EnMambaHSI: Enhanced Multi-Scale Spatial-Spectral Mamba Network for Hyperspectral Image Classification

Hyperspectral image classification has made significant progress in recent years with the introduction of the Mamba architecture, which learns long-range dependencies while maintaining linear computational complexity. However, existing Mamba-based hyperspectral image (MambaHSI) classification methods face limitations in handling multi-scale spatial processing and preserving spectral continuity. To address these limitations, this paper proposes a novel MambaHSI classification method, with three enhancements: (1) an enhanced multi-scale spatial Mamba module, which captures multi-scale spatial features while maintaining linear complexity; (2) an enhanced spectral Mamba module, which preserves spectral continuity through cross-group interactions; and (3) an adaptive fusion module, which learns the optimal fusion weights for spatial-spectral features. Extensive experiments on four benchmark datasets demonstrate the effectiveness of the proposed method. Specifically, the method achieves overall accuracies of 97.54%±0.81%, 95.13%±0.83%, 94.37%±0.64%, 96.59%±0.25%, 97.70%±0.64%, and 99.58%±0.23% on the Pavia University, Houston, Hanchuan, Honghu, Longkou, and Botswana datasets, respectively, which represent improvements of 1.80%, 0.67%, 4.16%, 2.01%, 1.21%, and 0.24% over MambaHSI. Comprehensive ablation studies reveal that the multi-scale spatial Mamba module provides the most significant contribution, with improvements of 88.68%, 85.19%, 86.19%, and 92.21% over the baseline on the first four datasets, while the adaptive fusion module provides an additional performance boost of 1.70%∼2.59% across all datasets. The proposed model achieves excellent performance in diverse hyperspectral classification scenarios, while retaining the linear computational complexity advantage of the Mamba architecture.