%0 Computer Program %A Abramowicz, Alice %A Whan, Kirien %A Garcia-Marti, Irene %A Monteiro, Isabel %D 2025 %T Code underlying: Bias prediction of microwave radiance observations using machine learning methods %U %R 10.4121/5d6619ee-0668-4f12-acb3-26bd8fa0c65f.v1 %K Data science %K Bias correction %K satellite %K observations %K HARMONIE-AROME %K Variational Bias Correction %X

HARMONIE-AROME is a regional Numerical Weather Prediction (NWP) model used for short- range weather forecasting in several operational centres across Europe. Microwave radiance observations from satellites are one of the main contributors to its forecast skill, but these observations contain biases that need to be corrected. Currently, in HARMONIE-AROME, bias correction is performed using a Variational Bias Correction method (VarBC), which consists of linear models including different bias predictors alongside their bias coefficients. However, this method is computationally intensive. This study investigates the potential of machine learning models to emulate the VarBC method by predicting the bias coefficients of its linear equations, offering a more computationally efficient alternative. This research uses data corresponding to the microwave radiance observations from the instruments AMSU-A, MHS and MWHS2. The machine learning models were trained on data from the DINI domain and subsequently tested on a new domain. The findings reveal that the machine learning models successfully emulate VarBC on a new, unseen domain and that they can be trained within seconds/minutes. Moreover, the predictions generated by these models are immediately usable without requiring a spin-up phase. The study further reveals that including bias predictors additional to those preselected by HARMONIE-AROME does not enhance the prediction accuracy. Furthermore, the research suggests that data from some instruments are helpful in predicting bias coefficients from other instruments. 

%I 4TU.ResearchData