利用改进SEaTH算法的面向对象分类特征选择方法  

Object-oriented Feature Selection Algorithms Based on Improved SEaTH Algorithms

作　　者：余晓敏[1,2] 湛飞并[1] 廖明生[3] 胡金星[2] 

机构地区：[1]武汉大学资源与环境科学学院,武汉市珞喻路129号430079 [2]中国科学院深圳先进技术研究院,深圳市学苑大道1068号518055 [3]武汉大学测绘遥感信息工程国家重点实验室,武汉市珞喻路129号430079

出　　处：《武汉大学学报（信息科学版）》2012年第8期921-924,共4页Geomatics and Information Science of Wuhan University

基　　金：国家863计划资助项目(2010AA122202);国家重大国际合作资助项目(2010DFA92720);深圳基础研究(重点项目)资助项目(JC201005270334A)

摘　　要：针对分离阈值法(SEaTH)仅从类间距离评价特征,没有考虑类内距离和特征之间相关性的不足,提出了一种改进的SEaTH算法——ISEaTH。该算法分别依据特征相关性、类间距离和类内距离对特征进行评价,然后综合利用多种评价结果获取最优的特征子集。采用新疆喀什地区的QuickBird数据进行了特征选择的实验。结果表明,该方法不但能降低特征维数,有效优化特征空间,还能提高分类精度。SEaTH algorithm is an object-oriented feature selection algorithm based on inter- class distance. We present an improved SEaTH--ISEaTH. It overcomes some limitations of existing SEaTH algorithms. The ISEaTH algorithm evaluates the features according to the relation between features, inter-class distance and intra-class distance respectively, and in- tegrates multiple estimation results to obtain the best feature subset. Compared SEaTI-I and ISEaTH with QuickBird data in Kashi area of Xinjiang province, the experimental results show that the ISEaTH algorithm not only reduces feature dimensions, but also improves the classification accuracy, thus is a much more efficient object-oriented feature selection algo- rithm.

关 键 词：面向对象分类 特征选择 SEaTH算法 

分 类 号：P237.3[天文地球—摄影测量与遥感]