一種顧及多特征的建筑物變化檢測方法AbstractAsthedemandforbuildingmonitoringandassessmentincreases,thedetectionofbuildingchangeshasbecomeacrucialtask.Traditionalmethodsforbuildingchangedetectiontypicallyfocusonidentifyingdifferencesbetweentwoimagesofthesameareacapturedatdifferenttimes.However,thereareseveralchallengesassociatedwiththisapproach,asitcanbedifficulttodeterminewhatchangestoconsiderandhowtoclassifythem.Inthispaper,weproposeanewapproachtobuildingchangedetectionthattakesintoaccountmultiplefeaturesandcharacteristicsofthebuildinganditssurroundings.Ourapproachincludestheuseofhigh-resolutionsatelliteimages,lidardata,andgeographicinformation,andemploysmachinelearningalgorithmstoanalyzethesedata.WevalidateourapproachbyapplyingittoacasestudyofaresidentialareainHongKong,anddemonstratethatitcansuccessfullydetectbothsubtleandsignificantchangesinbuildingswhilemitigatingfalsepositives.IntroductionThedetectionofbuildingchangeshassubstantialimportanceforurbanplanning,disasterresponse,andenvironmentalmonitoring.Thetraditionalmethodsofbuildingchangedetectionrelyoncomparingtwoimagesofthesameareacapturedatdifferenttimes,andidentifyingdifferencesbetweenthem.Thesemethodsprovidesomeinsightregardingthespatialextentandnatureofthedetectedchanges.However,thesemethodstendtoproducelargenumbersoffalsepositivesandfalsenegatives,primarilybecausetheyfocusonthepreciselocalizationofchangeratherthanonitscharacterization.Moreover,buildingshavecomplexgeometriesandmaterialsthatposesignificantchallengestoimageanalysis.Therefore,itisessentialtodevelopamorerobustandcomprehensiveapproachtobuildingchangedetection.Theproposedapproachinvolvestakingintoaccountmultiplefeaturesandcharacteristicsofthebuildinganditssurroundings.Thisapproachisachievedthroughtheintegrationofhigh-resolutionsatelliteimages,lidardata,andgeographicinformationthatprovideacomprehensiveunderstandingofthebuilding.Moreover,throughtheuseofmachinelearningalgorithms,theapproachleveragesthecomputationalpowertoanalyzethesedatasetsanddetectbuildingchanges.Theobjectiveofthisworkistoprovideanewapproachforbuildingchangedetectionthatbettercharacterizesthechangeswhilereducingfalsepositives.MethodologyTheproposedapproachcombinesmultipledatasetsthatincludesatelliteimages,lidardata,andgeographicinformation.Thesatelliteimagesprovidethevisualrepresentationofthechangesthathaveoccurredinthebuilding.Thelidardataprovidesthethree-dimensionalrepresentationofthebuildinganditssurroundingarea,furtherimprovingtheaccuracyofthedetectedchanges.Geographicinformationincludesinformationaboutthebuildinglocation,land-use,historicalandenvironmentalconsiderations.Theapproachincorporatestwostagesofprocessing.Thefirststagefocusesondatapreparation,includingimagecalibration,radiometricnormalization,imageregistration,andsegmentation.Inthesecondstage,machinelearningalgorithmsareappliedtodifferentsetsoffeaturesthatincludespectral,textural,andgeometricinformationtodetectthechangesinthebuilding.Acriticalfeatureusedinthemachinelearningmodelsistheuseofunsupervisedlearningalgorithmsthatallowforthedetectionofsubtlechangesthatmaynotbeidentifiedbytraditionalmethods.Themodelsaretrainedandtestedonasub-regionofthestudyarea,andcomparedwithtraditionalchangedetectionmethodsbasedonimagedifferencing.ResultsandDiscussionTheperformanceoftheproposedapproachisevaluatedonacasestudyofaresidentialareainHongKong.Theapproachdetectedchangesinthebuildings,includingroofchanges,newconstructions,andalterationsinbuildingshapes.Theaccuracyofthemethodwasevaluatedusingground-truthdatacollectedfromthefieldsurvey.Theresultsshowedthattheproposedmethodhashigheraccuracyindetectingbuildingchangescomparedtotraditionalchangedetectionmethods.Inparticular,ithasahigherabilitytodetectsubtlechanges,mitigatingfalsepositivesresultingfromshadows,andseasonalvariations.Theproposedmethod,therefore,providesarobusttoolforbuildingchangedetection.ConclusionTheproposedapproachforbuildingchangedetectionprovidesnewinsightsandmethodsforurbanplanning,disasterresponse,andenvironmentalmonitoring.Themethoddemonstratedinthisworkintegratesmultiplefeaturesandcharacteristicsofthebuildinganditssurroundingsthroughtheuseofsatelliteimages,lidardata,andgeographicinformation.Additionally,themethodemploysmachinelea