Hyperspectral band selection based on metaheuristic optimization approach

Abstract

Hyperspectral images generally contain hundreds of contiguous spectral bands, which can precisely discriminate the various spectrally similar land cover classes. However, such high dimensional data also contains highly correlated but irrelevant band information. Selection of useful bands is an effective approach to alleviate the curse of dimensionality of hyperspectral image classification. In this manuscript, band selection problem is formulated as a combinatorial optimization problem. Cuckoo search (CS) is one of the widely used, effective nature inspired algorithm in global optimization approaches. In spite of its efficiency, the homogeneous search behaviour of the standard CS leads to slower convergence and gets easily trapped into a local optimal solution. A modified CS (MCS) algorithm is proposed for dealing with these drawbacks and solving the problem of hyperspectral band selection. The proposed algorithm uses the Chebyshev chaotic map to initialize the nest locations at initial step. Meanwhile, the step size is adaptively adjusted based on fitness value and current iteration number. The experimental results on two standard benchmark datasets namely, Pavia University and Indian Pines, prove the superiority of the proposed method over standard CS approach as well as the other traditional approaches in terms of average accuracy, overall accuracy, Cohen’s kappa coefficient (κ), statistical significance assessment using McNemar’s test, and fitness curve analysis. The proposed technique has achieved the maximum overall accuracy of 95.10% for Pavia University dataset, and 86.92% for Indian Pines dataset.