基于核加权类对准则的高光谱影像特征提取

Spectral feature extraction of hyperspectral remote sensing images based on kernel class pair-weighted criterion

针对高光谱遥感影像中相似光谱的不同地物与野类同时存在时,提取有效的非线性可分性特征的问题,提出一种核加权类对准则。首先,推导出类对形式的核线性判别分析准则,即核类对准则,将核类间和类内散布矩阵均表示为类对形式。然后,提出核加权类对准则,依据核空间中各类对的可分性分别对各类对的核类间和类内散布矩阵进行加权,使得各类对的可分性均衡地保留在特征子空间中。采用K近邻分类器和最小距离分类器评估特征提取的效果。基于两个实测高光谱遥感影像的实验结果均表明:相比原空间法、核线性判别分析方法和kernelweightedpairwiseFisher准则,所提核加权类对准则在降维的同时,通过提高可分性小的类对的识别率来提高整体地物识别率。

To extract efficient nonlinear discriminant features when foreign objects,with similar spectra and outlier classes,are present in hyperspectral remote sensing images(HRSIs),a kernel class pairweighted(KCP-weighted)criterion is proposed.First,we derive a class pair form of the kernel linear discriminant analysis(KLDA)criterion,viz.the kernel class pair(KCP)criterion,in which the kernel-between-class and kernel-within-class scatter matrices are both expressed in the form of class pairs.Then,the KCP-weighted criterion is proposed to weight the kernel-between-class and kernel-within-class scatter matrices of each class pair according to their separability in a kernel space.The KCP-weighted criterion can ensure that the separabilities of class pairs are balanced in the KCP-weighted feature subspace.Finally,the K-nearest neighbor and minimum distance classifiers are used to evaluate the feature extraction performance.Experimental results of two real HRSIs show that,compared with the original space and KLDA methods as well as the kernel weighted pairwise Fisher criterion,the presented KCP-weighted criterion can effectively improve the overall terrain classification rate while reducing the dimensionality of the data.