Abstract: The existing band selection methods for hyperspectral images are limited by specific feature extraction techniques and band selection strategies, and fail to fully consider multi-scale features and information from different subspaces, resulting in incomplete or inaccurate band selection results. To address this issue, this study proposes a method based on multi-scale feature fusion and subspace partitioning. This method mainly consists of two parts, namely the Multiscale Feature Extraction Fusion Module (MSFEFM) and the Gaussian Mixture Model Based Subspace Partitioning (GMMSP). Firstly, the original hyperspectral image is processed using Principal Component Analysis (PCA), and then the multi-scale features are gradually extracted through Gaussian pyramid decomposition. Secondly, by using Graph Attention (GAT) networks, feature information of different scales can be effectively fused. Finally, the subspace partitioning method based on Gaussian mixture model is used to recombine the fused features, achieving fine segmentation of hyperspectral images and selecting the optimal band. Experiments were conducted on three datasets, Indian Pines, Pavia University, and Salinas, and classification was used to verify the band selection effect. The results showed that overall accuracy could achieve 79.96%, 93.96%, and 93.37%, respectively, which significantly improved the classification accuracy compared to other band selection methods.