Significance  Hyperspectral images (HSIs), which simultaneously contain abundant spectral signatures and spatial structural information, have been widely used in various fields such as remote sensing monitoring, environmental monitoring, and agricultural resource management. However, the high dimensionality of HSIs inevitably introduces substantial redundancy, leading to increased storage requirements and computational burden. Moreover, the strong inter-band correlations and the susceptibility of certain bands to noise further exacerbate the difficulty of HSI processing and analysis. Therefore, effectively reducing the dimensionality of hyperspectral data, minimizing redundant information, and improving computational efficiency have become key research areas in HSI processing. The goal of dimensionality reduction is to preserve as much useful information as possible while removing redundant components and lowering computational cost; therefore, it has become an indispensable step prior to practical HSI applications. Existing dimensionality reduction techniques mainly follow two routes: feature extraction and band selection. Feature extraction transforms the original data into a new low-dimensional feature space through mapping, which may alter the physical meaning of the original spectral variables. In contrast, band selection directly selects a representative subset of original bands, reducing dimensionality and redundancy while largely retaining the physical interpretability of spectra and the key information of the imagery. For this reason, band selection has long-term research value and strong practical demand in hyperspectral engineering applications, and developing efficient and accurate band selection algorithms remains one of the key problems in HSI studies.

Progress  A systematic review of hyperspectral band selection methods is organized around three aspects: evaluation criteria, methodological taxonomy, and emerging trends. First, from the perspective of dimensionality reduction, the differences and characteristics of feature extraction and band selection are clarified, and the advantages of band selection are highlighted in terms of preserving the physical interpretability of original bands, supporting subsequent spectral analysis, and enabling cross-task reuse. Second, from the methodological perspective, band selection approaches are categorized into supervised and unsupervised paradigms according to whether class label information is utilized. For supervised band selection, methods are further divided based on the degree of coupling between the selection procedure and the learner into three categories: filter methods, wrapper methods, and embedded methods. For unsupervised band selection, five classical families are summarized: 1) Ranking-based methods, which score and order bands mainly according to information content or redundancy measures; 2) Clustering-based methods, which group bands and select representative bands from each cluster to enhance diversity; 3) Search-based methods, which explore the combinatorial space via point-wise or group-wise strategies to obtain improved subsets; 4) Sparse-representation-based methods, which leverage sparsity constraints to capture inter-band structures and suppress redundancy; and 5) Hybrid schemes, which combine multiple criteria or modules to balance information preservation, redundancy reduction, robustness, and separability. Although these traditional approaches have achieved notable success to some extent, they remain limited when facing the complex nonlinear dependencies and high-dimensional characteristics of HSIs, because they largely rely on manually designed selection criteria and thus may fail to fully exploit the latent structure of the data. In recent years, the rise of deep learning has provided new approaches to band selection. Through deep learning models such as Convolutional Neural Networks (CNNs), Autoencoders (AEs), Attention mechanisms, and Graph Neural Networks (GNNs), complex nonlinear relationships within hyperspectral data can be effectively learned, and the most representative band subsets can be automatically selected. These deep learning models have shown strong potential in band selection tasks. For example, CNNs can learn spatial information from local features, AEs perform unsupervised learning to map nonlinear relationships between bands, and Attention mechanisms adaptively focus on the most important bands, improving both selection efficiency and accuracy. Additionally, GNNs have been employed to model the graph-based relationships between bands, further enhancing the performance of band selection.

Conclusions and Prospects  With the continuous advancement of deep learning technologies, deep neural network-based band selection methods have demonstrated significant advantages, particularly when processing high-dimensional, complex data. These methods are capable of automatically learning the complex relationships between bands and have shown superior performance in various tasks. However, deep learning methods still face challenges in computational efficiency and generalization ability, especially when dealing with large-scale datasets. Future research should focus on improving the computational efficiency of deep learning models, reducing their complexity, and enhancing their generalization ability across various tasks. Furthermore, with the ongoing development of technologies such as drone-based remote sensing and satellite remote sensing, the demand for hyperspectral images will continue to grow, and the research on band selection will increasingly focus on real-time performance, cross-platform applications, and more complex environments. Therefore, future band selection methods will not only need to improve computational efficiency but also be adaptable to meet diverse application requirements.