MambaMSF: A Mamba-based Multi-scale Feature Fusion Method for Hyperspectral Image Classification

Hyperspectral image (HSI) classification faces challenges due to complex spectral-spatial characteristics and limited labeled samples. While Mamba-based models have shown significant progress in HSI classification by effectively handling long-sequence data and modeling spatial-spectral relationships, existing approaches primarily focus on single-scale feature extraction, neglecting multi-scale information. To address this, we propose a Mamba-based multi-scale feature fusion (MambaMSF) method. MambaMSF integrates multi-scale pixel patches to capture both local and global features, followed by an adaptive fusion mechanism that optimizes interactions between pixel-level and patch-level representations, enhancing classification accuracy. Experiments on three HSI benchmarks demonstrated that MambaMSF exceeds the comparison models.