Interval scanning Mamba for efficient hyperspectral image classification

Hyperspectral image (HSI) classification is important for remote sensing and fine-grained material identification, but it remains challenging due to high dimensionality, strong spectral redundancy, and the mismatch between spectral and spatial resolutions. Deep neural networks have improved spectral–spatial feature learning, yet many methods are still computationally heavy and hard to deploy on limited hardware. Herein, we propose PLLG-MAMBA, a lightweight Mamba/state-space model (SSM) framework for cross-dimensional spectral–spatial modeling. PLLG-MAMBA uses lightweight spatial and spectral SSM blocks to capture long-range dependencies efficiently. It also introduces a redundancy-aware scanning scheme that combines serpentine spatial interval scanning and bidirectional spectral interval scanning. This design keeps structural continuity and cross-band context while reducing redundant interactions. We further use an adaptive gated fusion module to balance and merge spatial and spectral cues across different scenes, instead of relying on a fixed fusion rule. In addition, we adopt an RFF-NLPCA-based dimensionality reduction stage to obtain a compact spectral representation and improve efficiency. Experiments on Indian Pines, Pavia University, and Salinas show competitive performance and consistent improvements in many cases under identical settings against representative CNN-, Transformer-, and Mamba-based baselines. PLLG-MAMBA also achieves strong accuracy with fewer parameters and lower computation, which makes it suitable for resource-constrained HSI deployment.