Hyperspectral Remote Sensing Image Terrain Classification Based on M-ary Discriminant Analysis

Hyperspectral remote-sensing image has high data dimensionality and a small amount of labeled pixels, which causes the curse of dimensionality phenomenon. Therefore, feature extraction is needed ahead of recognition for reducing dimensionality and improving classification accuracy. A novel multiclass feature extraction method, i.e., M-ary discriminant analysis (M-ary DA), is presented for solving the problem. For a c-class feature extraction problem, M-ary DA performs the base 2 logarithm of c 2-class kernel linear discriminant analysis (KLDA) processes to extract features according to each bit of the gray codes of class labels. M-ary DA sufficiently discovers the potential feature extraction ability of the 2-class linear discriminant analysis (LDA), and reduces the subspace dimensionality from c minus one to the base 2 logarithm of c. The experimental results for three real datasets prove that, compared with the LDA and the KLDA method, the proposed M-ary DA algorithm can evidently enhance the terrain recognition accuracy.