FAHM: Frequency-Aware Hierarchical Mamba for Hyperspectral Image Classification

Convolutional neural networks and Transformers have garnered substantial attention in hyperspectral image (HSI) classification, and recently Mamba has made significant progress in modeling long sequences. However, existing Mamba-based approaches flatten 2-D images into 1-D sequences, inevitably disrupting 2-D local dependencies, thereby disregarding the distinctive difference between high-frequency and low-frequency components in the frequency domain. For HSI tasks, these methods suffer from inherent receptive field limitations and insufficient exploitation of spectral information, resulting in unclear classification of regional boundaries and poor generalizability. To address these issues, we present a novel frequency-aware hierarchical Mamba (FAHM) for HSI classification. Specifically, FAHM is comprised of three main parts: a spatial-spectral interaction block (SSIB), a frequency group embedding module (FGEM), and an adaptive bidirectional Mamba block (ABMB). Building on the observation that the high-frequency component contains detailed features while the low-frequency information provides abundant high-level semantics, we design FGEM to enhance feature representation in the frequency domain. To prepare spectral–spatial contextual tokens for FGEM, we incorporate SSIB to extract shallow features, and ABMB is designed to capture spectral dependencies by modeling spectral variation bidirectionally. We construct our Mamba in a hierarchical manner, and refine multilevel features using a feature affinity module for improved accuracy. Experiments conducted on four HSI benchmark datasets show that the proposed FAHM outperforms eleven existing approaches in classification accuracy and model generalization.