AERU-Net: Adaptive Edge Recovery and Attention U-Shaped Network for Remote Sensing Image Super-Resolution

Remote sensing image super-resolution (RSISR) has recently garnered significant attention, with major advances achieved through deep learning neural networks. Remote sensing images (RSIs) contain rich, recurring textural and structural patterns that vary substantially across different classes of RSIs. While state-of-the-art RSISR methods have begun incorporating attention mechanisms to explore and recover multi-scale global features, they face challenges in identifying the most relevant details and recovering complex edge patterns. In this research, we propose a novel Adaptive Edge Recovery and Attention U-shaped Network (AERU-Net), which integrates an Edge Recovery Block (ERB), Channel Attention Module (CAM) and Spatial Attention Module (SAM) within a three-level encoder-decoder architecture. Furthermore, we introduce a novel adaptive Cross-scale Interaction (CSI) module along with a Laplacian Edge Preserver (LEP) block to connect encoders and decoders across different network scales (levels). Our design facilitates the gradual acquisition of contextual information while enhancing the retention of high-frequency edge details. Performance analysis conducted on four RSI datasets (UCMerced, RSCNN7, AID, and WHU-RS19) in three up-sampling scales ( $\times 2$ , $\times 3$ , and $\times 4$ ) demonstrates that AERU-Net achieves significant improvements over state-of-the-art models in PSNR, SSIM, SCC, and SAM. Our model strikes an effective balance between accuracy and network complexity. The code will be released on GitHub and can be accessed at https://github.com/CUEE-MDAP/aeru-net.