Abstract

Precipitation is central to the global hydrological cycle, and its accurate monitoring is vital for preventing meteorological disasters. Traditional satellite retrieval methods fail to model nonlinear relationships and adapt to regional heterogeneity. Using China’s new-generation geostationary satellite FY-4B/AGRI, this study develops a two-step machine learning model—separating precipitation identification from intensity estimation—for the complex terrain of Anhui Province and further conducts experiments in the Huaibei Plain, Jianghuai Hills, and Jiangnan mountainous areas. This design separately addresses precipitation occurrence and rainfall intensity, which represent distinct classification and regression tasks. The model takes 37 features as input, including multispectral brightness temperatures, brightness temperature differences, spatiotemporal cloud-top temperature dynamics, secondary cloud parameters, and terrain. For identification, XGBoost at a 1:4 precipitation/non-precipitation ratio performed best, with POD, FAR, CSI, and ETS of 0.6961, 0.3676, 0.4956, and 0.4422, outperforming the FY-4B QPE product (0.5876, 0.5703, 0.3301, 0.2607). Subregional modeling further improved CSI to 0.4716, 0.5186, and 0.5210 for the three areas. For rainfall estimation, XGBoost trained with the original precipitation class ratio was optimal in all subregions, markedly surpassing the QPE product. Spatial aggregation of the three regional models yielded a correlation coefficient of 0.5304 and RMSE of 0.6274 mm, outperforming the unified model and QPE during the study period. This study provides a useful machine learning approach for precipitation retrieval, and the results demonstrate the efficacy of incorporating regional heterogeneity into machine-learning-based precipitation retrieval, leading to enhanced precipitation estimation during the June–July 2024 Meiyu period over Anhui Province.

IPC Classification

Keywords