Band selection (BS) is an effective approach to alleviate the curse of dimensionality in hyperspectral image (HSI). Despite the plethora of BS methods proposed, two critical issues persist. Firstly, many approaches fail to integrate local and global band characteristics when describing their adjacencies. Secondly, HSI is typically treated as a whole in spatial information extraction, ignoring the differences of spatial structure inherent in each band. To address these issues, this article proposes a novel spectral-spatial multi-view sparse self-representation model for hyperspectral BS. Firstly, a dynamic grouping strategy is designed to partition bands by incorporating the local and global adjacencies, promoting intra-group coherence and inter-group disparity. Accordingly, metrics such as local density, the difference between intergroup and information entropy are combined to evaluate band significance within each group, and ultimately selecting bands with high information and low redundancy to constitute a feature band subset. Secondly, a series of spatial similarity graphs of the feature bands is constructed to capture the spatial structure differences across multiple views. Simultaneously, a weighted adaptive multi-graph fusion strategy is developed to leverage the strengths of these graphs, yielding a unified similarity graph. This approach effectively exploits both local and global band adjacencies, so as to capture the spatial distribution differences of ground objects more precisely. Finally, experiments on two public datasets demonstrate the superiority of the proposed model.