Multiscale Sequence Aware Mamba for Hyperspectral Image Classification

Benefiting from the state-space model, Mamba-based networks can effectively model long-range sequential dependencies while maintaining linear complexity. However, for hyperspectral images (HSIs) with rich spectral and spatial information, Mamba-based networks struggle to capture multiscale spectral–spatial information and correlations between adjacent bands. Therefore, an end-to-end Mamba-based network is proposed, named multiscale sequence aware Mamba (MSAM), in which targeted structural designs are incorporated to address the aforementioned limitations of Mamba in HSIs classification, enabling flexible adaptation to diverse spectral–spatial scales of land cover structures. Particularly, to compensate for Mamba’s limitations in exploring multiscale and local features, dilated convolutions are introduced and integrated with Mamba to form a multiscale sequence aware module. In addition, considering the rich spectral information of HSIs, a spectral interaction embedding module is designed to learn interpixel features and capture correlations of adjacent bands. Finally, a hierarchical feature pyramid is constructed by sequentially stacking these modules, facilitating the progressive extraction of multilevel features. Comprehensive experiments carried out on four benchmark datasets validate that the proposed MSAM exhibits obvious superiority over state-of-the-art networks in HSI classification.