利用光学遥感数据_GIS及人工神_省略_网络模型分析区域滑坡灾害_英文_Bisw

地学前缘(中国地质大学(北京);北京大学)

EarthScienceFrontiers(ChinaUniversityofGeosciences,Beijing;PekingUniversity)Vol.14No.6

Nov.2007

UTILIZATIONOFOPTICALREMOTESENSINGDATA

ANDGISTOOLSFORREGIONALLANDSLIDEHAZARDANALYSISUSINGANARTIFICIALNEURALNETWORKMODEL

利用光学遥感数据、GIS及人工神经网络模型分析区域滑坡灾害

BiswajeetPradhan,　SaroLee

Lumpur,Malaysia

2.GeoscienceInformationCenter,KoreaInstituteofGeoscienceandMineralResources(KIGAM)30,Kajung-Dong,Yusung-Gu,Tae-jon,Korea

1

2,＊

1.CilixCorporation,LotL4-I-6,Level4,Enterprise4,TechnologyParkMalaysia,BukitJalilHighway,BukitJalil,57000,Kuala

BiswajeetPradhan,SaroLee.UtilizationofopticalremotesensingdataandGIStoolsforregionallandslidehazardanalysisusinganartificialneuralnetworkmodel.EarthScienceFrontiers,2007,14(6):143-152

Abstract:TheaimofthisstudyistoevaluatelandslidehazardanalysisatSelangorarea,MalaysiausingopticalremotesensingdataandaGeographicInformationSystem(GIS).Landslidelocationswereidentifiedinthestudyareafrominterpretationofaerialphotographsandfieldsurveys.Topographical,geologicaldataandsat-elliteimageswerecollected,processedandconstructedintoaspatialdatabaseusingGISandimageprocessing.Thereareabout10landslideoccurrencefactorsthatwereselectedas:topographicslope,topographicaspect,topographiccurvatureanddistancefromdrainage;lithologyanddistancefromlineament;landcoverfromTMsatelliteimages;thevegetationindexvaluefromLandsatsatelliteimages;precipitationdata.Thesefactorswereanalyzedusinganadvancedartificialneuralnetworkmodeltogeneratethelandslidehazardmap.Eachfactorsweightwasdeterminedbytheback-propagationtrainingmethod.Thenthelandslidehazardindiceswerecalculatedusingthetrainedback-propagationweights,andfinallythelandslidehazardmapwasgeneratedusingGIStools.Landslidelocationswereusedtoverifyresultsofthelandslidehazardmapandtheverificationresultsshowed82.92%accuracy.Theverificationresultsshowedsufficientagreementbetweenthepresump-tivehazardmapandtheexistingdataonlandslideareas.

Keywords:landslide;hazard;artificialneuralnetwork;GIS;Malaysia

CLCnumber:P2.22　Documentcode:A　ArticleID:1005-2321(2007)060143-10

摘　要:用光学遥感数据和地理信息系统(GIS)分析了马来西亚Selangor地区的滑坡灾害。通过遥感图像解译和野外调查,在研究区内确定出滑坡发生区。通过GIS和图像处理,建立了一个集地形、地质和遥感图像等多种信息的空间数据库。滑坡发生的因素主要为:地形坡度、地形方位、地形曲率及与排水设备距离;岩性及与线性构造距离;TM图像解译得到的植被覆盖情况;Landsat图像解译得到的植被指数;降水量。通过建

收稿日期:2007-09-24基金项目:韩国科学技术部基础研究项目

作者简介:BiswajeetPradhan,马来西亚Putra大学GIS与地球数学工程专业博士,现为马来西亚Cilix公司遥感/GIS部门经理。Tel:0060-3968260,Fax:0060-34179,E-mail:biswajeet@mailcity.com

通讯作者:SaroLee,韩国Yonsei大学博士,现为韩国地球科学与矿物资源研究所地球科学信息中心主任,已在国际上发表近30篇与GIS相关的SCI/SCIE文章,目前主要从事利用GIS分析滑坡灾害。Tel:0082-42-868-3057,Fax:0082-42-861-9714,E-mail:leesaro@kigam.re.kr

＊

　144　　

BiswajeetPradhan,SaroLee/

地学前缘(EarthScienceFrontiers)2007,14(6)

立人工神经网络模型对这些因素进行分析后得到滑坡灾害图:由反向传播训练方法确定每个因素的权重值,然后用该权重值计算出滑坡灾害指数,最后用GIS工具生成滑坡灾害图。用遥感解译和野外观测确定出的滑坡位置资料验证了滑坡灾害图,准确率为82.92%。结果表明推测的滑坡灾害图与滑坡实际发生区域足够吻合。

关键词:滑坡;灾害;人工神经网络;GIS;马来西亚

1　Introduction

LandslidepresentsasignificantconstraintondevelopmentinmanypartsofMalaysia.Damagesandlossesareregularlyincurred,because,histori-cally,therehasbeentoolittleconsiderationofthepotentialproblemsinlanduseplanningandslopemanagement.LandslidesmostlyoccurinMalaysiamainlyduetoheavyrainfall.Inrecentyearsgrea-terawarenessoflandslideproblemshasledtosig-nificantchangesinthecontrolofdevelopmentonunstableland,withtheMalaysiangovernmentandhighwayauthoritiesstressingtheneedforlocalplanningauthoritiestotakelandslideintoaccountatallstagesofthelandslidehazardmappingprocess.Toassisttheimplementationofthispoli-cy,theMalaysianCentreforRemoteSensing(MACRES)commissionedamajordemonstrationprojecttoassessthepotentialforlandslideandtoincorporatethisinformationinthestrategicplan-ningprocess.Sofar,fewattemptshavebeenmadetopredicttheselandslidesorpreventingthedam-agecausedbythem.Throughthispredictionmod-el,landslidedamagecouldbegreatlydecreased.

StudyareaThroughscientificanalysisoflandslides,landslide-2　

susceptibleareascanbeassessedandpredicted,andthuslandslidedamagecanbedecreasedTheeasternpartofSelangorstate,sufferedthroughproperpreparation.Toachievethisaim,frommuchlandslidedamagesfollowingheavylandslidehazardanalysistechniqueshavebeenap-rains,wasselectedasasuitablepilotareatoevalu-plied,andverifiedinthestudyareausingartificialatefrequencyanddistributionoflandslidesneuralnetwork.Inaddition,landslide-relatedfac-(Fig.1).Selangorisoneofthe13statesofthetorswerealsoassessed.FederationofMalaysia.TheSelangorareaisloca-TherehavebeenmanystudiesthathavebeentedonthesouthwestcoastoftheMalaysianpenin-carriedoutonlandslidehazardevaluationusingsular.Itisboundedtothenorthandeastbythe

[1]

GIS,forexample,Guzzettietal.(1999)sum-stateofPerak,PahangandtothesouthbyNegerimarizedmanylandslidehazardevaluationstudies.Sembilian,Melaka.Thestudyareaislocatedap-Recently,therehavebeenstudiesonlandslidehaz-proximatelybetween3°23′53.6″Eand3°45′18.05″ardevaluationusingGIS,andmanyofthesestud-Eand101°30′55.33″Nand101°3′36.3″N.Theland

[2-22]

ieshaveappliedprobabilisticmodels.Oneofuseatthestudyareaismainlypeatswampforest,thestatisticalmodelsavailable,thelogisticregres-plantationforest,inlandforest,scrub,grasslandsionmodel,hasalsobeenappliedtolandslidehaz-andex-miningarea.Thelandformofthearearan-[23-33]ardmapping,aswellasthegeotechnicalmod-gesfromveryflatterrain,especiallyforthepeatelandthesafetyfactormodel[34-41].Asanewap-swampforest,ex-mining,grasslandandscrubare-proachtolandslidehazardevaluationusingGIS,a,toquitehillyareaforthenaturalforestrangingdataminingusingfuzzylogic,safetyfactorandar-between0-420mabovesealevel.Basedondatatificialneuralnetworkmodelshavebeenap-plied[42-49].

LandslideoccurrenceareasweredetectedintheSelangorarea,Malaysiabyinterpretationofaerialphotographsandfieldsurveys.Alandslidemapwaspreparedfromaerialphotographs,incombinationwiththeGIS,andthiswasusedtoe-valuatethefrequencyanddistributionofshallowlandslidesinthearea.Topography,lithologyandprecipitationdatabaseswereconstructedandlinea-ment,landcoverandvegetationindexvaluewereextractedfromLandsatTMsatelliteimagefortheanalysis.Then,thecalculatedandextractedfac-torswereconvertedintoa10m×10mgrid(ARC/INFOGRIDtype).Artificialneuralnetworkwasappliedusingthedatabaseandlandslidehazardmapwascreated.Finally,themapwasverifiedandcomparedusingknownlandslidelocationsforquantitativeverification.

Inthisstudy,GeographicInformationSystem(GIS)software,ArcView3.3andArcGIS9.0versionsoftwarepackageswereusedasthebasica-nalysistoolsforspatialmanagementanddatama-nipulation.

BiswajeetPradhan,SaroLee/

地学前缘(EarthScienceFrontiers)2007,14(6)　　145　

fromtheMalaysianMeteorologicalServicesDe-partment,thetemperatureofnorthernpartof

Selangorisbetween29℃to32℃andmeanrelativehumidityof65%to70%.ThehighesttemperatureisbetweenApriltoJunewhiletherelativehumidi-tyislowestinJune,JulyandSeptember.Therainfallofabout58.6mmto240mmpermonthwasrecordedinthestudyarea(TanjungKarangweatherstationprovidedbyMalaysianMeteoro-logicalServicesDepartment).

ousthroughoutthePaleozoicandMesozoic,but

becauseoftheinstabilityofthebasin,thesedi-mentaryrecordisincomplete.Majorbreaksareap-parentbetweenthePaleozoic,MesozoicandCeno-zoicgroupsofrocks,whereaswithinandbetweenthesystemsthemselvesminorbreaksarealsopres-ented.Granitoidsoccupyalmosthalfofthestudyarea.Thesebodiescommonlyformtopographichighs,thelargestofwhichistheMainrangesitu-atedontheeasternflankofthearea.Althoughmanyofthegranitebodiesarealignedparalleltothestructuraltrend,theydonotalwaysoccupytheanticlinalridgesofthesedimentarycoversandsomeofthesmallerbodiesarefoundtocutacrossthestructuraltrends.RegionalmetamorphismiswidespreadandmostofthePaleozoicandMesozoicrocksshowslighttomoderatedeformation.Ingeneral,theolderrocksshowagreaterdegreeofmetamorphismthantheyoungerones.Contactmetamorphismisnotintense.Thecontactmeta-morphosedrocksgenerallyformnarrowaureolesa-roundtheigneousbodies.Thereareatleastfourmajorepisodesofgraniteemplacementanditisbe-lievedthatmuchoftheknownmineralizationoc-curredduringthelaterepisodesandiscommonlyassociatedwithfaulting.Faultingiscommoninalltypesofrock.Atleastthreesetsoffaultshavebeenrecognizedonstudyscale,theyoungestofwhichisatmostpost-EarlyCretaceousinage.Manylandslideshavebeenrecordedthroughthefieldworkalonghighwaysandsteepslopeareas.

3　Artificialneuralnetworkmodel

Fig.1　Landslidelocationmapwithhill

shadedmapofstudyarea

Tectonically,SelangorareaformsapartoftheSundaShield.Itsfold-mountainsystem,thedomi-nantregionaltrendofwhichisnortherlytonorth-northwesterly,isasoutherlycontinuationofthatextendingfromeasternBurmathroughThailand,peninsularMalaysia,theBankaandBillitonislandsandeastwardsintoIndonesianBorneo.Allthesys-tems,rangingfromtheCambriantotheQuaterna-ry,arerepresentedinPeninsularMalaysia.Thepre-Triassicrocksareessentiallymarineorigin,whereasthepost-Triassicrocksarecharacteristi-callynon-marineorigin.TheTriassicrocksthem-selvesarebothmarineandnon-marineoriginsbutingeneral,thenon-marinedepositswherepresent,occurintheUpperTriassic.WithinSelangorit-self,itisprobablethatsedimentationwascontinu-Anartificialneuralnetworkisa“computa-tionalmechanismabletoacquire,represent,andcomputeamappingfromonemultivariatespaceofinformationtoanother,givenasetofdatarepre-sentingthatmapping”.Theback-propagationtrainingalgorithmisthemostfrequentlyusedneu-ralnetworkmethodandisalsousedinthisstudy.Theback-propagationtrainingalgorithmistrainedusingasetofexamplesofassociatedinputandout-putvalues.Thepurposeofanartificialneuralnet-workistobuildamodelofthedata-generatingprocess,sothatthenetworkcangeneralizeandpredictoutputsfrominputsthatithasnotprevi-ouslybeenseen.Thislearningalgorithmisamulti-layeredneuralnetwork,whichcomprisesofaninputlayer,hiddenlayersandanoutputlayer.Thehiddenandoutputlayerneuronsprocesstheirinputsbymultiplyingeachinputbyacorrespond-ingweight,summingtheproductandthenpro-　146　　

BiswajeetPradhan,SaroLee/

地学前缘(EarthScienceFrontiers)2007,14(6)

cessingthesumusinganonlineartransferfunctiontoproducearesult.Anartificialneuralnetwork“learns”byadjustingtheweightsbetweentheneu-ronsinresponsetotheerrorsbetweentheactualoutputvaluesandthetargetoutputvalues.Attheendofthistrainingphase,theneuralnetworkpro-videsamodelthatshouldbeabletopredictatargetvaluefromagiveninputvalue.

Therearetwostagesinvolvedinusingneuralnetworksformulti-sourceclassification:thetrain-ingstage,inwhichtheinternalweightsareadjus-ted;andtheclassifyingstage.Typically,theback-propagationalgorithmtrainsthenetworkuntilsometargetedminimalerrorisachievedbetweenthedesiredandactualoutputvaluesofthenet-work.Oncethetrainingiscomplete,thenetworkisusedasafeed-forwardstructuretoproduceaclassificationfortheentiredata.

Aneuralnetworkcomprisesofanumberofin-terconnectednodes.Eachnodeisasimpleprocess-ingelementthatrespondstotheweightedinputswhichitreceivesfromothernodes.Thearrange-mentofthenodesisreferredtoasthenetworkar-chitecture(Fig.2).Thereceivingnodesumstheweightedsignalsfromallthenodesthatitiscon-nectedtointheprecedinglayer.Formally,thein-putthatasinglenodereceivesisweightedaccord-ingtoEquation(1).

layer.Thethirdlayeristheoutputlayerthatpres-entstheoutputdata.Eachnodeinthehiddenlayer

isinterconnectedtonodesinboththeprecedingandfollowinglayersbyweightedconnections.

Theerror,E,foraninputtrainingpattern,isafunctionofthedesiredoutputvector(d)andtheactualoutputvector(o),givenby:

1E=∑(dk-ok)(4)

2k

Theerrorispropagatedbackthroughtheneuralnetworkandisminimizedbyadjustingtheweightsbetweenlayers.Theweightadjustmentisex-pressedas:

wij(n+1)=η(δj·oi)+αΔwij(5)

where,ηisthelearningrateparameter(settoη=0.01inthisstudy),δjisanindexoftherateofchangeoftheerror,andαisthemomentumpa-rameter(settoα=0.01inthisstudy).

Thefactorδjisdependentonthelayertype.Forexample,

forhiddenlayers,δj=kwjkf′(netj)(6)∑δandforoutputlayers,δj=(dk-ok)f′(netk)

(7)

　　Thisprocessoffeedingforwardsignalsandback-propagatingtheerrorisrepeatediterativelyuntiltheerrorofthenetworkasawholeismini-mizedorreachesanacceptablemagnitude.

Usingtheback-propagationtrainingalgo-jijinet=∑w·o(1)rithm,theweightsofeachfactorcanbedeter-i

minedandmaybeusedforclassificationofdata(inputvectors)thatthenetworkhasnotseenbe-fore.Zhou(1999)describedamethodfordetermi-ningtheweightsusingbackpropagation[50].FromEquation(2),theeffectofanoutput,oj,fromahiddenlayernode,j,ontheoutput,ok,fromanoutputlayer(nodek)canberepresentedbythepartialderivativeofokwithrespecttoojasFig.2　Architectureofneuralnetwork

ok (netk)=f′(netk)·=f′(netk)·wjk(8)

　　where,wijrepresentstheweightsbetween oj ojnodesiandj,andoiistheoutputfromnodej,　　Equation(8)producesbothpositiveandnega-givenbytivevalues.Iftheeffect'smagnitudeisallthatof

oj=f(netj)(2)interest,thentheimportance(weight)ofnodej

　　Thefunctionfisusuallyanon-linearsigmoidrelativetoanothernodej0inthehiddenlayermayfunctionthatisappliedtotheweightedsumofin-becalculatedastheratiooftheabsolutevaluesde-putsbeforethesignalpropagatestothenextlayer.rivedfromEquation(8):Oneadvantageofasigmoidfunctionisthatitsde-| ok|| ok||f′(netk)·wjk||wjk|/==| oj|| oj0||f′(netk)·wj0k||wj0k|rivativecanbeexpressedintermsofthefunction

(9)itself:

f′(netj)=f(netj)(1-f(netj))(3)　　wj0kissimplyanotherweightinwjkotherthan

　　Thenetworkusedinthisstudycomprisedofwik.threelayers.Thefirstlayeristheinputlayer,Foragivennodeintheoutputlayer,there-wherethenodesaretheelementsofafeaturevec-sultsofEquation(9)showthattherelativeimpor-tor.Thesecondlayeristheinternalor“hidden”tanceofanodeinthehiddenlayerisproportionalBiswajeetPradhan,SaroLee/

地学前缘(EarthScienceFrontiers)2007,14(6)　　147　

totheabsolutevalueoftheweightconnectingthe

nodetotheoutputlayer.Whenthenetworkcom-prisesofoutputlayerswithmorethanonenode,thenEquation(9)cannotbeusedtocomparetheimportanceoftwonodesinthehiddenlayer.

J

1wj0k=·∑|wjk|(10)Jj=1

|wjk|J·|wjk|tjk==J(11)J1·|wjk|∑|wjk|j∑Jj=1=1

　　Therefore,withrespecttonodek,eachnodeinthehiddenlayerhasavaluethatisgreaterorsmallerthanunity,dependingonwhetheritismoreorlessimportant,respectively,thananaver-agevalue.Allthenodesinthehiddenlayerhaveatotalimportancewithrespecttothesamenode,givenby

Jj=1

∑t

jk

=J(12)

　　Consequently,theoverallimportanceofnode

jwithrespecttoallthenodesintheoutputlayercanbecalculatedby

K

1tj=·∑tjk(13)Kj=1

　　Similarly,withrespecttonodejinthehiddenlayer,thenormalizedimportanceofnodejintheinputlayercanbedefinedby

ij|ij||ω|ωsij==II·(14)I

1·|ωij|ij|∑|ωIi∑=1i=1

　　Theoverallimportanceofnodeiwithrespecttothehiddenlayeris

J

1si=·∑sij(15)

Jj=1

　　Correspondingly,theoverallimportanceofin-putnodeiwithrespecttooutputnodekisgivenby

J

1sti=·∑sij·tj(16)

Jj=1

4　DatausingGISandremotesens-ing

　　ThedatausedareshowninTable1.Accurate

Classification

GeologicalHazard

Sub-Classification

LandslideTopographicmapGeologicalmapDrainageLandcoverSoilmap

Vegetationindex(NDVI)Precipitationamountdetectionofthelocationoflandslidesisveryim-portantforprobabilisticlandslidehazardanalysis.VariousGISdatalayershavebeenillustratedinFig.3.Theapplicationofremotesensingmeth-ods,suchasaerialphotographsandsatelliteima-ges,areusedtoobtainsignificantandcost-effec-tiveinformationonlandslides.Inthisstudy,1∶25000～1∶50000-scaleaerialphotographswereusedtodetectthelandslidelocations.Thesepho-tographsweretakenduringtheperiodof1981—2000,andthelandslidelocationsweredetectedbyphotointerpretationandthelocationswereverifiedbyfieldwork.Recentlandslideswereobservedinaerialphotographsfrombreaksintheforestcano-py,baresoil,orothergeomorphiccharacteristicstypicaloflandslidescars,forexample,headandsidescarps,flowtracks,andsoilanddebrisdepos-itsbelowascar.Toassembleadatabasetoassessthesurfaceareaandnumberoflandslidesineachofthethreestudyareas,atotalof327landslides

2

weremappedinamappedareaof8179.28km.

Therewere10factorsthatwereconsidered,andthefactorswereextractedfromtheconstruc-tedspatialdatabase.Thefactorsweretransformedintoavector-typespatialdatabaseusingtheGIS,andlandslide-relatedfactorswereextractedusingthedatabase.Adigitalelevationmodel(DEM)wascreatedfirstfromthetopographicdatabase.Contourandsurveybasepointsthathadelevationvaluesfromthe1∶25000-scaletopographicmapswereextracted,andaDEMwasconstructedwitharesolutionof10m.UsingthisDEM,theslopean-gle,slopeaspectandslopecurvaturewerecalculat-ed.Inthecaseofthecurvature,negativecurva-turesrepresentconcave,zerocurvaturesrepresentflatandpositivecurvaturesrepresentsconvex.Thecurvaturemapwaspreparedusingtheavenuerou-tineinArcView3.2.Inaddition,thedistancefromdrainagewascalculatedusingthetopographicdatabase.Thedrainagebufferwascalculatedin100mintervals.Usingthegeologydatabase,thelithologywasextracted,andthedistancefromlin-eamentwerecalculated.Thelithologymapwasobtainedfroma1∶63300-scalegeologicalmap,

GISDataType

Pointcoverage

LineandPointcoveragePolygoncoverageLinecoverage

GRIDGRIDGRIDGRIDScale1∶250001∶250001∶633001∶2500030m×30m10m×10m10m×10m10m×10mTable1　Datalayerofstudyarea

BasicMap

　148　　

BiswajeetPradhan,SaroLee/

地学前缘(EarthScienceFrontiers)2007,14(6)

andthedistancefromlineamentmapwascalculatedin100mintervals.LandcoverdatawasclassifiedusingaLANDSATTMimageusinganunsupervisedclassificationmethodandfieldsur-vey.Thenineclassesidentifiedasurban,water,forest,agriculturalarea,tinmines,rubberandpalmoilplantationwereextractedforlandcovermapping.Finally,theNormalizedDifferenceVeg-etationIndex(NDVI)mapwasobtainedfromSPOTsatelliteimages.TheNDVIvaluewascal-culatedusingtheformulaNDVI=(IR-R)/(IR+R),whereIRvalueistheinfraredportionofthee-lectromagneticspectrum,andRvalueisthered

portionoftheelectromagneticspectrum.TheND-VIvaluedenotesareasofvegetationinanimage.Theprecipitationdatawasmadefrominterpolationofmeteorologicaldata.

Thefactorswereconvertedtoarastergridwith10m×10mcellsforapplicationoftheartifi-cialneuralnetwork.Theareagridwas14140rowsby12900columnsand327cellshadlandslideoc-currences.

BiswajeetPradhan,SaroLee/

地学前缘(EarthScienceFrontiers)2007,14(6)　　149　

Fig.3　Inputdatalayers

(a)Slope;(b)Aspect;(c)Curvature;(d)Distancefromdrainage;(e)Geology;(f)Distancefromlineament;(g)Soil;(h)Landcover;(i)Vegetationindex(NDVI)and(j)Precipitationamount

5　Landslidehazardanalysisusing

theartificialneuralnetwork

　　Beforerunningtheartificialneuralnetworkprogram,thetrainingsiteshouldbeselected.So,thelandslide-prone(occurrence)areaandtheland-slide-not-proneareawereselectedastrainingsites.Cellsfromeachofthetwoclasseswererandomlyselectedastrainingcells,with327cellsdenotingareaswherelandslidedidnotoccuroroccurred.First,areaswherethelandslidedidnotoccurwereclassifiedas“areasnotpronetolandslide”andare-aswherelandslidewasknowntoexistwereas-signedtoan“areaspronetolandslide”trainingFig.4　Landslidehazardmapbasedon

ArtificialNeuralNetworkmodelset.

Theback-propagationalgorithmwasthenap-pliedtocalculatetheweightsbetweentheinputsingtheMATLABsoftwarepackage.Here,“feed-layerandthehiddenlayer,andbetweenthehiddenforward”denotesthattheinterconnectionsbe-layerandtheoutputlayer,bymodifyingthenum-tweenthelayerspropagateforwardtothenextlay-berofhiddennodeandthelearningrate.Three-er.Thenumberofhiddenlayersandthenumberoflayeredfeed-forwardnetworkwasimplementedu-nodesinahiddenlayerrequiredforaparticular　150　　

BiswajeetPradhan,SaroLee/

地学前缘(EarthScienceFrontiers)2007,14(6)

classificationproblemarenoteasytodeduce.Inhazardmapwascreated(Fig.4).Thevalueswerethisstudy,a9×19×2structurewasselectedforclassifiedbyequalareasandgroupedintofourclas-thenetwork,withinputdatanormalizedinthesesforvisualinterpretation.Thepossibilitywasrangeof0.1-0.9.Thenominalandintervalclassclassifiedintofourclasses(highest10%,secondgroupdatawereconvertedtocontinuousvalues10%,third20%andreminding60%)basedonar-rangingbetween0.1and0.9.Therefore,thecon-eaforvisualandeasyinterpretation.Theminimumtinuousvalueswerenotordinaldata,butnominalvalueis0.0121andmaximumvalueis0.9976.Thedata,andthenumbersdenotetheclassificationofmeanvalueis0.3945andthestandarddeviationtheinputdata.valueis0.3060.

Thelearningratewassetto0.01,andtheini-tialweightswererandomlyselectedtovaluesbe-6　Verificationtween0.1and0.3.Theweightscalculatedfrom10testcaseswerecomparedtodeterminewhether

Thelandslidehazardanalysisresultwasveri-thevariationinthefinalweightswasdependenton

fiedusingknownlandslidelocations.Verification

theselectionoftheinitialweights.Theback-prop-wasperformedbycomparingtheknownlandslideagationalgorithmwasusedtominimizetheerror

locationdatawiththelandslidehazardmap.The

betweenthepredictedoutputvaluesandthecalcu-ratecurveswerecreatedanditsareasoftheunder

latedoutputvalues.Thealgorithmpropagatedthe

curvewerecalculatedforallcases.Therateex-errorbackwards,anditerativelyadjustedthe

plainshowwellthemodelandfactorpredictthe

weights.Thenumberofepochswassetto2000,

landslide.So,theareaunderthecurvecanassess

andtherootmeansquareerror(RMSE)valueused

thepredictionaccuracyqualitatively.Toobtainthe

forthestoppingcriterionwassetto0.01.Mostof

relativeranksforeachpredictionpattern,thecal-thetrainingdatasetsmetthe0.01RMSEgoal.

culatedindexvaluesofallcellsinthestudyarea

However,iftheRMSEvaluewasnotachieved,

weresortedindescendingorder.Thentheordered

thenthemaximumnumberofiterationswastermi-cellvaluesweredividedinto100classes,withac-natedat2000epochs.Whenthelattercaseoc-cumulated1%intervals.Therateverificationre-curred,themaximumRMSEvaluewas0.213.

sultsappearasalineinFig.5.Forexample,in

Thefinalweightsbetweenlayersacquiredduring

thecaseofallfactorsused,90%to100%(10%)

trainingoftheneuralnetworkandthecontribution

classofthestudyareawherethelandslidehazard

orimportanceofeachoftheninefactorsusedto

indexhadahigherrankcouldexplain35%ofall

predictlandslidehazardareshowninTable2.

thelandslides.Inaddition,the80%to100%

Table2　Weightsofeachfactorestimatedbyneural(20%)classofthestudyareawherethelandslide

hazardindexhadahigherrankcouldexplain58%networkconsideredinthisstudy

ofthelandslides.Tocomparetheresultquantita-Normalized

FactorWeight

tively,theareasunderthecurvewerere-calculatedWeight

asthetotalareais1,whichmeansperfectpredic-Slope(°)0.2293.132Aspect0.0981.340tionaccuracy.So,theareaunderacurvecanbeCurvature(unitless)0.1171.597usedtoassessthepredictionaccuracyqualitative-Distancefromdrainage(m)0.0951.297ly.Thearearatiowas0.8292andthepredictionGeology0.0811.107accuracyis82.92%.

Distancefromlineament(m)SoilLandcoverNDVIPrecipitation

0.1120.1020.0730.0950.099

1.5331.3931.0031.2970.166

　　Foreasyinterpretation,theaveragevalueswerecalculated,andthesevaluesweredividedbytheaverageoftheweightsofsomefactorthathadaminimumvalue.Thelandusevaluewasthemin-imumvalue,1.00,andtheslopevaluewasthemaximumvalue,3.123.Finally,theweightswereappliedtotheentirestudyarea,andthelandslideFig.5　Cumulativefrequencydiagramshowinglandslide

hazardindexrankoccurringincumulative

percentoflandslideoccurrence

BiswajeetPradhan,SaroLee/

地学前缘(EarthScienceFrontiers)2007,14(6)　　151　

tioninamulti-scalestudy,CentralItaly.Geomorphology,

7　Conclusionsanddiscussion

1999,31:181-216.

[2]　AkgǜnA,BulutF.GIS-basedlandslidesusceptibilityforAr-sin-Yomra(Trabzon,NorthTurkey)region.Environmental

Geology,2007,51:1377-1387.

[3]　DahalRK,HasegawaS,NonomuraS,etal.GIS-based

Landslidingpresentsasignificantconstraint

ondevelopmentinMalaysia,notablythroughthe

weights-of-evidencemodellingofrainfall-inducedlandslidesin

inadvertentreactivationofancientinlandland-smallcatchmentsforlandslidesusceptibilitymapping.Envi-slides.Aseriesofgovernment-fundedresearch

ronmentalGeology,2007,OnlineFirst.

projectshaveprovidedmuchbackgroundinforma-[4]　ClericiA,PeregoS,TelliniC,etal.

AGIS-basedautomated

tionandidentifiedsuitablemethodsfortheuseofprocedureforlandslidesusceptibilitymappingbytheCondi-landslidehazardinformationinlanduseplanning.tionalAnalysismethod:theBaganzavalleycasestudy(Ital-However,anumberofsignificantproblemsremainianNorthernApennines).EnvironmentalGeology,2006,overtheuseofthisinformation.Inthisstudy,a50:941-961.

dataminingapproachtoestimatingthesusceptible[5]　CevikE,TopalT.GIS-basedlandslidesusceptibilitymapping

foraproblematicsegmentofthenaturalgaspipeline,HendekareaoflandslidesusingGISandremotesensingis

(Turkey).EnvironmentalGeology,2003,44:949-962.presented.

GISmodelingofslopestabilityFromtheapplicationofartificialneuralnet-[6]　RowbothamD,DudychaDN.

inPhewaTalwatershed,Nepal.Geomorphology,1998,26:

work,therelativeimportance,weight,between

151-170.

factorswascalculated.Theslopeshowedthehigh-[7]　JibsonWR,EdwinLH,JohnAM.Amethodforproducingestvalue3.123,thenthecurvatureis1.593anddigitalprobabilisticseismiclandslidehazardmaps.Engineer-distancefromfaultis1.529.Fromtheresult,theingGeology,2000,58:271-2.

slopeisthemostimportantfactorwhichtwicethat[8]　LuziL,PergalaniF,TerlienMTJ.Slopevulnerabilitytooftheotherfactors,forlandslidehazardmapping.earthquakesatsubregionalscale,usingprobabilistictech-Usingtheweights,thelandslidehazardmapwasniquesandgeographicinformationsystems.EngineeringGe-ology,2000,58:313-336.createdandverified.Theresultofverification

Aseismiclandslidesusceptibilityratingofshowed82.92%predictionaccuracy.Theverifica-[9]　PariseM,JibsonWR.

geologicunitsbasedonanalysisofcharacteristicsoflandslides

tionresultisofahighvalue.

triggeredbythe17January,1994Northridge,Californiaearth-Landslidehazardmapsareofgreathelpto

quake.EngineeringGeology,2000,58:251-270.

plannersandengineersforchoosingsuitableloca-[10]　BaezaC,CorominasJ.Assessmentofshallowlandslidesuscepti-tionstoimplementdevelopments.Theseresultsbilitybymeansofmultivariatestatisticaltechniques.EarthSur-canbeusedasbasicdatatoassistslopemanage-faceProcessesandLandforms,2001,26:1251-1263.

mentandland-useplanning,butthemodelsused[11]　LeeS,MinK.StatisticalanalysisoflandslidesusceptibilityinthestudyarevalidforgeneralizedplanningandatYoungin,Korea.EnvironmentalGeology,2001,40:assessmentpurposes,althoughtheymaybeless1095-1113.

usefulatthesite-specificscale,wherelocalgeolog-[12]　TemesgenB,MohammedMU,KormeT.Naturalhazard

assessmentusingGISandremotesensingmethods,withpar-icalandgeographicheterogeneitiesmayprevail.

ticularreferencetothelandslidesintheWondogenetArea,E-Forthemodeltobemoregenerallyapplied,more

thiopia.Phys.Chem.Earth,2001,26:665-615.

landslidedataareneeded,aswellasapplicationto

[13]　ClericiA,PeregoS,TelliniC,etal.Aprocedureforland-moreregions.

slidesusceptibilityzonationbytheconditionalanalysismeth-

Acknowledgements

od.Geomorphology,2002,48:349-3.[14]　DonatiL,TurriniMC.Anobjectivemethodtoranktheim-portanceofthefactorspredisposingtolandslideswiththeGIS

methodology:applicationtoanareaoftheApennines(Valne-rina;Perugia,Italy).EngineeringGeology,2002,63:277-

ThisresearchwassupportedbytheBasicRe-searchProjectoftheKoreaInstituteofGeoscience

2.andMineralResources(KIGAM)fundedbythe

MinistryofScienceandTechnologyofKorea.[15]　LeeS,ChoiJ,MinK.Landslidesusceptibilityanalysisand

verificationusingtheBayesianprobabilitymodel,Environ-TheauthorswouldliketothanktheMalaysian

mentalGeology,2002,43:120-131.

CenterforRemoteSensingforprovidingvarious

[16]　LeeS,ChwaeU,MinK.Landslidesusceptibilitymapping

datasetsusedinthisanalysis.References:

[1]　GuzzettiF,CarrarraA,CardinaliM,etal.Landslidehazard

evaluation:areviewofcurrenttechniquesandtheirapplica-162.

[17]　ZhouCH,LeeCF,LiJ,XuZW.Onthespatialrelationship

betweenlandslidesandcausativefactorsonLantauIsland,

HongKong.Geomorphology,2002,43:197-207.bycorrelationbetweentopographyandgeologicalstructure:theJanghungarea,Korea.Geomorphology,2002,46:149-

　152　　

BiswajeetPradhan,SaroLee/

地学前缘(EarthScienceFrontiers)2007,14(6)

[18]　LeeS,ChoiU.DevelopmentofGIS-basedgeologicalhazardinformationsystemanditsapplicationforlandslideanalysisin

Korea.GeoscienceJournal,2003,7:243-252.

[19]　LeeS,ChoiJ,MinK.ProbabilisticLandslideHazardMap-

(settlement)regioninTurkey.EngineeringGeology,2000,55:277-296.

[35]　RomeoR.Seismicallyinducedlandslidedisplacements:apre-dictivemodel.EngineeringGeology,2000,58:337-351.

pingusingGISandRemoteSensingDataatBoeun,Korea.[36]　ReficeA,CapolongoD.Probabilisticmodelingofuncertain-InternationalJournalofRemoteSensing,2004,25:2037-tiesinearthquake-inducedlandslidehazardassessment.Com-2052.puter&Geosciences,2002,28:735-749.

[20]　LeeS,TalibJA.Probabilisticlandslidesusceptibilityand[37]　CarroM,DeAmicisM,LuziL,etal.Theapplicationof

factoreffectanalysis.EnvironmentalGeology,2005,47:predictivemodelingtechniquestolandslidesinducedbyearth-982-990.

[21]　Lee,S,DanNT.Probabilisticlandslidesusceptibilitymap-pingintheLaiChauprovinceofVietnam:focusontherela-tionshipbetweentectonicfracturesandlandslides.Environ-mentalGeology,2005,48:778-787.[22]　LeeS,LeeMJ.DetectinglandslidelocationusingKOMP-SAT1anditsapplicationtolandslide-susceptibilitymapping

attheGangneungarea,Korea.AdvancesinSpaceResearch,2006,38:2261-2271.

quakes:thecasestudyofthe26September1997Umbria-Marcheearthquake(Italy).EngineeringGeology,2003,69:

139-159.

[38]　ShouKJ,WangCF.AnalysisoftheChiufengershanland-slidetriggeredbythe1999Chi-ChiearthquakeinTaiwan.

EngineeringGeology,2003,68:237-250.[39]　ZhouG,EsakiT,MitaniY,etal.Spatialprobabilisticmod-elingofslopefailureusinganintegratedGISMonteCarlosimulationapproach.EngineeringGeology,2003,68:373-

[23]　AkgunA,DagS,BulutF.Landslidesusceptibilitymapping386.

foralandslide-pronearea(Findikli,NEofTurkey)bylikeli-[40]　LeeS.Comparisonoflandslidesusceptibilitymapsgeneratedhood-frequencyratioandweightedlinearcombinationmodels.throughmultiplelogisticregressionforthreetestareasinKo-EnvironmentalGeology,2007,OnlineFirst.

[24]　TunusluogluMC,GokceogluC,NefesliogluHA,Sonmez

H.Extractionofpotentialdebrissourceareasbylogisticre-gressiontechnique:acasestudyfromBarla,BesparmakandKapimountains(NWTaurids,Turkey).EnvironmentalGe-ology,2007,OnlineFirst.

[25]　LamelasMT,MarinoniO,HoppeA,RivaJ.Dolineproba-bilitymapusinglogisticregressionandGIStechnologyinthecentralEbroBasin(Spain).EnvironmentalGeology,2007,

OnlineFirst.

[26]　WangHB,SassaK.ComparativeevaluationoflandslidesusceptibilityinMinamataarea,Japan.EnvironmentalGeol-ogy,2005,47:956-966.[27]　SǜzenML,DoyuranV.AcomparisonoftheGISbasedland-slidesusceptibilityassessmentmethods:multivariateversus

bivariate.EnvironmentalGeology,2004,45:665-679.

[28]　AtkinsonPM,MassariR.Generalizedlinearmodelingof

rea.EarthSurfaceProcessesandLandforms,2007,32(14):

2133-2148.

[41]　LeeS,PradhanB.LandslidehazardmappingatSelangor,

Malaysiausingfrequencyratioandlogisticregressionmodels.Landslides,2007,4:33-41.[42]　XieM,EsakiT,CaiM.Atime-spacebasedapproachformappingrainfall-inducedshallowlandslidehazard.Environ-mentalGeology,2004,46:840-850.[43]　ErcanogluM,GokceogluC.Assessmentoflandslidesuscep-tibilityforalandslide-pronearea(northofYenice,NWTurkye)byfuzzyapproach.EnvironmentalGeology,2002,41:720-730.

[44]　PistocchiA,LuziL,NapolitanoP.Theuseofpredictive

modelingtechniquesforoptimalexploitationofspatialdata-bases:acasestudyinlandslidehazardmappingwithexpertsystem-likemethods.EnvironmentalGeology,2002,41:765-775.

susceptibilitytolandslidinginthecentralApennines,Italy.[45]　LeeS,RyuJH,MinK,etal.LandslideSusceptibilityAnal-Computer&Geosciences,1998,24:373-385.ysisusingGISandArtificialneuralnetwork.EarthSurface

[29]　DaiFC,LeeCF,LiJ,etal.Assessmentoflandslidesus-ceptibilityonthenaturalterrainofLantauIsland,HongKong.EnvironmentalGeology,2001,40:381-391.[30]　DaiFC,LeeCF.Landslidecharacteristicsandslopeinsta-bilitymodelingusingGIS,LantauIsland,HongKong.Geo-morphology,2002,42:213-228.[31]　OhlmacherGC,DavisJC.UsingmultiplelogisticregressionandGIStechnologytopredictlandslidehazardinnortheastKansa,USA.EngineeringGeology,2003,2157:1-13.[32]　LeeS.Applicationoflogisticregressionmodelanditsvalida-tionforlandslidesusceptibilitymappingusingGISandremotesensingdata.InternationalJournalofRemoteSensing,2005,26:1477-1491.

[33]　LeeS,SambathT.Landslidesusceptibilitymappinginthe

DamreiRomelarea,Cambodiausingfrequencyratioandlo-ProcessesandLandforms,2003,27:1361-1376.

[46]　LeeS,RyuJH,LeeMJ,etal.Landslidesusceptibilitya-nalysisusingartificialneuralnetworkatBoeun,Korea.Envi-ronmentalGeology,2003,44:820-833.[47]　LeeS,RyuJH,WonJS,etal.Determinationandapplica-tionoftheweightsforlandslidesusceptibilitymappingusinganartificialneuralnetwork.EngineeringGeology,2004,71:

2-302.

[48]　TangestaniMH.Landslidesusceptibilitymappingusingthe

fuzzygammaapproachinaGIS,Kakancatchmentarea,

southwestIran.AustralianJournalofEarthSciences,2004,51:439-450.

[49]　LeeS.Applicationandverificationoffuzzyalgebraicopera-torstolandslidesusceptibilitymapping,EnvironmentalGeol-ogy,2007,52:615-623.

gisticregressionmodels.EnvironmentalGeology,2006,50:[50]　ZhouW.Verificationofthenonparametriccharacteristicsof

847-855.backpropagationneuralnetworksforimageclassification.

[34]　GokceogluC,SonmezH,ErcanogluM.Discontinuitycon-trolledprobabilisticslopefailureriskmapsoftheAltindag

IEEETransGeosciencesRemoteSensing,1999,38:771-779.