A Supervised Band Selection Method for Hyperspectral Images Based on Information Gain Ratio and Clustering

Abstract

The high spectral dimension of hyperspectral images increases the computational complexity and processing time requirement for analysing such images. As a consequence, dimension reduction is an essential step to be performed prior to the processing of hyperspectral images. In this work we have presented a dimension reduction technique for hyperspectral images using a supervised band selection method. Initially significance of each band were evaluated by calculating information gain ratio. Then, by applying clustering technique, similar bands were grouped into k clusters, from each of which the band with the maximum information gain ratio value was selected as the representative band. In a subsequent band pruning step, this set of representative bands were further reduced by eliminating the bands with low information gain ratio. The results of the experiment showed that adequate accuracy could be achieved with relatively low number of bands selected by the proposed method.