Improved whale optimization based band selection for hyperspectral remote sensing image classification

Classification of hyperspectral images is one of the emerging areas of remote sensing. The high volume of the data, contiguous acquisition and correlation among the bands pose problems in the extraction of informative bands. To address the issues and to make use of the fast multi-resolution analysis technique for the remote sensing images, a new framework is designed with the inclusion of wavelet analysis. Here, a threefold strategy is implemented. In the first fold, a modified whale optimization algorithm by mimicking the hunting behavior of humpback whale is implemented for the selection of informative band set with nonlinear function and tournament selection. In the second fold, a fast and multi-resolution analysis technique, a three dimensional discrete wavelet transform, is implemented to elucidate the variation among the selected bands by convolving in three dimensions including spectral dimension. Then, a convolution neural network with three dimensional convolutions is trained by fusing the spectral and wavelet based spatial features in the third fold. The performance of the proposed architecture is tested with the state of the art methods on Indian Pines, University of Pavia and Salinas datasets. The classification maps of the proposed method show the effectiveness of the wavelet based approach and reported an overall accuracy of 99.44%, 99.85% and 99.83% on the three datasets respectively. Also, the Mean Spectral Divergence (MSD) measure values with the discrete wavelet transform on the datasets show low redundancy between the bands and hence improved the classification accuracy of the hyperspectral images.