Fusion Retrieval of Tropospheric Three-Dimensional Wind Fields through Integration of Advanced Geostationary Satellite Imager Observations and Numerical Model Products

Accurate and high spatiotemporal resolution three dimensional (3D) wind field information in near real-time is needed for severe weather monitoring and forecast; such important 3D wind observations are often not available for applications. To address this shortage, this study develops a 3D wind field fusion retrieval algorithm based on Multi-Layer Perceptron (MLP) framework that integrates observations from FengYun-4B Advanced Geostationary Radiation Imager (AGRI) and the 9-12 h numerical forecast products from the China Meteorological Administration Global Forecast System (CMA-GFS). The current implementation is limited to summertime over China and its surrounding areas (5~55°N, 70~145°E). Separate approaches for different cloud and wind speed regimes are performed to reduce the impacts of cloud presence and sample imbalance. Independent evaluations against ERA5 reanalysis and wind profiler radar observations show that 3D wind fusion retrievals from combined AGRI and CMA-GFS significantly reduce the wind forecast errors of CMA-GFS short-range forecast wind product at all vertical levels, with an overall improvement of 4.66% - 5.74% in wind speed and 2.36% - 2.75% in wind direction. Distinct correction effects are further exhibited under high-cloud and strong-wind conditions. In the case of Super Typhoon Doksuri, the combined AGRI and CMA-GFS retrieval method greatly improves CMA-GFS wind field in the clear‑sky environmental flow surrounding the typhoon, especially in the lower troposphere. SHAP interpretability analysis reveals that water vapor-sensitive infrared band observations from AGRI effectively compensate for the deficiency of CMA-GFS forecasts in middle and upper tropospheric wind prediction, with an obvious nonlinear relationship, while the CMA-GFS relevant predictors play a leading role in the lower troposphere and present a more linear correlation with wind variations.