Review on different band selection approaches for classification of hyperspectral images

Hyperspectral Imaging (HSI) has become a very necessary technology in remote sensing, agriculture, medical diagnostics, and environmental studies because of the imaging ability of hundreds of narrow spectral bands. Hyperspectral data is high dimensional, which causes spectral redundancy, expense of computation and Hughes phenomenon, where classification performance decreases as the number of features increases greatly relative to the available training samples. The given paper is a systematic literature review (SLR) of the hyperspectral band selection methods published in 2023-2025 that were retrieved through IEEE Xplore, ScienceDirect, Scopus, ACM Digital Library, and SpringerLink. The methods reviewed are classified into four methodological paradigms namely filter-based, wrapper-based, embedded and hybrid methods. Synthesis of 72% peer-reviewed works was done. Hybrid and deep-learning-enhanced designs, such as reinforcement learning (RL), multi-objective optimization (MOO), and transformer-based selection, are more accurate (OA ≥ 97%) and have competitive computational power (<300 ms). Benchmark analysis of the data of Indian Pines, Salinas, Pavia University, and Kennedy Space Centre (KSC) shows that Proximal Policy Optimization (PPO)-based RL has the largest OA (97.9%), and Discrete Wavelet Transform (DWT)-based fusion has the lowest cost-accuracy ratio. The areas of critical research gaps in the adaptive online selection, edge-deployable lightweight models, domain-specific spectral optimization, and explainability of deep-learning selectors remain. The research needs to focus its future efforts on interpretable, real-time and cross-domain adaptive band selection frameworks to enable hyperspectral imaging to be deployed to practice.