HRWKV-Net: A Hybrid Receptance Weighted Key–Value Network for Remote Sensing Image Super-Resolution

Abstract

Remote sensing image super-resolution (RSISR) remains a challenging task due to the complex spatial structures and diverse texture patterns inherent in aerial and satellite imagery. Existing deep learning-based SR methods often struggle with inadequate joint modeling of local and global information across hierarchical feature representations, restricting the recovery of fine structural details. Compounding this, insufficient disentanglement and propagation of low-frequency structural features and high-frequency textural details across network layers further lead to suboptimal reconstruction fidelity. To address these challenges, we propose the Hybrid Receptance Weighted Key–Value–based Super-Resolution Network (HRWKV-Net), a U-Net architecture that integrates Bidirectional RWKV with a window-based attention mechanism to achieve balanced recovery of global spatial context and local fine-grained details. The Bidirectional RWKV component effectively expands the effective receptive field by capturing long-range spatial dependencies, while the window-based attention mechanism complements it by modeling local structural patterns, together enabling more faithful reconstruction of complex geographical textures. Furthermore, we introduce an Adaptive Cross-Scale Frequency Decomposition (ACFD) module that bridges encoder and decoder features at multiple scales through progressive disentanglement of high-frequency textural and edge information from low-frequency structural representations, facilitating effective cross-scale feature fusion between corresponding encoder and decoder stages while reducing unnecessary architectural complexity and improving information flow. Extensive experiments conducted on four widely used remote sensing benchmark datasets, namely UCMerced, AID, RSSCN7, and WHU-RS19, at upscaling factors of $\times 2$ , $\times 3$ , and $\times 4$ demonstrate that HRWKV-Net consistently outperforms state-of-the-art SR methods in terms of both quantitative metrics and visual quality, confirming its effectiveness for remote sensing image super-resolution. The code is available on GitHub at https://github.com/cuee-mdap/HRWKV-Net