Language-Inspired Hierarchical Mamba Network With Superpixel-Structured Training for Hyperspectral Image Classification

Existing Mamba-based hyperspectral image classification (HSIC) methods typically decouple spectral and spatial modeling into two independent and parallel branches and fuse their features afterwards. Such physical separation risks information loss or feature conflicts when the fusion strategy is sub-optimal, ultimately degrading the classification accuracy. To address this issue, we propose a Language-Inspired Hierarchical Mamba network with Superpixel-structured Training (dubbed ‘L-HMambaS’) that unifies the spectral and spatial dimensions from the outset. Specifically, inspired by the structure of natural language, we regard the spectrum of a pixel as a “word” and the spatial arrangement of multiple pixels as a “sentence”, and devise a nested Mamba architecture. In L-HMambaS model, anInner Spectral Mambablock captures cross-band dependencies among pixel, while anOuter Spatial Mambablock explores local contextual relations among spatial patch. Furthermore, to suppress redundancy induced by overlapping patches in conventional sampling strategies, we introduce a superpixel-driven sampling strategy that guarantees high intra-patch homogeneity and minimal spatial overlap. Extensive experiments on five publicly available hyperspectral datasets, namely Pavia University, WHU-Hi-LongKou, WHU-Hi-HanChuan, WHU-Hi-HongHu, and Houston 2013, demonstrate that L-HMambaS achieves overall accuracies of 97.01%, 98.17%, 92.91%, 93.37%, and 96.79% with only 30 labelled samples per class, consistently surpassing state-of-the-art competitors. The codes will be available from the website: https://github.com/rachelitem/L-HMambaS.