GS-MambaNet: A Decoupled Global-Spatial and Grouped-Spectral Mamba Network for Hyperspectral Image Classification

Hyperspectral images (HSIs) inherently possess abundant spectral and spatial information, making them invaluable across diverse remote sensing applications. Recently, deep learning architecture has emerged as the dominant paradigm for HSI classification. However, existing methods face an inherent dilemma: Convolutional Neural Networks (CNNs) excel at extracting local spatial representations but fundamentally struggle to capture long-range dependencies. Conversely, Vision Transformers (ViTs) effectively capture global contexts but incur a prohibitive quadratic computational burden when processing long, continuous spectral sequences. To resolve this dual challenge, we propose GS-MambaNet, a novel decoupled architecture that synergizes Global-Spatial Attention with Grouped-Spectral State Space Modeling. Specifically, a Global-Spatial Attention Module (GSAM) is designed to capture macro-level spatial topologies, while a Grouped-Spectral State-Space Module (GSSM) adaptively partitions spectral channels into sequence groups, leveraging the linear complexity of the Mamba mechanism to efficiently model long-term spectral evolution dynamics. Extensive experiments on four benchmark datasets (Botswana, HongHu, HanChuan, and Pavia University) demonstrate that GS-MambaNet significantly outperforms state-of-the-art methods, achieving superior classification accuracy while maintaining computational efficiency.