遥感专业外文翻译高光谱遥感信息中的特征提取与应用研究.docx

遥感专业外文翻译高光谱遥感信息中的特征提取与应用研究.docx

《遥感专业外文翻译高光谱遥感信息中的特征提取与应用研究.docx》由会员分享，可在线阅读，更多相关《遥感专业外文翻译高光谱遥感信息中的特征提取与应用研究.docx（15页珍藏版）》请在冰点文库上搜索。

遥感专业外文翻译高光谱遥感信息中的特征提取与应用研究

2500单词，3900汉字

出处：

DuP,TaoF,HongT.SpectralFeaturesExtractioninHyperspectralRSDataandItsApplicationtoInformationProcessing[J].ActaPhotonicaSinica,2005,34

（2）:

293-298.

本科毕业设计（论文）

中英文对照翻译

院（系部）测绘与国土信息工程学院

专业名称测绘工程　

年级班级　

学生姓名　

指导老师

2012年6月3日

SpectralFeaturesExtractioninHyperspectralRSDataand

ItsApplicationtoInformationProcessing

OrientedtothedemandsofhyperspectralRSinformationprocessingandapplications,spectralfeaturesinhyperspectralRSimagecanbecategorizedintothreescales:

pointscale,blockscaleandvolumescale.Basedonthepropertiesandalgorithmsofdifferentfeatures,itisproposedthatpointscalefeaturescanbedividedintothreelevels:

spectralcurvefeatures,spectraltransformationfeaturesandspectralsimilaritymeasurefeatures.Spectralcurvefeaturesincludedirectspectraencoding,reflectionandabsorptionfeatures.SpectraltransformationfeaturesincludeNormalizedDifferenceofVegetationIndex（NDVI）,derivatespectraandotherspectralcomputationfeatures.Spectralsimilaritymeasurefeaturesincludespectralangle（SA）,SpectralInformationDivergence（SID）,spectraldistance,correlationcoefficientandsoon.Basedonanalysistothosealgorithms,severalproblemsaboutfeatureextraction,matchingandapplicationarediscussedfurther,anditprovedthatquaternaryencoding,spectralangleandSIDcanbeusedtoinformationprocessingeffectively.

1Introduction

HyperspectralRemoteSensingwasoneofthemostimportantbreakthroughsofEarthObservationSystem（EOS）in1990s.ItovercomesthelimitationsofconventionalaerialandmultispectralRSsuchaslessbandamount,widebandscopeandroughspectralinformationexpression,andcanprovideRSinformationwithnarrowbandwidth,morebandamountandfinespectralinformation,alsoitcandistinguishandidentifygroundobjectsfromspectralspace,sohyperspectralRShasgotwideapplicationsinresources,environment,cityandecologicalfields.BecausehyperspectralRSisdifferentfromconventionalRSinformationobviouslyinbothinformationacquisitionandinformationprocessing,therearemanyproblemsshouldbesolvedinpractice.OneofthemostimportantproblemsisaboutspectralfeaturesextractionandapplicationinhyperspectralRSdataincludinghyperspectralRSimageandstandardspectraldatabase.Nowadays,studiesonhyperspectralaremainlyfocusedonbandselectionanddimensionalityreduction,imageclassification,mixedpixeldecompositionandothers,andstudiesonspectralfeaturesarefew.Inthispaper,spectralfeaturesextractionandapplicationwillbetakenasourcentraltopicinordertoprovidesomeusefuladvicestohyperspectralRSapplications.

2FrameworkofspectralfeaturesinhyperspectralRSdata

Ingeneral,hyperspectralRSimagecanbeexpressedbyaspatial-spectraldatacube（Fig.1）.Inthisdatacube,everycoverageexpressedtheimageofoneband,andeachpixelformsaspectralvectorcomposedofalbedoofgroundobjectoneverybandinspectraldimension,andthatvectorcanbevisualizedbyspectralcurve（Fig.2）.Manyfeaturescanbeextractedfromspectralvectororcurve,andspectralfeaturesarethekeyandbasisofhyperspectralRSapplications.Alsoeachspectralcurveinspectraldatabasecanbeanalyzedwithsamemethod.Althoughtherearesomealgorithmstocomputespectralfeatures,theframeworkandsystemisstillnotobvious,sowewouldliketoproposeaframeworkforspectralfeaturesinhyperspectralRSdataincludinghyperspectralRSimageandstandardspectraldatabase.

Fig.1　HyperspectralimagedatacubeFig.2　Reflectancespectralcurveofapixel

2.1　Threescalesofspectralfeatures

Accordingtotheoperationalobjectsofextractionalgorithms,spectralfeaturescanbecategorizedintothreescales:

point-scale,block-scaleandvolume-

Scale.

Pointscaletakespixelanditsspectralcurveasoperationalobjectandsomeusefulfeaturescanbeextractedfromthisspectralvector（orspectralcurve）.Ingeneral,hyperspectralRSimagetakesspectralvectorofeachpixelasprocessingobject.

Blockscaleisorientedimageblockorregion.Blockisthesetofsomepixels,anditcanbehomogeneousorheterogeneous.Homogeneousregionsaregotbyimagesegmentationandpixelsinthisregionaresimilarinsomegivenfeatures;heterogeneousregionarethoseimageblockswithregularorirregularsize,andtheyarecutfromoriginalimagedirectly,forexample,animagecanbesegmentedaccordingtoquadtreemethod.InhyperspectralRSimage,blockscalefeaturescanbecomputedfromtwoaspects.Oneistocomputetexturefeatureofablockonsomecharacterizedbands,andtheotheristocomputespectralfeatureofablock.Iftheblockishomogeneousitsmeanvectorcanbecomputedfirstlyandthenspectralofthismeanvectorcanbeextractedtodescribetheblock.Iftheblockisheterogeneous,itcanbesegmentedtosomehomogeneousblocks.

Volumescalecombinesspatialandspectralfeaturesinawholeandextractsfeaturesin3D（row,columnandspectra）space.Here,some3Doperationalalgorithmsareneeded,forexample,3DwavelettransformationandhighorderArtificialNeuralNetwork（ANN）.Becausethistypeoffeaturesisdifficulttocomputeandanalyze,wedon′tresearchitincurrentstudies.

Inthispaper,wewouldliketofocusonpointscalefeature,orthosefeaturesextractedfromspectralvectorthatmaybespectralvectorofapixelormeanvectorofablock.

2.2　Threelevelsofpointscalefeatures

Fromoperationobject,algorithmprinciples,featureproperties,applicationmodesandotheraspects,wethinkitisfeasibletocategorizespectralfeaturesintothreelevels:

spectralcurvefeatures,spectraltransformationfeaturesandspectralsimilaritymeasurefeatures.Theyarecorrespondingtoanalysisonspectralcurvewithallbands,datatransformationandcombinationwithpartofallbandsandsimilaritymeasureofspectralvectors.Inourstudy,datafromOMISandPHIhyperspectralimage,USGSspectraldatabaseandtypicalspectradatainChinaisexperimentedandtwoexamplesaregiveninthispaper.OneistoselectthreeregionsfromPHIimage（RegionIisvegetation,RegionIIisbuilt-upland,andRegionIIIismixedregionofsomelandcovers）,andtheotherisspectralcurveofthreegroundobjectsfromUSGSspectraldatabase,amongthemS1isActinolite_HS22.3B,S2isActinolite_HS116.3BandS3isAlbite_HS66.3B,soS1andS2aresimilarandtheyaredifferentfromS3.

3　Spectralcurvefeatures

Spectralcurvefeaturesarecomputedbysomealgorithmsbasedonthespectralcurveofcertainpixelorgroundobject,anditcandescribeshapeandpropertiesofthecurve.Themainmethodsincludedirectencodingandfeaturebandanalysis.

3.1　Directencoding

Theimportantideaofspectralcurvefeatureistoemphasizespectralcurveshape,sodirectencodingisaveryconvenientmethod,andbinaryencodingisusedmorewidely.Itsprincipleistocomparetheattributevalueateachbandofapixelwithathresholdandassignthecodeof“0”or“1”accordingtoitsvalue.Thatcanbeexpressedby

Here,

iscodeoftheithband,

istheoriginalattributevalueofthisband,andTisthethreshold.Generally,thresholdisthemeanofspectralvector,anditcanalsobeselectedbymanualmethodaccordingtocurveshape,sometimesmedianofspectralvectorisprobablyused.

Onlyonethresholdisusedinbinaryencoding,sothedividedinternalislargeandprecisionislow.Inordertoimprovetheappoximatyandprecision,thequaternaryencodingstrategyisproposedinthispaper.Itsprimaryideaisasfollows:

（1）themeanofthetotalpixelspectralvectoriscomputedanddenotedbyT0,andtheattributeisdividedintotwointernalincluding[

]and[

];

（2）thepixelslocatedinthetwointernalsaredeterminedandthemeanofeachinternalisgotanddonatedby

andTR,sofourinternalsareformedincluding[

TL],[

TR]and[TR,

];（3）eachbandisassignedoneofthecodesets{0,1,2,3}accordingtotheinternalitislocated;（4）tocomputetheratioofmatchedbandsnumbertothetotalbandnumberasfinalmatchingratio.Itprovedthatquaternaryencodingcoulddescribethecurveshapemoreprecisely.

　Ifquarternaryencodingisused,theratioofthesameregionissmallerthanbinaryencoding,buttheratiobetweendifferentregionsdecreaseddramatically.Soquarternaryencodingismoreeffectiveinmeasuringthesimilaritybetweendifferentpixels.

Becausedirectencodingwilldispersethecontinuousalbedointodiscretecode,theencodingresultisaffectedbythresholdobviouslyandwillleadtoinformationloss.Althoughitsoperationisverysimple,itisonlyusedtosomeapplicationsrequiringlowprecision,andthethresholdshouldbeselectedaccordingtodifferentconditions.

3.2　Spectralabsorptionorreflectionfeature

Differingfromdirectencodinginwhichallbandsareused,spectralabsorptionorreflectionfeatureonlyemphasizesthosebandswherevalleysorapexesarelocated.Thatmeansthosebandswithlocalmaximumorminimuminspectralcurveshouldbedeterminedatfirstandthenfurtheranalysiscanbedone.Ingeneral,albedoisusedtodescribetheattributeofapixel,sothosebandswithlocalmaximumarereflectionapexandthosewithlocalminimumareabsorptionvalley.

Afterthelocationandrelatedparametersaregot,thedetailanalysiscanbedone.Ingeneraltwomethodsareused,oneistogivedirectencodingandanalysistofeaturebands,andtheotheristocomputesomequantitativeindexusingfeaturebandsandtheirparameters.

3.3　Encodingofspectralabsorptionorreflect