Non-overlapping classification of hyperspectral imagery with superpixel segmentation

Abstract

The spectral–spatial classification methods have improved the classification accuracy of hyperspectral imagery dramatically, however, the traditional random sampling method causes the high overlap between training samples and testing samples. So, the improvement of the accuracy is partly due to the overlap of these samples. In this paper, a non-overlapping classification framework for hyperspectral imagery based on the distance of regularized nearest points (NORNP) is proposed. First, a novel sampling method named controlled random sampling is adopted to generate the training sets. The controlled random sampling can reduce the overlap between the training and testing samples a lot. The training samples with the same label will be used to form a set. Then, the superpixel segmentation method is adopted to divide the hyperspectral imagery into many smoothed small regions. Finally, all sets are represented as regularized affine hulls (RAHs). Every two RAHs can automatically choose two regularized nearest points (RNP) by an iterative algorithm. The distance between RNPs is used as the sets distance. The label of testing set is determined by the minimum distance to each training set. Experimental results based on three real hyperspectral data sets demonstrate that the proposed set-to-set classification method can provide better classification results when training samples and testing samples are less overlapped.