Segmentation of multispectral images based on band selection by including texture and mutual information

Minimization of the dependent information (MDI) has been widely used for the selection of multispectral bands. It selects multispectral image bands using information theory concepts that consist of a relation between the joint entropy and the union of the conditional entropies of the considered set of image bands. In this paper, a new band selection method is proposed for segmentation of multispectral textures such that relevant texture information is maximized while reducing the number of spectral bands. Therefore, the proposed method could be viewed as an enhancement of MDI method. The idea is based on MDI, by using the co-occurrence matrices, under Classification Accuracy Rate (CAR) and computational requirement constraints. According to the optimal band subset selected by the proposed method, we make a segmentation of multispectral images and evaluate its accuracy. Experimental results show that the proposed band selection method compares favorably with the MDI.