一种特征选择的全极化雷达影像分类方法  

A fully polarimetric radar image classification method based on feature selection

作　　者：张继超[1] 刘宁 宋伟东[1] 李建飞 ZHANG Jichao;LIU Ning;SONG Weidong;LI Jianfei(School of Geomatics,Liaoning Technical University,Fuxin,Liaoning 123000,China;Liaoning Basic Institute of Surveying and Mapping,Liaoning Natural Resources Affairs Service Center,Jinzhou,Liaoning 121000,China)

机构地区：[1]辽宁工程技术大学测绘与地理科学学院,辽宁阜新123000 [2]辽宁省自然资源事务服务中心辽宁省基础测绘院,辽宁锦州121000

出　　处：《测绘科学》2022年第6期127-134,共8页Science of Surveying and Mapping

基　　金：国家“863”计划资助项目(2011AA120404);国家自然科学基金项目(42071343)

摘　　要：针对极化合成孔径雷达(PolSAR)影像面向对象分类过程中存在数据冗余、特征维数高导致分类精度降低的问题,该文提出了一种基于信息增益比和基于相关性的特征选择(CFS)算法的分类方法。该方法首先在经典的过滤式CFS算法基础上,引入信息增益比评估模型舍弃贡献小的特征。然后采用目前流行的封装式分类回归树(CART)算法做进一步筛选并分类。最后以GF-3不同场景和成像时间的影像数据为例进行实验,将该方法与信息增益比评估模型优化特征集、CFS算法优化特征集、全部特征集的CART分类结果进行对比。结果表明,该方法各项精度评价指标均优于其他对比方法,验证了该方法在PolSAR影像面向对象分类领域的可行性。In view of the problem of data redundancy and high feature dimension in the process of object-oriented classification of polarimetric synthetic aperture radar(PolSAR) images, a classification method based on information gain ratio CFS algorithm was proposed in this paper. Firstly, on the basis of the classical filtering correlation-based feature selection(CFS) algorithm, the information gain ratio evaluation model was used to remove the features with less contribution. Secondly, the popular encapsulated classification and regression tree(CART) algorithm was used to further screen and classify it. Finally, GF-3 satellite images with different scenes and imaging time were taken as examples to compare the proposed method with the CART classification results of the information gain ratio evaluation model, the optimized feature set of CFS algorithm and all feature sets. Experimental results showed that the accuracy evaluation index of this method was better than other comparative methods, which verified that the method was feasible in the field of object-oriented classification of PolSAR images.

关 键 词：极化分解 特征选择 信息增益比 CFS算法 CART分类 POLSAR 

分 类 号：P237[天文地球—摄影测量与遥感]