特征维数对支持向量机分类器性能影响的研究——以高光谱遥感影像为例  

Impacts of feature dimensionality to the support vector machine classifier for hyperspectral remote sensing image

作　　者：谭琨[1,2] 杜培军[1,2] 王小美[1,2] 

机构地区：[1]国土环境与灾害监测国家测绘局重点实验室(中国矿业大学),江苏徐州221116 [2]中国矿业大学测绘与空间信息工程研究所,江苏徐州221116

出　　处：《测绘科学》2011年第1期55-57,31,共4页Science of Surveying and Mapping

基　　金：国家自然科学基金项目(40401038);国家高技术研究发展计划863计划项目(2007AA12Z162);高等学校博士学科点专项科研基金资助课题(20070290516);江苏省普通高校研究生科研创新计划资助项目(CX08B_112Z);中央高校基本科研业务费专项资金资助(2010QNA18)

摘　　要：本文为验证SVM对高维特征的适应性和可靠性,针对不同特征提取方法与特征组合,以国产OMISⅡ传感器获得的北京昌平地区高光谱遥感据为例,对SVM分类器中特征维数对分类准确率的影响进行了试验,通过对主成分分析、最小噪声分离算法、相关系数分组后特征提取、导数光谱等的分析,表明SVM分类器的分类精度随着特征维数波动,其中主成分分析降维后提取的特征具有用于分类能够获得最高的准确率。通过与最大似然法和光谱角制图分类算法的比较,说明在同样的特征输入情况下SVM分类算法分类的准确率高于最大似然法和光谱角制图分类器。According to the SVM theory and the characteristics of hyperspectral remote sensing image, the classification model which is the characteristic extraction based on SVM was constructed in the paper, and the values of the penalty parameter of the error term and Radical Basis Function kernel parameter were gained by grid researching. In order to confirm SVM to the high dimension characteristic compatibility and the reliability, the different feature extraction methods and the characteristic combinations were considered, it experimented on the hyperspeetral image with the 64 bands OMIS data of Beijin Chang Pin. By experiment on the characteristic extraction algorithm of Principal Component Analysis （ PCA）, Minimum Noise Fraction （ MNF）, characteristic extraction after correlation coefficient grouping, derivative spectral and so on, it indicated that the support vector classification model fluctuated with characteristic dimension and PCA had the best precision. It was concluded that SVM classifier had more advantages under the same feature input after comparing with Maximum Likelihood classification and Spectral Angle Mapping classification model.

关 键 词：高光谱遥感 支持向量机 特征提取 分类 

分 类 号：TP753[自动化与计算机技术—检测技术与自动化装置]