoshnevisandChignell,1985).However,atthesametime,FLproductshavebeenaggressivelybuiltinEuropeandJapan(EET,1991;Smith,1993).Still,theUnitedStatesEPA(EnvironmentalProtectionAgency)hasinvestigatedfuzzycontrolforenergy-efﬁcientmotors(Clelandetal.,1992).NASA(NationalAeronauticsandSpaceAdministration)hasstudiedfuzzycontrolforautomatedspacedocking(OrtegaandGiron-Sierra,1995).Boe-ing,GeneralMotors,Allen-Bradley,Chrysler,Eaton,Whirlpool,andMaytaghavealsoworkedonfuzzylogicforuseinlow-powerrefrigerators,automateddishwashers,improvedautomotivetrans-missions,andenergy-efﬁcientelectricmotors.

Developmentoffuzzysystemshashelpedadvancetechniquesforhandlingimprecisioninsoftcomputing.In1992,theconceptofsoftcomputingwasintroduced(Zadeh,1992).Dr.Zadehenvisionedsoftcomputingasbeingconcernedwithmodesofcomputinginwhichprecisionistradedfortractability,robustnessandeaseofimplementation.Softcomputingservestohighlighttheemergenceofcomputingmethodologiesinwhichtheemphasisisonexploitingthetoleranceforimprecisionanduncertaintytoachievetractabil-ity,robustness,andlowsolutioncost.2.2.Artiﬁcialneuralnetworks

ANNsprovideawaytoemulatebiologicalneuronstosolvecomplexproblemsinthesamemannerasthehumanbrain.Formanyyears,especiallysincethemiddleofthelastcentury,inter-estinstudyingthemechanismandstructureofthebrainhasbeenincreasing.Thisincreasingresearchinteresthasledtothedevel-opmentofnewcomputationalmodels,connectionistsystemsorANNs,basedonthebiologicalbackgroundforsolvingcomplexproblemslikepatternrecognitionandfastinformationprocess-ingandadaptation.Intheearly1940s,McCullochandPitts(1943),pioneersintheﬁeld,studiedthepotentialandcapabilitiesofinter-connectingseveralbasiccomponentsbasedonthemodelofaneuron.Lateron,otherslikeHebb(1949),wereconcernedwiththeadaptationlawsinvolvedinneuralsystems.Rosenblatt(1958)coinedthename“Perceptron”anddevisedanarchitecture,whichhassubsequentlyreceivedmuchattention.MinskyandPapert(1969)laterintroducedarigorousanalysisofthePerceptron,ofwhichtheyprovedmanypropertiesandpointedoutlimitationsofseveralrelatedmodels.Inthe1970s,theworkofGrossberg(1976)cametoprominence.Hiswork,basedonbiologicalandpsycho-logicalevidence,proposedseveralnovelarchitecturesofnonlineardynamicsystems.Inthe1980s,Hopﬁeld(1982)appliedaparticu-larnonlineardynamicstructuretosolveproblemsinoptimization.AllofthemconductedpioneerstudiesonthetheoreticalaspectofANNs.

In1986,thePDP(ParallelDistributedProcessing)researchgrouppublishedaseriesofalgorithmsandresults(RumelhartandMcClelland,1986).Thispublicationcontainsanarticleenti-tled“LearningInternalRepresentationsbyErrorPropagation”(Rumelhartetal.,1986a).Thisarticlebroughtrecognitiontoan

Fig.2.ANNnetworkstructure.

ANNtrainingalgorithmnamedBackPropagation(BP),althoughithadalreadybeendescribedinWerbos(1974).ThisBPtrainingalgo-rithmimplementedwiththegeneraldeltarule(Rumelhartetal.,1986a,1986b)gaveastrongimpulsetosubsequentresearchandresultedinthelargestbodyofresearchandapplicationsinANNsalthoughmanyotherANNarchitecturesandtrainingalgorithmshavebeendevelopedandappliedsimultaneously.

Thearchitecture,training/learning,andimplementationofanANNareasimpliﬁedversionofthestructureandactivitiesofhumanbrain.Forproblemsolving,thehumanbrainusesawebofinterconnectedprocessingunits,neurons,toprocessinforma-tion.Eachoftheneuronsisautonomous,independent,andworksasynchronously.Thevastprocessingpowerinherentinbiologi-calneuralstructureshasinspiredthestudyofthestructureitselfasamodelfororganizinganddesigningman-madecomputingstructures.Comparedwithconventionaldataprocessingmeth-ods,ANNsprovideamodel-free,adaptive,parallel-processing,androbustsolutionwithfaultandfailuretolerance,learning,abilitytohandleimpreciseandfuzzyinformation,andcapabilitytogen-eralize.AnANNisabletomapprocessinputandoutputwithoutunderlyingassumptionaboutthedistributionofdata.ANNsarepowerfulindataprocessingandanalysis.Inmathematics,ANNsarethemodelsdeﬁningafunction:f(x)|X→Y.EachANNmodeldeﬁnesaclassofsuchfunctions.InANNs,thefunctionf(x)isdeﬁnedbyacombinationoffunctionsgi(x),whichcouldbefurtherdeﬁnedbyacombinationofotherfunctionsri(x),andsoon.ThisrecursivefunctiondeﬁnitioninanANNbringsupanetworkstructurefortheinterconnectionoffunctionunits.Thewidelyusedbinationisthenonlinearweightedsum:f(x)=Ac󰀃function󰀂ncom-󰀁whereAisapredeﬁnedactivationfunction,suchasi=wg(x),the1ii

c()hyper-bolictangentfunction(Werbos,1974;Rumelhartetal.,1986a,1986b).Fig.2showsatwohidden-layerfeedforwardnetworkbetweeninputxandoutputf.

Thefeedforwardnetworkwithasinglehiddenlayerthatcon-tainsaﬁnitenumberofhiddenneuronsaccompaniedwithanarbitraryactivationfunctionwasproventobeauniversalapproxi-matoronacompactsubsetofrealn-dimensionalEuclideanspaceRn(Horniketal.,19;Cybenko,19).ThistheoryassuresthatANNscanhandleengineeringproblems,whicharehighlycomplexandnonlinear.ANNsprovideapowerfulmethodforpracticallyaccuratesolutionsofpreciselyorimpreciselyformulatedproblemsandforphenomenathatareonlyunderstoodthroughexperimentaldataandﬁeldobservations.ANNshavebecomethemostpopularsoftcomputingmethodsforsolvingproblemsinengineering.2.3.Geneticalgorithms

EvolutionarycomputingisanArtiﬁcialIntelligence(AI)techniquetosolvecombinatorialoptimizationproblems.Itis

Y.Huangetal./ComputersandElectronicsinAgriculture71(2010)107–127111

Fig.3.FlowchartofGAoperation.

implementediniterativeprogresswiththegrowthinapopula-tionselectedinaguidedrandomsearchusingparallelprocessingtoreachthedesiredpoint.Thisoperationwasinspiredbybiologi-calmechanismsofevolution,whichissimilartoDarwin’stheoryonevolution(Darwin,1859).Thereareanumberofpeoplewholaidthetheoreticalfoundationsofevolutionalcomputing:LawrenceJ.FogelandJohnHenryHollandintheUnitedStates(Fogeletal.,1966;Holland,1975),andIngoRechenbergandHans-PaulSchwefelinGermany(Rechenberg,1973;Schwefel,1981).Fogelisregardedasthefatherofevolutionarycomputing.Hollandstudiedthismethodalittledifferentlyandnameditageneticalgorithm.Then,oneofHolland’sstudents,DavidGoldberg,solvedadifﬁ-cultprobleminvolvingthecontrolofgas-pipelinetransmissionusingthemethodofGAinhisdissertation(Goldberg,19).Gold-berg’sworkgreatlyinspiredsubsequentresearchandapplicationsofGA.

GAisanoptimizationandheuristicsearchtechniquethatusestechniquesinspiredbyevolutionarybiologysuchasinheritance,mutation,selection,andcrossover(alsocalledrecombination).GAworkssimultaneouslyonaset(population)ofpotentialsolutions(individuals)totheproblem.Thealgorithmstartswithasetofsolutions(representingchromosomes)calledasub-population.Theﬁtnesstowhichsolutionsmeetsomeperformancecriterionisevaluatedandusedtoselect“surviving”individualsthatwill“reproduce”anew,bettersub-population.Then,theindividualswillconductalterationssimilartothenaturalgeneticmutationandcrossover.Theselectionschememakestheprocesstowardshighperformancesolutions.Acarefulselectionofgeneticalgorithmstructureandparameterscanensureagoodchanceofreachingthegloballyoptimalsolutionafterareasonablenumberofiter-ations.Fig.3showsaﬂowchartofoperationalprocessesofGAs.GAsarecomputationallysimpleyetpowerfulenoughtoprovidearobustsearchfordifﬁcultcombinationalsearchproblemsincom-plexspaces,withoutbeingstuckinlocalextremes(Goldberg,19).Therefore,GAsarepowerfulalternativetoolstotraditionalopti-mizationmethods.GAshavebeensuccessfullyusedinmanyﬁelds,suchasscheduling(Wall,1996;LimandSim,2005),functionopti-mization(Houcketal.,1998),machinelearning(GoldbergandHolland,1988;Grefenstette,1994;Shapiro,1998)andhavebecomeanimportantmethodofsoftcomputing.

2.4.Bayesianinference

Probabilisticcomputingisasoftcomputingtechniquetoper-formprobabilisticreasoning.Theaimofprobabilisticreasoningistocombinethecapacityofprobabilitytheorytohandleuncertaintytomakeinferencewithbelief.Pearl(1988)madeanimportantsur-veyonthistopicwithanemphasisonBayesiannetworks.Grunwald(1997)madeausefulreviewalso.

Bayesianapproacheshavebeenstudiedandappliedexten-sively(Eddy,1982;Edwards,1982;GigerenzerandHoffrage,1995;Jaynes,1996;Lauritzen,2003)byusingBayes’stheorem(Bayes,1763)intheprocessofprobabilisticcomputation.Classicalinfer-encemodelsdonotpermittheintroductionofpriorknowledgeintothecalculations.However,theuseofpriorknowledgewouldsome-timesbeusefultotheprocessevaluation.BI(BayesianInference)isastatisticalinferenceincorporatingpriorknowledgeandpriorprobabilitydistributions.IntheprocessofBIevidenceorobser-vationsareusedtoupdatetheprobabilitythatahypothesismaybetrue.Conventionally,binaryhypothesistestingisusedtosta-tisticallydecidewhichhypothesisistruebetweentwohypotheses.Alternatively,aBayesiandecisionrule,suchasminimumBayesrisk,minimumprobabilityoferror,ormaximumaposterioriprobability(MAP),dependsonthepriorprobabilityofeachhypothesis(Dudaetal.,2001).

BayesianinferenceisbasedontheBayes’theorem(Bayes,1763)thatadjustsprobabilitiesgivennewevidence:P(H|E)=P(E|H)P(H)/P(E)whereisthespeciﬁchypothesis,P(H)isthepriorprobabilityofH,P(E|H)istheconditionalprobabilityofEwhenHhappenstobetrue,andP(E)isthemarginalprobabilityofE(newevidence).TheratioP(E|H)/P(E)representstheimpacttheevidencehasonthebeliefinthehypothesis.Whenthisratioislarge,alargerposteriorprobabilityofthehypothesisgiventheevidencewillbeproducedbymultiplyingthepriorprobabilityofthehypothesisbythisfactor.Whenthisratioissmall,asmallerposteriorprobabil-ityforHwillbeproduced.InBayesianinference,Bayes’theoremisusedtomeasurehowmuchnewevidenceshouldalterbeliefinahypothesis.

BayesianNetworkshavebeenacceptedastoolsfordecision-makingincomplexsituationswithinavarietyofdisciplines(Pearl,1999;Charniak,1991).BayesianNetworksareprobabilisticgraph-icalmodels.Eachofthemodelscharacterizesasetofvariablesandtheirprobabilisticindependencies.Thegraphical,probabilisticmodelsallowthestructuredrepresentationofacognitivepro-cessbasedonalinkandnodestructurewherethestateofparentnodepredictsthestateofthechildnode.ConditionalprobabilitytablesarethenusedtorelatetheparentandchildnodesusingBayesianstatisticalmethodstopredicttheirrelationship.DynamicBayesiannetworks(Ghahramani,1997)aretheBayesiannetworksthatrepresentsequencesofvariables.Thesesequencesareoftenvaluesoftimeseriesorsequencesofsymbols(forexampleproteinsequences).2.5.Decisiontree

DT(DecisionTree)isapopularmethodofmachinelearning.Itwasdevelopedinthe1960s(Magee,19)andcomprisesatree-structuredarrangementofasetofattributestoevaluateandpredicttheoutput.IntheoperationofaDT,thealgorithmisrecursivelylookingfortheattributewiththehighestinformationgain,whichisdeterminedtoevaluateﬁrst.DTscanbeusedtoidentifythestrategymostlikelytoreachagoal.Theycanalsobeusedasadescriptivemeansforcalculatingconditionalprobabilities.Inoper-ationsresearch,aDTisadecision(classiﬁcation)model(HillierandLieberman,2005).ADTcanbeusedtovisuallyandexplicitlyrep-resentdecisionsfordecisionsupport.Indataminingandmachinelearning,aDTisapredictivemodel(Teorey,1999;Flamig,2000;

112Y.Huangetal./ComputersandElectronicsinAgriculture71(2010)107–127

WittenandFrank,2000).TheDTdescribesdatabutnotdecisionsfortheresultingclassiﬁcationtreeasaninputfordecision-making.Withagraphicalrepresentation,aDTmodelmapsfromobserva-tionsaboutanitemtotheconclusionsaboutitstargetvalue.Inthestructureofthetree,leavesrepresentclassiﬁcationsandbranchesrepresentconjunctionsoffeaturesthatleadtothoseclassiﬁcations.Themachinelearningtechniqueforinducingadecisiontreefromdataiscalleddecisiontreelearning,or(colloquially)decisiontrees.UsingDTslargeamountsofdatacanbeanalyzedinarelativelyshorttimeforreal-timeapplications(Abrahametal.,2007).3.Applicationsofsoftcomputinginagriculturalandbiologicalengineering

Withthetheoreticaldevelopmentsofsoftcomputing,alargevarietyofsuccessfulapplicationstomanyindustrialsystemshavebeencreated.Duringthelastdecade,interestofapplyingsoftcomputingtechniquestosystemsinagriculturalandbiologicalengineeringhasbeengrowinggreatly.Asinotherﬁelds,softcom-putingplaysanespeciallyimportantroleinprovidingtechniquestointegratehuman-likevaguenessandreal-lifeuncertaintyintoconventionalcomputingprograms.Problemsinsoilandwater,cropmanagementandpostharvesting,precisionagriculture,foodprocessing,foodqualityandsafety,andagriculturalvehicleandroboticshavebeensolvedthroughsoftcomputing-basedclassi-ﬁcation,modelingandprediction,andoptimizationandcontrol.Thispaperreviewssigniﬁcantpapersforeachoftheseagriculture-relatedareas,organizedbythesoftcomputingtechniqueutilized.3.1.Fuzzylogicapplications

3.1.1.Overview

Insummaryof136relatedpapersandreports,FLhasbeenappliedinsolvingproblemsincropmanagement(17%),soilandwater(16%),foodqualityandsafety(14%),animalhealthandbehavior(10%),agriculturalvehiclecontrol(8%),precisionagricul-ture(7%),greenhousecontrol(7%),agriculturalmachinery(4%),foodprocessing(4%),airqualityandpollution(3%),agriculturalfacilities(2%),agriculturalrobotics(1%),chemicalapplication(1%),andothers(6%)suchasnaturalresourcesmanagementandagricul-turalproductdesign.TheseapplicationshavebeencreatedthroughFLmainlybycontrol(28%),modelingandprediction(24%),clas-siﬁcation(24%),fuzzyclustering(9%),rule-basedinference(7%),multisensordatafusion(4%),optimization(1%)andothers(3%)suchasthresholdingandpatterninference.

3.1.2.Cropmanagement

Forcroppestmanagement,Pydipatietal.(2005)usedthecolorco-occurrencemethodfortexturalanalysistodeterminewhetherclassiﬁcationalgorithmscouldbeusedtoidentifydiseasedandnormalcitrusleaves.Oneoftheclassiﬁcationstrategiesinvesti-gatedwasANNclassiﬁersbasedontheRBF(RadialBasisFunction)networkswithfuzzyoutputsthatindicateameasureofstrength.Theleveloffuzzinesswasdeterminedbysettinganyvalue<0.5asequivalentto0,andanyvalue>0.5asequivalentto1.ThisisatypicalwayofdealingwiththeoutputofclassiﬁerusingFL,whichhelpsdecidetheoutputclasslabelingwiththeoutputofdecimalnumbers.

Yangetal.(2000a,2003)reportedondevelopmentofanimagecapture/processingsystemtodetectweedsandafuzzylogicdecision-makingsystemtodeterminewhereandhowmuchher-bicidetoapplyinanagriculturalﬁeld.Asinformationconcerningeconomicthresholdsofweedimpactoncropproductivitycannoteasilybeadaptedtoagivenregionoreventoagivenfarm,afuzzylogicapproachwasappliedtoconvertimagedataintosprayercom-mandstoallowfarmerstouseexperiencetoclassifyweedstatus

atagivenlocationintheﬁeld.Thisresearchindicatedthatafuzzylogicsystemisabletounderstandandfacilitatestherepresentationandprocessingofhumanknowledgeincomputerandtheinputs,outputs,andrulesofFLareeasytomodify.

3.1.3.IrrigationandETcalculation

Forirrigationscheduling,Odhiamboetal.(2001a)examinedFLforestimatingdailyreferenceevapotranspiration(ET)withfewerparameterscomparedwiththestandardFAO(FoodandAgricultureOrganization)Penman–Monteithmethod,whichissophisticatedandrequiresseveralinputparameterssomeofwhichhavenoactualmeasurementsbutareestimatedfrommeasuredweatherparameters.Inthisstudy,twofuzzyETmodelsweredevelopedtoestimatereferencegrassET,usingtwoandthreeweatherparam-eters,respectively.TheresultsillustratedthatthefuzzyETmodelscouldyieldaccurateestimationofET.

Al-Farajetal.(2001)developedarule-basedFLcropwaterstressindex(CWSI)usinggrowthchamberdataandtestedthismethodontallfescuecanopiesgrowninagreenhouse.TheupperandlowertheoreticalCWSIbaselinesmayshiftaccordingtonetradiation,wind,andstomatalresistance.ThisstudyproposedfuzzylogictodevelopaFL-CWSIsystemtoovercometheuncertaintyofbaseline.TheFL-CWSIsystemhasjustthreeinputs:canopy-airtemperaturedifferential,vaporpressuredeﬁcit,andshortwaveradiation,andoneoutput:CWSI.Tocalculatethecropwaterstressindexunderdifferentlevelsofsolarradiationandvaporpressuredeﬁcit,150fuzzyruleswereestablishedtorelatethesysteminputsandtheoutput.TheuseofFL-CWSIeliminatedtheneedtocalculateorcal-ibrateempiricalortheoreticalbaselinelimits.TheFL-CWSIwassuccessfullytestedusingindependentgreenhousedata.

ThomsonandRossdevelopedacoupledsensor-andmodel-basedirrigationschedulingmethod(Thomsonetal.,1993;ThomsonandRoss,1996).BasedonthisworkThomson(1998)describedaconceptbywhichreadingsfromsensorscouldbeusedtoreparameterizeoradjustinputstoacropsimulationmodel.Readingsfromsensorsandthemodeledresultswereusedsyner-gisticallyinanadaptivelearningscheme.

3.1.4.Soilanalysis

SoilmappingusingFLhasbeenreportedbyZhuetal.(2001),Bragato(2004),Shietal.(2004),Qietal.(2006),andZhuetal.(inpress).Lark(2000)illustratedhowtodesignsamplinggridsfromimpreciseinformationonsoilvariabilitywithanapproachbasedonthefuzzykrigingvariance.Thisapproachcanderiveafuzzysetofgridspacingsthatwillachieveatargetkrigingvariancewhengivenafuzzyformulationofthesoilpropertyvariogram.Thisfuzzysetofgridspacingscanthenbedefuzziﬁedinamoreorlessconservativewaytodeﬁneasamplingscheme.Insupportofprecisionagri-culture,VanAlphenandStoorvogel(2000)developedafunctionalapproachtosoilcharacterizationinvolvedinwaterstress,nitrate-stress,nitrate-leachingandresidualnitrogen-contentatharvestforaprecisionagriculturedecisionsupportsystem.Inthestudy,afuzzyc-meansclassiﬁerwasusedtogroupthesoilproﬁlesintofunctionalclasses.Fergusonetal.(2003)evaluatedapproachesforsite-speciﬁcuseofthenitriﬁcationinhibitorbasedonslopeandsur-facetextureformanagementzonedeﬁnition.Inthisstudy,fuzzyclusteranalysiswasdemonstratedthepotentialtodeﬁnemanage-mentzonesforuseofnitriﬁcationinhibitorsfromeasilyobtainedspatialyieldorsoilECa,ratherthanexpensivegridsamplingofsoilchemicalandphysicalproperties.

3.1.5.Precisionagriculture

Ortizetal.(2008)usedfuzzyclusteringofelevationandslopeoftheterraintodelineaterootknotnematode(RKN)riskzonesforacomparisontestoftwonematicideapplicationratesonnema-todepopulationdensityandcottonlintyield.Theresultsfromthis

Y.Huangetal./ComputersandElectronicsinAgriculture71(2010)107–127113

studyclearlyshowedthatRKNcontrolandﬁnalyieldvariedwithrespecttothenematicidetypeandrateacrossriskmanagementzonesbasedonfuzzyclusteringofterrainelevationandslope,nor-malizeddifferencevegetationindex(NDVI)ofbaresoilreﬂectance,andapparentsoilelectricalconductivity.

XiangandTian(2007)usedfuzzyinferenceinanANNframeworktodevelopanartiﬁciallyintelligentcontrollerthatauto-maticallyadjustsmultispectralcameraparameters,suchas“gain”and“exposuretime”tocompensateforchangingnaturallight-ingconditionsandtoacquirewhite-balancedimages.Theartiﬁcialintelligencecontrolalgorithmdidrequirenomathematicalmodelofthesystem,andprovidedbettercontrolperformancecomparedtoconventionalcontrolmethods.

JonesandBarnes(2000)describedprecisioncropmanagementasamulti-objectivedecision-makingprocessthatmustincorpo-rateadiversityofdata,opinion,preferenceandobjective.Thispaperdevelopedfuzzycompositeprogramming,adistance-basedmulti-objectiveoptimizationproblemthatusesfuzzyrepresenta-tionofuncertaintytocombineremotesensingandcropmodelsfordecisionsupportinprecisioncropmanagement.Thisapproachallowsuserstoexpressindividualorcorporatevaluesandprefer-ences;highlightsthedegreeofimprecisionassociatedwitheachinformationsourceused(i.e.modelaccuracy,uncertaintyincostsandreturns,etc.);highlightsthedegreeofimprecisionassociatedwitheachalternative;facilitatesstructuringofthedecisionpro-cess;reducesseverallevelsofcomplexinformationintoasinglechart;allowsexaminationoftrade-offbetweenalternativesandinterests;andforcesexaminationofinter-relationshipsbetweeninterests.

Ambueletal.(1994)successfullydevelopedandusedafuzzylogicyieldsimulator.TwoexpertsystemmodelsweredevelopedusingFLrules.Inonemodelchemicalandphysicalcharacteristicsofthesoilweremeasuredandcombinedwithlocalmeteorologicaldataasinputparameters.Intheothermodelsoilpropertieswereestimatedratherthanmeasured.Modelpredictedyieldswerethencomparedwithmeasuredyieldsforthoseﬁelds.Theresultsindi-catethatonarelativebasis,predictedyieldsgenerallyagreedwithmeasuredyields.

3.1.6.Chemicalapplication

Giletal.(2008)usedlinearmultipleregressionandFLinferencemodelstoevaluatetheeffectsofmicrometeorologicalconditionsonpesticideapplicationfortwosprayqualities(ﬁneandveryﬁne).Spraylosseswerepredictedusingfuzzyinferencesystems.Interpretableruleswereestablishedforthecharacterizationofmicrometeorologicalparametersusingthetwosprays.Q.Chenetal.(2006)andY.Chenetal.(2006)developedanintelligentpes-ticidespraysystembasedonfuzzycontrolandmachinevisiontoperformsite-speciﬁcpesticideapplication.ChoandKi(1999)developedaFLcontrollerforautonomousoperationofaspeedsprayerinanorchard.Theoperationofthecontrollerwasgraph-icallysimulatedundertherealconditionoftheorchard.Machinevisionwasalsousedtodeterminevehicleheadingandfourultra-sonicsensorswereusedtodetectobstaclesduringtheoperation.3.2.Artiﬁcialneuralnetworkapplications

3.2.1.Overview

ANNshavethelargestbodyofapplicationsinagriculturalandbiologicalengineeringwhencomparedwithothersoftcomput-ingtechniques.Insummaryofrelated348papersandreports,ANNshavebeenappliedinsolvingproblemsinfoodqualityandsafety(35.34%),crop(22.7%),soilandwater(14.37%),preci-sionagriculture(6.61%),animalmanagement(5.17%),postharvest(2.59%),foodprocessing(2.3%),greenhousecontrol(2.01%),agri-culturalvehiclecontrol(1.15%),agriculturalmachinery(1.15%),

agriculturalpollution(1.15%),agriculturalbiology(1.15%),ecol-ogyandnaturalresources(1.44%),agriculturalrobotics(0.29%),chemicalapplication(0.29%),andothers(2.3%)suchasbioen-ergyandagriculturalfacilities.TheseANNapplicationshavebeencreatedmainlythroughclassiﬁcation(45.11%),modelingandpre-diction(43.97%),control(4.02%),andsimulation(2.59%),parameterestimation(2.01%),detection(1.15%),dataclustering(0.57%),opti-mization(0.29%)anddatafusion(0.29%)aswell.

Similartoapplicationsingeneralengineeringﬁelds,mostworksinagriculturalandbiologicalengineeringhavebeenaccomplishedusingamultilayerfeedforwardANNtrainedbythefamousBPalgorithm,whichwasinspiredbytheworkofRumelhartetal.(1986a,1986b).Among348collectedpapersandreports210werebasedonmultilayerfeedforwardneuralnetworkstrainedbytheBPalgorithmand83didnotexplicitlystateANNs’structureandtrainingalgorithm.OftheremaindertwelvewerebasedonPNN(ProbabilisticNeuralNetwork);elevenwerebasedonMLP(MultiLayerPerceptron)usingdifferenttrainingalgo-rithmssuchasLevenberg–Marquardtoptimizationprocedure,Broyden–Fletcher–Goldfarb–Shanno(BFGS)optimizationproce-dure,andGA;sevenwerebasedonKohonenSOM(SelfOrganizingMap),whichisapopularunsupervisedtrainingalgorithm;sixwerebaseduponotherunsupervisedtrainingalgorithmssuchasfuzzyART(Carpenteretal.,1991),ART2(CarpenterandGrossberg,1987),andAuto-Associativenetwork(Andersonetal.,1977);fourwerebasedonRBF,whichisaneuralnetworkgoodforfunctionapprox-imation;threewerebasedonLVQ(LearningVectorQuantization);twowerebasedonGRNN(GeneralizedRegressionNeuralNet-work).FinallytenwerebasedonothernetworkstructuressuchasCounter-Propagation(CP),andAdaptiveLogicNetwork(ALN).

3.2.2.Cropmanagement

Inacropﬁeldstudy,Meyeretal.(2004)conductedadigitalcameraoperationstudyforclassifyinguniformimagesofgrass,baresoil,cornstalksresidue,andwheatstrawresidueusingabariumsulfatereferencepanelbasedoncolor.Theclassiﬁcationswereconductedwithacombinationofleast-squaresestimationandBPtrainingtogenerateafuzzyinferencesystem.FuzzyinputvariablesincludedRGBorHSIaveragevaluesalone,withandwith-outmonochromaticlightsources,andwithandwithoutluminanceandbackgroundlightsourcecolortemperaturemeasurementsforeachfuzzyinferencemodel.TheANNsystemprovidedatoolforevaluatingdigitalcameraoperatingperformancebysettingupasimpleclass,supervisedlearningsystemforplant,soilresiduecolorimages.

Inordertodevelopamulti-spectralopticalsystemforremotesensingofnitrogencontentofcrops,Chenetal.(2007)determinedsigniﬁcantwavelengthsinimagedataforestimatingnitrogencon-tentofcabbageseedlingleavesbystepwisemulti-linearregressionanalysis.AfeedforwardANNmodelwithcross-learningschemewasfurtherdevelopedtoincreasepredictionaccuracy.TheresultsindicatedthattheANNmodelwithcross-learningusingspectralinformationat490,570,600,and680nmcouldbeusedtodevelopapracticalremotesensingsystemtopredictnitrogencontentofcabbageseedlings.

SuiandThomasson(2006)developedaBPtrainedfeedforwardANNtopredictnitrogenstatusincottonplantsbasedondatafromaground-basedsensingsystem.Thesystemconsistsofamulti-spectralopticalsensorforplantcanopysensing,anultra-sonicsensorforplantheightmeasurement,andadata-acquisitionandprocessingunit.Fieldtestsofthesystemover2yearsinvolvedmeasuringspectralreﬂectanceandplantheightsimultaneouslyinrealtimeinsitu.Resultsshowedthattheneuralnetworkswereabletopredictnitrogenstatusofthecottonplantsat90%accuracywithtwocategories:deﬁciencyandnon-deﬁciency.

114Y.Huangetal./ComputersandElectronicsinAgriculture71(2010)107–127

Karimietal.(2005)evaluateddiscriminantanalysisasatoolforclassifyingimageswithrespecttothenitrogenandweedmanage-mentpracticesappliedtotheexperimentalplots,andcomparedtheclassiﬁcationaccuracyofthistechniquewiththoseobtainedbyANNandDTalgorithmsonthesamedata.Foridentifyingweedandnitrogenstressesincorn,thediscriminantanalysiswasfoundtoprovidethebestclassiﬁcationaccuracyattheearlygrowthstage,whereasbetteraccuracywasobtainedwithANNmodelsatthetasselingandfullmaturitystages.

Tumboetal.(2002b)usedanon-the-gosystemforsensingchlorophyllstatusincornusingBPtrainedfeedforwardANNsandﬁber-opticspectrometrytoacquirespectralresponsepatterndataincornﬁelds.Theneuralnetworkmodelincorporatedintothemobilesystemwastrainedusingstaticallycollectedplant-centerspectraldataandchlorophyllreadingsacquiredbySPAD502chlorophyllmeteronthesomedayandinthesameﬁeldplots.Theneuralnetworkmodelshowedgoodcorrelationbetweenpredictedandactualchlorophyllreadingsofthecalibrationdataset.

Goeletal.(2001)comparedtheDTmethodandANNmethodtoclassifynitrogenstresswithinacornﬁeldbasedonhyperspectralimages.Goeletal.(2003)alsoevaluatedtheDTandANNclassi-ﬁcationalgorithmsfortheclassiﬁcationofhyperspectraldatatodiscriminatebetweendifferentgrowthscenariosinacornﬁeldtoidentifyweedstressandnitrogenstatusofcorn.Inbothstudies,theANNsobtainedslightlybetterresultsthantheDTs.

Tangetal.(2003)developedatexture-basedweedclassiﬁca-tionmethodconsistingofalow-levelGaborwavelets-basedfeatureextractionalgorithmandahigh-levelANN-basedpatternrecogni-tionalgorithm.Inthisresearch,threespeciesofbroadleafweeds(commoncocklebur,velvetleaf,andivyleafmorningglory)andtwograces,giantfoxtailandcrabgrass,whicharecommoninIllinois,werestudied.Afterprocessing40sampleimageswith20samplesfromeachclass,theresultsshowedthatthemethodwascapableofclassifyingallthesamplescorrectlywithhighcomputationalefﬁ-ciency,demonstratingitspotentialforpracticalimplementationunderreal-timeconstraints.

Choetal.(2002)usedanANNusingaregularizationmethodtoovercomeoverﬁttingwithbettergeneralizationfordistinguishingradishfromweedseffectively.Inthisstudy,amachinevisionsys-temusingachargecoupleddevicecamerafortheweeddetectioninaradishfarmwasdeveloped.Shapefeatureswereanalyzedwiththebinaryimagesobtainedfromcolorimagesofradishandweeds.Usingthediscriminantanalysis,thesuccessfulrecognitionratewas92%forradishand98%forweeds.Torecognizeradishandweedsmoreeffectively,anANNwasused.TheANNmodeldistinguishedtheradishfromtheweedswith100%.

Yangetal.(2002)developedaweedrecognitionimagingsys-tembasedonLVQnetworkstoassistintheprecisionapplicationofherbicidesincornﬁelds.Digitalimageswerecollectedandtheintensitiesofthecolorswerecomparedforeachpixeloftheimages.ThepixelintensitiesoftheimageswereusedastheinputsforLVQANNs.TheANNsweretrainedtodistinguishcornfromweeds,aswellastodifferentiatebetweenweedspecies.ThesuccessrateforasingleANNindistinguishingagivenweedspeciesfromcornwasashighas90%,andashighas80%indistinguishinganyoffourweedspeciesfromcorn.

Moshouetal.(2002)usedtheSOMANNinasupervisedwayforaclassiﬁcationtaskforaweedspeciesspectraldetector.TheneuronsoftheSOMbecomeassociatedwithlocallinearmappings.Errorinformationobtainedduringtrainingwasusedinanovellearningalgorithmtotraintheclassiﬁer.Theproposedmethodachievedfastconvergenceandgoodgeneralization.

BajwaandTian(2001)developedweeddensitymodelsusingafeedforward,BPtrainedMLPnetworktomapthespatialdistribu-tionofweeddensitybasedonaerialdigitalcolorinfraredremotesensingoverasoybeanﬁeld.Inthisstudy,theANNweeddensity

modelsresultedinR2valuesof0.83–0.83.ThismodelmappedthespatialdistributionofweeddensitywithaR2valueof0.58foraﬁeldnotusedinmodelingformodelvalidation.

Itisquitedifﬁculttousemachinevisiontodistinguishweedsfromthemaincropinrealtime,duetothesubstantialcomputa-tionalresourcesandthecomplicatedalgorithmsrequired.Yangetal.(2000b)developedANNstoovercomesomeofthesedifﬁcultiesbyinterpretingimagesquicklyandeffectivelytodistinguishyoungcornplantsfromweeds.TheANNswereone-hidden-layerfeedfor-wardnetworkstrainedwiththeBPalgorithm.Inthestudy,atotalof80imagesofcornplantsandweedswereusedfortrainingpur-poses.ForsomeANNs,thesuccessrateforclassifyingcornplantswasashighas100%,whereasthehighestsuccessrateforweedrecognitionwas80%.

Burksetal.(2000)usedthecolorco-occurrencetexturestatisticsasinputvariablesforaBPtrainedANNweedclassiﬁcationmodel.Thestudyevaluatedclassiﬁcationaccuracyasafunctionofnetworktopologyandtrainingparameterselection.Inaddition,trainingcyclerequirementsandtrainingrepeatabilitywerestudied.

El-Fakietal.(2000)developedandtestedANN-basedweeddetectionalgorithmscapableofdetectingtheleadingweedspeciescompetingwithwheatandsoybeancrops.ThisstudycomparedstatisticaldiscriminantanalysisandtwoANNclassiﬁers.Theseclassiﬁersweretrainedandtestedusingthreeweedspecies(John-songrass,redrootpigweed,andyellowfoxtail)withsoybeanandthreeweedspecies(wildbuckwheat,cheat,andﬁeldbindweed)withwheat.TheresultsshowedthatthestatisticaldiscriminantanalysisclassiﬁerwasmoreaccuratethantheANNclassiﬁersinclassiﬁcationaccuracy.

Suzukietal.(2009)developedadiscriminantmodelforauto-matedweedcontrolusingimageryfromalinescanhyperspectralimagingsensor.Priortodevelopingdiscriminators,explanatoryvariablesforthemodelsweregeneratedfromspectralbandsusingtwomethods:stepwiseselectionmethod(RAW)usingthemul-tivariateteststatisticWilks’LambdaandPrincipalComponentAnalysis(PCA).LinearDiscriminantAnalysis(LDA)andaneuralnetwork(NN)modelswereemployedfordevelopmentofthedis-criminator,resultinginfourdiscriminantmodelstotest:RAW-LDA,RAW-NN,PCA-LDA,andPCA-NN.AccuraciesoftheNNmodelsweresuperiortotheLDA,butsuccessratesforalldiscriminantmodelsweregreaterthan85%.RAWmethodwassuperiortoPCAforpro-cessingspeedbecausePCAselectedvariablescalculatedfromallwavebands.

Inastudyofcropleafareaindices,KollerandUpadhyaya(2005a)developedanANNmodelthatutilizedtheleafareaindex(LAI)val-uesderivedfromaerialimagestotrainandpredictLAIchangesonadailybasis.BasedonthisANNmodel,KollerandUpadhyaya(2005b)predictedtheprocessingtomatoyieldbasedonsoil,crop,andenvironmentalparameters.KollerandUpadhyaya(2005c)fur-therdevelopedaBPtrainedfeedforwardANNmodeltopredictdailyLAIvaluesbasedonsixsetsofmodiﬁedNDVIvaluesderivedfrombiweeklyaerialimagesobtainedduringthetomatogrowingsea-son.ThetrainednetworkwasabletopredictLAIvalueswithR2valuesof0.96orhigher.ThecumulativeLAIpredictedbytheANNmodelcorrelatedwellwiththemeasuredvalues(R2=0.83).

Walthalletal.(2004)comparedempiricalandANNapproachesforestimatingcornandsoybeanLAIfromLandsatETM+imagery.Inthisstudy,anevaluationofLAIretrievalmethodswascon-ductedusing(1)empiricalmethodsemployingNDVIandanewsoiladjustedindexthatusesgreenwavelengthreﬂectance,(2)ascaledNDVIapproachthatusesnocalibrationmeasurements,and(3)ahybridapproachthatusesanANNandaradiativetransfermodelwithoutsite-speciﬁccalibrationmeasurements.Comparedtoothermethods,theANN-basedapproachiscomputationallycomplexasmultipleanalyticalstepsmustbecompletedbeforeanestimatecanbeproduced.

Y.Huangetal./ComputersandElectronicsinAgriculture71(2010)107–127115

Inthestudyofcropdisease,Pydipatietal.(2005)usedthecolorco-occurrencemethodfortexturalanalysistodeterminewhetherclassiﬁcationalgorithmscouldbeusedtoidentifydiseasedandnormalcitrusleavesbasedonaMahalanobisminimumdistanceclassiﬁeraswellasANNclassiﬁersbasedontheBPalgorithmfortwo-hidden-layerfeedforwardnetworksandRBFnetworkswithfuzzyoutputsgivingameasureofstrength.TheresultsindicatedthatthattheMahalanobisclassiﬁerandtheBPtrainedANNclassi-ﬁerperformedequallywellwhenusinghueandsaturationtexturefeaturesselectedthroughastepwisevariablereductionmethod.

Inastudyofcropgrowth,Hsiehetal.(2001)adoptedadual-hidden-layerfeedforwardANNtrainedwiththeerrorBPalgorithmtoanalyzeexperimentaldataanddevelopstrategiesforadynamicgrowthmodeltosimulatetherelationshipbetweenenvironmentalfactors(temperature,watersupplyanddailyradiation)andcab-bageseedlingquality.Hsiehetal.(2003)continuedtoworkonthemodeldevelopmentofaBPtrained,multilayerfeedforwardANNmodeltoinvestigatetherelationshipbetweenthequalityofcabbageseedlingsandtheirgrowthenvironment.Thisstudydevel-opedandevaluatedthreedifferentANNmodels.Byintegrationofschemesforvariousgrowthstagesandthehistoricalgrowthfactor,themodelcontributesmarkedlyinpredictionability.Theerrorisdecreasedby77%whenthebestmodeldevelopedinthisworkwasused.

Tumboetal.(2002a)trainedasingle-hidden-layerfeedforwardANNusingtheBPalgorithmusingspectralchannelsofthehyper-spectralreﬂectanceresponsepatternsfromcornplantasinputsforpredictingchlorophyllvaluesastheoutput.TheBPtrainedANNmodelwasdevelopedusingspectralchannelsofthespectralreﬂectanceresponsepatternsasinputsandchlorophyllreadingsasanoutput.Themodelshowedstrongcorrelationbetweenpre-dictedandactualchlorophyllmeterreadingsfromthesamecornvarietyandsoiltypeasthetrainingset.

Moshouetal.(2001)usedtheSOMANNinasupervisedwayforclassiﬁcationofagriculturalplants.Theclassiﬁcationmethodwasappliedinaprecisionfarmingapplicationforclassiﬁcationofcropsandweedsusingspectralproperties.Theclassiﬁcationper-formanceoftheproposedmethodwasproventobesuperiorcom-paredwithotherstatisticalandneuralclassiﬁers,suchasanoptimalBayesianclassiﬁerintheformofaprobabilisticneuralnetwork.3.2.3.IrrigationandETcalculation

Elgaalietal.(2006)developedaone-hidden-layerfeedforwardANNmodelandaconsumptiveusemodelfortheregionofCol-orado’sArkansasRiverbasintoinvestigatethepossibleeffectsofregionalclimaticchangesonirrigationwatersupplyanddemand.Thetwomodelswereappliedtotheregiontoestimatetheeffectsofclimatechangeonirrigationwaterbalance.

Forselectionofthebestcompromiseirrigationplanningstrat-egyinthecasestudyofJayakwadiirrigationproject,Maharashtra,India,Rajuetal.(2006),basedmultiobjectivelinearprogrammingoptimization,employedKohonenneuralnetworkstosortnon-dominatedirrigationplanningstrategiesintosmallergroups.Theresultsindicatedthattheintegratedmodelmethodologywaseffec-tiveformodelingmultiobjectiveirrigationplanningproblems.

InETcalculation,Brutonetal.(2000)developedfeedforwardANNmodelstrainedwiththeBPalgorithmtoestimatedailypanevaporationusingmeasuredweathervariablesasinputsfromweatherdataofRome,Plains,andWatkinsville,Georgia,USA.Inthisstudy,dailypanevaporationwasalsoestimatedusingmultiplelin-earregressionandthePriestley–TaylormethodandwascomparedtotheresultsoftheANNmodels.TheANNmodelofdailypanevap-orationwithallavailablevariablesasinputswasthemostaccuratemodel.Inoverall,panevaporationestimatedwithANNmodelswasslightlymoreaccuratethanpanevaporationestimatedwithamultiplelinearregressionmodelorthePriestley–Taylorequation.

Odhiamboetal.(2001b)developedafeedforwardANNonaconceptualandstructuralbasisinastudyoneliminationoftrial-and-errorindeterminingtheshapeofthemembershipfunctionsinthefuzzycontrolrulesforestimatingdailyreferenceET.Theresultsofthestudyshowedthattheoptimizedfuzzy–neuralmodelisrea-sonablyaccurate,andiscomparabletotheFAOPenman–Monteithequation.ThisapproachcanprovideaneasyandefﬁcientmeansoftuningfuzzyETmodels.

3.2.4.Soilanalysis

Inthestudyofsoilproﬁles,Odhiamboetal.(2004)pre-sentedanapplicationofafuzzy–neuralnetworkclassiﬁerforunsupervisedclusteringandclassiﬁcationofsoilproﬁlesusingground-penetratingradarimagery.FreelandandOdhiambo(2007)usedatwo-layerperceptronneuralnetworkthatperformssupervisedclassiﬁcationtoexaminethefeasibilityofusingtexturalfeaturesextractedfromground-penetratingradarfornon-intrusivelymappingsubsurfacesoilconditions.TheANNclassiﬁerwasusedtoassigndatatotheknownsubsurfacecategories.Theresultsofsubsurfacecharacterizationusingextractedtexturalfea-tureswerefoundtobeincloseagreementwithresultsobtainedbycarefulvisualinterpretationofthedata.Thisapproachofground-penetratingradarimageryclassiﬁcationwastobeconsideredasanalternativemethodtotraditionalhumaninterpretationonlyintheclassiﬁcationofvoluminousdatasets,whereintheextensivetimerequirementwouldmakethetraditionalhumaninterpreta-tionimpractical.

Bajwaetal.(2004)usedamultilayerfeedforwardANNtrainedwithGAtooptimizetheANNtopologyandotherfourunsupervisedandsupervisedmethodstoidentifyaerialhyperspectralimagebandstocharacterizesoilelectricalconductivityandcanopycov-erageinagriculturalﬁelds.Bandselectionwasperformedwithbothunsupervisedandsupervisedapproaches.Fivemethods(threeunsupervisedandtwosupervised)areproposedandcomparedtoidentifyhyperspectralimagebandstocharacterizesoilelectricalconductivityandcanopycoverageinagriculturalﬁelds.Eachhyper-spectralimagebandwasrankedusingallﬁvemethodsandANNmeasurewasthemostusefulinselectingbandsspeciﬁctoatargetcharacteristicwithminimuminformationredundancy.

FerentinosandAlbright(2002)presentedafeedforwardANNmodelthatwastrainedwiththequasi-NewtonBPalgorithmandpredictspHandelectricalconductivitychangesintherootzoneoflettucegrowninadeep-troughhydroponicsystem.Themostsuit-ableandaccuratecombinationofnetworkarchitectureandtrainingmethodwasonehiddenlayerwithninehiddennodes,trainedwiththequasi-Newtonbackpropagationalgorithm.ThemodelprovedcapableofpredictingpHatthenext20-mintimestepwithin0.01pHunitsandelectricalconductivitywithin5␮Scm−1.Simplerpre-dictionmethods,suchaslinearextrapolationandthelazymanprediction(inwhichpredictionisthevalueoftheprevioustimestep),gavecomparableaccuracymuchofthetime.However,theyperformedpoorlyinsituationswherethecontrolactionsofthesys-temhadbeenactivatedandproducedrelativelyrapidchangesinthepredictedparameters.Inthosecases,theneuralnetworkmodeldidnotencounteranydifﬁcultiespredictingtherapidchanges.

Additionally,Fidêncioetal.(2001)appliedCPANNandRBFANNinclassiﬁcationofsoilsusingnear-infraredspectroscopy.InglebyandCrowe(2001)developedfeedforwardANNmodelswithareduced-memoryLevenberg–MarquardtBPtrainingalgorithmforpredictingorganicmattercontentinSaskatchewansoils.Altendorfetal.(1999)developedasetoffeedforwardANNswithBPtrainingtopredictsoilwatercontentatagivendepthasafunctionofsoiltemperature.

Inthestudyofsoilproperties,ZhangandKushwaha(1999)appliedANNstosimulatetheinteractionbetweensoilandtoolfortillageandsoilbehavior.Schaapetal.(1998)calibratedhierarchical

116Y.Huangetal./ComputersandElectronicsinAgriculture71(2010)107–127

ANNmodelsforpredictionofwaterretentionparametersandsaturatedhydraulicconductivityfrombasicsoilproperties.

Inthestudiesofsoiltemperature,Yangetal.(1997a)developedamodelbasedonafeedforwardANNtrainingwiththeBPalgorithmtosimulatedailysoiltemperaturesat100,500and1500mmdepthsofsoilfromOttawa,Ontario,Canada.Yangetal.(2004b)comparedamultivariateadaptiveregressionsplinesmodelwithfeedforwardANNsinsimulationofsoiltemperatureatdifferentdepths.Thecorrelationcoefﬁcientsoflinearregressionfrombothmultivari-ateadaptiveregressionsplinesandANNswerealwayshigherthan0.950.Theresultsdemonstratethepotentialofmultivariateadap-tiveregressionsplinestobeusedasaregressiontechnologyinagriculturalapplications.

AnumberofpapersdevelopingandusingANNshavebeenpub-lishedforsoilpropertystudiesinGeodermaandSSSAJinrecentyears.Anaguetal.(2009)developedsorptionmodelsasafunc-tionofbasicsoilpropertiesusingANNsforestimationheavymetalsorptioninGermansoils.TheresultsofthestudyindicatedthattheANNmodelsperformedbetterthanonmultiplelinearregressionMLRinallcasesfor9heavymetalsusing13soilpropertiesaswellassolutionphaseconcentrationsasinputsandsorbedphasecon-centrationsasoutput.Khalilmoghadametal.(2009)usedANNsinestimatingsoilshearstrengthfrommeasuredparticlesizedistribu-tion,topographicattributes,NDVI,soilorganiccarbon(SOC),andCaCO3.Threeneuralnetworksstructures:multilayerperceptron(MLP),generalizedfeedforward,andmodularfeedforwardnet-workswereusedandcomparedwithconventionalmultiple-linearregressionanalysis.InthecomparisontheimprovementswereidentiﬁedwithallthreeneuralnetworkmodelsoveraconventionalMLRmodels.Cockxetal.(2009)evaluatedANNsforextractionoftopsoiltexturalinformationfromthedepth-weightedEM38DD(anelectromagneticinductionsoilsensor)signalstopredictthetop-soilclaycontent.EarlierMerdunetal.(2006)comparedANNandregressionpedotransferfunctionsforpredictionofsoilwaterreten-tionandsaturatedhydraulicconductivityanddemonstratedthatANNproducedpromisingresults.

3.2.5.Precisionagriculture

XiangandTian(2007)developedaBPtrainedone-hidden-layerfeedforwardANNinanartiﬁciallyintelligentcontrollerforgainandexposuretimeautomationofamultispectralcameratocom-pensateforchangingnaturallightingconditionsandtoacquirewhite-balancedimages.ThisstudywasundertakentodevelopanartiﬁciallyintelligentcontrollerbasedonanANNwithanadaptiveneuro-fuzzyinferencesystem.Itwasshownthroughexperimentsthatthedevelopedalgorithmwasabletocompletemultispectralcameraparametercontrolwithinthreeiterationsforeachchannel.Theconvergencespeedwasfasterthanwithconventionalcontrolmethods.

Irmaketal.(2006)developedaBPANNmodeltopredictthespatialdistributionofsoybeanyieldsandtounderstandthecausesofyieldvariability.First,aBPtrainedANNmodelwasdevelopedbyrelatingsoybeanyieldtotopography,soil,weather,andsitefactorsandthemodelpredictionswereevaluatedforthesameﬁeldforindependentyears.Then,thepotentialuseoftheANNmodelwasalsoexploredforpredictingyieldsinindependentﬁelds.Finally,theabilityoftheANNtoattributeyieldlossesduetosoybeancystnematodes(SCN),soilpH,andweedswasevaluated.TheresultsshowedthattheANNmodelcouldpredictspatialyieldvariabilitywell.

Miaoetal.(2006)employedANNanalysistoevaluatetherela-tiveimportanceofselectedsoil,landscapeandseedhybridfactorsoncornyieldandgrainqualityintwoIllinois,USAﬁelds.TheresultsindicatedthattheresponsecurvesgeneratedbytheANNmodelsweremoreinformativethansimplecorrelationcoefﬁ-cientsorcoefﬁcientsinmultipleregressionequations.Unoetal.

(2005)developedyieldpredictionmodelsusingstatisticalandANNapproachesalongwithvariousvegetationindices.Themodelswereusedtopredictcornyieldfromcompactairbornespectrographicimagerdata.ThestudyshowedthatalthoughnocleardifferencewasobservedbetweenANNsandstepwisemultiplelinearregres-sionmodels,thehighpotentialusefulnessofANNswasconﬁrmed.Drummondetal.(2003)investigatedsupervisedfeedforwardANNsalongwithstepwisemultiplelinearregressionandprojectionpur-suitregressiontoidentifymethodsabletorelatesoilpropertiesandgrainyieldsonapoint-by-pointbasiswithintenindividualsite-years.Inthestudy,theANNsconsistentlyoutperformedstepwisemultiplelinearregressionandprojectionpursuitregressionandprovidedminimalpredictionerrorsineverysite-year.However,insite-yearswithrelativelyfewerobservationsandinsite-yearswhereasingle,overridingfactorwasnotapparent,theimprove-mentsachievedbyneuralnetworksoverstepwisemultiplelinearregressionandprojectionpursuitregressionweresmall.

Liuetal.(2001)designedafullyconnected,BPtrainedfeed-forwardANNtoapproximatethenonlinearyieldfunctionrelatingcornyieldtofactorsinﬂuencingyield.Factorsaffectingcropyields,suchassoil,weather,andmanagement,weresocomplexthattra-ditionalstatisticscouldnotgiveaccurateresults.Asanautomaticlearningtool,theANNisanattractivealternativeforprocessingthemassivedatasetgeneratedbyprecisionfarmingproductionandresearch.Inthisstudy,afeedforward,completelyconnected,back-propagationANNwasdesignedtoapproximatethenonlinearyieldfunctionrelatingcornyieldtofactorsinﬂuencingyield.3.2.6.Chemicalapplication

KrishnaswamyandKrishnan(2002)predictedthenozzlewearratesforfourfannozzlesusingregressionandANNtechniques.AcomparisonoftheregressionmodelandANNmodelshowedthattheregressionandtheneuralnetworkmodelsperformedequallywellforthepredictions.Yangetal.(1997b)usedfeedforwardANNsforsimulationofpesticideconcentrationsinagriculturalsoils.AnANNmodelcanbeexecutedinreal-time,whilethesprayeriswork-ingintheﬁeld,inordertoadjustapplicationratestotherealextentoftheproblem.Inthisstudy,anANNmodelwasbuiltandtrainedwithinputsof:accumulateddailyrainfall,soiltemperature,poten-tialET,aswellastillagepracticesandthenumberofdayselapsedafterpesticideapplication.3.3.Geneticalgorithmapplications

3.3.1.Overview

Basedonasummaryof83papersandreports,GAshavebeenappliedinsolvingproblemsincropmanagement(31%),waterman-agement(27%),foodqualityandsafety(11%),foodprocessing(6%),precisionagriculture(4%),agriculturalbiology(4%),agriculturalmachinery(2%),agriculturalfacilities(2%),animalbehavior(2%),andothers(11%)suchasagriculturalvehicle,robotics,andpol-lution.GAsarebasicallyanoptimizationandsearchmethod.Theapplicationsforoptimizationtakethelargestportionofthetotal,66%.GAshavebeenalsousedtoassistwithmodelingandpredic-tion(18%),classiﬁcation(12%),control(2%),dataclustering(1%)andvaluethresholding(1%).

3.3.2.Cropmanagement

Thorpetal.(2004)usedtheGA-basedimagesegmentationalgo-rithm(Tangetal.,2000)togeneratemeasurementsofpercentvegetationcoverfromthehyperspectralimagescollectedoverasoybeanﬁeld.Fangetal.(2003)estimatedcropleafareaindexbyintegratingacanopyradiativetransfermodelandtheGAoptimiza-tiontechnique.ThismethodwasusedtoretrieveleafareaindexfromﬁeldmeasuredreﬂectanceaswellasfromatmosphericallycorrectedLandsatETM+data.YaoandTian(2003)proposedand

Y.Huangetal./ComputersandElectronicsinAgriculture71(2010)107–127117

testedaGA-basedselectiveprincipalcomponentanalysismethodusinghyperspectralremotesensingdataandgroundreferencedatacollectedwithinacornﬁeldforchlorophyllcontent,plantpop-ulation,andvarioushybrids.TheGA-basedmethodwasusedtoselectasubsetoftheoriginalimagebands,whichcouldreducethedatadimension.Aprincipalcomponenttransformwassubse-quentlyappliedtotheselectedbands.Then,imageprocessingonthereducedfeaturespacecouldbeperformedwithimprovedaccu-racy.Tangetal.(2000)developedamachinevision-basedweeddetectiontechnologyunderoutdoornaturallightingconditions.Inimageprocessing,aGA-based,supervisedcolorimagesegmenta-tionmethodinthehue-saturation-intensitycolorspacewasusedtoseparateplants(cropandweeds)fromthebackground(soil,rocks,andresidue)forthisreal-time,machine-vision-based,in-ﬁeldvari-abilitymappingandselectiveherbicideapplicationsystem.Pabicoetal.(1999)presentedaconceptualframeworkforusingaGAtodeterminecultivarcoefﬁcientsofcropsimulationmodels.

Inastudyofweeddistributionincropﬁelds,Diazetal.(2005)presentedamethodforinducingamodelthatappropriatelypre-dictstheheterogeneousdistributionofwild-oatintermsofsomeenvironmentalvariables.Fromseveralexperiments,distinctrulesetshavebeenfoundbyapplyingaGAtocarryouttheautomaticlearningprocess.Therulesetextractedwasabletoexplainaboutmostofweedvariability.Noguchietal.(1998)developedanintelli-gentvisionsystemforautonomousvehicleﬁeldoperations.Fuzzylogicwasusedtoclassifythecropsandweeds.AGAwasusedtooptimizeandtunethefuzzylogicmembershiprules.

3.3.3.Irrigation

RajuandKumar(2004)presentedanapplicationofGAforirri-gationplanning.TheGAtechniquewasusedtoevolveefﬁcientcroppingpatternstomaximizebeneﬁtsforanirrigationprojectinIndia.Kumaretal.(2006)presentedaGAmodelforobtaininganoptimaloperatingpolicyandoptimalcropwaterallocationsfromanirrigationreservoir.ThemodelwasappliedtotheMalaprabhasingle-purposeirrigationreservoirinKarnatakaState,India.TheoptimaloperatingpolicyobtainedusingtheGAsissimilartothatobtainedbylinearprogramming.

WardlawandBhaktikul(2004)describedthedevelopmentofaGAtosolveanirrigationwaterschedulingproblem.Theimplemen-tationoftheGAwastooptimizetheutilizationofwaterresourcesinirrigationsystemsoperatingonarotationalbasisalongwithcon-straintsthatrelatetoﬁeldsoilmoisturebalancesaswellascanalcapacities.Caietal.(2001)describedstrategiesforsolvinglargenonlinearwaterresourcesmodelsmanagementwithacombina-tionofGAswithlinearprogramming.

Kuoetal.(2000)presentedamodelbasedonon-farmirriga-tionschedulingandasimpleGAoptimizationmethodfordecisionsupportinirrigationprojectplanning.TheproposedmodelwasappliedtoanirrigationprojectlocatedinDelta,Utahforoptimiz-ingeconomicproﬁts,simulatingthewaterdemand,cropyields,andestimatingtherelatedcropareapercentageswithspeciﬁedwatersupplyandplantedareaconstraints.SimulationresultsdemonstratedthatthemostappropriateparametersofGAsforthisstudywere:numberofgenerations,populationsizes,probabilityofcrossover,andprobabilityofmutation.

3.3.4.Soilanalysis

PachepskyandAcock(1998)developedstochasticimagingoftheavailablesoilwatercapacity(AWC)andasoybeancropmodelGLYCIMtosimulatevariabilityanduncertaintyincropyieldesti-matesasrelatedtosoilsamplingdensityandweatherpatternsusingGAs.Parasuramanetal.(2007)estimatedsaturatedhydraulicconductivityusinggeneticprogramming(GP),tree-likerepresen-tationsofGassolutiondomain.InthestudytheperformanceoftheGPmodelswerecomparedwiththeANNmodelsandGPappeared

tobeapromisingtoolforestimatingthesaturatedhydrauliccon-ductivity.

3.3.5.Precisionagriculture

Liuetal.(2001)optimizedﬁfteeninputfactorswithaGAtodeterminemaximumyieldaftertheANNwastrained.TheANNapproximatedthenonlinearyieldfunctionrelatingcornyieldtofactorsinﬂuencingyield,wastrained.

3.3.6.Chemicalapplication

Potteretal.(2000)usedGAtooptimizenozzletypeorreleaseheightusingadesireddepositionasmodelinputwithamodelcon-strainonotherfactorssuchasaircrafttypeandspraysystemforaerialpesticideapplicationoptimization.3.4.Bayesianinferenceapplications

3.4.1.Overview

Basedonastudyof46papersandreports,BIhasbeenappliedinsolvingproblemsincropmanagement(28%),foodqualityandsafety(28%),watermanagement(17%),foodprocessing(7%),pre-cisionagriculture(7%),andothers(13%)suchasbioenergyandanimalhealth.BIisbasicallyaclassiﬁerforpatternrecognition.Theapplicationsforclassiﬁcationtotal50%.BIalsohasbeenusedtohelpmodelingandprediction(35%),parameterestimation(11%),andoptimizationandcontrol(4%).

3.4.2.Cropmanagement

Inweeddetection,Mathankeretal.(2007)usedaBayesianclas-siﬁerforclassiﬁcationsusingshape,color,spectralimagefeatures.Theclassiﬁerconductedthreeclassiﬁcationsusingonlyshapefea-tures,onlygreencolorandspectralenergyfeatures,andallfeaturescombined.Theclassiﬁcationaccuracyusingthesefeaturesatdif-ferentresolutionsforweeddetectionvariedfrom80%to87%.Theperformancewasunaffectedbyimageresolutionandshowspoten-tialforﬁeldapplications.

Foroutdoorﬁeldplantdetection,TianandSlaughter(1998)developedanenvironmentallyadaptivesegmentationalgorithm.Theresultsoftheenvironmentallyadaptivesegmentationalgo-rithmwereusedtocreateaBayesianclassiﬁerasdecisionsurface.MarchantandOnyango(2003)comparedaBayesianclassiﬁerwithamultilayerfeedforwardANNforplant/weed/soildiscriminationincolorimageswithapreferenceofBayesianclassiﬁeroverANNinthiscase.

3.4.3.IrrigationandETcalculation

ClemmensandKeats(1992)appliedBItothereal-timefeedbackcontrolofabasinirrigationsystem.Bayesianinferenceisappliedtothereal-timefeedbackcontrolofabasinirrigationsystem.Esti-matesoftheKostiakovk(inﬁltrationparameter)andManningn(Roughnessparameter)wereobtainedduringwateradvancesothattheoptimumcutofftimecanbedeterminedfrom:(1)esti-matesfromobservationofadvancetimeanddistanceandsolutionofthezero-inertiaborderirrigationmodel;and(2)eitherhistoricalestimatesofparametersorsubjectiveestimatesmadebytheirriga-tor.ThestudyshowedthattheBIhassomepotentialforimprovingreal-timecontrolofsurfaceirrigationsystems.

Saydeetal.(2008)proposedananalyticalmethodtoreduceuncertaintyindeterminationofsoilwaterdepletionbyincorporat-ingestimatesoftheuncertaintiesofETandmeasuredsoilmoistureintotheanalysis.UsingaBayesiananalysistheprobabilitydistri-butionsofthetwoestimatorswereincorporatedintoaposterioriprobabilitydistributionofdepletionthatprovidesabetterbasisforirrigationdecisions.

118Y.Huangetal./ComputersandElectronicsinAgriculture71(2010)107–127

3.4.4.Precisionagriculture

Chinchuluunetal.(2007)developedarobustBayesianclassi-ﬁerwithnormaldistributionforcolorsegmentationinacitrusyieldmappingsystemonacanopyshakeandcatchharvester.TheBayesianclassiﬁerworkedrelativelywellonthenon-uniformimages.

3.4.5.Chemicalapplication

LeeandSlaughter(1999)usedBayesianclassiﬁersindevelop-mentofareal-timeintelligentroboticweedcontrolsystemforselectiveherbicideapplicationtoin-rowweedsusingmachinevisionandprecisionchemicalapplication.3.5.Decisiontreeapplications

3.5.1.Overview

Asummaryof31papersandreportsonDTapplicationsindicatesthatDThavebeenappliedinsolvingproblemsincropmanagement(26%),agriculturalsafety(16%),animalproductionandmanage-ment(13%),foodsafety(13%),water(13%),agriculturalpollution(6%),andothers(13%)suchasprecisionagricultureandbiologi-calcontrol.DTalsoisaclassiﬁer.Theapplicationsforclassiﬁcationaccountfor52%ofthetotal.DThasbeenalsousedtoconductalter-nativeinduction(23%),rulerepresentation(13%),patterninduction(6%),andothers(6%)suchasfunctionapproximationandcontrol.3.5.2.Cropmanagement

Karimietal.(2005)compareddiscriminantanalysis,ANNandDTforclassifyingimageswithrespecttothenitrogenandweedmanagementpracticesappliedtotheexperimentalplots.Foriden-tifyingweedandnitrogenstressesincorn,DTcouldnotcompetewiththediscriminantanalysisandANNforbetterresults.Yangetal.(2004a)usedDTstoclassifymultispectralimagesofexperimentalplotshavingdifferentcropandweedpopulations.DTsaredirectlysuitedtoclassiﬁcationsincedatarepresentingagivenindividualaresortedthroughthedecisiontreestructuretofalldirectlyintoapre-deﬁnedcategory.Inthisstudy,anaircraft-mountedpushb-roomimagingspectrometerwasusedtoobtainscansoftheplotsinoneblue,ﬁvegreen,ﬁvered,andthirteeninfraredbands.Threetypesofinputwere:absolutevaluesofradiancefromthe24wave-bands,vegetationindex(VI),whichconsistsof12inputs,andNDVIwhichconsistsof65inputs.Resultsshowedthatthemostcom-plexclassiﬁcationproblem(distinguishingbetween11crop/weedcombinations)wasbestresolvedusingtheNDVIinputs.

Goeletal.(2001)usedaDTmethodtoclassifynitrogenstresswithinacornﬁeldbasedonhyperspectralimages.Goeletal.(2003)comparedtheDTandANNclassiﬁcationalgorithmsfortheclas-siﬁcationofhyperspectraldatatodiscriminatebetweendifferentgrowthscenariosinacornﬁeldtoidentifyweedstressandnitro-genstatusofcorn.DTtechnologywasappliedtoclassifydifferenttreatmentsbasedonthehyperspectraldata.Varioustree-growingmechanismswereusedtoimprovetheaccuracyofclassiﬁcation.Misclassiﬁcationratesofdetectingallthecombinationsofdifferentnitrogenandweedcategorieswere43,32,and40%forhyperspec-traldatasetsobtainedattheinitialgrowth,thetasselingandthefullmaturitystages,respectively.However,satisfactoryclassiﬁcationresultswereobtainedwhenonefactor(nitrogenorweed)wascon-sideredatatime.Inthiscase,misclassiﬁcationrateswereonly22and18%fornitrogenandweeds,respectively,forthedataobtainedatthetasselingstage.

3.5.3.Precisionagriculture

Cohenetal.(2006)usedtheDTmethodfordevelopingaspatialdecisionsupportsystem(SDSS)forMedﬂy(theMediterraneanfruitﬂy)controlonfruitcropsinIsrael.Thisstudyestablishedthemainstepstoassistthecoordinatorsinthedecision-makingprocedure

inthedevelopmentoftheSDSS.Theoutputisamapthatclassiﬁesthecitrusplotsintooneofthefollowingblocks:spraying;sprayingisrecommended;sprayingisnotrecommended;no-spraying;nodata.

Yangetal.(2001)usedtheDTalgorithmtodistinguishbetweenmanureandchemicalfertilizertreatmentsforairbornehyperspec-tralimageryclassiﬁcationofagriculturalﬁelds.Thesuccessoftheclassiﬁcationratewasashighas91%fortheearlyplantingsea-son,99%forthemidplantingseason,and95%forthelateplantingseason.

3.5.4.Chemicalapplication

Murrayetal.(2005)usedtheDTmodeltomodelannualpas-tureproductionforanalysisofpotentialofutilizingvariablerateapplicationtechnologyfromaircraftforimprovedplacementoffertilizer.

4.Fusionofsoftcomputingmethodsinagriculturalandbiologicalengineering

Withthedevelopmentofsoftcomputinginagriculturalandbiologicalengineering,techniqueshaveintegrateddifferentsoftcomputingtechniquesinasynergisticway.ItisadvantageoustoemployANNs,fuzzysystems,andevolutionaryalgorithmsincom-binationinsteadofexclusively(Hoffmannetal.,2005).

Table1listspapersandreportsonfusionofsoftcomputingtech-niquesagriculturalandbiologicalengineeringapplications.ThelistindicatesthattheintegrationofFLandANNsisprobablythemostcommonmethodoffusioninsoftcomputing(thirteenoutoftwenty-ninecollectedpapersandreports).ANFIS(Adaptive-Network-basedFuzzyInferenceSystem)isafuzzyinferencesystemimplementedintheframeworkofadaptiveneuralnetworks(Jang,1993).Byusingahybridlearningprocedure,theANFIScancon-structaninput–outputmappingbasedonbothhumanknowledge(intheformoffuzzy‘If–Then’rules)andstipulatedinput–outputdatapairs.Ithasbeendevelopedandusedtosolveproblemsinagriculturalandbiologicalengineering.4.1.Cropmanagement

Forcropprotection,PearsonandWicklow(2006)developedaneuralnetworktoidentifyfungalspeciesthatinfectsinglekernelsusingprincipalcomponentsofthereﬂectancespectraasinputfea-tures.TheneuralnetworkwastrainedusingGAsasLestanderetal.(2003)indicatedthattheGAtrainingalgorithmmethodwasmuchlesslikelytooverﬁtthedata.

Meyeretal.(2004)usedfuzzyinferencesystemsbuiltwithsub-tractiveclusteringinanANFISforadigitalcameraoperationstudyforcolor-basedclassiﬁcationsofuniformimagesofgrass,baresoil,cornstalksresidue,andwheatstrawresidue.Forweeddetection,Netoetal.(2003)usedANFISforimageprocessingtoseparateweedfrombackground.AGAwasincorporatedtoadjustthefuzzymem-bershipfunctionstoreducemisclassiﬁcationandimproveimagesegmentation.4.2.ETcalculation

BasedonthepreviousresearchonfuzzyETmodels(Odhiamboetal.,2001a),Odhiamboetal.(2001b)furtherstudiedhowtoelim-inatetrial-anderrorindeterminingtheshapeofthemembershipfunctionsinthefuzzycontrolrulesforestimatingdailyreferenceETbyapplicationofafuzzy–neuralsystemthroughfusingfuzzylogicandANNonaconceptualandstructuralbasis.Theneuralcomponentprovidedsupervisedlearningcapabilitiesforoptimiz-ingthemembershipfunctionsandextractingfuzzyrulesfromaset

Y.Huangetal./ComputersandElectronicsinAgriculture71(2010)107–127

Table1

Papersandreportsonfusionofsoftcomputingtechniquesinagriculturalandbiologicalengineering.Year19921997199719972001b20012001200120012001200320032003200320032003200420042004200420062006200620062006200620062007

Author

Linkoetal.KimandChoMorimotoetal.NoguchiandTeraoOdhiamboetal.Liuetal.

JindalandSrisawasQuetal.Miu

MiuandPerhinschiMorimotoetal.JindalandSrithamAndriyasetal.Chtiouietal.Leeetal.Netoetal.

Odhiamboetal.Meyeretal.Goeletal.

JainandSrinivasuluMadeiroetal.Hashimotoetal.Oliveiraetal.

Ferentinosetal.HancockandZhangLakshmietal.

PearsonandWicklowXiangandTian

Fusiontype

ANNmodelingforfuzzycontrol

ANNmodelingplusfuzzycontrolsimulationANNmodelingplusGAparameteroptimizationforfuzzycontrol

ANNsimulationandGAoptimization

ConceptualandstructuralfusionofFLandANNANNmodelingplusGAparameteroptimizationPNNtrainedwithGAALNdecisiontreeFLandGAFLandGA

ANNmodelingplusGAparameteroptimizationPNNtrainedwithGA

FCMclusteringforRBFtrainingSOMwithFCMclusteringANFISmodelingANFISclassiﬁcation

Fuzzy–neuralnetworkunsupervisedclassiﬁcation

ANFISclassiﬁcation

Fuzzyc-meansclusteringforRBFtrainingANNtraininedbyGA

GAparametersearchforANN

ANNmodelingplusGAparameteroptimizationGAparametersearchforANN

ANNparameterizedbyGAANFISclassiﬁcation

ANNtrainingplusGAparameteroptimizationANNtrainedbyGA

ANNmodelingplusANFIStrainingofFLcontroller

Applicationarea

Extrusioncontrol

BreadbakingprocesscontrolFruitstoragecontrol

119

Pathplanningofanagriculturalmobilerobotbottomofform

ETmodeloptimizationSettingtargetcornyieldsClassiﬁcationofsnackfoods

Functionapproximationtoconvertthemeasurementsofanelectronicnoseintoodorconcentrations

Optimizationofdesignandfunctionalparametersofthreshingunits

Dynamicoptimizationfortomatocoolstoragetominimizewaterloss

Classiﬁcationofeggstodetecteggshellcracks

PredictionoftheperformanceofvegetativeﬁlterstripsColorimagesegmentationofediblebeansPredictionofmultiplesoilproperties

AdaptiveimagesegmentationforweeddetectionClassiﬁcationofsoils

Classiﬁcationofuniformplant,soil,andresiduecolorimages

PredictionofsedimentandphosphorousmovementthroughvegetativeﬁlterstripsRainfall–runoffmodeling

ApproximationofsugarcanematurationcurvesGreenhousecroppingcontrol

Forecastingofagronomicalperformanceindicatorsinsugarcaneharvest

Greenhousecultivationcontrol

Hydraulicvanepumphealthclassiﬁcation

AutomaticcalibrationofcomplexwatershedmodelsClassiﬁcationofcornkernelsfordetectionoffungiinfection

Outdoorautomaticcameraparametercontrol

ofinput–outputexamplesselectedtocoverthedatahyperspaceofthesitesevaluated.4.3.Soilanalysis

Inastudyofsoilproﬁles,Odhiamboetal.(2004)combinedanANNandafuzzysysteminseriesandappliedthefuzzy–neuralnetworkclassiﬁerforunsupervisedclusteringandclassiﬁcationofsoilproﬁlesusingground-penetratingradarimagery.AftertheANNclassiﬁessoilproﬁlestripsintoacertainnumberofclustersdeter-mineddynamically,thefuzzymembershipvaluesforeachproﬁlestripareevaluatedinthesetofclassiﬁedclusters.Topredictsed-imentandphosphorousmovement,Goeletal.(2004)usedfuzzyc-meanclusteringalgorithmsfortheRBFANNmodelthroughthedatafromvegetativeﬁlterstrips.

Tosimultaneouslypredictwatercontentandsalinityusingthefrequency-responsesoildata,Leeetal.(2003)developedpartialleast-squaresandANFIS(Adaptive-NetworkFuzzyInferenceSys-tem)modelsatdifferentwatercontentandsalinitylevelsmeasuredwiththefour-electrodeWennerarraymethod.4.4.Precisionagriculture

XiangandTian(2007)developedacompleteartiﬁciallyintelli-gentcontrollerbasedonanANNandanANFISforimplementingthecontrollertoautomaticallyadjustmultispectralcameraparam-etersforcompensationofchangesinnaturallightingandtoacquirewhite-balancedimages.Oliveiraetal.(2006)proposedatwo-

stepdecisionsupportsystemstartswithanANNfollowedbyGAsforheuristicsearchtorecommendsuitablesugarcaneareastobeharvested.Liuetal.(2001)designedafeedforward,completelyconnected,BP-trainedANNtoapproximatethenonlinearyieldfunctionrelatingcornyieldtofactorsinﬂuencingyield.Bystratiﬁedsamplingbasedonrainfall,someofthedatawereexcludedfromthetrainingsetandusedtoverifytheyieldpredictionaccuracyoftheANN.AftertheANNwasdevelopedandtrained,optimiza-tionoftheﬁfteeninputfactorswasstudiedwithaGAtodeterminemaximumyield.

5.Supportvectormachines

SVMsasanewsetofsupervisedgeneralizedlinearclassiﬁers,havebeenintroducedtosolveproblemsandhaveattractedgreaterinterestrecentlyinagriculturalandbiologicalengineering.SVMsarecloselyrelatedtoneuralnetworks.Infact,anSVMmodelusingsigmoidkernelfunctionisequivalenttoatwo-layerperceptronneuralnetwork.Usingakernelfunction,SVMsarealternativetrain-ingmethodsforpolynomial,radialbasisfunction,andmultilayerperceptronclassiﬁersinwhichtheweightsofthenetworkarefoundbysolvingaquadraticprogrammingproblemwithlinearconstraints,ratherthanbysolvinganonconvex,unconstrainedminimizationproblemasinstandardANNtraining.SVMshaveoftenhigherclassiﬁcationaccuraciesthanmultilayerperceptronANNs.

SVMs,asasetofsupervisedgeneralizedlinearclassiﬁers,haveoftenbeenfoundtoprovidehigherclassiﬁcationaccuraciesthan

120Y.Huangetal./ComputersandElectronicsinAgriculture71(2010)107–127

Fig.4.PublicationsonapplicationsofSVMsinagriculturalandbiologicalengineer-ing.

otherwidelyusedpatternclassiﬁcationtechniquessuchasMLPneuralnetworks.Inagriculturalandbiologicalengineering,thegrowinginterestinSVMsisillustratedinFig.4.Fig.4indicatesthatalthoughthenumberofpeerreviewedpublicationsonthistopic

Table2

PapersandreportsonapplicationsofSVMsinagriculturalandbiologicalengineering.Year2003

Author

FletcherandKong

ApplicationmethodSVMclassiﬁcation

didnotincreasesteadilyoverthepastfewyears,anincreasingtrendinallpaperssuggestincreasingusageofestablishedtech-niqueswhichdonotrepresentoriginalresearchdevelopments.ThesuccessfulimplementationsofSVMsinagriculturalandbiologicalengineeringarelistedinTable2.

ThegrowinginterestinSVMsaredueto:(1)theirintrinsiceffec-tivenesswithrespecttotraditionalclassiﬁers,whichresultsinhighclassiﬁcationaccuraciesandverygoodgeneralizationcapability;(2)thelimitedeffortrequiredforarchitecturedesign(i.e.,theyinvolvefewcontrolparameters);and(3)thepossibilityofsolvingthelearningproblemaccordingtolinearlyconstrainedquadraticprogrammingmethods(MelganiandBruzzone,2004).

Insoilanalysis,Lamorskietal.(2008)estimatedsoilhydraulicparametersfrommeasuredsoilpropertiesusingSVMs.Theresultsoftheresearchindicatedthatthethree-parameterSVMsperformedmostlybetterthanorwiththesameaccuracyastheeleven-parameterANNs.TheadvantageofSVMwasmorepronouncedatsoilmatricpotentialswherelargerrelativeerrorshavebeenencounteredandthecorrelationbetweenpredictedandmeasuredsoilwatercontentswaslower.Andmostrecently,Twarakavietal.(2009)developedSVMsforestimatingthehydraulicparame-tersdescribingthesoilwaterretentionandhydraulicconductivity.Inestimatingwatercontentsandsaturatedhydraulicconductiv-itiestheSVM-basedmethodpredictedthehydraulicparameters,

Applicationarea

Classifyingfeaturevectorsanddecidewhethereachpixelinhyperspectralﬂuorescenceimagesofpoultrycarcassesfallsinnormalorskintumorcategories

ClassiﬁcationofmilkbyanelectronicnoseClassiﬁcationforrecognitionofplantdiseaseClassiﬁcationofelectronicnosedata

ClassiﬁcationofmodiﬁedstarchesbyFouriertransforminfraredspectroscopy

Identiﬁcationofteavarietiesbycomputervision

Classiﬁcationforweedandnitrogenstressdetectionincorn

Detectionofunderdevelopedhazenutsfromfullydevelopednutsbyimpactacoustics

DiscriminationofscreeningofcompoundfeedsusingNIRhyperspectraldata

DiscriminationofwheatclasseswithNIRspectroscopy

Blackwalnutshellandmeatclassiﬁcationusinghyperspectralﬂuorescenceimaging

Simulationofdaily,weekly,andmonthlyrunoffandsedimentyieldfronawatershed

Classiﬁcationtodifferentiateindividualfungalinfectedandhealthywheatkernels.

QuantiﬁcationofvitaminCcontentinkiwifruitusingNIRspectroscopyClassiﬁcationofmeatwithsmalldataset

Predictionofdifferentconcentrationclassesofinstantcoffeewithelectronictonguemeasurements

EstimationofhydraulicparametersfrommeasuredsoilpropertiesClassiﬁcationofpaddyseedsbyharvestyear

Predictionofporkmeattotalviablebacteriacountwithhyperspectralimaging

On-lineassessinginternalqualityofpearsusingvisible/NIRtransmissionClassiﬁcationofforestdatacovertypes

IdentiﬁcationofvarietiesofChinesecabbageseedsusingvisibleandNIRreﬂectancespectroscopy

Ricewinecompositionpredictionbyvisible/NIRspectroscopyClassiﬁcationofintactandcrackedeggs

QuantiﬁcationofpearﬁrmnessusingNIRspectroscopy

Classrecognitionofriceblastwithmultispectralimagingtosupervisevariablespray

Quantitativeassessmentofamyloseandproteincontentinriceaftergammairradiationusinginfraredspectroscopyandchemometrics

Estimationofhydraulicparametersdescribingthesoilwaterretentionandhydraulicconductivity

200420042005200520062006200620062006200720072007200820082008200820082008200820082008200820092009200920092009

Brudzewskietal.Tianetal.

PardoandSberveglieriPiernaetal.Chenetal.Karimietal.Onaranetal.Piernaetal.WangandPaliwalJiangetal.Oommenetal.Zhangetal.Fuetal.Khotetal.Kovacsetal.Lamorskietal.Lietal.

PengandWangSunetal.

TrebarandSteeleWuetal.Yuetal.Dengetal.Fuetal.QiandMaShaoetal.Twarakavietal.

SVMneuralnetworkclassiﬁcationSVMclassiﬁcation

SVMwithRBFkernelofRBFSVMclassiﬁcationSVMclassiﬁcationSVMclassiﬁcationSVMclassiﬁcationSVMclassiﬁcation

Least-squaresSVMclassiﬁcationGaussiankernelbasedSVMclassiﬁcation

SVMmodelingandpredictionMulti-classSVMwithkernelofRBFneuralnetwork

Least-squaresSVMmodelingandprediction

SVMclassiﬁcation

SVMmodelingandpredictionSVMmodelingandpredictionLeast-squaresSVMclassiﬁcationLeast-squaresSVMmodelingandprediction

SVMmodelingandpredictionSVMclassiﬁcation

Least-squaresSVMclassiﬁcationLeast-squaresSVMmodelingandprediction

SVMclassiﬁcation

Least-squaresSVMmodelingandprediction

SVMclassiﬁcation

Least-squaresSVMmodelingandprediction

SVMmodelingandprediction

Y.Huangetal./ComputersandElectronicsinAgriculture71(2010)107–127121

whichmostlyimprovedcomparedwiththoseobtainedusingtheANN-basedmethod.

Inthestudiesofcropsforprecisionagriculture,Karimietal.(2006)evaluatedSVMasatoolforclassifyingairbornehyperspec-tralimagestakenoveracornﬁeld.Theclassiﬁcationwasperformedwithrespecttonitrogenapplicationratesandweedmanagementpractices,andtheclassiﬁcationaccuracywascomparedwiththoseobtainedbyanANNmodelonthesamedata.TheSVMmethodresultedinverylowmisclassiﬁcationratesascomparedtotheANNapproachforallcases.DetectionofstressesinearlycropgrowthstageusingtheSVMmethodcouldaidineffectiveearlyapplicationofsite-speciﬁcremediestotimelyin-seasoninterven-tions.Tianetal.(2004)usedtheSVMandchromaticitymomentforrecognitionofplantdiseasesbasedonthefeaturesofcolortextureimageofplantdisease.TheexperimentalresultsprovedthattheSVMmethodhadexcellentclassiﬁcationandgeneralizationabilityinsolvinglearningproblemwithsmalltrainingsetofsample,andcouldﬁtforclassiﬁcationofplantdisease.

TrebarandSteele(2008)employedaSVMforclassiﬁcationofforestdatacovertypes.Inthisstudy,alarge,imbalanceddatasetindifferentforestcovertypeclasseswastransformedintoanumberofnewdatasetsandaSVMwasusedtoconductabinaryclassiﬁcationofbalancedandimbalanceddatasetswithvarioussizes.TheuseofdistributedSVMarchitecturesbasicallyreducedthecomplexityofthequadraticoptimizationproblemofverylargedatasets.TheexperimentalresultsofdistributedSVMarchitecturesshowedtheimprovementoftheaccuracyforlargerdatasetsincomparisontoasingleSVMclassiﬁerandtheirabilitytoimprovethecorrectclassiﬁcationoftheminorityclass.

6.Comparisonandlimitationsofsoftcomputingtechniques

Therearelimitsonsoftcomputingweneedtobeawareofintheoreticalstudyandpracticalapplication.Tikketal.(2003)stud-iedtheapproximationbehaviorofsoftcomputingtechniques.Inthestudy,theauthorsconductedasurveyoftheresultsofuni-versalapproximationtheoremsachievedthusfarinvarioussoftcomputingareas.Thesetechniquescentermainlyinfuzzycontrolandneuralnetworks.Theauthorspointoutthatthesetechniqueshavecommonapproximationbehaviorinthesensethatanarbi-traryfunctionfromacertainsetoffunctions(usuallythesetofcontinuousfunction)canbeapproximatedwitharbitraryaccu-racyonacompactdomain.Thismeansthatfuzzysystemsandneuralnetworkshavetheabilitytoapproximateanyfunctiontoanarbitrarydegreeofaccuracy.However,fortheapproximation,unboundednumbersof“buildingblocks”(i.e.fuzzysetsorhiddenneurons)areneededtoachievetheprescribedaccuracy.Ifthenum-berofbuildingblocksisrestricted,itisprovedforsomefuzzyandneuralsystemsthattheuniversalapproximationpropertyislost.Therefore,itisreasonabletomakeatrade-offbetweenaccuracyandthenumberofthebuildingblocksindeterminingfunctionalrelationships.Atypicalpracticalapplicationofthisrecommenda-tionistodeterminethenumberofhiddenneuronsincrementallyinANNtrainingforacceptablefunctionapproximationandbettergeneralization.

Comparedtoclassicallogic,FLisnotalwaysaccuratebecausetheresultsareoftenperceivedasanestimate.Also,fuzzysys-temstypicallyrequirethedifﬁcultandtime-consumingprocessofknowledgeacquisitionalthoughtheyprovidetheunderstandableformofknowledgerepresentation.

ANNsarepowerfulcomputingtechniques,whicharedesignedtomimichumanlearningprocessesbyestablishinglinkagesbetweenprocessinputandoutputdata.Thesetechniqueshavebeenwidelyappliedwithadvanceddevelopmentwiththeirunique

advantages,suchasnounderlyingassumptionaboutthedistri-butionofdata,arbitrarydecisionboundarycapabilities,universalapproximationcapabilities,easyadaptationtodifferenttypesandstructuresofdata,abilitytofuzzyoutputvaluestoenhanceclassi-ﬁcation,andgoodgeneralizationcapabilities.However,ANNshavesomedisadvantagesincommon,whichneedtobeconsideredinpracticalapplication:

•Blackbox

ANNsareblackboxinnature.Therefore,iftheproblemistoﬁndtheoutputresponsetotheinputinsystemidentiﬁcation,ANNscanbeagoodﬁt.However,iftheproblemistospeciﬁcallyiden-tifycausal-effectiverelationshipbetweeninputandoutput,ANNshaveonlylimitedabilitytodoitcomparedwithconventionalstatisticalmethods.•Longcomputingtime

ANNtrainingneedstoiterativelydeterminenetworkstruc-tureandupdateconnectionweights.Also,datausedfortrainingmaycontaincertaindegreeofnoises.Therefore,ANNtrainingisatime-consumingprocess.Withatypicalpersonalcomputerorworkstation,theBPalgorithmwilltakealotofmemoryandmaytakeminutes,hours,daysandevenlongerbeforethenetworkconvergestotheoptimalpointwithminimummeansquareerror.Conventionalstatisticalregressionwiththesamesetofdata,onthecontrary,maygenerateresultsinsecondsusingthesamecomputer.•Localminima

Duetoadditionofhiddennodesandlayer(s)andthenonlinear-ityoftheactivationfunctionofeachhiddennodesand/oroutputnodesinnetworkstructure,ANNtraininghasthepossibilitytoproduceforcomplexerrorsurfaceswhichcontainmanyminimasuchasuseofBPalgorithmfortrainingMLPs.Sincesomemin-imaaredeeperthanothers,itispossiblethatthealgorithmwillnotﬁndaglobalminimum.Instead,thenetworkmayfallintolocalminima,whichrepresentsuboptimalsolutionsinsteadofoptimal.•Overﬁtting

Withtoomuchtrainingtime,toomanyhiddennodes,ortoolargetrainingdataset,thenetworkwilloverﬁtthedataandhaveapoorgeneralization,i.e.highaccuracyfortrainingdatasetbutpoorinterpolationoftestingdata.ThisisanimportantissuebeinginvestigatedinANNresearchandapplications(Huang,2009).AlthoughANNshavestrongcapabilitiesoflearningandadapta-tionwitha“blackbox”naturethatdealswithinputsandoutputs,theyrepresentknowledgeimplicitlyandmayproduceresultsthataredifﬁculttointerpret.ANNsalsomayrequirelongtrainingtimesonnoisydata.Therefore,ANNsmaynotbeneededwhentraditionalmethodsareappropriate.

GAsworkontheirowninternalrulesandaregoodforcomplexorlooselydeﬁnedproblemsusingtheirinductivenaturewithouttheneedtoknowanyrulesoftheproblem.However,withuseofthisinductiveabilityalone,thealgorithmsdonotnecessarilyevolvetotheoptimalsolution.GAsalwayshaveariskﬁndingasuboptimalsolution.BIusespriorprobabilities.However,priorprobabilitiesareintrinsicallysubjective,whichcanbedifferentfrompersonbyperson.ThismaybethefundamentallimitofBI.

Usually,learningbyDTsisfast,andtheresultiseasilyinter-pretedbyhumanspecialists.However,thelearningcouldproduceoverﬁtting,theoutputattributeofDTsneedtobecategorical,andeachdecisionislimitedtooneoutputattribute.Theinduc-tionprocessusuallyselectsonlyasmallnumberoftheavailableattributessothatinformationthatisdistributedonalargenumberofattributes(witheachattributecarryingonlylimitedinformationabouttheclass)cannotbehandledadequately,resultinginsubopti-malpredictionaccuracy.Insuchsituations,Bayesianclassiﬁersand

122Y.Huangetal./ComputersandElectronicsinAgriculture71(2010)107–127

ANNsareoftenusedinstead.SVMshavegainedpopularityinmanytraditionallyANNdominatedﬁelds.UseoftheSVMalwaysfoundtheminimumglobally,whicheliminatesthelocalminimumissuehappenedinANNtraining.However,SVMswereoriginallydevel-opedtosolvebinaryclassiﬁcationproblems.Howtoeffectivelyextenditformulti-classclassiﬁcationisstillanongoingresearchissue.Typically,amultipleclassclassiﬁercanbeconstructedbycombiningseveralbinaryclassiﬁers.Further,allclassescanbeconsideredatonce.HsuandLin(2002)gavedecompositionimple-mentationsfortwosuch“all-together”methods.Theycomparedtheperformancewiththreemethodsbasedonbinaryclassiﬁca-tions:one-against-all,one-against-one,andDirectedAcyclicGraphSVM(DAGSVM).Theexperimentsindicatedthattheone-against-oneandDAGmethodsaremoresuitableforpracticalusethantheothermethods.7.Thefuture

Asdescribedabove,eachsoftcomputingtechniquehasitsownlimitations.Thefusionoftwoorthreeofthesetechniqueswillcon-tinuetobeoneofthemajortrendsinsoftcomputingengineeringapplications.Fuzzyandneuro-fuzzysystemsrepresentknowledgeinanexplicitform,suchasfuzzyrules,ratherthaninanimplicitformasANNsalone.Accordingly,FLandANNsaremergedtoinherittheadvantagesofbothparadigmsandtoavoidtheirdrawbacks.

Recently,withthedevelopmentofelectronicsandinformationtechnologies,3S(remotesensing(RS),globalpositioningsystem(GPS),andgeographicinformationsystem(GIS))technologiesandvariableratetechnology(VRT)havestartedtobecomeavailabletoagriculturalproductionforcropseeding,irrigation,andchem-icalapplication.Foreffectiveapplicationsofthetechnologies,asystemintegrationisneeded.Forthesystemintegration,FLisacriticaltooltoestablish‘If–Then’rulestorepresentextensivehumanknowledgetobuildupaknowledge-basedsystemtocon-nectindependentcomponentstogether.Inthesystemaknowledgebaseobtainstheinformationfromvarioussourcessuchassci-entists,engineers,andfarmersandprovidesadvicesuitedtoageographiclocationbasedonlocalweather,soilandwaterdatafromGPS,GISandtraditionalknowledgearchives.Theweatherdataarecollectedfromthedifferentsensorsonadailybasisforsolarradiation,airtemperature,rainfall,andwindrun.Theinfer-enceengineanalyzesuserqueryandsendsinformationrequeststotheknowledgebaseandmatchesthemwiththestoredknowledgerulesthroughfuzzy‘If–Then’rulesoralgorithmsspeciﬁctothepar-ticulardomainordiscipline.Thenthecontrolsystemoutputsthedecisionoftheinferenceenginetothevariablerateequipmentforcropseedingandchemicalapplication.Acomprehensivemodelbaseincludescropgrowthmodel,healthmodel,perhapsothermodelsthatdescribestheconditionofthecropsothatinformationwillbeprovidedfordecisionsupporttoresultinmoreprecisetim-ingofseeding,moreaccurateuseofseed,fertilizersand,irrigatedwaterinordertoenhancecropproductionandenvironmentalprotectionaswell.Amethodbasecontainslinearandnonlinearmethodsofpatternrecognition,statistics,andsoftcomputingsuchasFL,ANNsandGAs.Inthemodelingprocess,oneormoremeth-odscanbeactivatedtoassembleandgeneratedesignatedmodelstructureandparameters.

ANNshavebeenusedasapowerfultoolinsolvingproblemsinscientiﬁcresearchandengineeringapplications.ANNshavetheirownlimitationstorestrictthemasasubstituteoftraditionalmeth-odssuchasstatisticalregression,patternrecognition,andtimeseriesanalysis.Inthenextdecadeorsowithadvanceddevelopmentofcomputerpower,ANNswillcontinuetodevelopnewapplica-tionsinvariousﬁeldsinagriculturalandbiologicalengineering.Asapowerfulalternativetoconventionalmethods,ANNswillbestud-iedmoretodevelopapproachestoovercomingproblemsofANNsingeneralorinaspeciﬁcresearcharea.Infoodscienceandengineer-ing,soilandwaterrelationshipsforcropmanagement,anddecisionsupportforprecisionagriculture,moreapplicationsofANNsmaybeexpected.ANNswillbecontinuouslyappliedstandaloneorfusionwithothersoftcomputingtechniques.Areasofstudymayactivelyinvolvefoodqualityandsafety,soilandwaterresourcesmanage-ment,croppestmanagement,andprecisionagriculture.

ForeffectiveresearchanddevelopmentitmaybepossibletogenerateguidelinesforpredeterminingoptimalANNstructuresandtrainingalgorithms.Forprocessstatisticalmodeling,additionalresearchtopicsmaybetoestablishgenericprocedurestoincludesigniﬁcantvariablestoandexcludenon-signiﬁcantonesfromANNmodelsandtoaddconﬁdencelimitsontheoutputpredictionsandparameterestimations.ThesedevelopmentswillpermitANNstoalsoutilizethetechniquesofconventionalstatistics.

SVMsappearedasanewtechniquewhichhasadvancedsoftcomputingdevelopment.Inpracticalapplications,SVMsoftenpro-ducedhighclassiﬁcationaccuraciesandverygoodgeneralizationcapabilitiescomparedwithANNs.Also,intheprocessofmodel-ingSVMsrequirelesseffortinsettingupcontrolparametersforarchitecturedesign.SVMmodelsareclosetoclassicalMLPneuralnetworks.Usingakernelfunction,SVMsareanalternativetrain-ingmethodforpolynomial,RBFandMLPclassiﬁersinwhichtheweightsofthenetworkarefoundbysolvingaquadraticprogram-mingproblemwithlinearconstraints,ratherthanbysolvinganon-convex,unconstrainedminimizationproblemasinstandardneuralnetworktraining.WiththeadvantagesofSVMsoverANNsandthegrowinginterestsofSVMs,itcanbeexpectedthatinthenextdecadeSVMswillbemoreactivelyusedinagriculturalandbiologicalengineering.

Nomatterwhichsoftcomputingmethodisused,adaptivelearn-ingcanbeapowerfultoolwhenalsoimplementedtoappropriatelyexploitthepotentialsynergybetweenmethods,somethatcanincludeinputfromsensors.Anexampleofhowthiscanworkinvolvesmodel-basedirrigationschedulingwithassistancefromsoilwatersensors.Thomson(1998)explainedamethodbywhichgranularmatrixsoilwatersensorswereusedtoprovide“feedback”orcorrectiontotwoinputparametersofthewaterbalancecom-ponentofthePnutgro1.02growthmodel(Booteetal.,19).Soilwatersensorsinferredrelativewateruptakeintherootzoneandcorrectedtherootgrowthfunctioninﬂuencingthedepthofwaterregulation.Thegoalwastoimprovemodelpredictionssothatthemodelcouldrunstand-aloneaftersensorsprovidedcorrectiondataduringseveralsoildryingcycles.Experimentsindicatedtemporalconvergenceofmodel-basedrepresentationsofsoilwaterpoten-tialonsensor-basedrepresentationsastwoparameters(soil–waterparameterDrainedUpperLimitorDUL)androotweightingfactorswereadjusted.

Inthesoilandwatercontextforcropmanagementanddecisionsupportforprecisionagriculturemoreapplicationsofsoftcom-putingtechniques,especiallySVMsstandaloneorfusedwithothersoftcomputingtechniquesandsensor-derivedinformation,maybeexpected.Areasofapplicationmightinvolveclassiﬁcationforagri-culturalsoilspatialdistribution,waterresourceoptimizationforirrigationplanning,detectionandclassiﬁcationofcropstressandpests(weeds,insectsanddiseases)detection,analysisofremotesensingimagery,studyofcropandyield,andﬁeldprescriptionsforvariableratechemicalapplication.

Anothertrendinsoftcomputingapplicationsislikelytobethefusionofsoftcomputingandhardcomputing.Althoughnosuccess-fulapplicationsofhardandsoftcomputingfusioninagriculturalandbiologicalengineeringcouldbefoundthusfar,thetechniqueshowsgreatpotentialforfutureresearchoverthenextdecade.Thefusionofsoftcomputingandhardcomputingshouldbeabletopro-videinnovativesolutionstotheproblemswithhigh-performance,cost-effective,andreliablecomputingsystems.Manypublications

Y.Huangetal./ComputersandElectronicsinAgriculture71(2010)107–127123

areusefultohelpdevelopsuchcomputingsystems.Ovaska(2005)editedabookthatcombinestheexperienceofmanyinternation-allyrecognizedexpertsinthesoft-andhard-computingresearchworldstopresentpracticingengineerswiththebroadestpossiblearrayofmethodologiesfordevelopinginnovativeandcompetitivesolutionstoreal-worldproblems.Eachofthechaptersillustratesthewide-rangingapplicabilityofthefusionconceptinsuchcriticalareasas•computersecurityanddatamining;

•electricalpowersystemsandlarge-scaleplants;•motordrivesandtoolwearmonitoring;

•userinterfacesandtheWorldWideWeb;and•

aerospaceandrobustcontrol.

Ovaskaetal.(2006)clariﬁedthepresentvaguenessrelatedtothefusionofsoftcomputingandhardcomputing.Differentfusionschemeswereclassiﬁedas12corecategoriesandsixsupplemen-tarycategories,andthecharacteristicfeaturesofsoftcomputingandhardcomputingconstituentsinpracticalfusionimplementa-tionswerediscussedaswell.SickandOvaska(2007)introducedamulti-dimensionalcategorizationschemeforfusiontechniquesandapplieditbyanalyzingseveralfusiontechniqueswherethesoftcomputingpartwasrealizedbyaneuralnetwork.Thecategoriza-tionschemefacilitatedthediscussionofadvantagesordrawbacksofcertainfusionapproaches,thussupportingthedevelopmentofnovelfusiontechniquesandapplications.References

Abraham,A.,Jain,R.,Thomas,J.,Han,S.Y.,2007.D-SCIDS:distributedsoftcomputing

intrusiondetectionsystem.JournalofNetworkandComputerApplications30,81–98.

Anagu,I.,Ingwersen,J.,Utermann,J.,Streck,T.,2009.Estimationofheavymetal

sorptioninGermansoilsusingartiﬁcialneuralnetworks.Geoderma152(1–2),104–112.

Al-Faraj,A.,Meyer,G.E.,Horst,G.L.,2001.Acropwaterstressindexfortallfescue

(FestucaarundinaceaSchreb.)irrigationdecision-making—afuzzylogicmethod.ComputersandElectronicsinAgriculture32,69–84.

Altendorf,C.T.,Elliott,R.L.,Stevens,E.W.,Stone,M.L.,1999.Developmentand

validationofaneuralnetworkmodelforsoilwatercontentpredictionwithcomparisontoregressiontechniques.TransactionsoftheASAE42(3),691–699.Ambuel,J.R.,Colvin,T.S.,Karlen,D.L.,1994.Afuzzylogicyieldsimulatorforpre-scriptionfarming.TransactionsoftheASAE37(6),1999–2009.

Anderson,J.A.,Silverstein,J.W.,Ritz,S.A.,Jones,R.S.,1977.Distinctivefeatures,cat-egoricalperception,andprobabilitylearning:someapplicationsofaneuralmodel.PsychologicalReview84,413–451.

Andriyas,S.,Negi,S.C.,Rudra,R.P.,Yang,S.X.,2003.Modellingtotalsuspendedsolids

invegetativeﬁlterstripsusingartiﬁcialneuralnetworks.AmericanSocietyofAgriculturalEngineers,St.Joseph,MI,ASAEpapernumber:032079.

Bajwa,S.G.,Tian,L.F.,2001.AerialCIRremotesensingforweeddensitymappingin

asoybeanﬁeld.TransactionsoftheASAE44(6),1965–1974.

Bajwa,S.G.,Bajcsy,P.,Groves,P.,Tian,L.F.,2004.Hyperspectralimagedatamining

forbandselectioninagriculturalapplications.TransactionsoftheASAE47(3),5–907.

Bayes,T.,1763.Anessaytowardssolvingaprobleminthedoctrineofchances.

PhilosophicalTransactionsoftheRoyalSociety53,370–418.

Bonissone,P.P.,1994.Fuzzylogiccontrollers:anindustrialreality.In:Zurada,J.M.,

MarksII,R.J.,Robinson,C.J.(Eds.),ComputationalIntelligence:ImitatingLife.IEEEPress,Piscataway,NJ,pp.316–327.

Boote,K.J.,Jones,J.W.,Hoogenboom,G.,Wilkerson,G.G.,Jagtap,S.S.,19.PNUTGRO

v.1.02.PeanutCropGrowthSimulationModel.User’sGuide.FloridaAgriculturalExperimentStationJournalNo.8420.UniversityofFlorida,Gainesville,FL,76pp.Bragato,G.,2004.Fuzzycontinuousclassiﬁcationandspatialinterpolationincon-ventionalsoilsurveyforsoilmappingofthelowerPiaveplain.Geoderma118(1–2),1–16.

Brudzewski,K.,Osowski,S.,Markiewicz,T.,2004.Classiﬁcationofmilkbymeansof

anelectronicnoseandSVMneuralnetwork.SensorsandActuatorsB98(2–3),291–298.

Bruton,J.M.,McClendon,R.W.,Hoogenboom,G.,2000.Estimationdailypanevapo-rationwithartiﬁcialneuralnetworks.TransactionsoftheASAE43(2),491–496.Burges,C.J.C.,1998.Atutorialonsupportvectormachinesforpatternrecognition.

DataMiningandKnowledgeDiscovery2,121–167.

Burks,T.F.,Shearer,S.A.,Gates,R.S.,Donohue,K.D.,2000.Backpropagationneural

networkdesignandevaluationforclassifyingweedspeciesusingcolorimagetexture.TransactionsoftheASAE43(4),1029–1037.

Cai,X.,McKinney,D.C.,Lasdon,L.,2001.Solvingnonlinearwatermanagementmod-elsusingacombinedgeneticalgorithmandlinearprogrammingapproach.AdvancesinWaterResources24(6),667–676.

Carpenter,G.A.,Grossberg,S.,1987.ART2:self-organizationofstablecategory

recognitioncodesforanaloginputpatterns.AppliedOptics26,4919–4930.Carpenter,G.A.,Grossberg,S.,Rosen,D.B.,1991.FuzzyART:faststablelearning

andcategorizationofanalogpatternsbyanadaptiveresonancesystem.NeuralNetworks4,759–771.

Charniak,E.,1991.Bayesiannetworkswithouttears.AIMagazine12(4),50–63.Chen,Q.,Zhao,J.,Cai,J.,Wang,X.,2006.Studyonidentiﬁcationofteausingcom-putervisionbasedonsupportvectormachine.ChineseJournalofScientiﬁcInstruments27(12),1704–1706.

Chen,Y.,Zheng,J.,Xiang,H.,Huang,S.,2006.Studyonanintelligentsystemforpreci-sionpesticideapplicationbasedonfuzzycontrolandmachinevision.AmericanSocietyofAgriculturalandBiologicalEngineers,St.Joseph,MI,ASABEpapernumber:061129.

Chen,C.T.,Chen,S.,Hsieh,K.W.,Yang,H.C.,Hsiao,S.,Yang,I.C.,2007.Estimation

ofleafnitrogencontentusingartiﬁcialneuralnetworkwithcross-learningschemeandsigniﬁcantwavelengths.TransactionsoftheASABE50(1),295–301.

Chinchuluun,R.,Lee,W.S.,Ehsani,R.,2007.Citrusyieldmappingsystemonacanopy

shakeandcatchharvester.AmericanSocietyofAgriculturalandBiologicalEngi-neers,St.Joseph,MI,ASABEpapernumber:073050.

Cho,S.I.,Ki,N.H.,1999.Autonomousspeedsprayerguidanceusingmachinevision

andfuzzylogic.TransactionsoftheASAE42(4),1137–1143.

Cho,S.I.,Lee,D.S.,Jeong,J.Y.,2002.Weed-Plantdiscriminationbymachinevision

andartiﬁcialneuralnetwork.BiosystemsEngineering83(3),275–328.

Chtioui,Y.,Panigrahi,S.,Backer,L.F.,2003.Self-organizingmapcombinedwitha

fuzzyclusteringforcolorimagesegmentationofediblebeans.TransactionsoftheASAE46(3),831–838.

Cleland,J.,Turner,W.,Wang,P.,Espy,T.,Chappell,P.J.,Spiegel,R.J.,Bose,B.,

1992.FuzzylogiccontrolofACinductionmotors.FuzzySystems8(12),843–850.

Clemmens,A.J.,Keats,J.B.,1992.Bayesianinferenceforfeedback-control.II.Sur-faceirrigationexample.JournalofIrrigation&DrainageDivisionASCE118(3),416–432.

Cockx,L.,VanMeirvenne,M.,Vitharana,U.W.A.,Verbeke,L.P.C.,Simpson,D.,Saey,

T.,VanCoillie,F.M.B.,2009.ExtractingtopsoilinformationfromEM38DDsensordatausinganeuralnetworkapproach.SoilScienceSocietyofAmericaJournal73,2051–2058.

Cohen,Y.,Cohen,A.,Timar,D.,Gazit,Y.,2006.Developingspatialdecisionsup-portsystemformedﬂycontrolinIsrael.AmericanSocietyofAgriculturalandBiologicalEngineers,St.Joseph,MI,ASABEpapernumber:061152.

Cristianini,N.,Taylor,J.S.,2000.AnIntroductiontoSupportVectorMachinesand

OtherKernel-BasedLearningMethods.CambridgeUniversityPress,NewYork.Cybenko,G.V.,19.Approximationbysuperpositionsofasigmoidalfunction,

MathematicsofControl.SignalsandSystems2(4),303–314.Darwin,C.,1859.OntheOriginofSpecies.JohnMurray,London,UK.

Deng,X.,Wang,Q.,Wu,L.,Gao,H.,Wen,H.,Wang,S.,2009.Eggshellcrackdetection

byacousticimpulseresponseandsupportvectormachine.AfricanJournalofAgriculturalResearch4(1),40–48.

Diaz,B.,Ribeiro,A.,Bueno,R.,Guinea,D.,Barroso,J.,Ruiz,D.,Fernadez-Quintanilla,

C.,2005.Modellingwild-oatdensityintermsofsoilfactors:amachinelearningapproach.PrecisionAgriculture6(2),213–228.

Drummond,S.T.,Sudduth,K.A.,Joshi,A.,Birrell,S.J.,Kitchen,N.R.,2003.Statistical

andneuralmethodsforsite-speciﬁcyieldprediction.TransactionsoftheASAE46(1),5–14.

Duda,R.O.,Hart,P.E.,Stork,D.G.,2001.PatternClassiﬁcation,2nded.Wiley,New

York.

Eddy,D.M.,1982.Probabilisticreasoninginclinicalmedicine:problemsandoppor-tunities.In:Kahneman,D.,Slovic,P.,Tversky,A.(Eds.),JudgementunderUncertainty:HeuristicsandBiases.CambridgeUniversityPress,Cambridge,UK.Edwards,W.,1982.Conservatisminhumaninformationprocessing.In:Kahneman,

D.,Slovic,P.,Tversky,A.(Eds.),JudgementunderUncertainty:HeuristicsandBiases.CambridgeUniversityPress,Cambridge,UK.

Eerikäinen,T.,Linko,P.,Linko,S.,Siimes,T.,Zhu,Y.H.,1993.Fuzzylogicandneural

networksapplicationsinfoodscienceandtechnology.TrendsinFoodScience&Technology4,237–242.

EET(ElectronicsEngineeringTimes),1991.Europegetsintofuzzylogic.Electronics

EngineeringTimes,November11,1991.

El-Faki,M.S.,Zhang,N.,Peterson,D.E.,2000.Weeddetectionusingcolormachine

vision.TransactionsoftheASAE43(6),1969–1978.

Elgaali,E.,Garcia,L.A.,Ojima,D.S.,2006.Sensitivityofirrigationwaterbalanceto

climatechangeinthegreatplainsofColorado.TransactionsoftheASABE49(5),1315–1322.

Fang,H.L.,Liang,S.L.,Kuusk,A.,2003.Retrievingleafareaindexusingageneticalgo-rithmwithacanopyradiativetransfermodel.RemoteSensingofEnvironment85(3),257–270.

Ferentinos,K.P.,Albright,L.D.,2002.PredictionneuralnetworkmodelingofpHand

electricalconductivityindeep-troughhydroponics.TransactionsoftheASAE45(6),2007–2015.

Ferentinos,K.P.,Arvanitis,K.G.,Tantau,H.J.,Sigrimis,N.,2006.Specialaspectsof

ITforgreenhousecultivation.294–312ofChapter5PrecisionAgriculture,inCIGRHandbookofAgriculturalEngineeringVolumeVIInformationTechnol-ogy.EditedbyCIGR—TheInternationalCommissionofAgriculturalEngineering;VolumeEditor,AxelMunack.ASABE.St.Joseph,MI.

124Y.Huangetal./ComputersandElectronicsinAgriculture71(2010)107–127

Ferguson,R.B.,Lark,R.M.,Slater,G.P.,2003.Approachestomanagementzonedeﬁ-nitionforuseofnitriﬁcationinhibitors.SoilScienceSocietyofAmericaJournal67,937–947.

Fidêncio,P.H.,Ruisánchez,I.,Poppi,R.J.,2001.Applicationofartiﬁcialneuralnet-workstotheclassiﬁcationofsoilsfromSãoPaulostateusingnear-infraredspectroscopy.Analyst126,2194–2200.

Flamig,B,2000.Discoverthevalueofdatamining:turningpilesofstagnantinfor-mationintogoldenopportunities.ReferenceSeries:HowComputersWork,PartII4(3),120–125.SmartComputing,Lincoln,NE.

Fletcher,J.T.,Kong,S.G.,2003.Principalcomponentanalysisforpoultrytumor

inspectionusinghyperspectralﬂuorescenceimaging.In:ProceedingsoftheInternationalJointConferenceonNeuralNetworks,vol.1,Portland,Oregon,pp.149–153.

Fogel,L.J.,Owens,A.J.,Walsh,M.J.,1966.ArtiﬁcialIntelligencethroughSimulated

Evolution.JohnWiley,NewYork.

Freeland,R.S.,Odhiambo,L.O.,2007.Subsurfacecharacterizationusingtexturalfea-turesextractedfromGPRdata.TransactionsoftheASABE50(1),287–293.

Fu,X.,Ying,Y.,Xu,H.,Yu,H.,2008.Supportvectormachinesandnearinfraredspec-troscopyforquantiﬁcationofvitaminCcontentinkiwifruit.AmericanSocietyofAgriculturalandBiologicalEngineers,St.Joseph,MI,ASABEpapernumber:085204.

Fu,X.,Ying,Y.,Xu,H.,2009.Quantitativeanalysisoffruitﬁrmnessbynearinfrared

spectroscopyandleast-squaressupportvectormachine.AmericanSocietyofAgriculturalandBiologicalEngineers,St.Joseph,MI,ASAEpapernumber:0975.

Ghahramani,Z.,1997.LearningdynamicBayesiannetworks.LectureNotesinCom-puterScience1387,168–197.

Gigerenzer,G.,Hoffrage,U.,1995.HowtoimproveBayesianreasoningwithout

instruction:frequencyformats.PsychologicalReview102,684–704.

Gil,Y.,Sinfort,C.,Guillaume,S.,Brunet,Y.,Palagos,B.,2008.Inﬂuenceofmicrom-eteorologicalfactorsonpesticidelosstotheairduringvinespraying:dataanalysiswithstatisticalandfuzzyinferencemodels.BiosystemsEngineering100,184–197.

Goel,P.K.,Prasher,S.O.,Patel,R.M.,Landry,J.A.,2001.Weedandnitrogenstress

detectionincornusinghyperspectralremotesensing.AmericanSocietyofAgri-culturalEngineers,St.Joseph,MI,ASAEpapernumber:01-1199.

Goel,P.K.,Prasher,S.O.,Patel,R.M.,Landry,J.A.,Bonnell,R.B.,Viau,A.A.,2003.Clas-siﬁcationofhyperspectraldatabydecisiontreesandartiﬁcialneuralnetworkstoidentifyweedstressandnitrogenstatusofcorn.ComputersandElectronicsinAgriculture39(2),67–93.

Goel,P.K.,Andriyas,S.,Rudra,R.P.,Negi,S.C.,2004.Modelingsedimentandphospho-rousmovementthroughvegetativeﬁlterstripsusingartiﬁcialneuralnetworksandGRAPH.AmericanSocietyofAgriculturalEngineers,St.Joseph,MI,ASAEpapernumber:042263.

Goldberg,D.E.,Holland,J.H.,1988.Geneticalgorithmsandmachinelearning.

MachineLearning3(2–3),95–99.

Goldberg,D.E.,19.GeneticAlgorithmsinSearch,Optimization,andMachine

Learning.Addison-Wesley,Reading,MA.Grefenstette,J.J.,1994.Geneticalgorithmsformachinelearning.

http://citeseerx.ist.psu.edu/viewdoc/summary?doi=10.1.1.17.1053.

Grossberg,S.,1976.Adaptivepatternclassiﬁcationanduniversalrecoding.1.Parallel

developmentandcodingofneuralfeaturedetectors.BiologicalCybernetics23,187–202.

Grunwald,P.,1997.Theminimumdescriptionlengthprincipleandnon-deductive

inference.In:ProceedingsoftheIJCAIWorkshoponAbductionandInductioninAI,Nagoya,Japan.

Hancock,K.M.,Zhang,Q.,2006.Ahybridapproachtohydraulicvanepumpcon-ditionmonitoringandfaultdetection.TransactionsoftheASABE49(4),1203–1211.

Hashimoto,Y.,Morimoto,T.,DeBaerdemaeker,J.,2006.Speakingplant/speaking

fruitapproaches.244-259ofChapter5PrecisionAgriculture,inCIGRHand-bookofAgriculturalEngineeringVolumeVIInformationTechnology.EditedbyCIGR—TheInternationalCommissionofAgriculturalEngineering;VolumeEditor,AxelMunack.ASABE.St.Joseph,MI.

Hebb,D.O.,1949.TheOrganizationofBehavior.Wiley,Hoboken,NJ.

Hellendoorn,H.,1993.DesignanddevelopmentoffuzzysystemsatsiemensR&D.

In:ProceedingsofSecondIEEEInternationalConferenceonFuzzySystems,SanFransisco,CA,USA,pp.1365–1370.

Hillier,F.S.,Lieberman,G.J.,2005.IntroductiontoOperationsResearch.McGraw-Hill,

Boston,MA.

Hirota,K.,1993.IndustrialApplicationsofFuzzyTechnology.Springer,Tokyo,Japan.Hoffmann,F.,Gao,X.Z.,Olhofer,M.,Satyadas,A.,2005.Editorial:applicationreviews.

AppliedSoftComputing5,261–2.

Holland,J.H.,1975.AdaptationinNaturalandArtiﬁcialSystems.UniversityofMichi-ganPress,AnnArbor,MI.

Holmblad,L.P.,Ostergaard,J.J.,1982.Controlofacementkilnbyfuzzylogic.In:

Gupta,M.M.,Sanchez,E.(Eds.),FuzzyInformationandDecisionProcesses.North-Holland,pp.3–399.

Hopﬁeld,J.J.,1982.Neuralnetworksandphysicalsystemswithemergentcom-putationalabilities.ProceedingsoftheNationalAcademyofSciences79(8),2554–2558.

Horikawa,S.,Furuhashi,T.,Uchikawa,Y.,1992.Onfuzzymodellingusingfuzzy

neuralnetworkswithbackpropagationalgorithm.IEEETransactionsonNeuralNetworks3(5),801–806.

Hornik,K.,Stinchcombe,M.,White,H.,19.Multilayerfeedforwardnetworksare

universalapproximators.NeuralNetworks2(5),359–366.

Houck,C.,Joines,J.,Kay,M.,1998.Ageneticalgorithmforfunctionopti-mization:aMatlabimplementation.TheGAOTToolboxforMatlab:http://www.ie.ncsu.edu/mirage/GAToolBox/gaot/gaotindex.html.

Hsieh,K.W.,Chen,S.,Chang,W.H.,Lee,M.T.,Chen,C.T.,2001.Adynamicsimulation

modelforseedlinggrowth.TransactionsoftheASAE44(6),1949–1954.

Hsieh,K.W.,Chen,S.,Lai,J.H.,Yang,I.C.,2003.Neuralnetworkanalysisofenvi-ronmentalconditionsinﬂuencingcabbageseedlingquality.TransactionsoftheASAE46(2),501–506.

Hsu,C.W.,Lin,C.J.,2002.Acomparisonofmethodsformulti-classsupportvector

machines.IEEETransactionsonNeuralNetworks13,415–425.

Huang,Y.,2009.Advancesinartiﬁcialneuralnetworks—methodologicaldevelop-mentandapplication.Algorithms2,973–1007.

Ingleby,H.R.,Crowe,T.G.,2001.Neuralnetworkmodelsforpredictingorganicmat-tercontentinSaskatchewansoils.CanadianBiosystemsEngineering43(7),1–5.

Irmak,A.,Jones,J.W.,Batchelor,W.D.,Irmak,S.,Boote,K.J.,Paz,J.O.,2006.Artiﬁcial

neuralnetworkmodelasadataanalysistoolinprecisionfarming.TransactionsoftheASABE49(6),2027–2037.

Jain,A.,Srinivasulu,S.,2004.Developmentofeffectiveandefﬁcientrainfall-runoff

modelsusingintegrationofdeterministic,real-codedgeneticalgorithmsandartiﬁcialneuralnetworktechniques.WaterResourceResearch40,w04302.Jang,R.J.S.,1993.ANFIS:adaptive-network-basedfuzzyinferencesystem.IEEE

TransactionsonSystems,ManandCybernetics23(3),665–685.

Jang,R.J.S.,Sun,C.T.,1995.Neuro–fuzzymodellingandcontrol.Proceedingsofthe

IEEE83(3),378–406.

Jaynes,E.T.,1996.Probabilitytheorywithapplicationsinscienceandengineering.

http://bayes.wustl.edu/etj/science.pdf.html.

Jiang,L.,Zhu,B.,Jing,H.,Chen,X.,Rao,X.,Tao,Y.,2007.Gaussianmixture

model-basedwalnutshellandmeatclassiﬁcationinhyperspectralﬂuorescenceimagery.TransactionsoftheASABE50(1),153–160.

Jindal,V.K.,Srisawas,W.,2001.Acoustictestingofsnackfoodtexture.American

SocietyofAgriculturalEngineers,St.Joseph,MI,ASAEpapernumber:016038.Jindal,V.K.,Sritham,E.,2003.Detectingeggshellcracksbyacousticimpulseresponse

andartiﬁcialneuralnetworks.AmericanSocietyofAgriculturalEngineers,St.Joseph,MI,ASAEpapernumber:036170.

Jones,D.,Barnes,E.M.,2000.Fuzzycompositeprogrammingtocombineremote

sensingandcropmodelsfordecisionsupportinprecisioncropmanagement.AgriculturalSystems65,137–158.

Karimi,K.,Prasher,S.O.,McNairn,H.,Bonnell,R.B.,Dutilleul,P.,Goel,P.K.,2005.

Classiﬁcationaccuracyofdiscriminantanalysis,artiﬁcialneuralnetworks,anddecisiontreesforweedandnitrogenstressdetectionincorn.TransactionsoftheASAE48(3),1261–1268.

Karimi,Y.,Prasher,S.O.,Patel,R.M.,Kim,S.H.,2006.Applicationofsupportvector

machinetechnologyforweedandnitrogenstressdetectionincorn.ComputersandElectronicsinAgriculture51(1–2),99–109.

Khalilmoghadam,B.,Afyuni,M.,Abbaspour,K.C.,Jalalian,A.,Dehghani,A.A.,Schulin,

R.,2009.EstimationofsurfaceshearstrengthinZagrosregionofIran—acom-parisonofartiﬁcialneuralnetworksandmultiple-linearregressionmodels.Geoderma153(1–2),29–36.

Khoshnevis,B.,Chignell,M.H.,1985.Aframeworkforartiﬁcialintelli-genceapplicationssoftwaredevelopment.ComputersinIndustry6(5),363–369.

Khot,L.R.,Panigrahi,S.,Woznica,S.,2008.Neural-network-basedclassiﬁcationof

meat:evaluationoftechniquestoovercomesmalldatasetproblems.BiologicalEngineering1(2),127–143.

Kim,S.,Cho,I.,1997.Neuralnetworkmodelingandfuzzycontrolsimulationfor

bread-bakingprocess.TransactionsoftheASAE40(3),671–676.

Koller,M.,Upadhyaya,S.K.,2005a.Relationshipbetweensoil-adjustedvegetation

indexandleafareaindexforprocessingtomatoes.AppliedEngineeringinAgri-culture21(5),927–933.

Koller,M.,Upadhyaya,S.K.,2005b.Predictionofprocessingtomatoyieldusingacrop

growthmodelandremotelysensedaerialimages.TransactionsoftheASAE48(6),2335–2341.

Koller,M.,Upadhyaya,S.K.,2005c.Relationshipbetweenmodiﬁednormalizeddif-ferencevegetationindexandleafareaindexforprocessingtomatoes.AppliedEngineeringinAgriculture21(5),927–933.

Kovacs,Z.,Kantor,D.B.,Fekete,A.,2008.Comparisonofquantitativedetermination

techniqueswithelectronictonguemeasurements.AmericanSocietyofAgricul-turalandBiologicalEngineers,St.Joseph,MI,ASABEpapernumber:084879.Krishnaswamy,M.,Krishnan,P.,2002.Nozzlewearratepredictionusingregression

andneuralnetwork.BiosystemsEngineering82(1),49–56.

Kumar,N.K.,Raju,S.,Ashok,B.,2006.Optimalreservoiroperationforirrigation

ofmultiplecropsusinggeneticalgorithms.JournalofIrrigationandDrainageEngineering132(2),123–129.

Kuo,S.,Merkey,G.P.,Liu,C.,2000.Decisionsupportforirrigationprojectplan-ningusingageneticalgorithm.AgriculturalWaterManagement45(3),243–266.

Lakshmi,G.,Sudheer,K.P.,Chaubey,I.,2006.Autocalibrationofcomplexwatershed

modelsusingsimulation-optimizationframework.AmericanSocietyofAgricul-turalandBiologicalEngineers,St.Joseph,MI,ASABEpapernumber:062126.Lamorski,K.,Pachepsky,Y.,Slawinski,C.,Walczak,R.T.,2008.Usingsupportvec-tormachinestodeveloppedotransferfunctionsforwaterretentionofsoilsinPoland.SoilScienceSocietyofAmericaJournal72,1243–1247.

Lark,R.M.,2000.Designingsamplinggridsfromimpreciseinformationonsoilvari-ability,anapproachbasedonthefuzzykrigingvariance.Geoderma98(1–2),35–59.

Y.Huangetal./ComputersandElectronicsinAgriculture71(2010)107–127

125

Lauritzen,S.L.,2003.Somemodernapplicationsofgraphicalmodels.In:Green,P.J.,

Richardson,S.,Hjort,N.L.(Eds.),HighlyStructuredStochasticSystems.OxfordUniversityPress,Oxford,UK,pp.13–28.

Lee,K.H,Zhang,N.,Das,S.,2003.Comparingadaptiveneuro-fuzzyinferencesystem

(ANFIS)topartialleast-squares(PLS)methodforsimultaneouspredictionofmultiplesoilproperties.AmericanSocietyofAgriculturalEngineers,St.Joseph,MI,ASAEpapernumber:033144.

Lee,W.S.,Slaughter,D.C.,1999.Roboticweedcontrolsystemfortomatoes.Precision

Agriculture(1),95–113.

Lestander,T.A.,Leardi,R.,Geladi,P.,2003.Selectionofnear-infraredwavelengths

usinggeneticalgorithmsforthedeterminationofseedmoisturecontent.JournalofNear-InfraredSpectroscopy11(4),433–446.

Li,X.,He,Y.,Wu,C.,2008.Leastsquaresupportvectormachineanalysisforthe

classiﬁcationofpaddyseedsbyharvestyear.TransactionsoftheASABE51(5),1793–1799.

Lim,C.,Sim,E.,2005.Productionplanninginmanufacturing/remanufacturingenvi-ronmentusinggeneticalgorithm.In:Proceedingsofthe2005ConferenceonGeneticandEvolutionaryComputation,Washington,DC,pp.2217–2218.

Linko,P.,Zhu,Y.H.,Linko,S.,1992.Applicationofneuralnetworkmodelinginfuzzy

extrusioncontrol.FoodandBioproductsprocessing.TransactionsIChemE70,131–137.

Liu,J.,Goering,C.E.,Tian,L.,2001.Aneuralnetworkforsettingtargetcornyields.

TransactionsoftheASAE44(3),705–713.

Madeiro,S.S.,Oliveira,F.R.,Alexandre,F.B.A.,Neto,F.B.,2006.Intelligentmodelingof

sugar-canematuration.In:Proceedingsofthe4thWorldCongressConferenceonComputersinAgricultureandNaturalResources,Orlando,FL,pp.2–8.Magee,J.F.,19.Decisiontreesfordecisionmaking.HarvardBusinessReviewJuly,

126–138.

Mamdani,E.H.,Assilian,S.,1975.Anexperimentinlinguisticsynthesiswithafuzzy

logiccontroller.InternationalJournalofMan-MachineStudies7,1–13.

Marchant,J.A.,Onyango,C.M.,2003.ComparisonofaBayesianclassiﬁerwitha

multilayerfeed-forwardneuralnetworkusingtheexampleofplant/weed/soildiscrimination.ComputersandElectronicsinAgriculture39,3–22.

Mathanker,S.K,Weckler,P.R.,Taylor,R.K.,2007.Effectivespatialresolutionforweed

detection.AmericanSocietyofAgriculturalandBiologicalEngineers,St.Joseph,MI,ASABEpapernumber:073049.

McCulloch,W.S.,Pitts,W.H.,1943.Alogicalcalculusoftheideasimmanentinner-vousactivity.BulletinofMathematicalBiophysics5,115–133.

Melgani,F.,Bruzzone,L.,2004.Classiﬁcationofhyperspectralremotesensingimages

withsupportvectormachines.IEEETransactionsonGeoscienceandRemoteSensing42(8),1778–1790.Merdun,H.,C¸ınar,Ö.,Meral,R.,Apan,M.,2006.Comparisonofartiﬁcialneuralnet-workandregressionpedotransferfunctionsforpredictionofsoilwaterretention

andsaturatedhydraulicconductivity.Geoderma90(1–2),108–116.

Meyer,G.E.,Hindman,T.W.,Jones,D.D.,Mortensen,D.A.,2004.Digitalcameraopera-tionandfuzzylogicclassiﬁcationofuniformplant,soil,andresiduecolorimages.AppliedEngineeringinAgriculture20(4),519–529.

Miao,Y.,Mulla,D.J.,Robert,P.C.,2006.Identifyingimportantfactorsinﬂuencing

cornyieldandgrainqualityvariabilityusingartiﬁcialneuralnetworks.PrecisionAgriculture7(2),117–135.

Minsky,M.,Papert,S.,1969.Perceptrons.MITPress,Cambridge,MA.

Mitra,S.,Pal,S.K.,1994.LogicaloperationbasedfuzzyMLPforclassiﬁcationandrule

generation.NeuralNetworks7(2),353–373.

Miu,P.I.,2001.Optimaldesignandprocessofthreshingunitsbasedonagenetic

algorithm.I.algorithm.AmericanSocietyofAgriculturalEngineers,St.Joseph,MI,ASAEpapernumber:013124.

Miu,P.I.,Perhinschi,M.G.,2001.Optimaldesignandprocessofthreshingunits

basedonageneticalgorithm.II.Application.AmericanSocietyofAgriculturalEngineers,St.Joseph,MI,ASAEpapernumber:013125.

Morimoto,T.,Tu,K.,Hatou,K.,Hashimoto,Y.,2003.Dynamicoptimizationusing

neuralnetworksandgeneticalgorithmsfortomatocoolstoragetominimizewaterloss.TransactionsoftheASAE46(4),1151–1159.

Moshou,D.,Vrindts,E.,DeKetelaere,B.,DeBaerdemaeker,J.,Ramon,H.,2001.A

neuralnetworkbasedplantclassiﬁer.ComputersandElectronicsinAgriculture31,5–16.

Moshou,D.,Ramon,H.,DeBaerdemaeker,J.,2002.Aweedspeciesspectraldetector

basedonneuralnetworks.PrecisionAgriculture3(3),209–223.

Murray,R.,Yule,I.,Lawrence,H.,2005.Economicandenvironmentalopportunities

fromutilizingVRATfromaircraftforimprovedplacementoffertilizer.AmericanSocietyofAgriculturalEngineers,St.Joseph,MI,ASAEpapernumber:051075.Neto,J.C.,Meyer,G.E.,Jones,D.D.,Surkan,A.J.,2003.Adaptiveimagesegmenta-tionusingafuzzyneuralnetworkandgeneticalgorithmforweeddetection.AmericanSocietyofAgriculturalEngineers,St.Joseph,MI,ASAEpapernumber:033088.

Nie,J.,Linkens,D.,1992.Neuralnetwork–basedapproximatereasoning:

principlesandimplementation.InternationalJournalofControl56(2),399–413.

Noguchi,K.,Terao,H.,1997.Pathplanningofanagriculturalmobilerobotbyneural

networkandgeneticalgorithm.ComputersandElectronicsinAgriculture18,187–204.

Noguchi,N.,Reid,J.F.,Zhang,Q.,Tian,L.F.,1998.Visionintelligenceforprecision

farmingusingfuzzylogicoptimizedgeneticalgorithmandartiﬁcialneuralnet-work.AmericanSocietyofAgriculturalEngineers,St.Joseph,MI,ASAEpapernumber:983034.

Odhiambo,L.O.,Yoder,R.E.,Yoder,D.C.,2001a.Estimationofreferencecropevapo-transpirationusingfuzzystatemodels.TransactionsoftheASAE44(3),543–550.

Odhiambo,L.O.,Yoder,R.E.,Yoder,D.C.,Hines,J.W.,2001b.Optimizationoffuzzy

evapotranspirationmodelthroughneuraltrainingwithinput–outputexamples.TransactionsoftheASAE44(6),1625–1633.

Odhiambo,L.O.,Freeland,R.S.,Yoder,R.E.,Hines,J.W.,2004.Investigationof

afuzzy–neuralnetworkapplicationinclassiﬁcationofsoilsusingground-penetratingradarimagery.AppliedEngineeringinAgriculture20(1),109–117.

Oliveira,F.R.,Pacheco,D.F.,Leonel,A.,Neto,F.B.,2006.Intelligentsupportdecision

insugarcaneharvest.In:Proceedingsofthe4thWorldCongressConferenceonComputersinAgricultureandNaturalResources,Orlando,FL,pp.456–462.Onaran,I.,Pearson,T.C.,Yardimci,Y.,Cetin,A.E.,2006.Detectionofunderdeveloped

hazenutsfromfullydevelopednutsbyimpactacoustics.TransactionsoftheASABE49(6),1971–1976.

Oommen,T.,Misra,D.,Agarwal,A.,Mishra,S.K.,2007.Analysisandapplicationof

supportvectormachinebasedsimulationforrunoffandsedimentyield.Ameri-canSocietyofAgriculturalandBiologicalEngineers,St.Joseph,MI,ASABEpapernumber:073019.

Ortega,G.,Giron-Sierra,J.M.,1995.Geneticalgorithmsforfuzzycontrolofautomatic

dockingwithaspacestation.In:ProceedingsofIEEEInternationalConferenceonEvolutionaryComputation,vol.1,pp.157–161.

Ortiz,B.V,Perry,C.,Sullivan,D.G.,Kemerait,B.,Ziehl,A.,Davis,R.,Vellidis,G.,Rucker,

K.,2008.Cottonyieldresponsetovariableratenematicidesaccordingtoriskzones.AmericanSocietyofAgriculturalandBiologicalEngineers,St.Joseph,MI,ASABEpapernumber:081026.

Ovaska,S.J.,Vanlandingham,H.F.,Kamiya,A.,2002.Fusionofsoftcomputingand

hardcomputinginindustrialapplications:anoverview.IEEETransactionsonSystems,ManandCybernetics32(2),72–79.

Ovaska,S.J.,2005.ComputationallyIntelligentHybridSystems:TheFusionofSoft

ComputingandHardComputing.Wiley-IEEEPress:JohnWiley&Sons,Inc.andtheIEEEPress.

Ovaska,S.J.,Kamiya,A.,Chen,Y.Q.,2006.Fusionofsoftcomputingandhardcom-puting:computationalstructuresandcharacteristicfeatures.IEEETransactionsonSystems,Man,andCybernetics,PartC:ApplicationsandReviews36(3),439–448.

Pabico,J.P.,Hoogenboom,G.,McClendon,R.W.,1999.Determinationofcultivarcoef-ﬁcientsofcropmodelsusingageneticalgorithm:aconceptualframework.TransactionsoftheASAE42(1),223–232.

Pachepsky,Y.,Acock,B.,1998.Stochasticimagingofsoilparameterstoassessvari-abilityanduncertaintyofcropyieldestimates.Geoderma85(2–3),213–229.Parasuraman,K.,Elshorbagy,A.,Si,B.C.,2007.Estimatingsaturatedhydrauliccon-ductivityusinggeneticprogramming.SoilScienceSocietyofAmericaJournal71,1676–1684.

Pardo,M.,Sberveglieri,G.,2005.Classiﬁcationofelectronicnosedatawithsupport

vectormachines.SensorsandActuatorsB107(2005),730–737.

Pearl,J.,1988.ProbabilisticReasoninginIntelligentSystems:NetworksofPlausible

Inference.MorganKaufmann,SanFrancisco,CA.

Pearl,J.,1999.BayesianNetworks.TechnicalReport(R-246),RevisionI,July1997,

MITEncyclopediaoftheCognitiveSciences,Cambridge,MA.

Pearson,T.C.,Wicklow,D.T.,2006.Detectionofcornkernelsinfectedbyfungi.Trans-actionsoftheASABE49(4),1235–1245.

Peng,Y.,Wang,W.,2008.Predictionofporkmeattotalviablebacteriacountusing

hyperspectralimagingsystemandsupportvectormachines.In:ProceedingsoftheFoodProcessingAutomationConference,Providence,RI,PublicationNum-ber701P0508cd.

Pierna,J.A.F.,Volery,P.,Besson,R.,Baeten,V.,Dardenne,P.,2005.Classiﬁcationof

modiﬁedstarchesbyFouriertransforminfraredspectroscopyusingsupportvectormachines.JournalofAgriculturalandFoodChemistry53(17),6581–6585.Pierna,J.A.F.,Baeten,V.,Dardenne,P.,2006.Screeningofcompoundfeedsusing

NIRhyperspectraldata.ChemeometricsandIntelligentLaboratorySystems84,114–118.

Potter,W.D.,Bi.,W.,Twardus,D.,Thistle,H.W.,Ghent,J.,Twery,M.,Teske,M.E.,2000.

Ageneticalgorithmforaerialsprayapplicationoptimization.AmericanSocietyofAgriculturalEngineers,St.Joseph,MI,ASAEpapernumber:001053.

Pydipati,Y.,Burks,T.F.,Lee,W.S.,2005.Statisticalandneuralnetworkclassiﬁersfor

citrusdiseasedetectionusingmachinevision.TransactionsoftheASAE48(5),2007–2014.

Qi,F.,Zhu,A.,Harrower,M.,Burt,J.E.,2006.Fuzzysoilmappingbasedonprototype

categorytheory.Geoderma136(3–4),774–787.

Qi,L.,Ma,X.2009.RiceBlastDetectionUsingMultispectralImagingSensorandSup-portVectorMachine.AmericanSocietyofAgriculturalandBiologicalEngineers,St.Joseph,MI,ASAEpapernumber:0951.

Qu,G.,Feddes,J.J.R.,Armstrong,W.W.,Coleman,R.N.,Leonard,J.J.,2001.Measuring

odorconcentrationwithanelectronicnose.TransactionsoftheASAE44(6),1807–1812.

Raju,K.S.,Kumar,D.N.,2004.Irrigationplanningusinggeneticalgorithms.Water

ResourcesResearch18(2),163–176.

Raju,K.S.,Kumar,D.N.,Duckstein,L.,2006.Artiﬁcialneuralnetworksandmulti-criterionanalysisforsustainableirrigationplanning.Computers&OperationsResearch33,138–153.

Rechenberg,I.,1973.Evolutionstrategie:OptimierungTechnisherSystemenach

PrinzipiendesBiologischenEvolution.Fromman-HozlboogVerlag,Stuttgart,Germany.

Rosenblatt,F.,1958.ThePerceptron:aprobabilisticmodelforinformationstorage

andorganizationinthebrain.PsychologicalReview65,386–408.

Rumelhart,D.E.,McClelland,J.L.,1986.ParallelDistributedProcessing:Explorations

intheMicrostructuresofCognition,vol.I.MITPress,Cambridge,MA.

126Y.Huangetal./ComputersandElectronicsinAgriculture71(2010)107–127

Rumelhart,D.E.,Hinton,G.E.,Williams,R.J.,1986a.Learninginternalrepresenta-tionsbyerrorpropagation.In:Rumelhart,D.E.,McClelland,J.L.(Eds.),ParallelDistributedProcessing:ExplorationsintheMicrostructuresofCognition,vol.I.MITPress,Cambridge,MA(Chapter8).

Rumelhart,D.E.,Hinton,G.E.,Williams,R.J.,1986b.Learningrepresentationsby

back-propagatingerrors.Nature323,533–536.

Santhanam,S.,Langari,R.,1994.Supervisoryfuzzyadaptivecontrolofabinarydis-tillationcolumn.In:ProceedingsoftheThirdIEEEInternationalConferenceonFuzzySystems,vol.2,Orlando,FL,pp.1063–1068.

Sayde,C.,Khoury,L.,Gitelman,A.,English,M.,2008.Optimizingestimatesofsoil

moistureforirrigationscheduling.AmericanSocietyofAgriculturalandBiolog-icalEngineers,St.Joseph,MI,ASABEpapernumber:084699.

Schaap,M.G.,Leij,F.J.,VanGenuchten,M.T.,1998.Neuralnetworkanalysisofhier-archicalpredictionofsoilhydraulicproperties.SSSAJournal62(4),847–855.Schwefel,H.P.,1981.NumericalOptimizationofComputerModel.JohnWiley&

Sons,NewYork.

Shao,Y.,Zhao,C.,He,Y.,Bao,Y.,2009.Applicationofinfraredspectroscopytech-niqueandchemometricsformeasurementofcomponentsinriceafterradiation.AppliedEngineeringinAgriculture52(1),187–192.

Shapiro,J.,1998.Geneticalgorithmsinmachinelearning.http://citeseerx.

ist.psu.edu/viewdoc/summary?doi=10.1.1.22.5158.

Shi,X.,Zhu,A.,Burt,J.E.,Qi,F.,Simonson,D.,2004.Acase-basedreasoning

approachtofuzzysoilmapping.SoilScienceSocietyofAmericaJournal68,885–4.

Sick,B.,Ovaska,S.J.,2007.Fusionofsoftandhardcomputing:multi-dimensional

categorizationofcomputationallyintelligenthybridsystems.NeuralComputingandApplications16(2),125–137.

Simpson,P.K.,Jahns,G.,1993.Fuzzymin–maxneuralnetworksforfunction

approximation.In:Proc.IEEEInt.Conf.onNeuralNetworks,vol.3,pp.1967–1972.

Smith,E.T.,1993.WhytheJapanesearegoinginforthis‘fuzzylogic’?BusinessWeek,

February20,39.

Sui,R.,Thomasson,J.A.,2006.Ground-basedsensingsystemforcottonnitrogen

statusdetermination.TransactionsoftheASABE49(6),1983–1991.

Sun,T.,Ying,Y.,Liu,K.,Xu,H.,2008.Comparisonofchemometricsmethodsfor

assessinginternalqualityofpearson-lineusingvisible/nearinfraredtransmis-siontechnique.In:ProceedingsoftheFoodProcessingAutomationConference,Providence,RI,PublicationNumber:701P0508cd.

Suzuki,Y.,Okamoto,H.,Kataoka,T.,2009.Developmentofdiscriminantmodelfor

weeddetectionusinghyperspectralimagery.In:Abrigo,L.G.,Ehsani,R.(Eds.),ISHSActaHorticulturae824:InternationalSymposiumonApplicationofPreci-sionAgricultureforFruitsandVegetables.InternationalSocietyforHorticulturalScience,pp.67–73.

Takagi,T.,Hayashi,I.,1991.NN-drivenfuzzyreasoning.InternationalJournalof

ApproximateReasoning5(3),191–212.

Tang,L.,Tian,L.,Steward,B.L.,2000.Colorimagesegmentationwithgeneticalgo-rithmforin-ﬁeldweedsensing.TransactionsoftheASAE43(4),1019–1027.Tang,L.,Tian,L.,Steward,B.L.,2003.Classiﬁcationofbroadleafandgrassweeds

usingGaborwaveletsandanartiﬁcialneuralnetwork.TransactionsoftheASAE46(4),1247–1254.

Tani,T.,Utashiro,M.,Umano,M.,Tanaka,K.,1994.Applicationofpracticalfuzzy-PIDhybridcontrolsystemtopetrochemicalplant.In:ProceedingsofThirdIEEEInternationalConferenceonFuzzySystems,vol.2,Orlando,FL,pp.1211–1216.Teorey,T.J.,1999.DatabaseModelingandDesign.MorganKaufmannPublishers,San

Francisco,CA.

Terano,T.,Asai,K.,Sugeno,M.,1994.AppliedFuzzySystems.AcademicPress,Inc.,

Boston,MA.

Thomson,S.J.,Peart,R.M.,Mishoe,J.W.,1993.Parameteradjustmenttoacropmodel

usingasensor-baseddecisionsupportsystem.TransactionsoftheASAE36(1),205–213.

Thomson,S.J.,Ross,B.B.,1996.Dynamicparameteradjustmentmethodforamodel-basedirrigationmanagementsystem.ComputersandElectronicsinAgriculture14,269–290.

Thomson,S.J.,1998.Expertsystemsforself-adjustingprocesssimulation.In:Curry,

R.B.,Peart,R.M.(Eds.),AgriculturalSystemsModelingandSimulation.MarcelDekker,NewYork,pp.157–195.

Thorp,K.R.,Tian,L.,Yao,H.,Tang,L.,2004.Narrow-bandandderivative-based

vegetationindicesforhyperspectraldata.TransactionsoftheASAE47(1),291–299.

Tian,L.F.,Slaughter,D.C.,1998.Environmentallyadaptivesegmentationalgorithm

foroutdoorimagesegmentation.ComputersandElectronicsinAgriculture21(3),153–168.

Tian,Y.,Zhang,C.,Li,C.,2004.Studyonplantdiseaserecognitionusingsupport

vectormachineandchromaticitymoments.TransactionsofChineseSocietyofAgriculturalMachinery35(3),95–98.

Tikk,D.,Koczy,L.T.,Gedeon,T.D.,2003.Asurveyonuniversalapproximationand

itslimitsinsoftcomputingtechniques.InternationalJournalofApproximateReasoning33(2),185–202.

Trebar,M.,Steele,M.,2008.ApplicationofdistributedSVMarchitecturesinclas-sifyingforestdatacovertypes.ComputersElectronicsinAgriculture63(2),119–130.

Tumbo,S.D.,Wagner,D.G.,Heinemann,P.H.,2002a.Hyperspectral-basedneuralnet-workforpredictingchlorophyllstatusincorn.TransactionsoftheASAE45(3),825–832.

Tumbo,S.D.,Wagner,D.G.,Heinemann,P.H.,2002b.On-the-gosensingofchloro-phyllstatusincorn.TransactionsoftheASAE45(4),1207–1215.

Twarakavi,N.K.C.,Simunek,J.,Schaap,M.G.,2009.Developmentofpedotrans-ferfunctionsforestimationofsoilhydraulicparametersusingsupportvectormachines.SoilScienceSocietyofAmericaJournal73,1443–1452.

Uno,Y.,Prasher,S.O.,Lacroix,R.,Goel,P.K.,Karimi,Y.,Viau,A.,Patel,R.M.,2005.

Artiﬁcialneuralnetworkstopredictcornyieldfromcompactairbornespec-trographicimagerdata.ComputersandElectronicsinAgriculture47(2),149–161.

VanAlphen,B.J.,Stoorvogel,J.J.,2000.Afunctionalapproachtosoilcharacterization

insupportofprecisionagriculture.SoilScienceSocietyofAmericaJournal,1706–1713.

Wall,M.B.,1996.AGeneticAlgorithmforResource-ConstrainedScheduling.Ph.D.

Dissertation.MIT,Cambridge,MA.

Walthall,C.,Dulaney,W.,Anderson,M.,Norman,J.,Fang,H.,Liang,S.,2004.Acom-parisonofempiricalandneuralnetworkapproachesforestimatingcornandsoybeanleafareaindexfromLandsatETM+imagery.RemoteSensingofEnvi-ronment92,465–474.

Wang,W.,Paliwal,J.,2006.Spectraldatacompressionandanalysestech-niquestodiscriminatewheatclasses.TransactionsoftheASABE49(5),1607–1612.

Wardlaw,R.,Bhaktikul,K.,2004.Applicationofgeneticalgorithmsforirrigation

waterscheduling.IrrigationandDrainage53,397–414.

Werbos,P.J.,1974.BeyondRegression:NewToolsforPredictionandAnalysisin

theBehavioralSciences.DoctoralDissertation.AppliedMathematics,HarvardUniversity,Boston,MA.

Whittaker,A.D.,Park,B.S.,McCauley,J.D.,Huang,Y.,1991.Ultrasonicsignalclas-siﬁcationforbeefqualitygradingthroughneuralnetworks.In:AutomatedAgricultureforthe21stCentury—Proc.1991Symp.ASAE,St.Joseph,MI,pp.116–125.

Wilkinson,R.H.,1963.Amethodofgeneratingfunctionsofseveralvariables

usinganalogdiodelogic.IEEETransactionsonElectronicComputers12,112–129.

Witten,I.H.,Frank,E.,2000.DataMining:PracticalMachineLearningToolsand

TechniqueswithJAVAImplementations.MorganKaufmannPublishers,SanFrancisco,CA.

Wu,D.,Feng,L.,He,Y.,Bao,Y.,2008.VarietyidentiﬁcationofChinesecabbageseeds

usingvisibleandnear-infraredspectroscopy.TransactionsoftheASABE51(6),2193–2199.

Xiang,H.,Tian,L.F.,2007.Artiﬁcialintelligencecontrollerforautomaticmul-tispectralcameraparameteradjustment.TransactionsoftheASABE50(5),1873–1881.

Yang,C.C.,Prasher,S.O.,Mehuys,G.R.,Patni,N.K.,1997a.Applicationofartiﬁcial

neuralnetworksforsimulationofsoiltemperature.TransactionsoftheASAE40(3),9–656.

Yang,C.C.,Prasher,S.O.,Sreekanth,S.,Patni,N.K.,Masse,L.,1997b.Anartiﬁcialneural

networkmodelforsimulatingpesticideconcentrationsinsoil.TransactionsoftheASAE40(5),1285–1294.

Yang,C.C.,Prasher,S.O.,Landry,J.A.,Perret,J.,Ramaswamy,H.S.,2000a.Recognition

ofweedswithimageprocessingandtheirusewithfuzzylogicforprecisionfarming.CanadianAgriculturalEngineering42(4),195–200.

Yang,C.C.,Prasher,S.O.,Landry,J.A.,Ramaswamy,H.S.,Ditommaso,A.,2000b.Appli-cationofartiﬁcialneuralnetworksinimagerecognitionandclassiﬁcationofcropandweeds.CanadianAgriculturalEngineering42(3),147–152.

Yang,C.C.,Prasher,S.O.,Whalen,J.,Goel,P.K.,2001.Applicationofdatamining

technologyforhyperspectralimageryclassiﬁcationinagriculturalﬁelds.Amer-icanSocietyofAgriculturalEngineers,St.Joseph,MI,ASAEpapernumber:013116.

Yang,C.C.,Prasher,S.O.,Landry,J.A.,Ramaswamy,H.S.,2002.Developmentofneural

networksforweedrecognitionincornﬁelds.TransactionsoftheASAE45(3),859–8.

Yang,C.C.,Prasher,S.O.,Landry,J.,Ramaswamy,H.S.,2003.Developmentofanimage

processingsystemandafuzzyalgorithmforsite-speciﬁcherbicideapplications.PrecisionAgriculture4(1),5–18.

Yang,C.C.,Prasher,S.O.,Goel,P.K.,2004a.Differentiationofcropandweedsby

decisiontreeanalysisofmulti-spectraldata.TransactionsoftheASAE47(3),873–879.

Yang,C.C.,Prasher,S.O.,Lacroix,R.,Kim,S.H.,2004b.Applicationofmultivariate

adaptiveregressionsplines(MARS)tosimulatesoiltemperature.TransactionsoftheASAE47(3),881–887.

Yao,H.,Tian,L.,2003.Ageneticalgorithm-basedselectiveprincipalcomponent

analysis(GA-SPCA)methodforhighdimensionaldatafeatureextrac-tion.IEEETransactionsonGeoscienceandRemoteSensing41(1),1469–1478.

Yasunobu,S.,Miyamoto,S.,1985.Automatictrainoperationsystembypredictive

fuzzycontrol.In:Sugeno,M.(Ed.),IndustrialApplicationsofFuzzyControl.North-Holland,pp.1–18.

Yu,H.,Niu,X.,Ying,Y.,Pai,X.,2008.Non-invasivedeterminationofenological

parametersofricewinebyVis-NIRspectroscopyandleastsquaressupportvectormachines.AmericanSocietyofAgriculturalandBiologicalEngineers,St.Joseph,MI,ASABEpapernumber:084875.

Zadeh,L.A.,1965.Fuzzysets.InformationandControl8,338–353.

Zadeh,L.A.,1973.Outlineofanewapproachtotheanalysisofcomplexsystemsand

decisionprocesses.IEEETransactionsonSystems,Man,andCyberneticsSMC-3,28–44.

Zadeh,L.A.,1981.Possibilitytheoryandsoftdataanalysis.In:Cobb,L.,Thrall,R.M.

(Eds.),MathematicalFrontiersoftheSocialandPolicySciences.WestviewPress,Boulder,CO,pp.69–129.

Y.Huangetal./ComputersandElectronicsinAgriculture71(2010)107–127

Zadeh,L.A.,1992.Foreword.In:ProceedingsoftheSecondInternationalConference

onFuzzyLogicandNeuralNetworks,Iizuka,Japan,pp.xiii–xiv.

Zhang,Z.X.,Kushwaha,R.L.,1999.Applicationofneuralnetworkstosimulate

soil–toolinteractionandsoilbehavior.CanadianAgriculturalEngineering41(2),119–125.

Zhang,H.,Paliwal,J.,Jayas,D.S.,White,N.D.G.,2007.Classiﬁcationoffungalinfected

wheatkernelsusingnear-infraredreﬂectancehyperspectralimagingandsup-portvectormachine.TransactionsoftheASABE50(5),1779–1785.

127

Zhang,Q.,Litchﬁeld,J.B.,1992.Advancedprocesscontrols:applicationsofadaptive,

fuzzyandneuralcontroltothefoodindustry.In:FoodProcessingAutomationII.ASAE,St.Joseph,MI,pp.169–176.

Zhu,A.,Hudson,B.,Burt,J.,Lubich,K.,Simonson,D.,2001.SoilmappingusingGIS,

expertknowledge,andfuzzylogic.SoilScienceSocietyofAmericaJournal65,1463–1472.

Zhu,A.,Yang,L.,Li,B.,Qin,C.,Pei,T.,Liu,B.,inpress.Constructionofmembership

functionsforpredictivesoilmappingunderfuzzylogic.Geoderma.

因篇幅问题不能全部显示，请点此查看更多更全内容

查看全文