HAM: Hierarchical Attention Mamba With Spatial–Frequency Fusion for Remote Sensing Image Super-Resolution

Vision Mamba shows significant potential for enhancing remote sensing image super-resolution (RSISR) performance due to its linear-complexity global modeling capabilities. However, existing Mamba-based RSISR models face two key limitations in handling remote sensing imagery. First, insufficient hierarchical feature interaction hinders the capture of multi-scale spatial structures, which are crucial for reconstructing diverse patterns and fine textures. Second, inadequate local-global correlation extraction weakens the restoration of high-frequency details in conjunction with broader contextual information, ultimately limiting spatial consistency and image fidelity. To address these issues, we propose the Hierarchical Attention Mamba (HAM) network with spatial-frequency feature enhancement for RSISR. Key innovations include the Hierarchical Aggregation Attention (HAA) module, which enhances the propagation and utilization of multi-level features, and the Spatial-Frequency Information Interaction Module (SFIIM), which facilitates the fusion of spatial-frequency features. The SFIIM employs the Dynamic Frequency Spectrum Enhancement (DFSE) technique for discriminative feature learning through Fourier amplitude-phase decoupling, along with the Hybrid Spatial Feature Aggregator (HSFA), which utilizes a spatial-domain dual-branch channel-wise gating mechanism. Together, these techniques significantly enhance the model’s ability for synchronous local-global feature extraction. Extensive evaluations across six RSISR benchmarks (including AID, DIOR, DOTA, etc.) demonstrate that HAM achieves state-of-the-art performance, delivering superior image reconstruction quality while maintaining computational efficiency.