Abstract
Precipitation downscaling aims to enhance coarse-resolution data to higher resolutions. Due to the similarity between downscaling and super-resolution (SR), deep learning-based SR approaches have been increasingly adopted in this domain. However, single-image super-resolution (SISR) methods applied to precipitation data face two main challenges: weak high-frequency signals and highly skewed distributions in precipitation datasets, which often lead to overly smooth reconstructions, failure to capture precipitation extremes, and loss of fine-scale variability with predictions biased toward mean values. To address these issues, we propose EdgeGeoDiff, a two-stage diffusion model for precipitation downscaling that leverages both edge information and geographical priors (e.g., terrain-related factors such as elevation). In the first stage, a residual network reconstructs an initial high-resolution precipitation field with preliminary structural details. In the second stage, edge features extracted using the Laplacian operator, together with geographical priors, guide a diffusion model to generate residuals that enhance fine-scale precipitation structures. Experimental results on real-world precipitation datasets show that EdgeGeoDiff effectively reconstructs fine-scale details while preserving large-scale patterns and outperforms conventional SISR methods in terms of its RMSE, PSNR, SSIM, and CSI, particularly demonstrating superior performance in the high-frequency region of the spectrum.