A Review of Band Selection in Hyperspectral Imaging: from Classical Heuristics to Deep Learning Challenges

Hyperspectral images, with their hundreds of narrow and highly correlated spectral bands, provide rich spectral information essential for accurate classification of land-cover and terrestrial objects. However, the high dimensionality of hyperspectral data introduces redundancy and incur in the curse of dimensionality, often reducing classification performance when limited training samples are available. To address these challenges, dimensionality reduction, particularly hyperspectral band selection, is an essential preprocessing step, aimed to retain discriminative spectral channels while significantly reducing the computational burden of downstream classification and mitigating the curse of dimensionality problem. Recent advances in deep learning have significantly enhanced hyperspectral band selection, enabling a more precise extraction of informative bands through advanced feature representation and optimization methods. This article presents a comprehensive review of hyperspectral band selection methods, discusses the technical challenges, including issues arising from non-differentiable band selection operations, and outlines potential solutions and future research directions. The survey serves as a guide for researchers and practitioners in selecting appropriate band selection methodologies, facilitating improved classification performance across diverse real-world hyperspectral imaging applications.