Feature Selection Based on Bhattacharyya Distance: A Generalized Rough Set Method

In allusion to the feature selection of multiclass problems, a generalized attribute reduction method based on between-class distance is proposed. In the new method, the interclass separability is described by the tolerance relation and distance matrix defined by Bhattacharyya distance. Some attribute reduction problems were investigated on generalized approximation space, which make use of the tolerance relation and the matrix. Based on the Boolean reasoning and the monotony of the distance function, a feasible algorithm for feature selection is presented. Compared with traditional principle component transformation (PCT) approach, practical experiment result shows that the proposed method can prominently improve the accuracy of classification and generalization on hypersepectral remote sensing images.