HSMMamba: Hierarchical Superpixel MoE With Mamba for Hyperspectral Image Classification

Le résumé fourni par la source

Hyperspectral image classification aims to fully exploit spatial and spectral information for fine-grained land-cover recognition in complex scenes. However, existing methods still face two major challenges. On the one hand, traditional convolutional networks are ineffective at modeling long-range spatial–spectral dependencies, whereas self-attention-based models usually suffer from high computational complexity. On the other hand, although hierarchical superpixels provide effective multiscale structural priors, most existing methods rely on fixed cross-level mapping relationships and shared feature transformation strategies, making them inadequate for handling the pronounced heterogeneity among features at different scales. To address these issues, this article proposes a Hierarchical Superpixel Mixture-of-Experts Mamba network for hyperspectral image classification, termed HSMMamba. The proposed framework unifies hierarchical structural priors, adaptive mixture-of-experts enhancement, and global–local joint state-space modeling within a single architecture. Specifically, a learnable hierarchical superpixel mapping module is first developed to adaptively optimize cross-level downsampling and upsampling relationships by introducing learnable residuals on top of fixed hierarchical mapping priors, thereby enhancing the consistency of multilevel feature propagation and reconstruction. Then, a hierarchical superpixel mixture-of-experts module is introduced to perform differentiated modeling of pixel-level and multiscale superpixel-level features, and to adaptively fuse them through a gating mechanism, thus improving the representation capability of heterogeneous multilevel features. Finally, a forward–backward–differential Mamba module is constructed to preserve the strength of long-range dependency modeling while enhancing the characterization of local boundaries, fine-grained variations, and class transition regions. Extensive experiments demonstrate that HSMMamba achieves overall accuracies (OAs) of 98.98%, 99.30%, and 95.65% on the three public HSI datasets of Pavia University, Salinas, and HongHu, respectively, verifying its effectiveness and superiority in complex scenarios.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.