A Binary Social Spider Optimization algorithm for unsupervised band selection in compressed hyperspectral images

Abstract

The presence of high spectral correlation in hyperspectral images has demanded dimensionality reduction. Selection of bands based on supervised techniques has been already developed for dimensionality reduction. In this manuscript, the selection of an optimal set of bands is formulated as an unsupervised problem. The images are initially compressed using discrete wavelet transform (DWT). Then, the details of the image are extracted in the transform domain using entropy which accounts for spatial information content and first spectral derivative (FSD) that exploits the temporal redundancy. These details are optimized using meta-heuristic algorithms to obtain the final set of bands. A new Binary Social Spider Optimization algorithm (BSSO) is proposed which maintains the exploration–exploitation balance and avoids premature convergence of solutions. The performance of BSSO is compared with other benchmark algorithms: Clonal algorithm (BCLONAL), Time-varying transfer function Binary Particle Swarm Optimization (TV-BPSO), Binary Genetic Algorithm (BGA), Improved Gravitational Search Algorithm (IBGSA) and New Binary Bat Algorithm (NBBA). A simulation study is carried on four standard hyperspectral images. The proposed BSSO have provided with 83% of overall accuracy for 24% bands in Salinas image set, 75% for 33% in Pavia University, 75% against 26% for Pavia Center and finally 91% in KSC with 32% bands retention. The achievement of higher accuracy at less than 50% bands is appreciable.