当前位置:文档之家 > 基于树种分类的高分辨率遥感数据纹理特征分析

基于树种分类的高分辨率遥感数据纹理特征分析

浙江农林大学学报,2012,29(2):210—217

JournalofZhejiangA&FUniversity

基于树种分类的高分辨率遥感数据纹理特征分析

王妮,彭世揆,李明诗

(南京林业大学森林资源与环境学院,江苏南京210037)

摘要:遥感图像尤其是高分辨率(1—4m)遥感图像在树种分类方面有着广阔的应用前景。利用主成分分析法对遥感数据去相关分析,然后通过对纹理提取过程的分析,探讨不同移动窗口大小对纹理特征的影响,以期为中山陵园风景区的森林调查提供依据,分类方法为经典的最大似然分类器。根据不同移动窗口大小的纹理因子相关性和对保持纹理信息丰富度的影响,来选择合适的窗12'大小及纹理因子组合,以对树种分类精度的提高程度为评价标准。研究结果表明,利用窗口大小为19×19下的纹理信息可有效提高分类精度,总精度达到66%,Kappa系数达到0.59。比单纯的光谱信息最大似然法图像分类精度高,其中均值与均匀性、对比度、偏斜度纹理因子组合为最佳纹理组合,能有效减少数据冗余。高分辨率遥感数据纹理信息的运用为树种分类识别时的特征选择提供了有利技术参考。图4表3参19

关键词:森林经理学;树种分类;移动窗口;纹理因子;总精度;灰度共生矩阵

中圈分类号:¥757:P237.4文献标志码:A文章编号:2095—0756(2012)02.0210-08

High—resolutionremotesensingoftexturalimagesfor

treespeciesclassification

WANGNi,PENGShi—kui,LIMing-shi

(CollegeofForestResourcesandEnvironment,NanjingForestryUniversity,Nanjing210037,Jiangsu,China)

Abstract:Remotesensingimagesshowaverypromisingperspectivefordistinguishingtreespecies,especial—lythosewiththevery.higllresolutionrangingfrom1to4m.However,thetraditionalmethodologyforclassify—inglandcovertypes,solelydependingonspectralfeatures,whiletextureandotherspatialinformationareneglected,hastheweaknesssuch.asinadequatelyutilizationofinformation,lOWaccuraciesofclassificationetc.Cunsideringtothetexturedifferencesamongforestspecies,itismoreimportantforspatialinformationdiscriptionofhigh-resolutionremotesensingimagetoimprovetheprecisionoftexturalfeatureschoosing.Inthisstudy,thefactorstoinfluencetheninetexturalfeatureschoosingwereanalyzedandtheresultsshowedthatthemovingwindowsizewasthemainfactortoaffecttheobtainingprocessesoftexturalfeaturesbasedonthegraylevelCO—occurrencematrix(GLCM)method,andtheimagerywasthenclassifiedcombiningthemax—imum1ikelihoodclassification(MLC)methodwiththeoriginalspectralvaluesandtexturefeatures.First,thisstudyutilizedacorrelationanalysisoftheimagesfromaprincipalcomponentanalysis.Second.throughmulti—pleinformationsources,includingtextualfeaturesderivedfromthedata.Forthehigh—resolutionremote¥ens—ingimage,themostpropermovingwindowsizewasdeterminedfrom3×3to31X31.Classificationofthemajortreespeciesthroughoutthestudyarea(theSunYat—SenMausoleuminNanjing)wasundertakenusingtheMLC.Third,toaidforestresearch,classificationaccuracywasimprovedusingtheGLCM.Accordingtocor-relationsamongtexturesandrichnessofthedata,GLCMprovidedthebest
windowsizeandtexturalparmne—ters.Resultsindicatedthatthetexturecharacteristicswereaddinthespectralcharacteristicstoimprovetheprecisionoftheresultsoftheclassification,19X19windowforbestwindow.Thetotalprecisioncanreach

Abstract:Remotesensingimagesshowaverypromisingperspectivefordistinguishingtreespecies,especial—lythosewiththevery.higllresolutionrangingfrom1to4m.However,thetraditionalmethodologyforclassify—inglandcovertypes,solelydependingonspectralfeatures,whiletextureandotherspatialinformationareneglected,hastheweaknesssuch.asinadequatelyutilizationofinformation,lOWaccuraciesofclassificationetc.Cunsideringtothetexturedifferencesamongforestspecies,itismoreimportantforspatialinformationdiscriptionofhigh-resolutionremotesensingimagetoimprovetheprecisionoftexturalfeatureschoosing.Inthisstudy,thefactorstoinfluencetheninetexturalfeatureschoosingwereanalyzedandtheresultsshowedthatthemovingwindowsizewasthemainfactortoaffecttheobtainingprocessesoftexturalfeaturesbasedonthegraylevelCO—occurrencematrix(GLCM)method,andtheimagerywasthenclassifiedcombiningthemax—imum1ikelihoodclassification(MLC)methodwiththeoriginalspectralvaluesandtexturefeatures.First,thisstudyutilizedacorrelationanalysisoftheimagesfromaprincipalcomponentanalysis.Second.throughmulti—pleinformationsources,includingtextualfeaturesderivedfromthedata.Forthehigh—resolutionremote¥ens—ingimage,themostpropermovingwindowsizewasdeterminedfrom3×3to31X31.Classificationofthemajortreespeciesthroughoutthestudyarea(theSunYat—SenMausoleuminNanjing)wasundertakenusingtheMLC.Third,toaidforestresearch,classificationaccuracywasimprovedusingtheGLCM.Accordingtocor-relationsamongtexturesandrichnessofthedata,GLCMprovidedthebestwindowsizeandtexturalparmne—ters.Resultsindicatedthatthetexturecharacteristicswereaddinthespectralcharacteristicstoimprovetheprecisionoftheresultsoftheclassification,19X19windowforbestwindow.Thetotalprecisioncanreach

收稿日期:2011-06-13;修回13期:2011-09-10

基金项目:江苏省普通高校研究生科研创新计划项目(CX09B一188Z);南京林业大学优秀博士例新基金项目作者简介:王妮,博士研究生,从事林业遥感与地理信息系统研究。.E—mail:wnstrive@163.cortl

万方数据

TOP相关主题