A Dual-Stream Transformer-Augmented UNet framework for Accurate Satellite Quantitative Precipitation Estimation

Meteorological satellite remote sensing offers a crucial solution to the issue of insufficient ground-based observation coverage in regions with complex terrains. However,being an indirect method for measuring meteorological elements,current satellite-based precipitation estimation still suffers from limited accuracy. This paper proposes a novel Dual-Stream Transformer-Augmented UNet (DSTA-UNet) framework for satellite quantitative precipitation estimation (QPE). It generates high-accuracy precipitation analysis fields in near real-time by using the satellite cloud imagery from China's new-generation geostationary satellite,Fengyun-4A (FY-4A). Extensive comparisons against operational products (FY-4A QPE,GPM IMERG,and CFSv2) and deep learning-based precipitation estimation methods (Attention-Unet and UCTransNet) demonstrate that DSTA-UNet achieves competitive performance. When evaluated against ERA5 as the benchmark,DSTA-UNet reduces the root mean square error (RMSE) by 52.38% for FY-4A QPE,52.27% for GPM IMERG,27.02% for CFSv2,33.65% for Attention-Unet,and 0.71% for UCTransNet. Using station observations as the benchmark,DSTA-UNet also achieves significant RMSE reductions: 25.80% for FY-4A QPE,14.23% for GPM IMERG,8.51% for CFSv2,and 3.72% for ERA5.