<p>Hyperspectral images, characterized by rich spatial and spectral information, have been increasingly applied across numerous fields. However, the high dimensionality inherent in hundreds of spectral bands results in the “curse of dimensionality,” significantly complicating data processing. Therefore, dimensionality reduction techniques, especially band selection, have become a pivotal research focus in the hyperspectral imaging community. Recently, leveraging its powerful feature representation capabilities, deep learning has revolutionized the development of band selection approaches. This study comprehensively reviews recent advances dedicated to deep learning for hyperspectral band selection and delineates a universal fundamental technical framework of deep learning-based band selection, providing a global perspective on this methodology. Furthermore, it develops a multi-perspective analytical scheme that thoroughly elaborates on deep learning-based band selection techniques from multiple aspects, including learning paradigms, neural network architectures, selection strategies, and optimization methods for non-differentiable operations. On this basis, an integrated perspective is provided to sort out the overall evolutionary trajectory of this field and analyze the cross-perspective correlation, compatibility, and synergy. From this comprehensive synthesis, the study further explores critical challenges in this domain and outlines potential prospects, providing valuable guidelines for future research.</p>