Support+vector+machines+in+remotesensing+A+review

ISPRSJournalofPhotogrammetryandRemoteSensing()–ContentslistsavailableatScienceDirectISPRSJournalofPhotogrammetryandRemoteSensingjournalhomepage::AreviewGiorgosMountrakis∗,JunghoIm,CaesarOgoleDepartmentofEnvironmentalResourcesEngineering,SUNYCollegeofEnvironmentalScienceandForestry,1ForestryDr,Syracuse,NY13210,USAarticleinfoArticlehistory:Received6June2010Receivedinrevisedform17September2010Accepted1November2010AvailableonlinexxxxKeywords:SupportvectormachinesReviewRemotesensingSVMSVMsabstractAwiderangeofmethodsforanalysisofairborne-andsatellite-derivedimagerycontinuestobeproposedandassessed.Inthispaper,wereviewremotesensingimplementationsofsupportvectormachines(SVMs),apromisingmachinelearningmethodology.Thisreviewistimelyduetotheexponentiallyincreasingnumberofworkspublishedinrecentyears.SVMsareparticularlyappealingintheremotesensingfieldduetotheirabilitytogeneralizewellevenwithlimitedtrainingsamples,acommonlimitationforremotesensingapplications.However,theyalsosufferfromparameterassignmentissuesthatcansignificantlyaffectobtainedresults.Asummaryofempiricalresultsisprovidedforvariousapplicationsofoveronehundredpublishedworks(asofApril,2010).ItisourhopethatthissurveywillprovideguidelinesforfutureapplicationsofSVMsandpossibleareasofalgorithmenhancement.©2010InternationalSocietyforPhotogrammetryandRemoteSensing,Inc.(ISPRS).PublishedbyElsevierB.V.Allrightsreserved.1.IntroductionRemotely-senseddataareusedinnumerousapplications.Typ-ically,animageclassificationprocessisinitiatedtoconvertdataintomeaningfulinformation.Unfortunately,imageclassificationisnotatrivialtask.AsnotedbyChietal.(2008),classificationofremotesensingdataisparticularlydauntingbecausemostofthesupervisedlearningschemesrequiresufficientlylargeamountoftrainingsamples,yetdefinitionandacquisitionofreferencedataisoftenacriticalproblem.Variousclassificationtechniques,bothparametricandnon-parametric,havebeendevelopedandusedindifferentcontexts—remotesensinginclusive.Previousreviews,suchasthatbyPlazaetal.(2009),focusedonrecentdevelopmentsinmethodologiesforprocessingaspecifictypeofimagery,forexamplehyperspectralimages.Thereviewprovidedinthispaperfollowsthealgorithmicperspectiveratherthanimagecharacteristics.Morespecifically,wefocusonapplicationsofsupportvectormachines(SVMs)inremotesensing.Themotivationtocarryoutthisstudycomesfromdifferentsources.First,SVMsarenotaswell-knownasotherclassifiers(e.g.,decisiontrees,variantsofneuralnetworks)inthegeneralremotesensingcommunity,yettheycanmatchifnotexceedtheperformanceofestablishedmethods.Second,theirperformance∗Correspondingaddress:DepartmentofEnvironmentalResourcesEngineering,SUNYCollegeofEnvironmentalScienceandForestry,419BakerHall,1ForestryDr,Syracuse,NY13210,USA.Tel.:[phone];fax:[phone].E-mailaddress:[e-mail](G.Mountrakis).URL:(G.Mountrakis).gainsseemwell-suitedforremotesensingapplications,wherealimitedamountofreferencedataisoftenprovided.Third,eventhoughthemethodisnotwidelypopular,inrecentyearstherehasbeenasignificantincreaseinSVMworksonremotesensingproblemssuggestingthisreviewiscurrentandappropriate.Thisreviewfocusesonrecentresearchpapers(availablebyApril,2010)publishedineightmajorjournalsofremotesens-ing,namely,ISPRSJournalofPhotogrammetryandRemoteSens-ing,RemoteSensingofEnvironment,PhotogrammetricEngineer-ing&RemoteSensing,IEEETransactionsonGeoscienceandRe-moteSensing,IEEEGeoscienceandRemoteSensingLetters,Inter-nationalJournalofRemoteSensing,CanadianJournalofRemoteSensingandGIScienceandRemoteSensing.Alimitednumberofresearchpapersrelevanttothethematicpointandthusincludedinthisreviewcamefromadditionalsources.Theselectedpapersrepresentawiderangeof:(i)applicationsfromcoalreservedetec-tiontourbangrowthmonitoring,(ii)resolutionsfromsub-metertoseveralkilometerspixelsize,(iii)spectralresolutionfromsingletohundredsofbands,and(iv)comparativemethodsfrommax-imumlikelihoodclassifierstoneuralnetworks.Forcompleteness,wefirstrecaponthebasicsofSVMmethodologybeforedivingintospecificworks.Relevantpapersarethensummarized,whilejuxta-positionofgeneralpatternsenablesustoderiveconclusionsandrecommendationsforfurtherinvestigations.2.OverviewofsupportvectormachinesSupportvectormachines(SVMs)isasupervisednon-parametricstatisticallearningtechnique,thereforethereisnoassumption0924-2716/$–seefrontmatter©2010InternationalSocietyforPhotogrammetryandRemoteSensing,Inc.(ISPRS).PublishedbyElsevierB.V.Allrightsreserved.doi:10.1016/j.isprsjprs.2010.11.0012G.Mountrakisetal./ISPRSJournalofPhotogrammetryandRemoteSensing()–SupportvectorsMarginwidthMisclassifiedinstancesSVMhyperplaneFig.1.Linearsupportvectormachineexample.Source:adaptedfromBurges(1998).madeontheunderlyingdatadistribution.Initsoriginalformula-tion(Vapnik,1979)themethodispresentedwithasetoflabeleddatainstancesandtheSVMtrainingalgorithmaimstofindahyper-planethatseparatesthedatasetintoadiscretepredefinednumberofclassesinafashionconsistentwiththetrainingexamples.T