RSSMamba: A Region Spectral-Spatial Mamba Network with Bidirectional Spiral Scanning and Cross-Attention for Hyperspectral Image Classification

Hyperspectral image (HSI) classification aims to assign a land cover label to each pixel by jointly exploiting spectral signatures and spatial contextual information. Recently, Mamba-based state-space models have shown potential for long-range dependency modeling with improved efficiency compared with Transformer-based architectures. However, directly applying standard Mamba structures to patch-based HSI classification still faces three major limitations: Noisy pixel-level tokens may disturb state-space propagation, conventional scanning paths are not well aligned with central-pixel prediction, and fine-grained spectral signatures may be weakened in deep feature representations. To address these issues, this paper proposes a Region Spectral–Spatial Mamba network with Bidirectional Spiral Scanning and Cross-Attention, termed RSSMamba. Specifically, a Dynamic Spatial–Spectral Saliency Tokenization (DSST) module is first designed to suppress redundant spatial–spectral responses and generate semantically consistent regional tokens. Then, a Bidirectional Spiral Selective Scanning (B3S) mechanism serializes the feature map along clockwise and counter-clockwise spiral paths, enabling contextual information to be aggregated toward the central target pixel. Finally, a Spectral Cross-Attention (SCA) module injects the shallow central spectral feature into the deep center-oriented descriptor through a lightweight, head-partitioned cross-layer fusion, thereby alleviating spectral feature attenuation. Under the adopted random pixel-wise evaluation protocol, experiments on four benchmark HSI datasets, including Indian Pines, WHU-HI-Longkou, Pavia University, and WHU-HI-HanChuan, yield competitive aggregate results relative to the included CNN-, Transformer-, and Mamba-based baselines. In the ten-run reporting format, RSSMamba obtains OA values of 95.37±0.42%, 98.47±0.21%, 97.72±0.31%, and 97.64±0.27% on the four datasets, respectively. It reports 0.33M parameters and 8.84G FLOPs on Indian Pines.