Fengyun Satellite-Based Bias Correction of Model Precipitation Forecasts

In recent years, demand for high-precision weather forecasts across diverse sectors of social production and daily life has grown steadily. Extreme weather events have become increasingly frequent. Consequently, the limitations of traditional numerical prediction methods are more apparent than ever, posing significant challenges. To address these challenges, this study proposes a bias correction approach based on multi-source data fusion and deep learning. We integrate Fengyun-4B satellite observations, ECMWF model forecasts, and ERA5 reanalysis data into a dual-stream, pre-trained deep neural network, introducing a spatiotemporal fusion strategy to refine forecast errors. The core innovation lies in effectively fusing satellite and numerical prediction data. This fusion significantly enhances precipitation forecast accuracy and spatial continuity, particularly in capturing extreme precipitation events. Experimental results demonstrate that the proposed method outperforms traditional numerical models across multiple evaluation metrics and offers a new direction for meteorological data correction technologies.