基于多特征核的高光谱遥感影像分类方法

Multi-Feature Kernel Based Classification for Hyperspectral Remote Sensing Image

支持向量机(SVM)的核函数选取是制约其分类性能的重要因素,而当前的核函数大多以光谱距离作为构核元素,而忽略了光谱角度这一光谱特征。文章提出一种均衡化光谱距离与光谱角多特征组合核(ESAD)的SVM分类器,对2003年意大利帕维亚大学的ROSIS高光谱数据作分类处理,并对影像的分类精度作评价分析。实验结果表明:ESAD核SVM整体分类精度相较于光谱距离核SVM和光谱角核SVM分别提升8.88%和11.03%,分类精度理想,一定程度上抑制了“同谱异物”现象。

Kernel function is an important element which can absolutely affect the ability of Support Vector Machine(SVM),most of kernel functions are made of spectral distance only today,and they often ignore the spectral angle which is an important feature of an image.This paper proposes a SVM classifier who’s kernel is made of multi-feature Equalized Spectral Angle and Distance(ESAD),and classes Pavia University,Italy with ROSIS hyperspectral data which was gotten in 2003 with that.It also evaluates the accuracy of the classified image.It turns that the ESAD kernel SVM gets the overall accuracy increased by 8.88%and 11.03%comparing with spectral angle kernel SVM and spectral distance kernel SVM,the accuracy is optimistic and the method can solve the problem of“different objects with same spectral curve”during image classifying.