PRGD: Proximal Residual-Guided Diffusion with pattern-aware experts for remote sensing image super-resolution

High-resolution remote sensing imagery is essential for accurate Earth observation and analysis. While diffusion models show promise for image super-resolution (SR), their application to remote sensing is often hindered by fidelity loss from error accumulation and by inadequate guidance from simple low-resolution (LR) conditioning. To address these dual challenges, we introduce the Proximal Residual-Guided Diffusion (PRGD) model, a novel framework for high-fidelity remote sensing SR. PRGD is built upon a novel holistic paradigm designed to synergistically mitigate error accumulation throughout the diffusion process. This paradigm combines a Residual Shifting strategy, which simplifies the generation path by anchoring it to the LR image, with a measurement-consistent Proximal Sampling technique that enforces strict data fidelity at each iterative step. To provide superior guidance, PRGD introduces an innovative Mixture-of-Experts (MoE) feature extraction module, powered by our Pattern-Aware Deformable Convolution (PADC). This novel convolution dynamically groups feature channels based on their activation patterns to extract rich, adaptive, multi-scale priors. Extensive experiments on benchmark datasets demonstrate that PRGD achieves a new state-of-the-art, establishing superior performance across both fidelity and perceptual quality. For instance, on the Toronto (Formula presented) dataset, PRGD not only surpasses the leading regression-based method in PSNR but also reduces the Fréchet Inception Distance (FID) by 11.96% compared to the next-best diffusion model. The resulting images exhibit superior fidelity to the ground truth, enhanced visual realism, and remarkable detail preservation, thereby enhancing the potential of diffusion models to serve as a reliable data source for downstream remote sensing applications.