MS 2 DB-Mamba: a spatial-spectral dual-branch Mamba based on multi-scale spiral scanning for hyperspectral image classification

Hyperspectral image (HSI) classification is a key task in remote sensing. The Mamba architecture provides a lightweight approach to HSI classification with linear computational complexity. Existing Mamba methods mostly adopt row-column or multi-directional scanning to model spatial dependencies. However, these patterns may disrupt spatial adjacency or introduce redundant modeling, while also lacking an explicit representation of the central pixel. Furthermore, applying multi-scale convolution after spatial-spectral fusion undermines the independence of spatial modeling, causing multi-scale feature extraction to be disturbed by spectral information. We therefore propose MS 2 DB-Mamba: a spatial and spectral dual-branch Mamba model based on multi-scale spiral scanning. The model employs a spatial-spectral dual-branch architecture to fully exploit spatial and spectral information while maintaining a lightweight design via a single Mamba layer. Specifically, the Multi-scale Spiral Scanning Spatial Mamba (MS-SpaM) module is designed, which integrates lightweight multi-scale convolutions with center-first spiral scanning to effectively extract spatial features. The Adaptive Enhancement Fusion Module (AEFM) is used to achieve complementary enhancement and fusion of the two branches. The proposed method yields superior results on four HSI datasets: Indian Pines, Pavia University, Houston 2013 and Salinas. The overall accuracy reaches 97.19%, 98.56%, 96.71%, and 98.77% on these datasets, respectively, validating the effectiveness of the proposed approach.