Band selection based on a new seperability measure for hyperspectral images classification

Classification is an important application of hyperspectral images. In terms of classification, separability measures are usually used for band selection, and Bhattacharyya distance (BHD) is a conventional separability measure. In this paper, a new separability measure, weighted Bhattacharyya distance (W-BHD) is introduced and a band selection algorithm based on it is proposed. The proposed algorithm composes of four steps: initial subset selection, weights calculation, W-BHD calculation and final subset selection. W-BHD calculation is crucial part of the proposed algorithm. W-BHD assigns different weights to each class pair, which improves classification accuracy by stressing importance on hard-to-separate class pairs. Numerical experiments are conducted on two hyperspectral data respectively, and results show that W-BHD greatly outperforms other separability measures based on BHD.