Direct assimilation of ground-based microwave radiometer observations with machine learning bias correction based on developments of RTTOV-gb v1.0 and WRFDA v4.5

The application of ground-based microwave radiometers (GMWRs), which provide high-quality and continuous vertical atmospheric observations, has traditionally focused on the indirect assimilation of retrieved profiles. This study advanced this application by developing a direct assimilation capability for GMWR radiance observations within the Weather Research and Forecasting Data Assimilation (WRFDA) system, along with a bias correction scheme based on the random forest technique. The proposed bias correction scheme effectively reduced the observation-minus-background ( O − B ) biases and standard deviations by 0.83 K (97.1 %) and 1.63 K (64.6 %), respectively. A series of 10 d experiments demonstrated that assimilating GMWR radiances improves both the initial conditions and the forecasts, with additional benefits from higher assimilation frequencies. In the initial conditions, hourly assimilation significantly enhanced low-level temperature and humidity fields, reducing the root-mean-square error (RMSE) for temperature by 6.32 % below 1 km and for water vapor mixing ratio by 1.98 % below 5 km. These improvements extended to forecasts, where 2 m temperature and humidity showed sustained benefits for over 12 h, and precipitation forecasts exhibited improvements to a certain extent. The time-averaged Fractions Skill Score (FSS) for 3 h accumulated precipitation within the 24 h forecasts increased by 0.02–0.04 (3.9 %–10.2 %) for thresholds of 3–6 mm.