UPD-Diff: a unified precipitation downscaling method based on multi-stream elucidating diffusion model

Abstract Precipitation downscaling plays a pivotal role in improving the resolution of coarse-grained weather and climate datasets, offering significant benefits for the issuance of localized precipitation warnings, which are indispensable for agricultural operations and economic planning in vulnerable regions. However, current downscaling approaches face several critical challenges: many deep learning methods, often adapted from computer vision, struggle to accurately model the long-tailed distribution of precipitation, leading to physically inconsistent outputs; reliance on regional a priori information severely limits model transferability, while inadequate merging of multivariate data yields ambiguous results; and prohibitive training costs and computational loads hinder real-time, large-scale deployment. To address these limitations, we propose the UPD-Diff, a unified multi-stream elucidating diffusion model (EDM) framework designed for robust and efficient precipitation downscaling. Firstly, to promote physical consistency by meteorological variable conditioning and to more accurately model the long-tailed distribution of precipitation, the core EDM architecture of UPD-Diff employs refined noise scheduling and learns precipitation bias based on various physical auxiliary variables, thereby enhancing its alignment with meteorological dynamics. Secondly, to enhance model transferability and facilitate the effective fusion of multi-modal data inputs, each stream integrates a newly designed mixed-attention UNet (MA-UNet). This MA-UNet synergistically blends channel attention and local-importance attention, effectively capturing shared atmospheric features while preserving fine-grained local details across different geographical areas. Finally, to address the challenges associated with data and computational requirements, our method leverages its inherent architectural efficiency and bias-learning strategy to achieve superior structural similarity index (SSIM) and Corr scores compared to baseline models, even when trained on a relatively small dataset, thereby substantially reducing computational overhead. Trained in this globally sampled dataset, our method outperforms the leading state-of-the-art baselines, showcasing competitive results with a peak signal-to-noise ratio of 29.02, a SSIM of 0.7942, a critical success index of 0.867 (0.1 mm threshold) and improved event skill at higher thresholds (50/100 mm). Notably, our model supports both training and inference on a single GPU like RTX 4090, with local-region inference at about 5.1 s per 224×224 tile (100 steps) on the same GPU.