英文文獻(xiàn)科技類原文及譯文33

Multi-texture-modelforWaterExtractionBasedonRemoteSensingImageHuaWANG,LiPAN,HongZHENGSchoolofRemoteSensingandInformation&Engineering,WuhanUniversity129LuoyuRoad,Wuhan430079,P.R.ChinaSchoolofElectronicInformation,WuhanUniversity129LuoyuRoad,Wuhan430079,P.R.ChinaAbstract:Inthispaper,amulti-texture-modelforwaterextractionbasedonremotesensingimageryisproposed.Themodelisappliedtoextractinlandwater(includingwideriver,lakeandreservoir)fromhigh-resolutionpanchromaticimages.Firstlydirectionalvarianceisusedtofindriverregions,andthengraintableisadoptedtoavoidnoiseincludingobjectsthathavesimilardirectionalvariancecharacteristicaswatersurfaces.Theexperimentresultshowsthattheproposedmethodprovidesaneffectivewayforwaterextraction.IntroductionTherecognitionofwaterfromremotesensingimagehasdrawnconsiderableattentioninrecentyeas.Alargenumberofpublicationsaboutwaterextractionappearedandvariousapproachesforwaterextractionhavebeenproposed.Zhoudevelopedadescriptivemodelforautomaticextractionofwaterbasedonspectralcharacteristics[1].Bartonappliedchannel4forNOAA/AVHRRtoextractwatery.DuproposedaapproachforwaterextractionfromSPOT-5basedondecisiontreealgorithm[3].LirecognizedandmonitoredclearwaterfromMODIS[4].WuextractedwaterfromQuickBirdimageandusedactivecontourmodeltoobtainaccuratepositionofriverbank[5];Inordertoextractwaterfromhigh-spatialremotesensingimages,Heusedwavelettechniquetoexpendtheinformationandcleanedmainnoiseoftheimages,andthenpresentedmulti-windowlinearityreservetechniquetoconservelinearwater[6].Recently,mostresearchworkonwaterextractionwasforcedonautomaticrecognitionofwaterfromremotesensingimagesbasedonspectralcharacteristics.However,therearesomedisadvantagesofthesemethods:(1)Theresolutionofimageusedforwaterextractionislow.Theminimumsizeofrecognizableobjectisdependedonthespatialresolutionofsensor.Thereforeitisdifficulttoobtainaccuratepositionofwaterboundary.(2)Duetothecharacteristicofwateritselfandthesensorapplied,incertainchannelsthespectralfeaturesofdifferentobjectsareequilibrated.Theequilibrationleadstothephenomenaof“differentobjectssameimage"d”iffeorentimagessameobject”,whichresultsinnoiseobjectsincludedinextractionresult.Inthispaper,amulti-texture-modelforwaterextractionbasedonremotesensingisproposed.Themodelisappliedtoextractinlandwater(includingwideriver,lakeandreservoir)fromhigh-resolutionpanchromaticimage.Firstlydirectionalvarianceisappliedtofindriverregions,andthengraintableisadoptedtoavoidnoiseincludingobjectsthathavesimilardirectionalvariancecharacteristicaswatersurfaces.Theexperimentresultshowsthattheproposedmethodprovidesaneffectivewayforwaterextraction.Thispaperisorganizedasfollows.InSection2,thedirectionalvariancemodeladoptedisintroduced.Then,fusionofproposedgraintablemodelwithdirectionalvariancemodelisdiscussedinSection3.Theexperimentalresultsoftheproposedmulti-texture-modelandcomparativestudieswithsinglemodelsaregiveninSection4.WeconcludethispaperinSection5.DirectionalVarianceModel2Theaimofourresearchistoextractwaterlargerthan100mfrompanchromaticimages.AsshowninFigure2(a),theresearchobjectscanbedividedintothreeclasses:wideriver,lakeandreservoir,whichallrepresentasregioninhigh-resolutionimageries.Theobjectsofbackgroundcanbedividedintotwoclasses:buildingandcropland,whichalsorepresentasregion.Inpanchromaticimagery,wideriverhasasimilargrayleveltobuildingandcropland,thoughthemeangrayoflakeandreservoirismuchlowerthanthebackgroundobjects.Conventionalmethodsforwaterextractionbasedonspectralcharacteristicsarenoteffectiveinthesituation.Inthemeantime,waterbodydefineshomogeneousareaswhereasbuildingandcroplandcorrespondtoheterogeneousregions.Therefore,wetakeintoaccountthehomogeneityoftheimagetoseparatewideriver,lakeandreservoirfrombackgroundinstead.Tocharacterizethedifferenceofhomogeneitybetweenwaterbodyandtheothertypesofareas,weuseatextualoperator:thedirectionalvariance.TheDirectionalVarianceOperatorTheoperatorisderivedfromthosedefinedbyGuerin&MaitreandAirault&Jamet[10].AsshowninFigure1,thedirectionalvarianceconsistsincomputing,foreachpixelMoftheimage,thevarianceofthegraylevelsoftheimageonseveraldirectionofacirclewhosecenterisMandradiusisR.Then,thedirectionwiththehighestvariancevalueiskept.Itsdirectiondefinesthedirectionforwhichimageisthemostheterogeneous,locally.ItsvariancevalueisthedirectionalvariancevalueofthepixelM.ExtractionofwaterbasedondirectionalvarianceAccordingtothedefinitionoftheoperator,theminimumacreageofrecognizablewaterbodyisdependedonthelengthofradiusR.Wehavechosenalengthof10pixelsfor1mresolution.Thedirectionalvariancesofthefivetypicaltrainingsamples(wideriver,lake,reservoir,buildingandcropland)havebeencomputedandthestatisticalcomparisonissummarizedinTable1.Theoverallaverageofwaterdirectionalvarianceislowerthantheobjectsofbackground.Nevertheless,thedirectionalvarianceofcroplandissimilartowideriverwithoverlappingpotionover90%.Inhigh-resolutionpanchromaticimagery,detailsinsidewideriver,suchasboat,wave,etc,arerepresentedclearlywhichresultintheheterogeneousofwater.Inthemeantime,thetexturesofpartsofbuilding(forexample,roof)andcroplandareratherfine.Inasmallwindow,thesepotionsdefinehomogeneousareaswithsimilardirectionalvarianceaswideriver.Theresultisimprovedifwechosenalengthof100pixels.ThestatisticalcomparisonisshowninTable2.Ifthelengthofradiusislargeenough,directionalvarianceofbuildingishigherthanotherobjectsobviouslywithnooverlappingportion;thedifferencebetweencroplandandwideriverisincreasedwhiletheoverlappingpotionisdecreased.However,increasingtheradiusleadstotwoproblemswhichareoutlinedasfollow:Thesizeofrecognizablewaterbodyincreases;thereforewaterwhichhassmallacreage(forexamplenarrowriver)cannotbedetected.Thepositionofwaterbankisnotaccuratealthoughthespatialresolutionofimageryisratherhigh.Hence,inthispaper,amulti-texture-modelispresentedandtwotexturemodelsarefusedtoextractwaterfrompanchromaticimages.Firstly,wechosearadiusof10pixelstoextractwaterbasedondirectionalvariance;andthen,graintableisadoptedtoavoidnoiseincludingpartsofbuildingandcroplandthathavesimilardirectionalvariancecharacteristicaswatersurface.Multi-texture-modelInhigh-resolutionimagery,croplandandbuildingrepresentsstructuralcharacteristic.Accordingtothischaracteristic,grainanalysisisadoptedforfurtherresearchontheoriginalextractionbasedondirectionalvariance.Thegraintablehistogramisabletorepresentstructuralcharacteristicoftheresearchobject,whichcanbeappliedtorecognizemanykindsofdifferentobjects[12].3.1.ExtractionofwaterfusedbygraintableThegraintablehistogramsofthefivetypicaltrainingsamples(wideriver,lake,reservoir,buildingandcropland)arecomputedandcorrelationcoefficientsbetweenthemaresummarizedinTable3.Correlationcoefficientsbetweenwaterclassesareover85%,however,correlationcoefficientsbetweenwaterclassesandbackgroundclassesarelowerthan65%.Hence,wecomparethecorrelationcoefficientsofregionsinextractionimagebaseondirectionalvariancewiththreewatersamplesandtwobackgroundsamplesrespectively.Iftheregionhasahighercorrelationcoefficientwithbackgroundclasses,itwillbemarkedbackgroundandwipedoff[13].ExperimentalResultsWerunthealgorithmonseveralhigh-resolutionpanchromaticimages.InFigure2.(a),wehavebeenconsideringanaerialphotograph(6126x4800)ofaregioninWuhan,China,theresolutionofwhichis1m,includingbuilding,cropland,wideriver(Changjiangriver),lake,reservoirandcropland.Theresultsofextractionbasedondirectionalvariancewithradiusof10pixelsisdisplayedinFigure2.(b),andclearly,waterhasbeendetectedcompletely,whereaspartsofbuildingandcroplandareincludedasnoiseobjectsintheresult.Waterextractionusingdirectionalvariancewithradiusof100pixelsisdisplayedinFigure2.(c)withcorrectnessover95%,however,smalllakesaremissedandthepositionofbankisnotasaccurateasFigure2.(b).Finally,inFigure2.(d),theresultofFigure2.(b)isfusedbygraintableanalysis,sothatthecorrectnessandcompletenessofextractionarebothover90%.ConclusionsBasedontexturalanalysisofwaterinhigh-resolutionpanchromaticimagery,amulti-texture-modelispresentedforwaterextraction.Theexperimentalresultsprovedthattheapproachisefficientforinlandwater(includingwideriver,lakeandreservoir)extraction.Asthecomplexityanddiversityofwater,therateofrecognitionofouralgorithmfluctuates.Furthermore,themethodissupervisedwhichneedsalotofhumaninterferencetoobtaintrainingsamples.Therefore,thereareproblemstobesolvedinfuture:Ourfurtherworkshouldbeextensibletomultispectralremotesensingimages.Todecreasehumaninterference,oldvectorwillbeappliedtoobtaintrainingsamplesinstead.AcknowledgmentsTheworkwassupportedbytheNationalKeyTechnologyR&DProgramofChinaundergrantNo.2006BAB10B01.根據(jù)遙感圖象的多紋理模型相關(guān)的水抽取HuaWANG,LiPAN,HongZHENGSchoolofRemoteSensingandInformation&Engineering,WuhanUniversity129LuoyuRoad,Wuhan430079,P.R.ChinaSchoolofElectronicInformation,WuhanUniversity129LuoyuRoad,Wuhan430079,P.R.China文摘：在本文中，提議了一個(gè)多紋理模型為根據(jù)遙感成像的水提取。而且運(yùn)用模型從高分辨率泛色圖象提取內(nèi)陸水域(包括寬河、湖和水庫)。首先定向變化用于發(fā)現(xiàn)河地區(qū)，五谷桌然后被采取避免噪聲包括有相似的定向變化特征當(dāng)水表面的對(duì)象。實(shí)驗(yàn)結(jié)果表示，提出的方法為水提取提供有效方式。1.簡介在最近這些年，水的圖像測量識(shí)別是作為一個(gè)相當(dāng)值得關(guān)注的問題。有大量的關(guān)于水的抽提的出版物涌現(xiàn)，以及關(guān)與水的抽提的多種方法被大量的提出。周朝曾經(jīng)發(fā)展過一個(gè)基于光譜特征的描述型的自動(dòng)抽取模型。巴頓在NOAA/AVHRR應(yīng)用下用來抽取液體。Du被提議于從SPOT-5基于決策樹運(yùn)算法則來逐步進(jìn)行水的抽取。Li認(rèn)可了并且從MODIS監(jiān)測出了清楚的水。Wi利用從飛的很快的鳥的圖象中抽取水的流體特性并運(yùn)用運(yùn)動(dòng)等高模型來獲取河岸的準(zhǔn)確位置。為了從高空間遙感圖象提取水的特性，他使用小波技術(shù)擴(kuò)展信息并且消除了圖象的主要噪聲，然后利用被提出的多窗口線性預(yù)留技術(shù)保存線性水。最近，在水抽取的多數(shù)研究工作是牽強(qiáng)的在水的自動(dòng)識(shí)別從根據(jù)光譜特性的遙感圖象的過程。然而，有這些方法的有些缺點(diǎn)：(1)用于水抽取的圖象的決議是低的?？赊q識(shí)的對(duì)象的極小的大小取決于傳感器的空間分辨率。所以獲得水邊界限的準(zhǔn)確位置是難的。(2)由于水自身的特征和傳感器被用于在某些渠道應(yīng)用的不同的對(duì)象的特點(diǎn)被均衡化。這種平衡帶來了“不同的對(duì)象現(xiàn)象同樣圖象”或“不同的圖象同樣對(duì)象”的現(xiàn)象，導(dǎo)致在提取結(jié)果包括了噪聲對(duì)象。在本文，一個(gè)基于細(xì)微的感覺的水抽取的多紋理模型被最終提議。運(yùn)用模型從高分辨率泛色圖象提取內(nèi)陸水域（包括寬河、湖和水庫）的范圍內(nèi)進(jìn)行。首先定向變化的趨勢(shì)被應(yīng)用于發(fā)現(xiàn)河地區(qū)，然后顆粒表格的方法被采取于避免噪聲包括有相似的定向變化特征當(dāng)水表面的對(duì)象。實(shí)驗(yàn)結(jié)果表示，這些被提出的方法都很好的為水提取提供有效方式。本文接下來要討論的問題有：在第2部分，將介紹被采取的定向變化模型。然后，在部分3會(huì)提出的顆粒表格法模型的融合與定向變化模型在提出的多紋理模型.其中實(shí)驗(yàn)性結(jié)果會(huì)被談?wù)?，并且在?部分將給出比較研究同唯一的模型。最后，我們?cè)诘?部分結(jié)束本文。2.定向變化模型我們的研究的目標(biāo)大于100m2from泛色圖象將提取水。如圖2（a）所顯示，研究對(duì)象可以被劃分成三類：寬河、湖和水庫，所有在高分辨率成像都會(huì)被代表作為區(qū)域。背景對(duì)象可以被劃分成二類：大廈和農(nóng)田，也將會(huì)代表作為區(qū)域。在泛色成像，寬河有相似的灰級(jí)級(jí)別的圖像，并且農(nóng)田，這個(gè)背景對(duì)象可能會(huì)比湖和水庫的灰度低。常規(guī)方法為：根據(jù)光譜特性的水抽取是不是有效的情況下所決定。同時(shí)，水本體也定義了同類的區(qū)域，而修造和農(nóng)田則可以對(duì)應(yīng)于各種不同的地區(qū)。所以，我們考慮到圖象的同質(zhì)性則從背景上改為分離成寬河、湖和水庫。要描繪同質(zhì)性區(qū)別在水本身和其他區(qū)域類型之間的不同，我們使用一個(gè)本文里講到操作：定向變化。定向變化的操作這些操作的過程都是源于從Gierin&Maitre和Airailt&Jamet定義的那些結(jié)論中找到的。如Figirel所顯示，定向變化在計(jì)算其中救包括為圖象，圖象的灰級(jí)的變化的每個(gè)映像點(diǎn)M都是以m為圓心，并且半徑是R的弧線的幾個(gè)不同指導(dǎo)方向上進(jìn)行的。然后，方向以最高的變化價(jià)值被保留。它的方向當(dāng)?shù)囟x了圖象是最異向的方向。它的變化價(jià)值就是映像點(diǎn)M.的定向變化價(jià)值?；诙ㄏ蜃兓碚撋系某槿∷^程根據(jù)操作流程的定義，可辯識(shí)的水本體的極小的面積也取決于半徑R的長度。我們選擇了10個(gè)映像點(diǎn)的長度為1m決議。五個(gè)典型的測試樣品（寬河、湖、水庫、大廈和農(nóng)田）的定向變化將會(huì)在Table1被計(jì)算了和統(tǒng)計(jì)來進(jìn)行比較總結(jié)。整體平均水定向變化將有可能低于背景對(duì)象。然而，農(nóng)田的定向變化對(duì)于寬河是相似的，并且重疊的部分超過90%。在高分辨率泛色導(dǎo)致不同種的水的成像，細(xì)節(jié)在寬河里面顯示出來，例如小船、波浪等等，都可以清楚地表示出來。同時(shí)，大廈的部分紋理（例如，屋頂）和農(nóng)田是相當(dāng)完整的。在一個(gè)小窗口里，這些部分定義了同類的區(qū)域以相似的定向轉(zhuǎn)變成寬河。這個(gè)改進(jìn)結(jié)果就是如我們選上的100個(gè)映像點(diǎn)的長度。統(tǒng)計(jì)的比較已經(jīng)在Table2顯示了。如果半徑的長度足