Assimilation of FY-3G Precipitation Data Using a Machine Learning-Based Observation Operator in the CMA-MESO Regional Model

Abstract

Accurate assimilation of satellite-derived precipitation data remains a critical challenge in regional numerical weather prediction (NWP), particularly for convective-scale rainfall. Conventional observation operators rely on radiative transfer models or simplified moist physics, introducing substantial uncertainty at convective scales. This study develops a machine learning-based observation operator and implements a “one-dimensional variational (1D-Var) + three-dimensional variational (3D-Var)” framework to assimilate FY-3G/PMR precipitation data into the CMA-MESO. Results show the machine learning-based operator demonstrates high accuracy for light, moderate, and heavy rain (BIAS < 0.5 mm/h, RMSE < 2.5 mm/h) and exhibits good generalization capability across different weather systems. Through the two-step assimilation, the retrieved humidity profiles improve initial moisture fields. Both the single case study and the one-month continuous cycling experiment consistently show that the assimilation yields measurable improvements in short-term precipitation forecasts, particularly for extreme precipitation events, with a maximum TS improvement of 19.9% for severe torrential rain. This work demonstrates that the machine learning-based observation operator exhibits potential in precipitation data assimilation, offering a viable pathway to enhance NWP.