shoreline extraction using satellite imagery · shoreline extraction using satellite imagery...

BEACHMED-e/OpTIMAL BeachErosionMonitoring

81

Shoreline extraction using satellite imagery

Michalis Lipakis, Nektarios Chrysoulakis, Yiannis Kamarianakis

Shorelinechangeisconsideredtobeoneofthemostdynamicprocessesinthecoastalarea.Ithasbecomeimportanttomapshorelinechangesasaninputdataforcoastalhazardassessment.Inrecentyears,satelliteremotesensingdatahasbeenusedinshorelineextractionandmapping.Theaccuracyofimageorthorectification,aswellastheaccuracyofimageclassification,arethemostimportantfactorsaffectingtheaccuracyoftheextractedshoreline.Inthisstudy,theshore-lineoftheareaofGeorgioupoliswasmappedfortheyears1998and2005usingaerialimageryandIkonosdata,respectively.Ikonosdatawereorthorectifiedandafeatureextractiontechniquewasthenusedtoextracttheshoreline.Thistechniqueemployedmachine-learningalgorithmswhichexploitboththespectralandspatialinformationoftheimage.Resultswerevalidatedwithin situmeasurementsusingDifferentialGPS.Theanalysisshowedthattherehavenotbeenseverechangesintheshorelinebetween1998and2005,exceptinsomelocationswherechangewassubstantial.

IntroductionThecoastalareaisahighlydynamicenvironmentwithmanyphysicalprocesses,suchastidalflood-ing,sea level rise, landsubsidence,anderosion-sedimentation.Thoseprocessesplayan importantroleinshorelinechangeanddevelopmentofthecoastallandscape.Multi-yearshorelinemappingisconsideredtobeavaluabletaskforcoastalmonitoringandassessment.Theshorelineisdefinedasthelineofcontactbetweenlandandabodyofwater.Itiseasytodefinebutdifficulttocapture,sincethewaterlevelisalwayschanging.Therefore,aproblemexistsinthemappingcommunitybecausediffer-entpublicorprivateentitieshavecompiledandpublishedshorelinedelineationsthatarebasedondifferentshorelinedefinitions.Thishascreatedconfusionanduncertaintyforthosewhouseshorelineinformationdailyforthesakeofdecisionmaking,resourceplanning,emergencypreparednessetc.IntheUSAforexample,NOAAusesthetide-coordinatedshoreline,whichistheshorelineextractedfromaspecifictidewaterlevel.TheMLLW(MeanLowerLowWater)andMHW(MeanHighWater)areusedinthiswaytomapshorelinesthatcanbegeo-referenced.BoththeMLLWandMHWarecalculatedfromaveragesoveraperiodof18.6lunaryears(Lietal.,2001).Incontrast,theU.S.GeologicalSurvey(USGS)compilesshorelinedataforthe1:24.000scaletopographicbasemapseriesfromdigitalorthophotoquadranglescreatedfromphotographsthatarenottidecoordinated,therebymakingtheshorelineasnapshotintime(Scottetal.,2003).Itisthereforeobviousthatsinceshorelinehasadynamicnature,itsdefinition,mappingandmonitoringarecomplicatedtasks.Differentapproachestoshorelinemappingandchangedetectionhavebeenusedinthepast.Tradi-tionalshorelinemappinginsmallareasiscarriedoutusingconventionalfieldsurveyingmethods.ThemethodusedtodaybytheAmericanNationalGeodeticSurveytodelineatetheshorelineisanalyticalstereophotogrammetryusingtide-coordinatedaerialphotographycontrolledbykinematicGPStech-niques(Dietal.,2003).Landvehicle-basedmobilemappingtechnologyhasbeenproposedtotrace

Lipakisetal. Shorelineextractionusingsatelliteimagery

82

watermarksalongashorelineusingGPSreceiversandabeachvehicle.LiDARdepthdatahavealsobeenusedtomapshorelines(ShawandAllen,1995;Li,1997).Automaticextractionofshorelinefeaturesfromaerialphotoshasbeeninvestigatedusingneuralnet-worksandimageprocessingtechniques(Ryanetal.,1991).Photogrammetrictechniqueshavebeenemployedtomapthetide-coordinatedshorelinefromtheaerialimagesthataretakenwhenthewaterlevelreachesthedesiredlevel.Aerialphotographstakenatthesewaterlevelsaremoreexpensivetoobtainthansatelliteimagery.Besidesaerialimagery,space-bornradarandespeciallySyntheticApertureRadar(SAR)haveproventobeavaluabletoolforcoastalmonitoring.SARimageryhasalsobeenusedtoextractshorelinesatvariousgeographiclocations(Erteza,1998;ChenandShyu,1998;Trebossenetal.,2005;WuandLee,2007).SARisaverypromisingtechnology,especiallyforEuropesincetheEuropeanSpaceAgency(ESA)isrecognizedasaworldleaderinSARmissions(ERS1,ERS2,Envisat,GMES-Sentinel-1).Inrecentyears,opticalsatelliteremotesensingdatahavebeenusedinautomaticorsemi-automaticshorelineextractionandmapping.BraudandFeng(1998)evaluatedthresholdlevelslicingandmulti-spectralimageclassificationtechniquesfordetectionanddelineationoftheLouisianashorelinefrom30-meterresolutionLandsatThematicMapper(TM)imagery.TheyfoundthatthresholdingTMBand5wasthemostreliablemethodology.FrazierandPage(2000)quantitativelyanalysedtheclassificationaccuracyofwaterbodydetectionanddelineationfromLandsatTMdataintheWaggaregioninAus-tralia.TheirexperimentsindicatedthatthedensityslicingofTMBand5achievedanoverallaccuracyof96.9percent,which isassuccessfulasthe6-bandmaximumlikelihoodclassification.Scottetal.(2003)proposedasemi-automatedmethodforobjectivelyinterpretingandextractingtheland-waterinterface,whichhasbeendevisedandusedsuccessfullytogeneratemultipleshorelinedataforthetestStatesofLouisianaandDelaware.ThismethodwasbasedontheapplicationofTasseledCaptrans-formationcoefficientsderivedbytheEROSDataCenterforETM+dataasdescribedbyHuangetal.(2002).TheTasseledCaptransformationwaschosenoverothermethodsprimarilybecauseoftheob-jectiveandconsistentmannerinwhichitclassifiespixelsandbecauseitsuseallowedthecreationofotherusefulrasterbyproductfiles.Inoperation,theTasseledCaptransformationrecombinedspectralinformationofthe6ETM+bandsinto3principalviewcomponentsthroughtheuseofcoefficientsde-rivedbysamplingknownlandcoverspectralcharacteristics.Ofthethreeprincipalviewcomponentscreated,i.e.,Brightness,Greenness,andWetness,theWetnesscomponentisexploitedtodifferentiatelandfromwater.Zakariyaetal.(2006)triedtodetectshorelinechangesfortheTerengganurivermouthandrelatedcoastalarea.LandsatdatawereusedtogetherwithGIScapabilitytodetermineshoreline,sandyareaandthechangesthatoccurespeciallyonsedimentmovementfrom1996to2002.RGBtoIHSimageryconversionanalysisISODATA(IterativeSelf-OrganizingDataAnalysis)classificationwereemployed.LiuandJezek(2004),aswellasKarantzalosandArgialas(2007)automatedtheextractionofcoastlinefromsatelliteimagerybycannyedgedetectionusingdigitalnumber(DN)threshold.Lietal.(2001)comparedshorelinesofthesameareathatwereextractedusingdifferenttechniques,evaluated their differences and discussed the causes of possible shoreline changes. The differentshorelineproductshadbeengeneratedusingdifferenttechniques:bydigitisingfromaerialortho-photos, intersectingadigitalwatersurfacewithacoastalterrainmodelandextractingfromstereosatelliteimages.Inaddition,existingshorelinesdigitisedfromUSGSmapsandNOAAT-Sheetswereincludedintheiranalysis.Withthedevelopmentofremotesensingtechnology,satellitescancapturehighresolutionimagerywiththecapabilityofproducingstereoimagery.Thenewgenerationofveryhighspatialresolutionsatelliteimagingsystems,suchasIkonosandQuickbird,opensaneweraofearthobservationanddigitalmapping.Theyprovidenotonlyhigh-resolutionandmulti-spectraldata,butalsothecapabilityforstereomapping.Becauseoftheirhighresolutionandtheirshortrevisitrate(~3days),IkonosandQuickbirdsatelliteimagesareveryvaluableforshorelinemappingandchangedetection,thereforetheirdatahavebeenusedinseveralpaststudies.Wangetal.(2003)investigatedanovelapproachforautomaticextractionofshorelinefromIkonosimagesusingameanshiftsegmentationalgorithm.Dietal.(2003)investigatedanovelapproachforautomaticextractionofshorelinesfromIkonosimagery:


83

4mand1mresolutionIkonosimagesalongtheLakeErieshorewereused.Inthefirststeptheimagesweresegmentedintohomogeneousregionsbymeanshiftsegmentation.Then,themajorwaterbodywasidentifiedandaninitialshorelinewasgenerated.Thefinalshorelinewasobtainedbylocalrefine-mentwithintheboundariesofthecandidateregionsadjacenttotheinitialshoreline.Lietal.(2003)usedIkonosstereoimageryinshorelineextraction.TheypresentedtheresultsofanexperimentinwhichtheyattemptedtoimproveIkonosRationalFunctions(RF)forabettergroundaccuracyandtoemploytheimprovedRFfor3-Dshorelineextractionusing1-meterpanchromaticstereoimagesinaLakeEriecoastalarea.Inthismethod,a2DshorelineisextractedbymanualdigitisingononeIkonosimage;thencorrespondingshorelinepointsontheotherimageofthestereo-pairareautomaticallyextractedbyimagematching.The3Dshorelineiscomputedusingphotogrammetrictriangulation.Chalabietal.(2006)hadusedpixel-basedsegmentationonIkonosimageusingDNthreshold.Thepar-titionofthelandandseaboundarywasdoneusingpseudo-colourwhichexhibitsastrongcontrastbetweenlandandwaterfeatures.Shorelinechangeisconsideredtobeoneofthemostdynamicprocessesincoastalarea.Ithasbecomeimportanttomaptheshorelinechangeasaninputdataforcoastalhazardassessment.Therearemanychange detection techniques currently in use including visual interpretation, spectral-value-basedtechnique (differencing, image regression, DN value analysis), multi-data composites, and changevectoranalysis.Visualinterpretationofmulti-temporalimagesforcoastalmonitoringwaspresentedbyMazianetal.(1989)andElkoushyandTolba(2004).BagliandSoille(2003)analysedDNvalueus-ingslicingoperationforchangemonitoring.Inaddition,WhitheandElAsmar(1999)introducedanalgorithmfunctionandDNanalysis todeviate thewater fromthe land.TheDNvalueanalysishasalsobeenappliedonLandsatimages,e.g.byFrazierandPage(2000)andMarfai(2003).Fromardetal.(2004)identifiedcoastalchangesthattookplaceoverthelast50years,andrelatedthemtonaturalprocessesofturnoverandreplenishmentofmangroveforests.Theyusedacombinationofremotesensingtechniques (aerialphotographsandSPOTsatellite images)andfieldsurveys intheareaoftheSinnamaryEstuary,FrenchGuiana.Millsetal.(2005)introducedtheintegrationofthegeomat-icstechniquestoformaccuraterepresentationsofthecoastline.AhighlyaccurateDigitalElevationModel(DEM),createdusingkinematicsGPS,wasusedascontroltoorientatesurfacesderivedfromtherelativeorientationstageofphotogrammetryprocessing.MostafaandSoussa(2006)haveappliedGISandremotesensingtechniquetomonitorthelakeofNasserincludingtheshorelinedynamics.ThreesatelliteLandsatimagesforNasserLakewereavailableinatimeseries(1984,1996,and2001).Topographymapofscale1:50.000,thatissuitabletotheresolutionofLandsatimages,wasusedfordevelopingDEM.Chalabietal.(2006)assessedmulti-datasourcesformonitoringshorelineinKualaTerengganu,MalaysiausingIkonosandaerialphotographs.Resultsoftimeseriesdatawerecombinedtoeachothershowingspatialchangeofshoreline.Marfaietal.(2007)illustratedtheshorelinedynam-icsinacoastalareaofSemarang-Indonesiausingmultisourcespatialdata.Inspiteofthetechniqueandapproachtoshorelinemonitoringanddelineation,nosinglemethodhasbeenimplementedthatisfreefrommajordisadvantages.Therefore,shorelinesof thesameareamaybeextractedatdifferent timesusingsatellitedataandrepresentchangesthatappearedindifferenceperiodsthatareillustratedasdifferencesamongthem.Therearetwopossibleinterpretationsoftheshorelinedifferences.Oneisthattheshorelineindeedchanged in the realworld.Theotherpossibility is that thedifferencesare introducedbyshorelinemappingerrors.Theaccuracyoftheshorelinederivedfrom1meterIkonosimageryshouldbeabout2m-4m(ZhouandLi2000;Lietal.,2001,GrodeckiandDial,2003),consideringthefactthattheac-curacyof3Dgroundcontrolpoints(GCPs)reaches2m–3m,withGCPsandtheaccuracyofidentifyingandlocatingconjugateshorelinepointsisabout1.5pixels(1m-2m).Anoptimisticestimationoftheshorelineaccuracyderivedfromthe4-meterIkonosimagesinthisspecificcaseisabout8.5m(Lietal.,2001).Inmostoftheaforementionedmethods,theshorelineextractionusingIkonosorthoimageryisbasedonlandcoverclassificationtodiscriminatethepixelscorrespondingtowaterbodiesfromthosecor-respondingtoland.Following,theresultingthematicimageisconvertedtovectorcoverage,usually


84

apolygonshapefile(ESRI,2005)containingthepolygonscorrespondingtoeachclass.Theshorelineisfinallyextractedfromthepolygonthatcorrespondstowaterbyemployingautomaticorsemi-auto-maticGISprocedures.Thus,theaccuracyoftheimageorthorectification,aswellastheaccuracyoftheimageclassification,isthemostimportantfactorsaffectingtheaccuracyoftheextractedshoreline.Theorthorectificationaccuracywasdiscussedabove.Concerningclassificationaccuracy,itdependsonthespatial,spectralandradiometricresolutionoftheimage,aswellasontheclassificationmethod.Numerousstudieshavebeencarriedoutusingsatelliteimagestoextractlandcovertypes(Congalton,1991;RiddandLiu,1998;Martinetal.,1988;GongandHowarth,1990;Chrysoulakis,2003;Gallego,2004).Themajorityofthepaststudiesrelyonremotesensingdatatoclassify landcovertypesus-ingeitherrawDNorcalibratedradiancevalues.However,ifveryhighspatialresolutiondatasuchasIkonosimagesareused,thelandcoverclassificationofcoastalareasmaybeproblematicbecauseoftheheterogeneityandsmallspatialsizeofthesurfacematerials,whichleadstosignificantsub-pixelmixing(Foody,2000;Kontoesetal.,2000).Therefore,thespatialcontextshouldbetakenintoaccountinimageclassificationandobjectorientedalgorithmsshouldbeused.Improvementsintheaccuracyofclassificationhavebeenachievedusingavarietyofsophisticatedapproachesincludingtheuseofneuralnetworks (Berberogluetal.,2000), fuzzy logic (Bastin,1997;ZangandFoody1998;), textureanalysis(Stuckensetal.,2000),machinelearning(VLS,2007)andincorporationofancillaryspatialdataintheclassificationscheme(HarrisandVentura,1995;Vogelmannetal.,1998,Stefanovetal.,2001).

MethodologyTheIkonossatelliteprovidesglobal,accurate,high-resolutionimageryformapping,monitoring,anddevelopment.Thepanchromaticsensorwith82cmresolutionandan11.3kmwideswathatnadirpro-videshighresolution,intelligence-qualityimagery.Themultispectralsensor,simultaneouslycollect-ingblue,green,red,andnearinfraredbandswith3.28mresolutionatnadir,providesnatural-colourimageryforvisualinterpretationandcolour-infraredimageryforremotesensingapplications.Com-biningthemultispectralimagerywiththehighresolutionpanchromaticresultsin1-metercolourimages(pan-sharpenproduct),whichcanbeorthorectifiedafterwards.Theorthorectificationisneed-edtoeliminatethegeometricdistortions,whichwillbeexplainedbelow,sothatimagefeatureshavecorrectplanimetriccoordinates.Quantitativeestimationssuchasshorelinedetectionareperformedusingorthorectifiedimages.Apartfromthedifferenttechniquesthatcanbeappliedforshorelineextractionandmonitoringfromhighresolutionsatelliteimages,theprocessingchainconsistsofthefollowingbasicsteps:· acquisitionofimagesandpre-processing;· acquisitionofthegroundControlPoints(GCPs)withimagecoordinatesandmapcoordinates;· computationoftheunknownparametersofthemathematicalfunctionsusedforthegeometric

correctionmodel;· imageorthorectificationusinganappropriateDEM;· automatic,semi-automaticormanualshorelineextractionfromtheorthorectifiedimagery;· monitoring of shoreline changes by repeating the above steps at predefined time periods and

comparingtherelativepositionsoftheextractedshorelines.Thus,beforetheapplicationofanyalgorithmforautomaticextractionofshorelinefrommultispectralsatelliteimages,theseimagesshouldbeothorectifiedtotakeintoaccountthegeometricdistortionsduringimageacquisition,aswellastheeffectoftopography.Eachimageacquisitionsystemproducesuniquegeometricdistortionsinitsrawimagesandconsequentlythegeometryoftheseimagesdoesnotcorrespondtotheterrainorofcoursetoaspecificmapprojection.Obviously,thegeometricdis-tortionsvaryconsiderablywithdifferentfactorssuchastheplatform,thesensorandalsothetotalfieldofview.However,asithasbeendescribedbyToutin(2004),itispossibletomakegeneralcat-egorisationsofthesedistortions.Thesourcesofdistortioncanbegroupedintotwobroadcategories:theobserverortheacquisitionsystem(platform,imagingsensorandothermeasuringinstruments,suchasgyroscope,stellarsensors,etc.)andtheobserved(atmosphereandEarth).Inadditiontothesedistortions,thedeformationsrelatedtothemapprojectionhavetobetakenintoaccountbecause


85

theterrainandmostGISend-userapplicationsaregenerallyrepresentedandperformedrespectivelyinatopographicspaceandnotinthegeoidorareferencedellipsoid.Mostofthesegeometricdistor-tionsarepredictableorsystematicandgenerallywellunderstood.Someofthesedistortions,espe-cially those related to the instrumentation, are generally corrected at ground receiving stations orbyimagevendors.Others,forexamplethoserelatedtotheatmosphere,arenottakenintoaccountandcorrectedbecausetheyarespecifictoeachacquisitiontimeandlocationandinformationontheatmosphereisrarelyavailable.Theremaininggeometricdistortionsrequiremodelsandmathematicalfunctionstoperformgeometriccorrectionsofimagery:eitherthrough2D/3Dempiricalmodels(suchas2D/3Dpolynomialor3DRF)orwithrigorous2D/3Dphysicalanddeterministicmodels.With2D/3Dphysicalmodels,whichreflectthephysicalrealityoftheviewinggeometry(platform,sensor,Earthandsometimesmapprojection),geometriccorrectioncanbeperformedstep-by-stepwithamath-ematicalfunctionforeachdistortion/deformation,orsimultaneouslywithacombinedmathematicalfunction.2D/3Dphysicalfunctionsusedtoperformthegeometriccorrectiondiffer,dependingonthesensor,theplatformanditsimageacquisitiongeometry(Toutin,2004):· instantaneousacquisitionsystems,suchasphotogrammetriccameras,MetricCameraorLargeFor-

matCamera;· rotatingoroscillatingscanningmirrors,suchasLandsat-MSS,TMandETM+;· push-broomscanners,suchasSPOT-HRV,IRS-1C/D,IkonosandQuickbird;and· SARsensors,suchasJERS,ERS-1/2,RADARSAT-1/2andEnvisat.Whateverthegeometricmodelused,evenwiththeRFsomeGCPshavetobeacquiredtocompute/refinetheparametersofthemathematicalfunctionsinordertoobtainacartographicstandardaccu-racy.Generally,aniterativeleast-squareadjustmentprocessisappliedwhenmoreGCPsthanthemini-mumnumberrequiredbythemodel(asafunctionofunknownparameters)areused.ThenumberofGCPsisafunctionofdifferentconditions:themethodofcollection,sensortypeandresolution,imagespacing,geometricmodel,studysite,physicalenvironment,GCPdefinitionandaccuracyandthefinalexpectedaccuracy.Theaerialtriangulationmethodhasbeendevelopedandappliedwithdifferentopticalandradarsatellitedatausing3Dphysicalmodels(Toutin,2003a,b),aswellaswithIkonosdatausing3DRFmodels(Fraseretal.,2002a,b).Allmodelparametersofeachimage/striparedeterminedbyacommonleast-squaresadjustmentsothattheindividualmodelsareproperlytiedinandanentireblockisoptimallyorientedinrelationtotheGCPs.Asithasbeenalreadymotioned,shorelineextractionneedsorthorectifiedimages.Torectifytheorigi-nalimageintoamapimage,therearetwoprocessingoperations:· ageometricoperationtocomputethecellcoordinatesintheoriginalimageforeachmapimage

cell,eliminatingthegeometricdistortionsaspreviouslyexplained;and· aradiometricoperationtocomputetheintensityvalueorDNofthemapimagecellasafunctionof

theintensityvaluesoforiginalimagecellsthatsurroundthepreviously-computedpositionofthemapimagecell.

Thegeometricoperationrequirestheobservationequationsofthegeometricmodelwiththeprevi-ouslycomputedunknowns,andsometimeselevationinformation.The3DmodelstakeintoaccountelevationdistortionandDEMsisthusneededtocreatepreciseorthorectifiedimages.DEMsimpactontheorthorectificationprocess,bothintermsofelevationaccuracyforthepositioningaccuracyandofgridspacingforthelevelofdetails.Thislastaspectismoreimportantwithhigh-resolutionimagesbecauseapoorgridspacingwhencomparedtotheimagespacingcouldgenerateartefactsforlinearfeaturessuchasshorelines.Foranymapcoordinates(x,y),withthezelevationextractedfromaDEMwhen3Dmodelsareused,theoriginalimagecoordinates(columnandline)arecomputedfromthetworesolvedequationsofthemodel.However,thecomputedimagecoordinatesofthemapimagecoordinateswillnotdirectlyoverlayintheoriginalimage;inotherwords,thecolumnandlinecom-putedvalueswillrarely,ifever,beintegervalues.Sincethecomputedcoordinatevaluesintheoriginalimagearenotintegers,onemustcomputetheDNtobeassignedtothemapimagecell.Inordertodothis,theradiometricoperationusesaresamplingkernelappliedtooriginalimagecells:eithertheDN


86

oftheclosestcell(callednearestneighbourresampling)oraspecificinterpolationordeconvolutionalgorithmusingtheDNsofsurroundingcells(Toutin,2004).InordertoaccuratelycreateorextractgeographicinformationfromrawIkonosimagery,theImageGeometryModel(IGM)mustaccompanytheimagery.TheIGMconsistsofseveralmetadatafileswhichcontainRPCs(rationalpolynomialcoefficients).TheRPCsareaseriesofcoefficientsthatdescribetherelationshipbetweentheimageasitexistedwhencapturedandtheEarth’ssurface.Althoughtheydonotdescribesensorparametersexplicitly,RFsaresimpletoimplementandperformtransforma-tionsveryrapidly.WiththeavailabilityofRPCs,theIkonosinteriorandexteriororientationsareveryaccurate.ThereforeIkonosimagerycanbeorthorectifiediftheIGM,anaccurateDEMandsomeGCPsare available by employing any photogrametric software such as Orthoengine (PCI, 2003) or LeicaPhotogrammetrySuite(Leica,2005).Thenextstepforshorelineextractionisthewater-landseparation;thereforetheorthorectifiedimageshouldbeclassifiedorapolygoncorrespondingtowater(orland)areashouldbeextracted.Takingintoaccounttheaforementionedlandcovermappingconstraintsforveryhighspatialresolutionsatel-litedata,amachinelearningclassifierapproachseemsthebestsolutionforIkonosmultispectralimageclassification.Thistypeofclassifierusesaninductivelearningalgorithmtogenerateproductionrulesfromtrainingdata.Aswithaneuralnetwork,thereareseveraladvantagestousingamachine-learn-ingapproach.Sinceancillarydatalayersmaybeusedtohelpimprovediscriminationbetweenclasses,fewerfieldsamplesaregenerallyrequiredfortraining.Thismachinelearningmodelisnon-parametricanddoesnotrequirenormally-distributeddataorindependenceofattributes.Itcanalsorecognizenonlinearpatternsintheinputdatathataretoocomplexforconventionalstatisticalanalysesortoosubtletobenoticedbyananalyst.FeatureAnalystsoftware(VLS,2007)wasselectedforshorelineextractionfromIkonosimagery,sinceitemploysmachine-learningtechniqueswhichhavethepoten-tialtoexploitboththespectralandspatialinformationoftheimage.Itprovidesaparadigmshifttoautomatedfeatureextractionsinceit:(a)utilisesspectral,spatial,temporal,andancillaryinformationtomodel the featureextractionprocess, (b)provides theability to removeclutter, (c) incorporatesadvancedmachinelearningtechniquestoprovideunparalleledlevelsofaccuracy,and(d)providesan

exceedingly simple interfaceforfeatureextraction.Itworksby taking a small and sim-ple set of training examples,learnsfromtheexamples,andclassifiestheremainderoftheimage. When classifying thecontentsofimagery,thereareonly a few attributes acces-sible to human interpreters.Foranysinglesetof imagerytheseare:Shape,Size,Colour,Texture,Pattern,Shadow,andAssociation. Traditional im-age processing techniquesincorporate only colour(spectral signature) and per-haps texture or pattern intoan involved expert workflowprocess.Theshorelineextrac-tionstepsusingFeatureAna-lystareshowninFigure1.

Figure 1 - Shoreline extraction work flow (adapted from VLS, 2007).


87

Case Study: Shoreline extraction in the area of Georgioupolis, CreteTheShorelineextractionfortheareaofGeorgioupoliswasperformedfortheyears1998and2005usingaerial imageryandIkonosdata,respectively.Theresultsvalidatedwith in situmeasurementswithDifferentialGPS(DGPS)providedbyOANAK.TheHellenicGeodeticReferenceSystemof1987(EGSA87)wasusedinallcases.

Shoreline extraction using an aerial image acquired in 1998AnorthorectifiedaerialimageandaDEMofthebroaderareaofGeorgioupoliswereavailabletoOANAKaspastprojectproducts.Theaerialimagehasthespatialresolutionof1m,whereasitspositionalac-curacywasbetterthan2m(RMSExy<2m).ItisshowninFigure2,wheretheroadnetwork(producedbyDGPSmeasurements)oftheareahasbeensuperimposed(redlines).FeatureAnalystwasemployedtoclassifytheaerialorthoimageinto2classes:landandwater.Theextractionoftheshorelinewasafter-wardsstraightforwardfromthepolygoncorrespondingtowaterclass,usingtheArcGISsoftware(ESRI,2005).ThelatterwillbealsoappliedtoIkonosimage,soitwillbedescribedinmoredetailbelow.The1998shorelineasextractedfromtheaerialimageisalsoshowninFigure2(yellowline).

Shoreline extraction using an Ikonos image acquired in 2005The 2005 shoreline was ex-tracted from an Ikonos mul-tispectral image. ERDAS Im-agine (Leica, 2005) was usedtopre-processtheimage;LPS(Leica, 2005) was used to or-thorectify the image; FeatureAnalyst (VLS, 2007) was usedto extract the shoreline andArcGIS(ESRI,2005)wasusedtofine-tunetheextractedshore-line.The processing chain in-cludedthefollowingsteps:

Step 1: Acquisition of images and pre-processing.Apanchromatic(PAN)andafourband(R,G,B,NIR)multispectralIkonosimagetogetherwithIGMandRPCswasprovidedbytheimagevendorasRawGeoProduct.ThefourchannelsareshowninFigure3.ERDAS Imaginewasused tomerge thepanchromaticandthemultispectral images toprovidea

four-band pan-sharpen productwithspatialresolutionof1m.TheWavelet Resolution Merge func-tion (Leica, 2005) was used.Theresultingpan-sharpenproductisshowninFigure4asapseudocol-ourcompositionRGB:4-3-2.

Figure 3 - Ikonos spectral channels of the 2005 acquisition over the area of Georgioupolis: a) blue band; b) green band; c) red band; d) near infrared band. The strong absorption of water in the near infrared band is obvious in d), therefore this band is the most use-ful for the discrimination between land and water.

Figure 2 - Orthorectified aerial image acquired in 1998. The road network (red) and the extracted shoreline (yellow) have been superimposed.


88

Step 2: Acquisition of GCPs.AfieldcampaignwasorganizedbyOANAKand15GCP’swereselectedusingadifferentialGPS.Thelocationofthesepoints(circles)superimposedtotheroadnetworkofthestudyareaisshowninFig-ure5.Theirpositionalaccuracywasbetterthan1m(RMSExy<1m).SeveralotherGCPswererandomlyselectedtobeusedforvalidationoftheorthorectificationprocedure.

Figure 5 - Location of the meas-ured GCPs superimposed to the road network of the study area.

Step 3: Computation of the unknown parameters of the mathematical functions used for the geometric correction model.TheinteriorandexteriororientationswereautomaticallycomputedusingLPSsincetheIGMandtheGCP’swereavailable.Theaerial triangulationprocedure (Leica,2005)was followed.Thisprocedureeliminatedthegeometricerrorswhichhavebeenpreviouslydiscussed,except fortheeffectof to-pography.ThiseffectwastakenintoaccountbyusingaDEMofthestudyareaavailableatOANAK’s.ThisDEMhadbeenproducedphotogrammetricallyusingastereo-pairofaerialimagesfromthe1998campaign.TheverticalaccuracyoftheDEMwasbetterthan1m(RMSEz<1m).TheDEMwiththeroadnetworkoftheareasuperimposedisshowninFigure6.

Figure 6 - The photogrammetri-cally derived DEM with the road network and the shoreline ex-tracted from the aerial image superimposed.

Step 4: Orthorectification of images.SinceanaccurateDEMwasavailableandtheaerialtriangulationparameterswerecomputedbyLPS,theorthorectificationprocedurewasalsostraightforwardasageometricandaradiometricoperation,


89

asithasbeenalreadydescribed.TheorthorectificationresultisshowninFigure7asapseudocolourcompositionRGB:3-2-1,withtheroadnetworkofthearea(redlines)andtheshorelineextractedfromtheaerialimage(yellowline)superimposed.Thespatialresolutionoftheothorectifiedimageis1m,whereasitspositionalaccuracyisbetterthan2m(RMSExy<2m),therefore,asithasbeenexplained,theexpectedaccuracyoftheshorelinetobeextractedisaround4m.

Figure 7 - Pseudocolour composi-tion RGB: 3-2-1 of the orthorecti-fied Ikonos pan-sharpened im-age with the road network (red lines) and the shoreline extract-ed from the aerial image (yellow line) superimposed. Differences in shoreline between 1998 and 2005 are obvious.

Step 5: Semi-automatic shoreline extraction using Feature Analyst.TheextractionoftheshorelineusingFeatureAnalystisasemi-automaticprocedure,asshowninFigure1,involving:· thecreationofanewsetcontainingtwotrainingclasses;· thepreparationofamulti-classinputlayer;· thesetupofthelearningparametersandtheexecutionofthealgorithm;· thesplittingofthetwoclassesintoafinalpolygon;· thepostprocessing,forsmoothingoftheresultpolygon;· theconversionofthepolygonintoaline.Thefirstactionistocreatethetrainingsetofpolygons.Thisisthemostimportantaction,becausethequalityoftheresultsisdependantonthequalityofthetrainingsets.Thetrainingsetwillconsistoftwoclasses,onethatcorrespondstowaterelementsandanotheronethatcorrespondstolandelements.At thispoint, thesetwoclassesare indicatedas twoseparatesetsofpolygonsandthencombinedintoamulti-classinputlayer.Creatingthewatersetofpolygons,includesanumberoftrials,tofinallydeicideandconcludetothemostrepresentativeset.Theresultofthesetrialsisasetoffourpolygons,

three of which are locatedat the edge of the watermass, extending throughthe inshore, as shown inFigure8 (purple).Thesameprocedureisfollowedtode-cideaboutthelandset.Theresultisacorrespondingsetofsixpolygons(blue).

Figure 8 - Pseudocolour composi-tion RGB: 3-2-1 of the orthorecti-fied Ikonos pan-sharpened image with the water (purple) and land (blue) training polygons super-imposed.


90

Inordertorunamulti-classextraction,thetwodifferentclassesofwaterandlandhavetobecom-binedintoasingle,multi-classlayer.Thereisabuilt-infunctionofFeatureAnalystforthis.Thepro-ducedtwo-classlayeristheinputofthealgorithm.Havingtheinput,thenextactionistosetthealgorithm’sparameters(VLS,2007).Allfourimagebandsareabouttobeincludedandtheimageresolutionissetto1m.Thespectralinformationisderivedfromthismultispectralimage,whereasthespatialcontextforeachpixelistakenintoaccountbyad-justingtheinputrepresentationoftheclassificationalgorithm.Theinputrepresentation,thatdeter-mineshoweachpixelislookedatinrelationtoitsneighbours,issetto“Manhattan”representation(VLS,2007).Manhattanisapre-definedinputpattern,usedmainlyforwatermassandlandcoverfea-tures,likeoceans,lakes,floods,wetland,impermeablesurfaces,etc.Thepatternwidthissetto5.Thismeansthatconsideringthatthe“decisionpixel”(redinFigure9)isinthecentreofa5×5grid,accordingtotheManhattaninputrepresentation,thealgorithmwilltakeintoaccount13pixels(blueinFigure9),locatedasshowninFigure9.Computing13pixelsofeachband,thereisatotalof52pixelsthatwillbecomputedtomakeadecisionforasinglepixel.Theminimumaggregateareaissetto500pixels.Thatistomaintainrelativefeaturecharacteristics,whiletryingtofindalargearea.

Figure 9 - The ‘Manhattan’ input representation that is used, with pattern width set to 5.

TheclassificationresultisshowninFigure10.Itisagainamulti-classlayer,whichneedstobesplit.Althoughtheborderofbothresultclassesisthesameline(theshoreline),thewaterclassischosenfortheextraction.Thebuilt-infunctioninFeatureAnalystthatsplitsoutclasseswasusedtosplittheclassesandkeeponlythewaterclassasaseparateset.Sinceonlyonepolygonhasbeenleftitsborderwassmoothedinordertoextracttheshoreline.The“BezierSmoothAlgorithm”(VLS,2007)wasused,withthefollowingparameters:thenumberofverticestoeachsidesetto2andthemaximumdistanceeachvertexisallowedtomovesetto3m.TheresultofthesmoothedpolygonisshowninFigure11.

Figure 10 - The land – water clas-sification result: a two class result corresponding to the two-class input. Both spectral and spatial information have been taken into account.


91

Figure 11 - The extracted – smoothed polygon correspond-ing to water.

Finally,theborderofthewaterpolygonwasautomaticallyextractedandconvertedtoalineshapefilefollowingastandardGISprocedure(VLS,2007;ESRI,2005).Thislineshapefileisthefinalresultoftheshorelineextractionprocedureas it isshowninFigure12,wheretheextractedshorelinehasbeensuperimposedontheorthorectifiedIkonosimage.

Figure 12 - Pseudocoloured composition RGB: 3-2-1 of the orthorectified Ikonos image with the extracted shoreline (yellow line) superimposed.

Step 6: Comparisons - Validation.The produced shorelines (1998 Shoreline, and 2005 Shoreline) were compared and a Root MeanSquareError(RMSE)wascomputedtoreflecttheshorelinechangeduringthis7yearsperiod.RMSEisaglobalmeasure,thusthemaximumchangewashighlightedanditispresentedbelow.TheRMSEwasusedasaquantitativeevaluationoftheextractedshorelinesaccuracy.RMSEencompassesbothsystematicandrandomerrorsandisdefinedas[1]:

[1]

Where:δsi=minimumdistanceofthetwolinesatapre-definedlocationi(x,y),n=numberofmeasuringlocationsinthisstudy.


92

Finally,boththeimage-derivedshorelineswerecomparedwithanin situderivedshorelinewhichwasproducedbyOANAKwiththeuseofdifferentialGPS.ThethreelinesareshowninFigure13,whereapartofthestudyareahasbeenextracted(scale1:2.000):The in situderivedshorelineispresentedinred,the1998Shorelineispresentedingreenandthe2005Shorelineispresentedinyellow.TheRMSEcalculatedusingthein situlineasbaseline,aswellastheRMSEbetweentheex-tractedshorelinesfor1998and2005,areshowninTable1.

Figure 13 - Part of the study area as a pseudocolour composition RGB: 3-2-1 of the orthorectified Ikonos image (scale 1:2000). The in situ derived shoreline is shown in red, the 1998 Shoreline is shown in green, whereas the 2005 Shoreline is shown in yellow.

Theanalysisshowedthattherewerenotseverechangesinshorelinebetween1998and2005,ex-ceptinsomelocationswherethechangewassubstantial.ThemostimportantchangeisshowninFigure14,whereapartofthestudyareaaroundtheriverwhichisclosetothetownofGeorgioupolishasbeenextracted(scale1:2.000):bothlineshavebeensuperimposedontheIkonosorthorectifiedimage.The1998Shorelineispresentedingreenandthe2005Shorelineispresentedinyellow,asbefore.ItisobviousfromFigure14thattheoutfalloftheriverhasbeenmodifiedduringthis7yearperiod.Ashiftofabout50mtotheESEdirectioncanbeobservedinFigure14.

Table 1 - RMSE between a) in situ and aerial image derived shoreline (1998); b) in situ and Ikonos derived shoreline (2005); c) aerial image and Ikonos derived shoreline.

Pair of Lines RMSE (m)

Insitu–1998Shoreline 3.01

Insitu–2005Shoreline 5.65

1998Shoreline–2005Shoreline 6.46


93

Figure 14 - Part of the study area as a pseudocolour composition RGB: 3-2-1 of the orthorectified Ikonos image (scale 1:2000). The 1998 Shoreline is shown in green and the 2005 Shoreline is shown in yellow. A shift of about 50m to the ESE direction of the river outflow is observed.

ReferencesBagliS.andSoilleP.(2003)-Morhological automatic extraction of Pan-European coastline from Landsat

ETM+ images.InternationalSymposiumonGISandComputerCartographyforCoastalZoneMan-agement.Genova,Italy.

BastinL.(1997)Comparison of fuzzy c-means classification, linear mixture modeling and MLC probabilities as tools for unmixing coarse pixels.InternationalJournalofRemoteSensing,18:3629–3648.

BerberogluS.,LloydC.D.,AtkinsonP.M.andCurranP.J.(2000)-The integration of spectral and textural information using neural networks for land cover mapping in the Mediterranean.Computers&Geo-sciences,26:385–396.

BraudD.H.andFengW.(1998)- Semi-automated construction of the Louisiana coastline digital land/wa-ter boundary using Landsat Thematic Mapper satellite imagery.DepartmentofGeography&Anthro-pology,LouisianaStateUniversity.LouisianaAppliedOilSpillResearchandDevelopmentProgram,OSRAPDTechnicalReportSeries97-002.

ChalabiA.,Mohd-LokmanH.,Mohd-SuffianI.,KaramaliK.,KarthigeyanV.andMasitaM(2006)-Monitor-ing shoreline change using Ikonos image and aerial photographs: a case study of Kuala Terengganu area, Malaysia.In:ProceedingsoftheISPRSMid-termSymposiumProceeding.Enschede,TheNetherlands.

ChenL.C.andShyuC.C.(1998)-Automated Extraction of Shorelines from Optical and SAR Image.In:Pro-ceedingofAsianConferenceonRemoteSensing(http://www.gisdevelopment.net/aars/acrs/1998/ps3/ps3013a.asp).

ChrysoulakisN.(2003)-Estimation of the all-wave urban surface radiation balance by use of ASTER mul-tispectral imagery and in situ spatial data.JournalofGeophysicalResearch,108:4582(DOI:10.1029/2003JD003396).

CongaltonR.G.(1991)-A Review of Assessing the Accuracy of Classifications of Remotely Sensed Data.RemoteSensingofEnvironment,37:35-46.


94

DiK.,WangJ.,MaR.andLiR.(2003)-Automatic shoreline extraction from high-resolution Ikonos satellite imagery.InproceedingoftheAnnualASPRSConference.Anchorage,Alaska.

ElkoushyA.A.andTolbaR.A.(2004)-Prediction of shoreline change by using satellite aerial imagery.In:ProceedingoftheXXISPRSCongress.Istanbul,Turkey.

ErtezaI.A.(1998)-An automatic coastline detector for use with SAR images.SANDIAReportSAND98-2102.SandiaNationalLaboratories,California.

ESRI(2005)-ArcGIS 9. What is ArcGIS 9.1?EnvironmentalSystemsResearchInstitute,California.FoodyG.M.(2000)-Estimation of sub-pixel land cover composition in the presence of untrained classes.

Computers&Geosciences,26:469–478.FraserC.S.,HanleyH.B.andYamakawaT.(2002a)-Three-dimensional geopositioning accuracy of Ikonos

imagery.PhotogrammetricRecord,17:465–479.FraserC.S.,BaltsaviasE.andGruenA.(2002b)-Processing of Ikonos imagery for sub-metre 3D positioning

and building extraction.ISPRSJournalofPhotogrammetryandRemoteSensing,56,177–194.FrazierP.S.andPageK.J. (2000)-Water body detection and delineation with Landsat TM data.Photo-

grammetricEngineeringandRemoteSensing,66:147–167.FromardF.,VegaaC.andProisyC.(2004)-Half a century of dynamic coastal change affecting mangrove

shorelines of French Guiana. A case study based on remote sensing data analyses and field surveys.Ma-rineGeology,208:265–280.

GallegoF.J. (2004) -Remote sensing and land cover area estimation. InternationalJournalofRemoteSensing,25:3019-3047.

GongP.andHowarthP.J.(1990)-The use of structural information for improving land-cover classification accuracies at the rural – urban fringe.PhotogrammetricEngineeringandRemoteSensing,56:67-73.

GrodeckiJ.andDialG.(2003)-Block adjustment of highresolution satellite images described by rational polynomials.PhotogrammetricEngineeringandRemoteSensing,69:59-68.

HarrisP.M.andVenturaS.J.(1995)-The integration of geographic data with remotely sensed imagery to improve classification in an urban area.PhotogrammetricEngineeringandRemoteSensing,61:993–998.

HuangC.,WylieB.,HomerC.,YangL.andZylstraG.(2002)-Derivation of a tasseled cap transformation based on Landsat 7 ETM at-satellite reflectance. InternationalJournalofRemoteSensing,23:1741-1748.

KarantzalosK.andArgialasD.(2007)-Automatic shoreline mapping from panchromatic satellite images.In:Proceedingsof8thPan-HellenicGeographicConference.GreekGeographicalSociety(inGreek).

KontoesC.C.,RaptisV.,LautnerM.andOberstadlerR.(2000)-The potential of kernel classification tech-niques for land use mapping in urban areas using 5 m-spatial resolution IRS-1C imagery.InternationalJournalofRemoteSensing,21:3145–3151.

Leica(2005)-ERDAS Field Guide.LeicaGeosystems.Atlanta,Georgia,USA.LiR.,DiK.andMaR.(2001)-A Comparative Study of Shoreline Mapping Techniques.In:Proceedingof

the4thInternationalSymposiumonComputerMappingandGISforCoastalZoneManagement.Halifax,NovaScotia,Canada.

LiR.,DiK.andMaR.(2003)-3-D Shoreline Extraction from Ikonos Satellite Imagery.MarineGeodesy,26:107–115.

LiR. (1997) -Mobile mapping: An emerging technology for spatial data acquisition.PhotogrammetricEngineeringandRemoteSensing,63:1165-1169.

LiuH.andJezekK.C.(2004)-Automated extraction of coastline from satellite imagery by integrating can-ny edge detection and locally adaptive thresholding methods.InternationalJournalofRemoteSens-ing,25:937–958.

MarfaiM.A.,AlmohammadH.,DeyS.,SusantoB.andKingL.(2007)-Coastal dynamic and shoreline mapping: multi-sourcesspatial data analysis in Semarang Indonesia.EnvironMonitoringandAssess-ment,DOI10.1007/s10661-007-9929-2.

MarfaiM.A.(2003)-Monitoring of the coastal zone dynamics by means of multi-temporal Landsat TM.In:ProceedingsoftheannualscientificmeetingXII,Indonesianremotesensingsociety.Bandung.


95

MartinL.R.G.,HowarthP.J.andHolderG.(1988)-Multispectral classification of land use at the rural-urban fringe using SPOT data.CanadianJournalofRemoteSensing,14:72–79.

MazianH.I.,AzizI.andAbdullahA.(1989)-Preliminary evaluation of photogrammetric-remote sensing approach in monitoring shoreline erosion.In:Proccedingsofthe10thAsianConferenceonRemoteSensingProceeding.KualaLumpur.

MillsJ.P.,BuckleyS.J.,MitchellH.L.,ClarkeP.J.andEdwardsS.J. (2005)-A geomatics data integration technique for coastal change monitoring.EarthSurfaceProcessesandLandforms,30:651–664.

MostafaM.M.andSoussaH.K. (2006) -Monitoring of Lake Nasser using remote sensing and GIS tech-niques.In:ProceedingsoftheISPRSMid-termSymposium.Enschede,TheNetherlands.

PCI(2003)-OrthoEngine User Guide.PCIGeomatics.Ontario,Canada.RiddM.K.andLiuJ.(1998)-A comparison of four algorithms for change detection in an urban environ-

ment.RemoteSensingofEnvironment,63:95–100.RyanT.W.,SementilliP.J.,YuenP.andHuntB.R.(1991)-Extraction of shoreline features by neural nets and

image processing.PhotogrammetricEngineeringandRemoteSensing,57:947-955.ScottJ.W.,MooreL.R.,HarrisW.M.andReedM.D.(2003)-Using the Landsat 7 Enhanced Thematic Mapper

Tasseled Cap Transformation to Extract Shoreline.Open-FileReportOF03-272U.S.GeologicalSurvey.ShawB.andAllenJ.R.(1995)-Analysis of a dynamic shoreline at Sandy Hook, New Jersey, using a geo-

graphic information system.In:ProceedingsofASPRS/ACSM.pp.382-391.StefanovW.L.,RamseycM.S.andChristensenP.R.(2001)-Monitoring urban land cover change: An ex-

pert system approach to land cover classification of semiarid to arid urban centers.RemoteSensingofEnvironment,77:173–185.

StuckensJ.,CoppinP.R.andBauerM.E.(2000)-Integrating contextual information with per-pixel classifi-cation for improved land cover classification.RemoteSensingofEnvironment,71:282–296.

ToutinTh.(2004)-Review article: Geometric processing of remote sensing images: models, algorithms and methods.InternationalJournalofRemoteSensing,25,1893–1924

ToutinTh.(2003a)-Block bundle adjustment of Ikonos in-track images.InternationalJournalofRemoteSensing,24:851–857.

ToutinTh.(2003b)-Error tracking in Ikonos geometric processing using a 3D parametric modelling.Pho-togrammetricEngineeringandRemoteSensing,69:43–51.

TrebossenH.,DeffontainesB.,ClasseauN.,KouameJ.andRudantJ.-P.(2005)-Monitoring coastal evo-lution and associated littoral hazards of French Guiana shoreline with radar images.ComptesRendusGeosciences,337:1140-1153.

VogelmannJ.E.,SohlT.andHowardS.M.(1998)-Regional characterization of land cover using multiple sources of data.PhotogrammetricEngineeringandRemoteSensing,64:45–57.

VLS (2007) - Feature Analyst Version 4.1 for Imagine.ReferenceManual.VisualLearningSystems Inc.Missoula,USA.

WangK.D.J.,MaR.andLiR.(2003)-Automatic shoreline extraction from high resolution Ikonos satellite imagery.In:ProceedingsoftheASPRSAnnualConference.Anchorage,Alaska.

WhitheK.andElAsmarH.M.(1999)-Monitoring changing position of coastline using thematic mapper imagery, an example from the Nile Delta.Geomorphology,29:93–105.

Wu T.D. and Lee M.T. (2007) - Geological Lineament and Shoreline Detection in SAR Images.Proceedings of IEEE Geosciences and Remote Sensing Symposium (IGARSS 2007). Barcelona,Spain.

ZakariyaR.,RosnanY.,SaidinS.A.,YahayaM.H.,KasawaniI.andLokmanH.(2006)-Shoreline detection and changes for Terengganu river mouth from satellite imagery (Landsat 5 and Landsat 7).JournalofSustainabilityScienceandManagement,1:47-57.

ZangJ.andFoodyG.M.(1998)-A fuzzy classification of sub-urban land cover from remotely sensed im-agery.InternationalJournalofRemoteSensing,19:2721-2738.

ZhouG.andLiR.(2000)-Accuracy Evaluation of Ground Points from IKONOS High-Resolution Satellite Imagery.PhotogrammetricEngineeringandRemoteSensing,66:1103-1112.

shoreline extraction using satellite imagery · shoreline extraction using satellite imagery...

Documents