A Machine Learning-Based Bias Correction Scheme for the All-Sky Assimilation of AGRI Infrared Radiances in a Regional OSSE Framework

Most bias correction (BC) schemes based on a linear fitting function have undesirable effects on the all-sky assimilation of satellite radiances from infrared bands. This study introduces a newly nonlinear BC method for the all-sky assimilation of Fengyun-4A (FY-4A) Advanced Geosynchronous Radiation Imager (AGRI) infrared radiances. The proposed BC method uses a machine learning technology of random forest (RF) to emulate the fitting relationship between the observation-minus-background (OMB) departures and BC predictors. The effectiveness of this BC algorithm is verified in an idealized case, where the sources of the systematic bias and the real states of the atmosphere are assumed to be known. The OMB departures here were artificially produced including the predictor-dependent systematic biases and the Gauss errors. Meanwhile, the so-called "truth" was simulated from natural run forecasts in a regional observing system simulation experiment (OSSE) framework. As expected, it is demonstrated that the RF BC method has the ability to remove linear and lower degree nonlinear biases of all-sky AGRI infrared observations whether caused by a single source or multiple sources. Another advantage of the RF BC method is that meteorological signals are potentially reserved after BC when the predictors have been properly selected according to feature importance scores in the RF model. Henceforth, assimilating the bias-corrected AGRI observations is conducive to decreasing the erroneous increments, followed by more accurate analyses of water vapor and cloud ice in the middle and upper troposphere.