DATAFUSIONTECHNIQUESFOROBJECTSPACECLASSIFICATION USINGAIRBORNELASERDATA ANDAIRBORNEDIGITALPHOTOGRAPHS By JOONGYONGPARK ADISSERTATIONPRESENTEDTOTHEGRADUATESCHOOL OFTHEUNIVERSITYOFFLORIDAINPARTIALFULFILLMENT OFTHEREQUIREMENTSFORTHEDEGREEOF DOCTOROFPHILOSOPHY UNIVERSITYOFFLORIDA 2002 Copyright2002 by JOONGYONGPARK Tomywife,SunghwaYoun, Mylovelydaughter,Shinhwo, Andmyparents,ByungJoParkandSoonJaLee ACKNOWLEDGMENTS Iwouldliketoexpressmygratitudetomysupervisor,ProfessorGradyTuell,forhis continuoussupport,encouragement,andguidancethroughoutmystudies.Heintroduced metotheworldofdatafusion,whichIhadneverbeenthoughtbefore.Healwayshelps metosolveproblemsandencouragesgoingtoahigheracademiclevel.Heisagreat teacherformeacademicallyandspiritually.IamsoproudtobehisfirstPh.D.student. IwishtothankProfessorRameshShresthaandWilliamCarter,whoalsogaveme continuoussupportandespeciallyfortheirpositiveinfluenceforyearstocome.They introducedmetotheentranceofLIDAR.Theymaintainaworldclassresearchteamof airbornelasermappingtechnology,whichtheygavemeachancetobeinvolvedin. Ialsowishtothankmyothercommitteemembers.ProfessorDavidBloomquistand PaulZwick,fortheircomments.Ialsowishtoexpressthankstomypastandpresent colleaguesintheresearchgroup. IV TABLEOFCONTENTS gage ACKNOWLEDGMENTS iv LISTOFTABLES viii LISTOFFIGURES x ABSTRACT xiii CHAPTER 1 INTRODUCTION 1 ObjectiveofResearch 4 DataFusion 5 AnObjectRecognitionExample 7 PotentialAdvantagesinIntegratingMultipleSensors 9 TheMethodsofDataFusion 9 Signal-LevelMethods 10 Pixel-LevelMethods 11 Feature-LevelMethods 12 Decision-LevelMethods 14 DataAcquisition 17 Outline 19 2 CLASSIFICATIONUSINGASUPERVISEDSTATISTICALPATTERN RECOGNITIONTECHNIQUE 22 DataFusionofDigitalColorImagery,ALSMIntensity,andDEM 22 Methodology 25 AnalysisandResults 25 Discussion 37 Summary 42 3 CLASSIFICATIONUSINGANEXPERTSYSTEMAPPROACH 44 Rule-basedClassification 44 AnalysisofStudyArea 47 V 1 Methods 49 Results 58 Discussion 62 Summary 62 4 CLASSIFICATIONUSINGTHEDEMPSTER-SHAFEREVIDENTIALTHEORY64 TheDempster-ShaferEvidentialTheory 64 TheConceptofEvidentialReasoning 65 Dempster’sRuleofCombination 68 Experiment 71 PriorProbabilityfromtheMaximumLikelihoodDecisionRule 71 DataFusionProcessingUsingtheDempster-ShaferEvidentialTheory 72 Results 80 Summary 8 5 CONCLUSIONS 84 6 DISCUSSIONANDRECOMMENDATIONFORFURTHERWORK 90 APEENDD( A DESCRIPTIONOFTHEDATAACQUISITIONSYSTEM 93 SystemComponents 93 AirborneLaserSwathMappingSystem 93 AirborneDigitalPhotographyImagingSystem 97 B DIRECTDIGITALIMAGEGEOREFERENCING 99 AirborneGPSPositioningandGeoreferencing 99 GeoreferencingofImageSensors 103 DirectGeoreferencingMethod 103 PrinciplesofDirectImageGeoreferencing 105 PositionalOffsetandMisalignmentAngleCalibration 107 TimeAlignment 112 AngleAnalysisforBack-Projection 112 CalibrationofMisalignmentAngle 116 StudyAreaandDataAcquisition 119 Results 120 Discussion 123 C SPECIFICSOFINTENSITYDATAFROMALSMSYSTEM 124 VI IntensityofReflectedLaserPulse 124 CorrelationBetweenALSMIntensityandTargetReflectance 126 D GAUSSIANMAXIMUMLIKELIHOODCLASSIFICATION 130 E TECHNIQUEFORASSESSINGTHEACCURACYOFCLASSIFICATIONS 135 SampleDesign 135 SampleUnit 135 SampleSize 136 SamplingSelection 136 EvaluationofErrorMatrices 138 KappaAnalysis 140 LISTOFREFERENCES 142 BIOGRAPHICALSKETCH 151 vii 1- 2- LISTOFTABLES Table page 1.LaserSystemParameters 17 1.ConfusionMatrixBetweenClassifiedDataandReferenceDataUsinganRGB Image(Pixels) 27 2-2.ConfusionMatrixBetweenClassifiedDataandReferenceDataUsinganHIImage (Pixels) 28 2-3.ConfusionMatrixBetweenClassifiedDataandReferenceDataUsinganRGI Combination(Pixels) 30 2-4.ConfusionMatrixBetweenClassifiedDataandReferenceDataUsinganRGH Combination(Pixels) 31 2- 2-5.ConfusionMatrixBetweenClassifiedDataandReferenceDataUsinganRHI 3- Combination(Pixels) 33 2-6.ConfusionMatrixBetweenClassifiedDataandReferenceDataUsinganRGHI Combination(Pixels) 34 2-7.ConfusionMatrixBetweenClassifiedDataandReferenceDataUsinganRBHI Combination(Pixels) 35 2-8.ConfusionMatrixBetweenClassifiedDataandReferenceDataUsinganRGBHI Combination(Pixels) 36 2-9.OverallAccuracyofEightBandCombinations(%) 38 10.StatisticsSummaryoftheMaximumLikelihoodClassificationfortheRGBHI BandCombination 38 1.Level-1ClassificationSchemeUsedinanRGBImageandanHIImage 52 3-2.RulesofLevel-1ClassificationfortheHIImage 53 3-3.RulesofLevel-1ClassificationforDigitalColorImage 54 3-4.Level-2ClassificationSchemeafterDataFusionUsingProductionRules 55 viii 33--5.Level-2ProductionRulesforDataFusionofALSMandADPLevel-1Classification 4- Data 55 6.ConfusionMatrixofRule-basedClassification(Pixels) 61 1.Dempster’sCombiningRule[fromThomopoulos(1990)] 69 4-2.TheCorrelationBetweenDataandInformationofClassesforanRGBDigitalColor Image 73 4-3.SummaryoftheMaximumLikelihoodClassificationforanRGBImage 73 44--4.TheCorrelationBetweenDataandInformationofClassesforanHIImage 76 45--5.StatisticsSummaryoftheMaximumLikelihoodClassificationforanHIImage 76 4-6.Dempster’sCombinationRuleofanRGBImageandanHIImage 78 4-7.TheMassAssignmentbyDempster’sCombinationRulefromtheALSMandADP Data 79 8.ConfusionMatrixofDataFusionClassificationbyDempster-ShaferEvidential Theory(Pixels) 82 1.TheComparisonofEachClassandOverallAccuracyamongThreeDataFusion Techniques(%) 87 A-1.SpecificationsoftheOptechModelALTM1210 96 B-1.PositionalOffsetsBetweenSensorstoIMULeverArm 110 B-2.PositionalOffsetsBetweenSensorstoGPSLeverArm 110 B-3.TheOffsetBetweenSensorsandIMUMisalignment 118 B-4.OffsetBetweenGPSDataandCalculationafterApplyingSubsystemParameters119 B-5.TheComparisonBetweenGPSGroundSurveyDataandtheSamePointsonthe Pseudo-orthorectifiedImage 123 C-1.ClassificationofALSMDataGroup 127 E-1.TheSamplePolygoninEachClass 139 E-2.ExampleErrorMatrix 141 ix LISTOFFIGURES Figure page I-l.Generalarchitectureofthedatafusionprocessinimageclassification 6 11--2.Thediscriminationoffourdifferentobjectsusingcomplementaryinformationfrom 2- twosensors 8 1-3.Data-,feature-,anddecision-levelfusion 13 1-4.TheintensityimagefromtheALSMdata 18 1-5.TheelevationimagefromtheALSMdata 18 6.DigitalcolorimagefromADPsystemfromtheALSM 19 1.Feature-leveldatafusion 23 2-2.Six-trialbandcombinationfromanRGBimageandanHIimage 24 2-3.Bandcombinationimages:(a)RGI(red,greenandintensity);(b)RHI(red,height, andintensity)combinations 24 2-4.GroundtruthreferencedataonanRGBcolorimage 26 2-5.Maximumlikelihoodclassification:(a)RGB(red,green,andblue);(b)HI(height andintensity)bandcombinations 29 2-6.Maximumlikelihoodclassification:(a)RGI(red,green,andintensity);(b)RGH (red,green,andheight)bandcombinations 32 2-7.Maximumlikelihoodclassification:(a)RHI(red,heightandintensity;(b)RGHI (red,green,height,andintensity)bandcombinations 35 2-8.Maximumlikelihoodclassification:(a)RBHI(red,blue,heightandintensity);(b) RGBHI(red,green,blue,height,andintensity)bandcombinations 37 2-9.Diagramofnormaldistributionofeightclassesonaredband 39 2-10.Diagramofnormaldistributionofeightclassesonagreenband 39 2-11.Diagramofnormaldistributionofeightclassesonablueband 40 X