摘要
Hyperspectralimageprovidesabundantspectralinformationforremotediscriminationofsubtledifferencesingroundcovers.However,theincreasingspectraldimensions,aswellastheinformationredundancy,maketheanalysisandinterpretationofhyperspectralimagesachallenge.Featureextractionisaveryimportantstepforhyperspectralimageprocessing.Featureextractionmethodsaimatreducingthedimensionofdata,whilepreservingasmuchinformationaspossible.Particularly,nonlinearfeatureextractionmethods(e.g.kernelminimumnoisefraction(KMNF)transformation)havebeenreportedtobenefitmanyapplicationsofhyperspectralremotesensing,duetotheirgoodpreservationofhigh-orderstructuresoftheoriginaldata.However,conventionalKMNForitsextensionshavesomelimitationsonnoisefractionestimationduringthefeatureextraction,andthisleadstopoorperformancesforpost-applications.Thispaperproposesanovelnonlinearfeatureextractionmethodforhyperspectralimages.Insteadofestimatingnoisefractionbythenearestneighborhoodinformation(withinaslidingwindow),theproposedmethodexplorestheuseofimagesegmentation.Theapproachbenefitsbothnoisefractionestimationandinformationpreservation,andenablesasignificantimprovementforclassification.Experimentalresultsontworealhyperspectralimagesdemonstratetheefficiencyoftheproposedmethod.ComparedtoconventionalKMNF,theimprovementsofthemethodontwohyperspectralimageclassificationare8and11％.Thisnonlinearfeatureextractionmethodcanbealsoappliedtootherdisciplineswherehigh-dimensionaldataanalysisisrequired.
出版日期
2017年04月14日（中国期刊网平台首次上网日期，不代表论文的发表时间）
