HL-Mamba: A High–Low Frequency Interaction Mamba Network for Hyperspectral Image Classification

Deep-learning-based methods have achieved remarkable success in hyperspectral image (HSI) classification tasks due to their promising ability. However, the high dimensionality and spectral-spatial correlations of HSIs usually lead to information redundancy and feature entanglement, limiting the classification performance. To address these issues, we propose a novel high-low frequency interaction Mamba network, called HL-Mamba, which achieves effective decoupling and interaction between global structures and edge details of HSIs in the frequency domain, thereby improving spectral-spatial representation for HSI classification. Specifically, a high-low frequency decomposition Mamba module is designed to decompose the HSI into low-frequency structural and high-frequency edge detail components, which allows the model to learn global structures and fine-grained details, enhancing classification performance. By employing two parallel Mamba branches to model long-range dependencies across different frequency components, the network achieves efficient global modeling while mitigating information redundancy. Furthermore, a cross-frequency interaction module is designed to establish complementary information flow between high- and low-frequency features through a dynamic attention mechanism. In this way, low-frequency structural features guide the aggregation of high-frequency details, whereas high-frequency textures refine global structural representations, yielding more discriminative spectral-spatial features for HSI classification. In addition, a frequency alignment loss is designed to enhance the consistency and complementarity between high- and low-frequency features, further improving classification performance. Extensive experiments on four public benchmark datasets (i.e., Indian Pines, Pavia University, WHU-Hi-HanChuan, and Houston datasets) demonstrate that the proposed HL-Mamba significantly outperforms eight comparison methods, achieving an overall accuracy of 94.07%, 93.82%, 95.28%, and 87.32%, respectively. Ablation studies further verify the effectiveness of core component within the network.