Neighborhood mutual information and its application on hyperspectral band selection for classification

Band selection is considered to be an important processing step in handling hyperspectral data. In this work, we combined Shannon's information entropy with neighborhood rough set and proposed a new measure, called neighborhood mutual information. With the proposed measure which can evaluate the significance of bands for classification, a forward greedy search algorithm for band selection was constructed. To assess the effectiveness of the proposed band selection technique, two classification models (Extreme Learning Machine and Random Forests) were built. The proposed algorithm was compared to neighborhood dependency measure based algorithm, genetic algorithm and uninformative variable elimination algorithm on three (soybean, maize and rice) hyperspectral datasets between 400 nm and 1000 nm wavelengths. Experimental results show that the proposed method can effectively select key bands and obtain satisfactory classification accuracy. (C) 2016 Published by Elsevier B.V.