Table Of ContentHYPERSPECTRAL DATA
PROCESSING
HYPERSPECTRAL DATA
PROCESSING
Algorithm Design and Analysis
Chein-IChang
UniversityofMaryland,BaltimoreCounty(UMBC),Maryland,USA
Copyright#2013byJohnWiley&Sons,Inc.Allrightsreserved
PublishedbyJohnWiley&Sons,Inc.,Hoboken,NJ
PublishedsimultaneouslyinCanada
Nopartofthispublicationmaybereproduced,storedinaretrievalsystem,ortransmittedinanyformorbyanymeans,
electronic,mechanical,photocopying,recording,scanning,orotherwise,exceptaspermittedunderSection107or108of
the1976UnitedStatesCopyrightAct,withouteitherthepriorwrittenpermissionofthePublisher,orauthorization
throughpaymentoftheappropriateper-copyfeetotheCopyrightClearanceCenter,Inc.,222RosewoodDrive,Danvers,
MA 01923, (978) 750-8400, fax (978) 750-4470, or on the web at www.copyright.com. Requests to the Publisher for
permissionshouldbeaddressedtothePermissionsDepartment,JohnWiley&Sons,Inc.,111RiverStreet,Hoboken,NJ
07030, (201) 748-6011, fax (201) 748-6008, or online at http://www.wiley.com/go/permission.
LimitofLiability/DisclaimerofWarranty:Whilethepublisherandauthorhaveusedtheirbesteffortsinpreparingthisbook,
theymakenorepresentationsorwarrantieswithrespecttotheaccuracyorcompletenessofthecontentsofthisbookand
specificallydisclaimanyimpliedwarrantiesofmerchantabilityorfitnessforaparticularpurpose. Nowarrantymaybe
createdorextendedbysalesrepresentativesorwrittensalesmaterials.Theadviceandstrategiescontainedhereinmaynot
besuitableforyoursituation.Youshouldconsultwithaprofessionalwhereappropriate.Neitherthepublishernorauthor
shallbeliableforanylossofprofitoranyothercommercialdamages,includingbutnotlimitedtospecial,incidental,
consequential,orotherdamages.
Forgeneralinformationonourotherproductsandservicesorfortechnicalsupport,pleasecontactourCustomerCare
DepartmentwithintheUnitedStatesat(800)762-2974,outsidetheUnitedStatesat(317)572-3993orfax(317)572-4002.
Wileyalsopublishesitsbooksinavarietyofelectronicformats.Somecontentthatappearsinprintmaynotbeavailablein
electronic formats. For more information about Wiley products, visit our web site at www.wiley.com.
LibraryofCongressCataloging-in-PublicationData:
Chang,Chein-I.
Hyperspectraldataprocessing:algorithmdesignandanalysis/Chein-IChang.
p.cm.
Includesbibliographicalreferencesandindex.
ISBN978-0-471-69056-6(hardback)
1. Imageprocessing–Digitaltechniques. 2. Spectroscopicimaging. 3. Signalprocessing. I. Chang,Chein-I.
Hyperspectralimaging. II. Title.
TA1637.C47762012
621.39’94–dc23
2011043896
PrintedintheUnitedStatesofAmerica
10 9 8 7 6 5 4 3 2 1
Thisbookisdedicatedtomembersofmyfamily,specifically
mymotherwhoprovidedmewithhertimelesssupport
andencouragementduringthecourseofpreparing
thisbook.Itisalsodedicatedtoallofmystudents
whohavecontributedtothisbook.
Contents
PREFACE xxiii
1 OVERVIEWANDINTRODUCTION 1
1.1 Overview 2
1.2 IssuesofMultispectralandHyperspectralImageries 3
1.3 DivergenceofHyperspectralImageryfromMultispectralImagery 4
1.3.1 Misconception:HyperspectralImagingisaNaturalExtension
ofMultispectralImaging 4
1.3.2 Pigeon-HolePrinciple:NaturalInterpretationofHyperspectralImaging 5
1.4 ScopeofThisBook 7
1.5 Book’sOrganization 10
1.5.1 PartI:Preliminaries 10
1.5.2 PartII:EndmemberExtraction 12
1.5.3 PartIII:SupervisedLinearHyperspectralMixtureAnalysis 13
1.5.4 PartIV:UnsupervisedHyperspectralAnalysis 13
1.5.5 PartV:HyperspectralInformationCompression 15
1.5.6 PartVI:HyperspectralSignalCoding 16
1.5.7 PartVII:HyperspectralSignalFeatureCharacterization 17
1.5.8 Applications 17
1.5.8.1 Chapter30:ApplicationsofTargetDetection 17
1.5.8.2 Chapter31:NonlinearDimensionalityExpansiontoMultispectral
Imagery 18
1.5.8.3 Chapter32:MultispectralMagneticResonanceImaging 19
1.6 LaboratoryDatatobeUsedinThisBook 19
1.6.1 LaboratoryData 19
1.6.2 CupriteData 19
1.6.3 NIST/EPAGas-PhaseInfraredDatabase 19
1.7 RealHyperspectralImagestobeUsedinthisBook 20
1.7.1 AVIRISData 20
1.7.1.1 CupriteData 21
1.7.1.2 Purdue’sIndianaIndianPineTestSite 25
1.7.2 HYDICEData 26
1.8 NotationsandTerminologiestobeUsedinthisBook 29
vii
viii Contents
I:PRELIMINARIES 31
2 FUNDAMENTALSOFSUBSAMPLEANDMIXEDSAMPLEANALYSES 33
2.1 Introduction 33
2.2 SubsampleAnalysis 35
2.2.1 Pure-SampleTargetDetection 35
2.2.2 SubsampleTargetDetection 38
2.2.2.1 AdaptiveMatchedDetector(AMD) 39
2.2.2.2 AdaptiveSubspaceDetector(ASD) 41
2.2.3 SubsampleTargetDetection:ConstrainedEnergyMinimization(CEM) 43
2.3 MixedSampleAnalysis 45
2.3.1 ClassificationwithHardDecisions 45
2.3.1.1 Fisher’sLinearDiscriminantAnalysis(FLDA) 46
2.3.1.2 SupportVectorMachines(SVM) 48
2.3.2 ClassificationwithSoftDecisions 54
2.3.2.1 OrthogonalSubspaceProjection(OSP) 54
2.3.2.2 Target-ConstrainedInterference-Minimized
Filter(TCIMF) 56
2.4 Kernel-BasedClassification 57
2.4.1 KernelTrickUsedinKernel-BasedMethods 57
2.4.2 Kernel-BasedFisher’sLinearDiscriminantAnalysis(KFLDA) 58
2.4.3 KernelSupportVectorMachine(K-SVM) 59
2.5 Conclusions 60
3 THREE-DIMENSIONALRECEIVEROPERATINGCHARACTERISTICS(3DROC)
ANALYSIS 63
3.1 Introduction 63
3.2 Neyman–PearsonDetectionProblemFormulation 65
3.3 ROCAnalysis 67
3.4 3DROCAnalysis 69
3.5 RealData-BasedROCAnalysis 72
3.5.1 HowtoGenerateROCCurvesfromRealData 72
3.5.2 HowtoGenerateGaussian-FittedROCCurves 73
3.5.3 HowtoGenerate3DROCCurves 75
3.5.4 HowtoGenerate3DROCCurvesforMultipleSignalDetectionand
Classification 77
3.6 Examples 78
3.6.1 HyperspectralImaging 79
3.6.1.1 HyperspectralTargetDetection 79
3.6.1.2 LinearHyperspectralMixtureAnalysis 80
3.6.2 MagneticResonance(MR)BreastImaging 83
3.6.2.1 BreastTumorDetection 84
3.6.2.2 BrainTissueClassification 87
3.6.3 Chemical/BiologicalAgentDetection 91
3.6.4 BiometricRecognition 95
3.7 Conclusions 99
Contents ix
4 DESIGNOFSYNTHETICIMAGEEXPERIMENTS 101
4.1 Introduction 102
4.2 SimulationofTargetsofInterest 103
4.2.1 SimulationofSyntheticSubsampleTargets 103
4.2.2 SimulationofSyntheticMixed-SampleTargets 104
4.3 SixScenariosofSyntheticImages 104
4.3.1 PanelSimulations 104
4.3.2 ThreeScenariosforTargetImplantation(TI) 106
4.3.2.1 ScenarioTI1(CleanPanelsImplantedintoCleanBackground) 106
4.3.2.2 ScenarioTI2(CleanPanelsImplantedintoNoisyBackground) 107
4.3.2.3 ScenarioTI3(GaussianNoiseAddedtoCleanPanels
ImplantedintoCleanBackground) 108
4.3.3 ThreeScenariosforTargetEmbeddedness(TE) 108
4.3.3.1 ScenarioTE1(CleanPanelsEmbeddedinCleanBackground) 109
4.3.3.2 ScenarioTE2(CleanPanelsEmbeddedinNoisyBackground) 109
4.3.3.3 ScenarioTE3(GaussianNoiseAddedtoCleanPanels
EmbeddedinBackground) 110
4.4 Applications 112
4.4.1 EndmemberExtraction 112
4.4.2 LinearSpectralMixtureAnalysis(LSMA) 113
4.4.2.1 MixedPixelClassification 114
4.4.2.2 MixedPixelQuantification 114
4.4.3 TargetDetection 114
4.4.3.1 SubpixelTargetDetection 114
4.4.3.2 AnomalyDetection 122
4.5 Conclusions 123
5 VIRTUALDIMENSIONALITYOFHYPERSPECTRALDATA 124
5.1 Introduction 124
5.2 ReinterpretationofVD 126
5.3 VDDeterminedbyDataCharacterization-DrivenCriteria 126
5.3.1 EigenvalueDistribution-BasedCriteria 127
5.3.1.1 ThresholdingEnergyPercentage 127
5.3.1.2 ThresholdingDifferencebetweenNormalizedCorrelation
EigenvaluesandNormalizedCovarianceEigenvalues 128
5.3.1.3 FindingFirstSuddenDropintheNormalizedEigenvalue
Distribution 128
5.3.2 Eigen-BasedComponentAnalysisCriteria 128
5.3.2.1 SingularValueDecomposition(SVD) 128
5.3.2.2 PrincipalComponentsAnalysis(PCA) 129
5.3.3 FactorAnalysis:Malinowski’sErrorTheory 129
5.3.4 InformationTheoreticCriteria(ITC) 130
5.3.4.1 AIC 131
5.3.4.2 MDL 131
5.3.5 GershgorinRadius-BasedMethods 131
5.3.5.1 ThresholdingGershgorinRadii 134
5.3.5.2 ThresholdingDifferenceGershgorinRadiibetweenRL(cid:1)L andKL(cid:1)L 134
x Contents
5.3.6 HFCMethod 135
5.3.7 DiscussionsonDataCharacterization-DrivenCriteria 138
5.4 VDDeterminedbyDataRepresentation-DrivenCriteria 140
5.4.1 OrthogonalSubspaceProjection(OSP) 140
5.4.2 SignalSubspaceEstimation(SSE) 142
5.4.3 DiscussionsonOSPandSSE/HySime 143
5.5 SyntheticImageExperiments 144
5.5.1 DataCharacterization-DrivenCriteria 144
5.5.1.1 TargetImplantation(TI)Scenarios 145
5.5.1.2 TargetEmbeddedness(TE)Scenarios 146
5.5.2 DataRepresentation-DrivenCriteria 149
5.6 VDEstimatedforRealHyperspectralImages 155
5.7 Conclusions 163
6 DATADIMENSIONALITYREDUCTION 168
6.1 Introduction 168
6.2 DimensionalityReductionbySecond-OrderStatistics-BasedComponentAnalysis
Transforms 170
6.2.1 EigenComponentAnalysisTransforms 170
6.2.1.1 PrincipalComponentsAnalysis 170
6.2.1.2 StandardizedPrincipalComponentsAnalysis 172
6.2.1.3 SingularValueDecomposition 174
6.2.2 Signal-to-NoiseRatio-BasedComponentsAnalysisTransforms 176
6.2.2.1 MaximumNoiseFractionTransform 176
6.2.2.2 Noise-AdjustedPrincipalComponentTransform 177
6.3 DimensionalityReductionbyHigh-OrderStatistics-BasedComponentsAnalysis
Transforms 179
6.3.1 Sphering 179
6.3.2 Third-OrderStatistics-BasedSkewness 181
6.3.3 Fourth-OrderStatistics-BasedKurtosis 182
6.3.4 High-OrderStatistics 182
6.3.5 AlgorithmforFindingProjectionVectors 183
6.4 DimensionalityReductionbyInfinite-OrderStatistics-BasedComponentsAnalysis
Transforms 184
6.4.1 Statistics-PrioritizedICA-DR(SPICA-DR) 187
6.4.2 RandomICA-DR 188
6.4.3 InitializationDrivenICA-DR 189
6.5 DimensionalityReductionbyProjectionPursuit-BasedComponentsAnalysis
Transforms 190
6.5.1 ProjectionIndex-BasedProjectionPursuit 191
6.5.2 RandomProjectionIndex-BasedProjectionPursuit 192
6.5.3 ProjectionIndex-BasedPrioritizedProjectionPursuit 193
6.5.4 InitializationDrivenProjectionPursuit 194
6.6 DimensionalityReductionbyFeatureExtraction-BasedTransforms 195
6.6.1 Fisher’sLinearDiscriminantAnalysis 195
6.6.2 OrthogonalSubspaceProjection 196
6.7 DimensionalityReductionbyBandSelection 196
Contents xi
6.8 ConstrainedBandSelection 197
6.9 Conclusions 198
II:ENDMEMBEREXTRACTION 201
7 SIMULTANEOUSENDMEMBEREXTRACTIONALGORITHMS(SM-EEAs) 207
7.1 Introduction 208
7.2 ConvexGeometry-BasedEndmemberExtraction 209
7.2.1 ConvexGeometry-BasedCriterion:OrthogonalProjection 209
7.2.2 ConvexGeometry-BasedCriterion:MinimalSimplexVolume 214
7.2.2.1 Minimal-VolumeTransform(MVT) 214
7.2.2.2 ConvexConeAnalysis(CCA) 214
7.2.3 ConvexGeometry-BasedCriterion:MaximalSimplexVolume 215
7.2.3.1 SimultaneousN-FINDR(SMN-FINDR) 216
7.2.3.2 IterativeN-FINDR(IN-FINDR) 216
7.2.3.3 VariousVersionsofImplementingIN-FINDR 218
7.2.3.4 DiscussionsonVariousImplementationVersionsofIN-FINDR 222
7.2.3.5 ComparativeStudyAmongVariousVersionsofIN-FINDR 222
7.2.3.6 AlternativeSMN-FINDR 223
7.2.4 ConvexGeometry-BasedCriterion:LinearSpectralMixtureAnalysis 225
7.3 Second-OrderStatistics-BasedEndmemberExtraction 228
7.4 AutomatedMorphologicalEndmemberExtraction(AMEE) 230
7.5 Experiments 231
7.5.1 SyntheticImageExperiments 231
7.5.1.1 ScenarioTI1(EndmembersImplantedinaCleanBackground) 232
7.5.1.2 ScenarioTI2(EndmembersImplantedinaNoisyBackground) 233
7.5.1.3 ScenarioTI3(NoisyEndmembersImplantedinaNoisy
Background) 234
7.5.1.4 ScenarioTE1(EndmembersEmbeddedintoaClean
Background) 235
7.5.1.5 ScenarioTE2(EndmembersEmbeddedintoaNoisy
Background) 235
7.5.1.6 ScenarioTE3(NoisyEndmembersEmbeddedintoaNoisy
Background) 236
7.5.2 CupriteData 237
7.5.3 HYDICEData 237
7.6 Conclusions 239
8 SEQUENTIALENDMEMBEREXTRACTIONALGORITHMS(SQ-EEAs) 241
8.1 Introduction 241
8.2 SuccessiveN-FINDR(SCN-FINDR) 244
8.3 SimplexGrowingAlgorithm(SGA) 244
8.4 VertexComponentAnalysis(VCA) 247
8.5 LinearSpectralMixtureAnalysis-BasedSQ-EEAs 248
8.5.1 AutomaticTargetGenerationProcess-EEA(ATGP-EEA) 248
8.5.2 UnsupervisedNonnegativityConstrainedLeast-Squares-EEA
(UNCLS-EEA) 249
xii Contents
8.5.3 UnsupervisedFullyConstrainedLeast-Squares-EEA(UFCLS-EEA) 250
8.5.4 IterativeErrorAnalysis-EEA(IEA-EEA) 251
8.6 High-OrderStatistics-BasedSQ-EEAS 252
8.6.1 Third-OrderStatistics-BasedSQ-EEA 252
8.6.2 Fourth-OrderStatistics-BasedSQ-EEA 252
8.6.3 CriterionforkthMoment-BasedSQ-EEA 253
8.6.4 AlgorithmforFindingProjectionVectors 253
8.6.5 ICA-BasedSQ-EEA 254
8.7 Experiments 254
8.7.1 SyntheticImageExperiments 255
8.7.2 RealHyperspectralImageExperiments 258
8.7.2.1 CupriteData 258
8.7.2.2 HYDICEData 260
8.8 Conclusions 262
9 INITIALIZATION-DRIVENENDMEMBEREXTRACTION
ALGORITHMS(ID-EEAs) 265
9.1 Introduction 265
9.2 InitializationIssues 266
9.2.1 InitialConditionstoTerminateanEEA 267
9.2.2 SelectionofanInitialSetofEndmembersforanEEA 267
9.2.3 IssuesofRandomInitialConditionsDemonstratedbyExperiments 268
9.2.3.1 HYDICEExperiments 268
9.2.3.2 AVIRISExperiments 270
9.3 Initialization-DrivenEEAs 271
9.3.1 InitialEndmember-DrivenEEAs 272
9.3.1.1 FindingMaximumLengthofDataSampleVectors 272
9.3.1.2 FindingSampleMeanofDataSampleVectors 273
9.3.2 EndmemberInitializationAlgorithmforSM-EEAs 274
9.3.2.1 SQ-EEAs 274
9.3.2.2 Maxmin-DistanceAlgorithm 275
9.3.2.3 ISODATA 275
9.3.3 EIA-DrivenEEAs 275
9.4 Experiments 278
9.4.1 SyntheticImageExperiments 278
9.4.2 RealImageExperiments 281
9.5 Conclusions 283
10 RANDOMENDMEMBEREXTRACTIONALGORITHMS(REEAs) 287
10.1 Introduction 287
10.2 RandomPPI(RPPI) 288
10.3 RandomVCA(RVCA) 290
10.4 RandomN-FINDR(RN-FINDR) 290
10.5 RandomSGA(RSGA) 292
10.6 RandomICA-BasedEEA(RICA-EEA) 292
10.7 SyntheticImageExperiments 293
10.7.1 RPPI 293
Description:Hyperspectral Data Processing: Algorithm Design and Analysis is a culmination of the research conducted in the Remote Sensing Signal and Image Processing Laboratory (RSSIPL) at the University of Maryland, Baltimore County. Specifically, it treats hyperspectral image processing and hyperspectral sign