Contextual Classification of Hyperspectral Remote Sensing Images Using SVM-PLR

In this paper, we propose a novel contextual classification of hyperspectral data. We use probabilistic label relaxation (PLR) process to incorporate context information into the spectral pixelwise classification procedure. In conventional PLR procedure, first a maximum likelihood classification is performed and class probabilities are computed by using multivariate normal models. However this method is not efficient for hyperspectral data with limited training samples. In this paper we suggest to use support vector machine (SVM) in order to initial classification and also use class probability estimates which are obtained from SVM classification for PLR postprocess. We call this proposed method as SVM-PLR. Experimental results are presented for two types of hyperspectral images, agricultural and urban data. The proposed method improves dramatically classification accuracies, when compare to spectral pixelwise classification. Moreover our proposed method can improve performance of conventional PLR postprocess for hyperspectral data.