Advancing Passive Microwave Retrievals of Precipitation Using CloudSat and GPM Coincidences: Integration of Machine Learning with a Bayesian Algorithm

Integration of machine learning with a classic Bayesian algorithm is investigated for passive microwave precipitation retrievals using coincidences from the Global Precipitation Measurement core satellite and the CloudSat cloud profiling radar (CPR). Among several machine learning models, the extreme gradient-boosting decision tree (XGBDT), equipped with a weighted cross-entropy loss function, exhibits the highest accuracy in the detection of precipitation occurrence and phase with a true positive rate greater than 94 (98)% and a false positive rate smaller than 1 (1)% for rainfall (snowfall) over land and oceans with no frozen surfaces. Bayesian retrievals in the embedding space of a fully connected multilayer perception (MLP), equipped with a focal loss function, provide the most accurate estimates of the rates with a mean absolute error of less than 1.80 (0.15) mm h²¹ for rainfall (snowfall). Mutual information analysis unravels that beyond the near-surface air temperature, the 37 and 183 6 7(3) GHz are the most informative channels for phase detection over the ocean (land). The physical consistency of the retrievals and new explanations of the precipitation passive microwave signatures are provided through partial dependence analysis and annual comparison with the reanalysis data.