Haoran Li, Shixue Li: Coarse-to-Fine: Multi-source Fusion for Bias-Corrected Precipitation Downscaling
Precipitation is a key component of the global water cycle, and high spatial and temporal resolution precipitation datasets are critical for understanding extreme weather events and supporting applications such as flood risk assessment, agriculture, and urban planning. However, existing precipitation products often suffer from sparse spatial coverage (e.g., rain gauges and radar observations) or low confidence and coarse resolution (e.g., satellite and reanalysis datasets). In addition, most publicly available precipitation datasets are limited to daily temporal resolution, which is insufficient for capturing short-duration extreme rainfall events and flash floods. In this project, we propose MS3-Diffusion, a Multi-Scale, Multi-Source, and Multi-Stage Diffusion model for precipitation downscaling and refinement. The framework integrates satellite precipitation products (IMERG), the convection-permitting regional reanalysis dataset BARRA-C2, weather radar measurements, and rain gauge observations. Satellite and reanalysis data provide continuous spatial coverage but contain higher uncertainty, while radar and rain gauges provide more accurate yet spatially sparse observations. By combining these complementary data sources, the proposed framework aims to generate precipitation estimates with improved consistency and finer structural detail. MS3-Diffusion adopts a two-stage training strategy. In Stage 1, a diffusion model performs super-resolution by downscaling multi-scale IMERG and BARRA-C2 data to learn high-resolution precipitation patterns. In Stage 2, radar and station observations are incorporated to constrain precipitation values and improve local accuracy. The goal of this project is to construct the first sub-hourly precipitation dataset for Australia with finer spatial resolution (~0.05°), enabling improved analysis of extreme precipitation and supporting downstream climate and environmental applications.