哨兵光学及SAR卫星影像协同分类研究  

Research on cooperative classification of sentinel optical and sar satellite images

作　　者：雷俊杰 杨武年[1] 李红 王芳[1] 杨鑫 LEI Junjie;YANG Wunian;LI Hong;WANG Fang;YANG Xin(College of Earth Science,Chengdu University of Technology,Chengdu 610059,China;College of Surveying and Planning,Shangqiu Normal University,Shangqiu 476000,China;School of Geomatics and Geoinformation,Chongqing Vocational Institute of Engineering,Chongqing 402260,China)

机构地区：[1]成都理工大学地球科学学院,四川成都610059 [2]商丘师范学院测绘与规划学院,河南商丘476000 [3]重庆工程职业技术学院测绘地理信息学院,重庆402260

出　　处：《现代电子技术》2022年第2期135-139,共5页Modern Electronics Technique

基　　金：国家自然科学基金资助项目(41671432);重庆市教育委员会科学技术研究项目(KJQN201803402)。

摘　　要：针对光学遥感影像中云及云阴影以及微波影像中叠掩、阴影、透视收缩等导致的地物信息缺失对分类、定量反演造成的不利影响,以光学卫星哨兵二号(S2)、SAR卫星哨兵一号(S1)影像为例,将R语言随机森林(RF)、支持向量机(SVM)、人工神经网络(ANN)机器算法引入光学微波遥感图像协同分类研究中。研究表明:协同S1,S2各自优势,可充分协同地物的光谱特征、微波后向散射、极化分解等信息,有效减轻S2云、云阴影以及S1影像地物信息缺失对地类识别造成的不利影响。采用RF交叉验证函数(RFCV)和IncNodepurity(节点纯度总增加)值确定RF最优分类因子,采用varImp函数以及按分类因子重要性逐个添加法可选取ANN、SVM最优分类因子。RF算法分类精度较高,适用于样点分布随机、均匀的研究区,研究区不同地类面积从大到小依次为灌草、其他、落叶林、草地、常绿林、混交林、水体。协同分类结果表明,光学、微波影像协同法、最优因子选择法以及R语言遥感数据处理方法在影像分类中应用效果较好。The cloud and shadow in the optical images,as well as the overlay,shadow and perspective shrinkage in the microwave images,etc.lead to the loss of ground information,which has a negative impact on the image classification and quantitative inversion.By taking the image of optical satellite sentinel 2(S2)and SAR satellite sentinel 1(S1)as examples,the R random forest(RF),support vector machine(SVM)and artificial neural network(ANN)machine algorithm are introduced into the cooperative classification research of optical microwave remote sensing images.The results show that:the spectral characteristics,microwave backscattering,polarization decomposition and other information of ground features can be cooperated in cooperation with the respective advantages of S1 and S2,so that the adverse impacts of S2 cloud,cloud shadow and the lack of ground feature information in S1 image on ground class recognition can be reduced.The RF cross validation function(RFCV)and incnodepurity(total increase in node purity)are used to determine the RF optimal classification factor,and the optimal classification factors of ANN and SVM can be selected by using varimp function and adding one by one according to the importance of classification factors.RF algorithm has high classification accuracy and is suitable for the study area with random and uniform sample distribution.The areas of different land types in the study area from large to small are:shrub and grass,others,deciduous forest,grassland,evergreen forest,mixed forest and water body.The optical and microwave image cooperation method,optimal factor selection method and R language remote sensing data processing method have good application effects in image classification.

关 键 词：遥感影像 协同分类 光学遥感 微波遥感 影像信息缺失 最优因子选择 

分 类 号：TN911.74-34[电子电信—通信与信息系统] TP75[电子电信—信息与通信工程]