Volume11,No.6,November-December2022
InternationalJournalofAdvancedTrendsinComputerScienceandEngineering
AvailableOnlineathttp://www.warse.org/IJATCSE/static/pdf/file/ijatcse041162022.pdf https://doi.org/10.30534/ijatcse/2022/041162022
AutomaticChangeDetectiononSatelliteImagesusing
PrincipalComponentAnalysis,ISODATAandFuzzy
C-MeansMethods
BEKKOUCHEIbtissem1,FIZAZIHadria2
1 DepartmentofComputerSciences,UniversityofScienceandTechnologyofOran,Algeria, ibtissem.bekkouche@univ-usto.dz
2 DepartmentofComputerSciences,UniversityofScienceandTechnologyofOran,Algeria, hadria.fizazi@univ-usto.dz
ReceivedDate:October17,2022AcceptedDate:November23,2022 PublishedDate:December06,2022
ABSTRACT
Changedetectionistheprocessofcomparingtwoormore imagesandidentifyingthepartswhereachangehasoccurred. Differencedetectionprocessingbetweensimpledigital images,suchasphotographicimages,iseasytoimplement. Whereasforsatelliteimages,whichcomposeofseveral images’grayscaleandbands,thisrequiresamethodological approachtoimageprocessingappropriatetotheexploitation ofthesedatabecausethiswillallowtofollowtheevolution overtimeofaregionofinterestthroughchangedetection techniques,sotheseimagesareatoolofchoiceinthe managementofnaturalresources.So,inthispaper,we proposeahybridautomaticchangedetectionapproachfor multi-temporalsatelliteimages.Itisbasedonseveral algorithms:ISODATAforautomaticthresholding,Principal ComponentAnalysisastransformationtechnique,Fuzzy C-Meansasclassificationtechnique.Experimentswere performedandassessedbytheiroverallaccuracyandresults validatedtheeffectivenessandefficiencyoftheproposed approach,namedISOFAP.
Keywords: Changedetection,Fuzzyc-meansclustering, ISODATA,Principalcomponentanalysis.
1. INTRODUCTION
Changedetectionisadigitalprocessthatcanbeperformedby traditionalmethodsandusingremotesensingtechnologies. ThebasicthisprocessistomeasurethechangeontheEarth's surfacebyjointlyanalyzingtwoormoretwotemporally separatedimages,inordertolocateandquantify (automatically)thechangesexistingbetweentheseimages [1].Thisisaveryactivesubjectduetopreoccupationabout theconsequencesofglobalandlocalchangesintheearth.
Therearemanychangedetectiontechniquesintheliterature: AlgebraicmethodssuchasImagedifferencingandChange vectoranalysis,…;TransformationmethodssuchasTasseled CapTransformation,PrincipalComponentAnalysis…; ClassificationmethodsasArtificialNeuralNetworks, Comparisonafterclassification…;GeographicInformation
System(GIS)asIntegratedMethodofGISandRemote Sensing,GISApproach…;VisualanalysisusingVisual interpretation;HybridapproachinCombinationofmethods; AdvancedmodelslikeSpectralMixingModel,Li-Strahler ReflectanceModel…[2][3]
Inrecentyears,thistechniquehasbecomeoneofthemost interestingsubjectsintheextractionofinformationfrom satelliteimagesandseveralresearchershaveopted,Theselast years,forthehybridizationofmethodssuchas:in2020, NeelamRuhiletal,havesuggestedanunsupervisedchange detectionmethodbasedonwaveletfusionandtheKohonen Hybrid FCM-σ [4], in the same year, Mohan Singh et al, have proposedanimagefusionusingimagefusionusingimage normalizationandradiometriccalibrationandParticleSwamp OptimizationFuzzyC-Means(PSOFCM).Inthisarticle,an unsupervisedchangeobservationtechniquebasedonthe PSOFCM[5],andin2022,AbdelkrimMaariretal,have proposedanunsupervisedmethodofdetectingchangein satelliteimagesbyfollowingtwomainsteps:Thefirststep focusesondatareductionusingtheIndependentComponent Analysis(ICA)algorithmtoimprovetheefficiencyofthe classifier.ThesecondstageforprocessingusestheFuzzy C-Meansclassificationmethodtofindspecifiedclusters[3].
So,forourpaper,sincethereareseveralchangedetection techniques,themostinterestingattitudewouldbetotryto combinethesetechniquesanddevelopahybridmethod,sofor thatweused:ISODATAforautomaticthresholding,Principal ComponentAnalysis(PCA)astransformationtechnique, FuzzyC-Means(FCM)asclassificationtechnique.
Afterhavingtestedseveralthresholdingalgorithmssuchas: Binarythresholdingonthemean,OTSUthresholdingandEM algorithm[6],wechoseISODATA.Thisisthealgorithmthat isusedforautomaticthresholding,wehavechosenitbecause itiseasytoimplementmorethanitgivesgoodresults.
PrincipalComponentAnalysis(PCA)consistsof transformingvariables,interconnected,intonewvariables unsquaredfromeachotherfordimensionreduction.So,we chosebecauseithasbeenwidelyusedforchangedetection
[7].Ithastheabilitytoprojectthemulti-dimensionaloriginal anditislesssusceptibletoimageoverlapping.
Wefoundthatthemostusedmethodsforchangedetection usedalottheK-means,sowetesteditonourdatabutwe chosetousetheimprovedFuzzyC-Means(FCM)version. FCMhasreducedcomplexityandgivesbestresultfor overlappingdatasetsandcomparativelybetterthank-means algorithm.
2.METHODSUSED
2.1.Isodata
Itisanimagesegmentationtechniquebyclusteringanditis animprovedversionofthek-meansalgorithm,whichwas firstintroducedbyVelascoin1980[8]asaclassicalgorithm whichmakesitpossibletocarryoutaverygood categorizationandwhichgivessignificantresults.[9]
Thethresholdingofanimagecanbedonebymanual thresholdingorbyautomaticthresholding.
Manualimagethresholdinginvolves4steps[9]:
1. Observingthehistogramoftheimage
2. Choiceofthresholdsinthevalleys
3. Definitionoftheclassesoftheregionsbycolorrange
4. Pixelclassification
ISODATAthresholdingisglobalthresholding,whereasingle thresholdisusedacrosstheentireimagetodivideitintotwo clusters.Itallowstofindthevalueofthesoughtthresholdin anautomaticwayfollowingitssteps[10]:
1. ChooseaninitialthresholdT,forexampleT=the averageintensity.
2. DividetheimageintotwogroupsG1andG2usingT.
3. CalculatetheaveragevaluesofeachregionR1andR2
4. Calculatethevalue =( 1+ 2)/2
5. Repeatsteps2to4untiltheTvaluedoesnotchange.
2.2.PrincipalComponentAnalysis
Principalcomponentanalysis(PCA)isamathematical techniqueusedfordataredundancyreductionbyJacksonand Bund,1983[12].Itisusefulwhenyouhaveobtaineddataona numberofvariables(perhapsalargenumberofvariables), andthereissomeredundancyinthesevariables.[12]
PCAisoneofthemostpopularmultivariateanalysis algorithmsforchangedetectionstudiesandcanbeperformed onoriginalornormalizeddata[7].Withthistechniquethe digitalimagesacquiredbyremotesensing,wecanreduceits dimensionalitysuchthatthemultispectralbandsarethe variablestobeintroduced.
TherearecertainstepstofollowtoimplementPCA[7]:
Takeanoriginaldatasetandcalculatethemeanofthe dataset. Subtractthemeanforeachdimension.
Calculatethecovariancematrix.
Calculatetheeigenvectorandtheeigenvalueofthe covariancematrix.
Extractthediagonalofthematrixasavector.
Variancesortingindescendingorder.
Choosecomponentsandformafeaturevector.
Derivationofthenewdataset.
Attheend,thenumberofPCislessthanthenumberof variancesintheoriginalimage.InCDstudies,the consequenceofthislinearizationisthattheunchangedpixels orcommoninformationsharedbyapairofimagesare assumedtobeinanarrowandelongatedspace.Clusteralong aprincipalaxisequivalenttothefirstcomponent(PC1).On thecontrary,pixelscontainingachangewouldbemore uniqueintheirspectralappearanceandshouldliefarfromthis axis(PC2).[13][14]
2.3.FuzzyC-Means
FuzzyC-Means(FCM),isanunsupervisedfuzzy classificationalgorithm.IssuedfromtheC-meansalgorithm, developedbyDunnin1973[15]andimprovedbyBezdekin 1981[16],itintroducedthenotionoffuzzysetinthe definitionofclasses:eachpointinthedatasetbelongstoeach clusterwithacertaindegree,andallclustersarecharacterized bytheircenterofgravity[17].
ThegoalofFuzzyC-Meansclusteringistofindtheminimum ofthefollowingfunction: (1)
where m isanyrealnumbergreaterthan1, uij isthedegreeof membershipof xi inthecluster j.Asarule,foreachpixel,sum ofallmembershipvaluebelongingtoallclassesmustbe1. [3][14]
3.METHODOLOGY
TheFigure1showsandsummarizestheorganizationofthe processingstepsadoptedforthedetectionofchangeson satelliteimagesbyourISOFAPapproachandthisaccording tothebasicprocedureofanimagechangedetection processingsystem[18].
Thisschemeiscomposedofseveralimplementationphases:
1ststep: Thisisadatapreparationstepandwemust:
Dataacquisitionandpreparation: Thisisthestageof collectingandassemblingdata,whichcanbe satelliteimagesandfieldinvestigations.Thenwe cangotopre-processing,forexampleimagecutting ifnecessary.
Geometriccorrection: theverificationofthe geometricaccuracyisessentialforthedetectionof changesbesidesabadgeoreferencedofmorethan onepixelwouldcauseabnormalresultsforanalyzes pixelbypixel.
Datanormalization: Datamustbenormalized, especiallyforsatelliteimages,inordertoreducethe variabilitybetweenmulti-dateimagesoverthesame geographicalarea.
2ndstep: Thisisadigitalprocessingstepforthedetectionof changesandforthiswewill:
Applythedifferenceimagemethod
Afterthat,theobtainedimageispartitionedintoH*H blocks,thenthecreationofeigenspacespaceusing PCAandprincipalcomponentsareachieved.
ApplytheISODATAalgorithmforthresholdingonX. AfterapplyingPCAandISODATA,createthefeature vectorspaceusingtheeigenspacespace. Todeterminetheareasthathavechangedandnot changed,weusedtheFCMalgorithmtogeneratethe twoclasses(k=2)andassigneachfeaturevectorto theclassclosesttoeithermodifiedpixelsor unchangedpixels.
3thstep: Thisisageneralizationofthefinalproduct,whichis amapoflandcoverchangesatascaleequivalenttothatofthe inputdata.
4thstep: Onceamodelhasbeendeterminedand implemented,thelaststepistoanalyzeandinterpretthe resultstoestablishthequalityofthismodel.Therearevarious evaluationmeasuresthatcanbeusedandchosencarefully, sincethechoiceofmeasurecaninfluencehowperformanceis assessedandinterpreted.Forthiswehavechosentouse:
Visualinterpretation: theuseofthehumanvisual systemasaqualityjudgmenttoolisnottobe neglectedbutnecessarytoverifythequalityofthe imagesobtainedbytheclassification.Toevaluate thisapproach,wealsousedvisualanalysisaccording tothegroundtruthavailableinthearea.
ConfusionMatrix: Oneofthemostpopularwaysto measuretheperformanceofaclassificationmodel. Eachlinecorrespondstoanactualclassandeach columncorrespondstoanestimatedclassandit includesthefollowingvalues[19]:
o TruePositive,TP,whentheactualclassandthe estimatedclassarebothpositive
o TrueNegative,TN,whentheactualclassandthe estimatedclassarebothnegative
o FalsePositive,FP,whentheactualclassis negativebuttheestimatedclassispositive.This iscalledaType1error.
o FalseNegative,FN,whentherealclassis positivebuttheestimatedclassisnegative.This iscalledaType2error.
Itcanbeusedformorein-depthmeasurementstogeta betterassessmentofthequalityofthemodel.Amongthe classificationmeasuresusedareaccuracy,precision, errorandspecificity.[19]
Accuracyisthenumberofcorrectpredictionsmadeby themodel.
(2)
ThismeasureisusedwhenthenumberofTruePositives andTrueNegativesarethemostimportant. Error=1– Accuracy (3) Precisionisthenumberofcorrectelementsrenderedby themodel.
(4)
ThismetricisusedwhenthenumberofFalsePositives ishighest.
Specificityisthenumberofnegativeclassespredicted bythemodel.
(5)
4.IMAGESUSED
Imagesdatabase: AsetofRGBimagesof650x650foundin adatabaseusedforthedetectionofchangesinsatellite imageryusingdeep-learning,producedbyHéloïse BAUDHUINandAntoineLAMBOT.[20]
Weappliedourapproachtotwoexamples,presentedin Figure2andFigure3,ofthisdatabase: (a) (b)
(b)
Figure3:Images2(a)beforechangeand(b)afterchange.
ImagesofBoumerdes: Figure4showsahigh-resolution satelliteimageofaLandsat5TMearthquakeandboth acquiredin2003,providedbytheCenterNationaldes TechniquesSpatialesd'ArzewandacquiredbyQuickBird.
Thecharacteristicsofthetwoimagesare:naturalcomposition imagewiththreechannels:TM1,TM2andTM3bandsand theirsizeinpixelsis1002x1002.Theycontaindifferent classeswhichare:asphalt,soil,vegetationandshade,andarea ofdamagetothepost-disasterimage.
(a)
(b)
Figure4:Images3(a)beforedisasterand(b)afterdisaster.
5.EXPERIMENTRESULTSANSDISCUSSION
Thefirstthingforourwork,wedidastudyonaknown changedetectionmethodtounderstanditsconceptandwe chosePCA-FCM.Afterthestudyourproposalwastochange theclassificationalgorithmandreplaceitwithanother,so afterseveraltries,wechosetoworkwiththeFuzzyC-means (FCM)whichisanimprovedversionoftheK-means algorithm.ThenwedidothertestswithPCA-FCMbuteach timewehadtoinitializethethresholdmanuallywiththe variationofthethresholdparametersforeachimage.So,for thatwehaveintegratedanISODATAthresholdingalgorithm tohaveanautomaticthreshold.Intherestofthearticle,some resultsofourproposedmethod"ISOFAP"incomparisonwith PCA-K-means.
Toanalyzeandvalidatetheproposedapproachweusedthree differentdataimagesdescribedinsection4below.
ForthePCA-K-meansparameters,wevariedtheparameters asfollows:numberofclassesis2classes,suchasclass1for changedpixelsandclass2forunchangedones,numberof iterationsbetween70and100,numberofblocks:between2 and5,andThreshold:between10and80.
AndforISOFAPwevariedtheparameterslikethe PCA-K-meansexceptthethresholdbyISODATA.
Figure5showsthefirstperformancetest,wecreatedtwo examplesofsimpleartificialimagestodoourtests,butwe reducedthenumberofiterationsto20andthethresholdto10 becausetheyaresimpleimages.
Bythevisualinterpretation,wenoticethatPCA-K-meanshas someconfusionsandthatthedetectionisnotsocorrect. However,ISOFAPgaveusabetterresultthanPCA-K-means althoughtherearealsoconfusionsbutitisminimal.
Afterconfirmingthecorrectoperationofthetwoalgorithms onasimpleartificialimage,weusedtheimages1ofthe database(presentedinFigure2)andweappliedforthetwo methodsPCA-K-meansandISODATA.Forinitializationof thePCA-kmeansmethod,afterseveraltests,thethresholdwas manuallyinitializedto60,thenumberofiterationsto80and wevariedjustthenumberofblocks(h*h)betweenhequalto 2and5.Wenoticethathedetectedthechangesforhequal2.
ThesameimageappliedtoourmethodISOFAP,thenumber ofiterationsat75,thevariationalsointhenumberofblocks between2and5,andknowingthatISODATAhasinitialized thethresholdto43.fromtheresultswealsonoticethatath equalto2theresultisbetterthantheothers. (a) (b)
Figure5:(a)artificialimage,(b)artificialimagechanged,change mapresults(c)withPCA-K-meansand(d)withISOFAP.
Figure6:Visualinterpretationofthebestresultsby(a)ISOFAPand (b)PCA-K-meansonimages1.
ThepreviousFigures6presentthebestresultsgivenby ISOFAPandPCA-K-meansappliedtoimages1.Fromour visualinterpretationforcomparingthebestresultsofthetwo methodstotheoriginalchangeimage,wenoticethatevenif ISOFAPhadconfusionsanddetectionerrorscausedbythe conflictbetweenthetwochangedandunchangedclasses,it therearesomedetailswerebetterdetectedcomparedto PCA-K-means.
Forimages2,wedidthesametestsasimages1.For PCA-Kmeans,afterseveraltestsweinitializedthethreshold at75,thenumberofiterationsat85andvariedthenumberof blocks.wenotethattheseresultsreinforcethechange detectionresultsonimages1becausethebestresultisgiven onthenumberofblocksequalto2.
ForISODATAthebestresultforimages2isthenumberof blocksat2becauseitgivesmoredetaileddetectionandfewer conflicts,suchasforthisimagethethresholdat20andthe numberofiterationsat93.
Thecomparisonbetweenthebestresultgivenbythetwo methodsapprovesthepreviousresultbecausewecansaythat thevisualinterpretation,presentedinFigure7,isthesameand thatISOFAPhasbetterdetectedthechangeevenifthereare conflictsanddetectionerrorsbuttheyarelessthan PCA-K-means.
Forthelasttests,wechosethebestparameters:numberof blocks(h*h)hat2,numberofiterationsat90andweapplied themtothesamepartoftheimages3(InFigure8).Wenotice thatthePCA-K-meansonlygaveustwoclasses:changedand unchanged,butISOFAPgaveathirdclassofpixelswiththe graycolor,fortheprogramitisconflictsandatthesametime wenoticethatitisnottotallyunchangedbutthechangeisnot great. (a)(b) (c)(d)
Figure8:(a)imagebefore,(b)artificialafter,changemapresults(c) withISOFAPand(d)withPCA-K-means.
FromthevisualinterpretationoftheresultsofourISOFAP approachonthethreegroupsofimages,wenoticethatthere aredetectionerrorsintheimages,whichcanbecausedby severalreasonsamongthemtheresolutionoftheimage becausesometimeswhentheimageisoflowerqualityposes conflictsbetweenthepixelsthereforegiveserrors.
Table1: Evaluationmetricresultsforimages1byISOFAPandPCA-K-means
Method Accuracy Error Precision Specificity Executiontime
ISOFAP 0.8556 0.1444 0.8918 0.1082 65.283269s
PCA-K-means 0.7624 0.2376 0.6881 0.3119 57.882101s
Table2: Evaluationmetricresultsforimages2byISOFAPandPCA-K-means
Method Accuracy Error Precision Specificity Executiontime
ISOFAP 0.6256 0.3744 0.7488 0.2512 75.243269s
PCA-K-means 0.5824 0.4176 0.6181 0.3819 71.874101s
Table3: Evaluationmetricresultsforimages3byISOFAPandPCA-K-means
Method Accuracy Error Precision Specificity Executiontime
ISOFAP 0.7356 0.2644 0.7918 0.2082 55.267169s
PCA-K-means 0.6424 0.3576 0.5841 0.4159 47.817301s
Wefinishourstudybyapplyingtheconfusionmatrix,tothe originalimagesandtheimagesofthebestresults,forextract theinformationthatinterestsusandherearetheresults:
Afterthecomparisonbytheevaluationmetric(inTable1, Table2andTable3)andthevisualinterpretation,wenotice thatourproposedISOFAPapproachgivesgoodresultsto detectchangesandthevaluesofaccuracyandprecisionare highforthreedifferentdataimages,soweconcludethat ISOFAPisbetterthanPCA-K-meanswiththeexceptionof theexecutiontime,ittakeslongerthanPCA-K-means.
6.CONCLUSION
Inthiswork,wehaveaddressedoneoftheimageprocessing operatorswhichisthedetectionofchangesinsatelliteimages. Wehaveproposedahybridmethodbasedontwotechniques todothistreatment.
Ourapproachisbasedonamethodalreadyusedforthe detectionofchanges(PCA-K-means)andwehavetriedto improveit.WeusedISODATAtomakethethresholding automaticandkeptthePCAbecauseitisthemostusedfor changedetectionandtheleastsensitivetoimageoverlap, whilewechoseFCMbecauseithasreducedcomplexityandit isanimprovedversionofthek-meansalgorithm.
Afterthetestsandtheadditions,wearrivedatthe implementationofourISOFAPmethod,whichiscompared withthePCA-K-means,theresultsallowedustoconclude thatourmethodcandetectthechangeandthatitgivesbetter results.
REFERENCES
1. S.Achour. Détectiondechangementenimagerie satellitairemulti-temporellepourlesuivietla surveillancedel'environnement,Ph.D.dissertation, Dept.Elect.Eng.,DjillaliLiabèsUniv,SidiBelAbbes, Algeria,2022.
2. D.LuandP.Mausel,E.Brondizio,andE.Moran. Changedetectiontechniques,Changedetection techniques,InternationalJournalofRemoteSensing,vol. 25,pp.2365-2401,2004.
3. A.MaarirandE.S.Azougaghe,andB.Bouikhalene. AutomaticChangeDetectionBasedonthe IndependentComponentAnalysisandFuzzy C-MeansMethods,In:InternationalConferenceon BusinessIntelligence,Springer,Cham,pp.178–187, 2022.
4. N.RuhilandM.Singh,D.Mitra,A.Singh,andK.K. Singh. DetectionofchangesfromSatelliteImages UsingFusedDifferenceImagesandHybridKohonen FuzzyC-MeansSigm,Elsevier,ProcediaComputer Science,vol.167,pp.431-439,2020.
5. M.Singh,K.D.Tyagi,A.Singh,andK.K.Singh. DetectionofchangesinLandsatImagesusingHybrid PSO-FCM,Elsevier,ProcediaComputerScience, vol.167,pp.423-430,2020.
6. B.Cayla. Traitementd’images (partie3:Seuillage d’image), availableat https://www.datacorner.fr/image-processing-3/
7. M.SaravananandM.A.SanthoshSivan. Efficient& analysisofPCAbasedimagechangedetection algorithms,InternationalJournalofAdvanced
TechnologyinEngineeringandScience,vol.5,pp. 235-243,2017.
8. F.R.D.Velasco. ThresholdingusingtheISODATA Clusteringalgorithm,IEEETransactiononsystem, ManandCybernetics,vol.10,pp.771-774,1980.
9. T.Baakek. Segmentationetclassificationintélligente desimagescérébrales,Ph.D.dissertation,Dept. ElectricalandElectronicEng.,AbouBekrBelkaidUniv, Tlemcen,Algeria,2015.
10. J.Neira,CourseProgram, Lesson1:Thresholding, UniversityofZaragoza.
11. B.B.JacksonandB.Bund. MultivariateDataAnalysis: AnIntroduction,McGraw-Hill,1983.
12. R.D.Jeanson. Détectiondechangementsur l’occupationdusolàl’aidedelaméthodeACPetla logiquefloue,Memory,AntananarivoUniv.,R.D. Madagascar, 2015.
13. S.Panthan,andP.Thakre. FuzzyClusteringTechnique andPCABasedUnsupervisedChangeDetection MethodinMultitemporalSARImages,Vandana Publications,InternationalJournalofEngineeringand ManagementResearch,vol.7,pp.345-348,2016.
14. M.H.Kesikoglu,U.H.Atasever,andC.Ozkana. Unsupervisedchangedetectioninsatelliteimages usingfuzzyC-meansclusteringandprincipal componentanalysis,InternationalArchivesofthe Photogrammetry,RemoteSensingandSpatial InformationSciences,Vol.7,P;W2,2013.
15. J.C.Dunn. AFuzzyRelativeoftheISODATAProcess andItsUseinDetectingCompactWell-Separated Clusters,Taylor&Francis,Journalofcybernetics,vol3, pp.32-57,1973.
16. J.C.Bezdek,R.Ehrlich,andW.Full, FCM:The fuzzy c-meansclusteringalgorithm,Elsevier, Computers&Geosciences,vol.10,pp.191-203,1984.
17. Z.Wang,N.Zhao,W.Wang,R.Tang,andS.Li. AFault DiagnosisApproachforGasTurbineExhaustGas TemperatureBasedonFuzzyC-MeansClustering andSupportVectorMachine,Hindawi,Mathematical problemsinEngieering,vol.2015,2015.
18. T.T.H.Pham. Détectiondeschangementsde l'occupationdesterresdelazonelittorale:casdu districtdeTiênHai(Viêt-nam),M.S.thesis, SherbrookeUniv.,Montréal,Canada,2005.
19. A.A.AlAnalyst. ApprentissageSuperviséet classification,availableat https://fr.linedata.com/apprentissage-supervise-et-classif ication
20. H.BaudhuinandA.Lambot; LamboiseNet, Master ThesisaboutChangeDetectioninSatelliteImageryusing DeepLearning,availableat https://www.datacorner.fr/image-processing-3/