Evaluation of CNN Models Using Deep Reinforcement Learning for Band Selection on Hyperspectral Image Classification

Abstract

Along with the high spectral rich information it provides, one of the difficulties in processing a hyperspectral image is the need for expert knowledge and high-spec hardware to process very high-dimensional data. The use of the most relevant bands in the hyperspectral image is quite decisive in deep CNN networks without loss of information and loss of accuracy. It is crucial to classify hyperspectral images with faster and less hardware-requiring models by creating subset groups by choosing a limited number of optimal bands. In this study, a comparative analysis about the effect of deep reinforcement learning (DRL)-based hyperspectral band selection on the classification performance of deep learning networks is presented. 3D CNN, 3D + 1D CNN and Multiscale 3D deep convolutional neural network (M3D-DCNN) algorithms were used for hyperspectral image classification. By choosing the most effective bands determined by DRL, it is aimed to perform classification with high accuracy with fewer bands instead of all bands. All tests were performed on popular hyperspectral datasets, Indian Pines, Salinas, and Pavia Center. The 3D + 1D approach reached 92.28% OA in the IP dataset. In Salinas, 94.87% OA with 3D CNN and 94.62% OA with M3D-DCNN was obtained. 3D + 1D CNN has 98.64% OA in PaviaC.