Superpixel-Integrated Dual-Stage Mamba for Hyperspectral Image Classification

Hyperspectral image (HSI) classification plays a key role in remote sensing applications such as land cover mapping and environmental monitoring. Mamba is a state-space-based model that has recently gained attention for its efficiency and strong capability in capturing long-range dependencies in sequence modeling. However, existing approaches often ignore spatial structure and Mamba’s position sensitivity, leading to spectral–spatial discontinuity and information loss. To address these limitations, we propose Superpixel-integrated dual-stage Mamba (SIDS-Mamba), a novel framework that integrates superpixel-guided spatial priors with a dual-stage sorting (DSS) strategy, thereby preserving spatial–semantic continuity and aligning discriminative pixels with Mamba’s memory dynamics (Mamba’s memory dynamics refers to the phenomenon where later tokens dominate the state updates, causing information from earlier tokens to decay over time and be more easily forgotten). Unlike prior methods that heavily rely on handcrafted preprocessing or purely spectral modeling, SIDS-Mamba explicitly bridges spatial structure and sequential modeling within a unified design. Extensive experiments demonstrate that SIDS-Mamba consistently outperforms representative state-of-the-art baselines, achieving high accuracy and robust performance across diverse scenarios, even under limited supervision. These results highlight the framework’s effectiveness, scalability, and potential for practical HSI analysis.