Band selection using oppositional whale optimization for hyperspectral image classification

This paper presents an innovative methodology for band selection pertinent to hyperspectral remote sensing imagery. The substantial volume of data, alongside its inherent redundancy and limited training samples, adversely influences the classification precision of these images. The discernment of informative, non-redundant, and uncorrelated bands from hyperspectral imagery represents a principal aim of the hyperspectral research community. In this study, we have proposed a pioneering band selection technique that emulates the hunting strategy of Whales, incorporating opposition learning to leverage the alternative candidate solutions. Subsequently, the intrinsic features are extracted from the oppositional whale bands and subjected to training via a three-dimensional convolutional neural network for classification, referred to as Modified Whale Optimization (MWO). The MWO is assessed against leading-edge methodologies across three benchmark datasets – Indian Pines, University of Pavia, and Salinas, both qualitatively and quantitatively. The reported classification accuracies are 98.67 %, 99.81 %, and 99.98 % respectively across the three datasets, achieved with a minimal number of bands. This methodology proves to be effective for applications in Land Use and Land Cover as well as Mineral identification.