Deep Reinforcement Learning With Mutual Information for Selecting Bands in Hyperspectral Image Classification

Uyperspectral imaging (USI) is a critical technology in remote sensing that provides extensive spectral information for material identification and classification on Earth's surface. Despite its benefits, the high dimensionality of USI data introduces significant challenges, such as elevated computational costs and extensive storage needs due to band redundancy. This research proposes an innovative approach to hyperspectral band selection using Deep Reinforcement Learning (DRL) with Mutual Information as a reward mechanism, aiming to more effectively capture inter-band dependencies than traditional methods. This approach selects the most informative bands that provide significant contributions to achieving better classification results. The study also introduces an advanced classifier array incorporating Convolutional Neural Networks (CNNs) alongside traditional machine learning classifiers like K-Nearest Neighbors (KNN) and Support Vector Machines with Radial Basis Function (SVM-RBF) to leverage both spectral and spatial data relationships. The DRL model, framed as a Markov Decision Process, learns an action-value function to select optimal band combinations, enhancing classification accuracy and reducing data redundancy. Experiments conducted on the Indian Pines and Botswana datasets demonstrate that the Mutual Information-based DRL model significantly outperforms traditional entropy-based methods across all used classifiers, confirming the efficacy of the proposed approach. This integration of Mutual Information and a DRL framework offers promising advancements for efficient and effective hyperspectral imaging systems, setting a new benchmark in the field.