多策略结合的高光谱图像波段选择新方法

*TheNationalNaturalScienceFoundationofChinaunderGrantNo.60673141();theKeyProjectsintheNa-tionalScienceandTechnologyPillarProgramduringtheEleventhFive-yearPlanPeriodofChinaunderGrantNo.2006BAB04A13().Received2009-11,Accepted2010-01.ISSN1673-9418CODENJKYTA8E-mail:[e-mail]mputerScienceandTechnology(05)-0464-09Tel:[phone]DOI:10.3778/j.issn.1673-9418.2010.05.009*1,1+,2,11.,2100982.,100038Multiple-strategyCombinationBasedApproachtoBandSelectionforHyper-spectralImageClassification*WUHao1,LIShijin1+,LINLin2,WANDingsheng11.SchoolofComputerandInformationEngineering,HohaiUniversity,Nanjing210098,China2.NetworkInformationCenter,InstituteofWaterResourcesandHydropowerResearch,Beijing100038,China+Correspondingauthor:E-mail:[e-mail],LIShijin,LINLin,etal.Multiple-strategycombinationbasedapproachtobandselectionforhy-perspectralimageclassification.JournalofFrontiersofComputerScienceandTechnology,2010,4(5):464-472.Abstract:Withthedevelopmentandpopularizationoftheremote-sensingimagingtechnology,therearemoreandmoreapplicationsofhyperspectralimageclassificationtasks,suchastargetdetectionandlandcoverinvestigation.Itisaverychallengingissueofurgentimportancethathowtoselectaminimalandeffectivesubsetfrommassofbands.Anovelbandselectionstrategyisputforwardbasedonconditionalmutualinformationbetweenadjacentbandsandbranchandboundalgorithmforthehighcorrelationbetweenthebands.Inaddition,geneticalgorithmandsupportvectormachineareemployedtosearchforthebestbandcombination.Experimentalresultsshowthattheproposedapproachisverycompetitiveandrobust.Keywords:hyperspectralremotesensing;bandselection;conditionalmutualinformation;adaptivebranchandboundalgorithm;supportvectormachine;geneticalgorithm465,,,,;;;;;ATP391.411,,[1],,,,,,,,10nm,[2],,,filterwrapper3[3],;filter,,filter,,,,,,,Wrapper[4],,,filter,,,,,,,(geneticalgorithmandsupportvectormachine,GA-SVM),,1,GA-SVM,,,,,GA-SVM,466JournalofFrontiersofComputerScienceandTechnology2010,4(5)Fig.1StructureofthehyperspectralimageclassificationalgorithmbasedonGA-SVMandAB&B1,,,,(adaptivebranchandbound,AB&B),,,[5],,,,[6],,[7]2;3(GA-SVM);4(AB&B);5WashingtonDCMall,;,22.1,,[8],X,p(x),x∈,X[8]()()log2()xHXpxpxΦ∈=−∑(1)XY,p(x,y),x∈,y∈,XY[8](,)(,)log2(,)xΦyΨHXYpxypxy∈∈=−∑∑(2)YX(|)(,)log2(|)xΦyΨHXYpxypxy∈∈=−∑∑(3),XY[9](,)()()(,)()(|)IXYHXHYHXYHXHXY=+−=−(4)XY,C,YCX(,|)(|)(|,)(,)(,,)(,)[(,)(,|)]ICXYHXYHXCYICYICXYICXIXYIXYC=−=−=−−(5)2.2,467,,,,,,,,,,,,HYDICEWashingtonDCMall,191,22,,,1,,,1823,Fig.2ConditionalmutualinformationofWashingtonDCMalldataset2WashingtonDCMall2,3,,,,,N,K,,Table1Originalbandgroupingresultbasedonconditionalmutualinformation1WashingtonDCMall123456781~3738~5657~7273~8687~102103~133134~140141~191Table2Bandgroupingresultwithadditionalgroupsbasedonconditionalmutualinformation2WashingtonDCMall12345678910111~1819~3738~5657~7273~8687~102103~133134~140141~158159~175176~191468JournalofFrontiersofComputerScienceandTechnology2010,4(5),,,,,,K,,,30,0.2,0.8,100,1,RBF14(branchandbound,BB),,,,,BB(),,,,BB,BB,GA-SVM,,[10](AB&B)GA-SVM,HYDICEWashingtonDCMall,(AB&B)1~8,3Fig.3Fitnessofthebandcombinationafterpruning3,,7,,72911,8(3)2,,8,Table3Bandgroupingresultafterpruning3WashingtonDCMall12345671~3738~5657~7273~8687~102103~133134~191469,,9,,35HYDICEWashingtonDCMall0.40~2.40μm,210,,1917,[11]137,,,,45.11,:,191(Non-Sel);,[1](GS)535,(,18,CGGS),,GS,2191−1,CGGS2.7×1010,GS1047,,CGGS5.22(AB&B),811(12)7(3),GA-SVM,310,6Table4Numberoftrainingandtestsamplesfortheexperiments41234567184918472111413699974198523210381426922523Table5Classificationaccuracyofthreegeneticalgorithms53/(%)/(%)/(%)Non-Sel19193.4093.4093.40GS24~2888.2895.0492.14CGGS895.7298.5797.26470JournalofFrontiersofComputerScienceandTechnology2010,4(5)Table6Classificationaccuracyofthreegroupingnumbers63/(%)/(%)/(%)1895.7298.5797.2621194.9996.5596.02797.3798.7998.25Table7Classificationaccuracyofdifferentalgorithms7/(%)/(%)/(%)11996.8096.8096.8023893.4893.4893.48797.3798.7998.256,,GA-SVM,,GA-SVM,5.33,(1911938,105,),,(7),,[12],,,(NWFE)2614,98.9%98.5%[12]798.25%,6,(AB&B)GA-SVM,,,,,GA-SVM,,,,GA-SVM,471DavidLandgrebeWashingtonDCMallReferences:[1]ZuoLi,ZhengJing,WangFang,etal,Ageneticalgo-rithmbasedwrapperfeatureselectionmethodforclassi-ficationofhyperspectraldatausingsupportvectorma-chine[J].GeographicalResearch,2008,27(3):493−501.[2]LiuChunhong,ZhaoChunhui,ZhangYanling.Anewmethodofhyperspectralremotesensingimagedimen-sionalreduction[J].JournalofImageandGraphics,2005,10(2):218−224.[3]ZhaoChunhui,LiuChunhong.Researchandanalysisofhyperspectralremotesensingimagedimensionalreduc-tion[J].ChineseSpaceScienceandTechnology,2004,10(5):28−36.[4]ZhangLixin,WangJiaqin,ZhaoYannan,etal,Featureselectioninmachinelearning[J].ComputerScience,2004,31(11):180−184.[5]SunLinxin,GaoWen.Selectiontheoptimalclassificationbandsbasedonroughsets[J].PatternRecognitionandArtificialIntelligence,2000,13(2):181−186.[6]WangLiguo,GuYanfeng,ZhangYe.Bandselectionmethodbasedoncombinationofsupportvectormachinesandsubspatialpartiti