A Binary Grey Wolf Optimization based Hybrid Convolutional Neural Network (BGWOHCNN) framework for hyperspectral image classification

In recent years deep learning (DL) models have obtained great success in hyperspectral image classification (HSIC) with commendable performance and especially convolutional neural networks (CNNs) have attracted huge attention due to the exceptional performance demonstrated in this area. However, most of the CNN-based models have to suffer with low classification performance due to non-availability of abundant training samples. Recently, hybrid-CNN models, utilizing both spectral and spatial features unitedly, have exhibited remarkable classification accuracy. Yet, the hybrid-CNN models have been adopted in very limited research works due to a very high computational complexity. Furthermore, huge dimensional HSIs contain highly correlated but irrelevant bands. The selection of most relevant spectral bands affects the performance and computational complexity of HSIC models very profoundly. To address these issues, the authors have proposed a binary grey wolf optimization-based hybrid CNN (BGWOHCNN) framework for HSIC, in this paper. In the hybrid framework, a 3-D CNN has been employed to exploit spatial and spectral features jointly along with a 2-D CNN utilizing the spatial features and reducing the computational complexity of the overall design. In addition, binary grey wolf optimization (BGWO) technique has been adapted to select the most relevant spectral bands for HSIC. As per the best of authors' knowledge and belief, the BGWO technique has been investigated for the first time with hybrid CNN for HSIC in the current work. Further, the experiments are conducted to investigate the performance and superiority of the proposed framework over three very popular datasets viz. Salinas, Indian Pines and Pavia University. The obtained results are compared with the six state-of-art DL models, in terms of average accuracy, overall accuracy and Cohen's kappa coefficient, to show that the proposed framework provides very promising results to HSIC tasks.