FS-CrossMamba: Dual-Branch Frequency-Guided Spatial and Spectral-Attention-Driven Spectral Cross-Mamba Network for Hyperspectral Image Classification

Hyperspectral Image Classification (HSIC) is intrinsically constrained by insignificant inter-class discrepancies and high spectral redundancy. Although Mamba-based methods have appeared recently, current techniques still struggle to properly address the minor differences between classes and the substantial spectral redundancy present in HSIC. To overcome these issues, we introduce the Dual-Branch Frequency-Guided Spatial and Spectral-Attention Driven Spectral Cross-Mamba Network (FS-CrossMamba). Unlike conventional methods that estimate state transition parameters solely from the input sequence, our method explicitly injects auxiliary priors into the SSM state transition parameters through two specialized branches: the Frequency-domain-guided Spatial Cross-Mamba (FSCM) branch, exploiting amplitude and phase priors to guide spatial selective scanning and preserve structural coherence, and the Spectral-attention-driven Spectral Cross-Mamba (SASCM) branch, utilizing global spectral attention to reweight spectral responses and strengthen long-range dependency modeling. Furthermore, we design a Triple-branch Convolutional Adaptive Spectral–Spatial Coupling Module (TCASSCM) to model rich multi-scale spectral–spatial relationships as the initial embedding. Comprehensive experiments on five benchmark datasets demonstrate that FS-CrossMamba reliably achieves state-of-the-art results in comparison with topperforming CNN-, Transformer-, and Mamba-based approaches.