• 遥感科学与应用技术 •

### 遥感影像混合像元的分解——基于加权后验概率的支持向量机分类算法

1. 首都师范大学资源环境与旅游学院, 北京100048
• 收稿日期:2012-11-08 修回日期:2013-02-18 出版日期:2013-04-25 发布日期:2013-04-18
• 通讯作者: 李小娟(1965-),女,博士,教授,博士生导师,主要从事资源环境遥感动态监测、土地利用数据建库与更新、资源环境信息系统等基础与应用研究。E-mail:xiaojuanli@263.net E-mail:xiaojuanli@263.net
• 作者简介:许菡(1979-),女,博士研究生,主要研究方向为遥感图像处理。E-mail:xuhan2009@gmail.com
• 基金资助:

国家“973”计划前期研究专项课题“城市典型区域不均匀沉降信息获取与机理研究”(2012CB723403)。

### Unmixing of Remote Sensing Images Based onWeighted Posterior Probability Support Vector Machines

XU Han, SUN Yonghua, LI Xiaojuan

1. College of Resource Environment and Tourism, Capital Normal University, Beijing 100048, China
• Received:2012-11-08 Revised:2013-02-18 Online:2013-04-25 Published:2013-04-18

Abstract: Considering a lot of mixed image pixels are contained in remote sensing images, weighted posterior probability support vector machines are introduced to deal with the remote sensing images unmixing. Weighted posterior probability support vector machines aremotivated by statistical learning theory and is the further research result of regular SVM. Each ground posterior probability can be computed when training samples are determined by the proposed method. Non-linear unmixing mixed pixels model precision are improved by the method because weighted posterior probability support vector machines can avoid the effect on classifier than SVM. In order to solve the multi-class problem, two-class classification methods has been extended to multi-class classification methods and many algorithms have been developed. There are three classes can be described as following, one to many combination model, one to one combination model, and SVM decision tree. With weighted posterior probability support vector machines used on sub-pixel unmixing on remote sensing images, the classifier number are depressed than remote sensing images unmixing without weighted posterior probability support vector machines. The classification result based on weighted posterior probability support vector machines are more accurate according to empirical knowledge. Sample weighting is the main reason avoiding the negative effect of ambiguous ground class. The ground object endmembers can be determined by the proposed method and the posterior probability also be count out at the same time. The result of posterior probability is considered as the percent of each ground object belong to a pixel of remote sensing images. Multi-channel remote images data are used to validate the proposed method in this paper. The experiment results show that the unmixing model based on weighted posterior probability support vector machines has been improved over support vector machines algorithms. The precision of unmixing result obtained based on the proposed method is better than those of support vector machines algorithms.