Self-attention augmented blind super-resolution with advanced Sobel loss for remote sensing images

Super-resolution (SR) is one of the core means to obtain high-spatial-resolution images for remote sensing Earth observation. Existing SR methods rely on paired multi-resolution images acquired under identical spatiotemporal conditions, and their reconstruction performance degrades significantly on cross-sensor remote sensing data, which severely limits the practical application of SR. This paper proposes EBSR, an enhanced blind super-resolution model based on the RealESRGAN framework. Integrating the self-attention mechanism and improved Sobel loss function, the model jointly improves the reconstruction of fine structural details. A multi-sensor remote sensing dataset GFtoBJ3A is further constructed for 4× super-resolution reconstruction, upgrading 2 m spatial-resolution imagery to 0.5 m. Compared with four mainstream blind super-resolution algorithms for remote sensing images, EBSR achieves an average relative improvement of 24.85% in the Perception Index (PI) over all baselines. When tested against two mainstream deep learning models, MHAN and SwinIR, EBSR still yields an average PI gain of 12.52%, demonstrating its outstanding reconstruction capability. Moreover, the model possesses favorable generalization performance and is applicable to various practical remotesensing tasks, helping reduce Earthobservation costs and extend effective observation coverage.