Hyperspectral　images　can　provide　dozens　to　hundreds　of　continuous　spectral　bands,　greatly　enhancing　the　richness　of　information.　However,　the　redundancy　among　adjacent　bands　leads　to　increased　data　processing　complexity.　Despite　the　recent　introduction　of　numerous　band　selection　methods,　there　has　been　limited　focus　on　incorporating　context　information　from　the　entire　spectral　range　into　this　task.　Furthermore,　researchers　have　primarily　concentrated　on　band　informativeness　and　sparse　representation　for　reconstructing　all　bands,　often　overlooking　the　separability　of　classes　in　downstream　tasks.　To　address　these　challenges,　we　propose　a　hyperspectral　band　selection　network　combining　Siamese　Network　Local-Global　Attention　(SLGA).　This　approach　first　segments　the　hyperspectral　image　into　homogeneous　regions　and　constructs　sample　pairs　based　on　a　random　elimination　strategy.　Next,　it　utilizes　the　Local-Global　Attention　(L-GA)　mechanism　to　obtain　band　weights　that　capture　both　local　and　global　spectral　structures.　These　reweighted　bands　are　then　fed　into　a　twin　network　to　obtain　their　high-dimensional　representations,　compute　loss　values,　and　update　network　parameters.　Finally,　extensive　classification　experiments　using　SVM,　KNN,　and　LDA　classifiers　are　conducted　on　the　Indian　Pines　and　Botswana　hyperspectral　image　datasets.　The　results　from　these　experiments　on　benchmark　datasets　demonstrate　that　the　proposed　SLGA　method　performs　exceptionally　well,　outperforming　state-of-the-art　algorithms.　©　2023　IEEE.