A Machine Learning‐Based Observation Operator for FY‐4B GIIRS Brightness Temperatures Considering the Uncertainty of Label Data

Abstract The increasing volume of satellite data, particularly hyperspectral infrared data, combined with the real‐time monitoring requirements of numerical weather prediction (NWP) systems, has heightened the demand for computational efficiency and accuracy in radiative transfer models (RTM). Machine learning (ML) offers a promising approach, and numerous studies on ML‐based RTM have emerged recently. However, existing ML‐based RTMs for hyperspectral infrared were not end‐to‐end. Moreover, since the label data do not represent truth, models trained with loss functions like mean squared error (MSE) or mean absolute error (MAE) fail to account for its uncertainty. This limitation can lead to suboptimal model parameters, as training may assign higher weights to labels with larger errors. This study construct an end‐to‐end ML‐based RTM focused on clear sky conditions over the ocean for the FengYun‐4B satellite (FY‐4B) Geostationary Interferometric Infrared Sounder (GIIRS), using maximum likelihood estimation (MLE) and MSE for training, respectively. MLE accounts for the uncertainty of labels. The results indicate both models achieve high accuracy, with mean errors within 0.1 K (K) and standard deviation (STD) of errors within 0.04 K compared to the labels. The model trained with MLE exhibits a mean error closer to 0 and a STD similar to the error STD of labels, suggesting better parameter configurations to reflect the actual error distribution of the labels. Additionally, the temperature and water vapor Jacobian computed by both models are comparable to those obtained from RTTOV, highlighting their potential for application as observational operator in satellite data assimilation for hyperspectral infrared sounder.