HIA: Hybrid Interactive Attention for Efficient Remote Sensing Image Super-Resolution

Despite the strong capability of Transformer-based methods in modeling long-range dependencies for Remote Sensing Image Super-Resolution (RSISR), they often suffer from large model sizes and high computational complexity. Furthermore, most existing approaches tend to focus exclusively on either spatial modeling or channel modeling. This overemphasis on local details or global semantics limits the expressive power of learned features, thereby hindering further improvements in reconstruction performance. To address these challenges, we propose an innovative Hybrid Interactive Attention Module (HIAM), which adopts a tri-branch architecture comprising a Channel Attention Module (CAM), a Spatial Attention Module (SAM), and a Permuted Attention Module (PAM). Specifically, CAM captures cross-channel semantic dependencies, SAM extracts local spatial details, and PAM focuses on modeling long-range spatial interactions. Through efficient integration and cross-module interaction among CAM, SAM, and PAM, our method achieves complementary and coordinated enhancement across both feature dimensions (spatial and channel) and spatial scales (local and global). Moreover, to mitigate the computational burden of large-window attention. PAM introduces a dimension permutation mechanism that preserves the benefits of a large receptive field and global awareness while significantly reducing the model’s complexity and parameter count. Our model strikes a favorable balance between lightweight design and reconstruction quality, markedly enhancing the detail fidelity and structural consistency of the reconstructed images. Extensive experiments on multiple benchmark remote sensing datasets demonstrate that our method achieves higher PSNR and SSIM scores, validating its effectiveness and superiority in RSISR tasks.