Impacts of Offline Nonlinear Bias Correction Schemes Using the Machine Learning Technology on the All-Sky Assimilation of Cloud-Affected Infrared Radiances

Abstract

Bias correction (BC) of the cloud-affected infrared (IR) radiances is one of the most difficult challenges in the all-sky data assimilation. This study introduces an offline nonlinear bias correction model based on the machine learning (ML) technology of Random Forest to enhance the impacts of Fengyun-4A Advanced Geostationary Radiation Imager (AGRI) all-sky radiance data assimilation. The effects of the developed model were comprehensively evaluated through sensitivity experiments based on the NoBC, BC and modified BC schemes for two super typhoon cases. Among them, the modified BC scheme is designed to extract the features of cloud-affected systematic biases, which are more prevalent in the all-sky IR radiance assimilation. Results showed that the modified BC scheme outperforms other schemes in terms of removing the cloud-impacted systematic bias while retaining the useful meteorological signal. Whereas, those biases were improperly corrected by the original BC scheme when the inputs of a grid point were handled by the ML model site by site without the feature extraction, leading to a non-Gaussian error distribution. Assimilating those better-corrected IR radiances in the modified BC experiments would lead to a greater improvement in the analysis of the humidity and cloud ice. Based on the improved initial condition, the positive effects of the modified BC scheme are also evident in the forecasts of atmospheric variables and typhoon systems.