Machine Learning-Based Bias Correction Scheme for FY-4A AGRI Radiance Assimilation and Its Impact on Typhoon Forecasting

Data assimilation generally assumes that observation-minus-background (OMB) departures follow an unbiased Gaussian distribution. Therefore, bias correction (BC) is essential for the effective assimilation of satellite radiances. This study applies four machine learning (ML) approaches to reduce systematic biases in clear-sky observations from the water-vapor-sensitive channels 9 and 10 of the Fengyun-4A (FY-4A) Advanced Geosynchronous Radiation Imager (AGRI). These methods include random forest (RF), extreme gradient boosting (XGBoost), multilayer perceptron (MLP), and the Kolmogorov–Arnold network (KAN). The machine-learning-based bias correction (MLBC) methods are evaluated for Typhoon In-Fa (2021) using the Weather Research and Forecasting Data Assimilation (WRFDA) system. Experiments without BC and with conventional variational bias correction (VarBC) are designed for comparison. Compared with VARBC, the four MLBC methods better capture nonlinear OMB biases and improve the statistical distributions of the departures. They produce more symmetric OMB distributions for channel 9 and distributions closer to the Gaussian distribution for channel 10, with smaller OMB STDVs and RMSEs than VARBC over both land and ocean for channel 10. After assimilation, KANBC produces the smallest OMA STDV and RMSE for both channels. For the Typhoon In-Fa case examined here, KANBC yields the smallest track errors among the experiments. In particular, KAN-based BC reduces the anomalously strong positive 500 hPa geopotential height increments over the continental high pressure and western North Pacific subtropical high pressure. The resulting initial conditions improve the environmental steering flow and reduce the northwestward track bias. Channel sensitivity experiments were further conducted for water-vapor-sensitive channels 9 and 10 of the FY-4A AGRI using the KAN-based BC method. The results show that single-channel assimilation performs better than dual-channel assimilation. Assimilating channel 10 alone produces the smallest track errors and lower analysis RMSEs relative to ERA5 than assimilating both channels.