LRCS-Mamba: Low-Rank Mamba With Content-Aware Sparse Scanning for Hyperspectral Image Classification

Abstract

Recently, state space models (SSMs), particularly the Mamba architecture, have emerged as a powerful paradigm for hyperspectral image (HSI) classification due to their linear complexity and ability to capture long-range dependencies. However, existing Mamba-based approaches mainly rely on predefined geometric scanning patterns that ignore spatial heterogeneity, leading to spatial–semantic misalignment. Moreover, standard full-rank parameterization in SSMs fails to exploit the intrinsic low-rank structure of HSI. Without structural constraints, redundant state dynamics may propagate noisy responses caused by misaligned input sequences, further limiting the robustness of spectral–spatial representation learning. To address these coupled issues, we propose LRCS-Mamba, a dual-branch Mamba network with content-aware sparse scanning (CASS) and low-rank state-space modeling for HSI classification. Specifically, we design a CASS mechanism that constructs input sequences via center-neighbor spectral affinity and Top- $k$ selection, reducing heterogeneous boundary-pixel contamination before state evolution. To reduce noisy responses during state evolution, we further develop a low-rank SSM (LRSSM) that factorizes the core evolution matrices and constrains state dynamics within a compact latent subspace. Finally, a lightweight convolutional neural network (CNN) branch and an adaptive dual-branch sparse gating (ADSG) module are integrated to balance global semantics and local textures. Extensive experiments on four benchmark datasets demonstrate that LRCS-Mamba consistently outperforms state-of-the-art methods and achieves robust HSI classification performance.