Towards Trusting Synthetic Microwave Retrievals: Introducing Uncertainty Quantification in a New Bayesian Convolutional Neural Network Architecture

Microwave satellite brightness temperatures (TBs) provide valuable insight into microphysics, precipitation characteristics, and convective organization, especially in data sparse areas like over the open ocean. However, due to physical constraints, microwave TBs are only retrieved in narrow swaths with relatively coarse resolution when compared to current geostationary infrared capabilities, and the revisit times of microwave swaths are insufficient for tracking rapidly evolving extreme weather phenomena like tropical cyclones. But what if microwave TBs did not have to be limited to swaths? In this study, we develop a new Bayesian convolutional neural network (BNN) architecture to investigate the feasibility of using machine learning to generate synthetic microwave TBs over the entire ocean-only portion of GOES-16 full-disk domain at higher spatial and temporal resolutions than current microwave retrievals. The advantage of this new BNN architecture is that it provides both synthetic microwave TBs and metrics of "trustworthiness" for each TB, which allows users to discriminate between confident retrievals versus uncertain guesses. Finally, we also investigate characteristics of the synthetic microwave TB skill with respect to different microwave frequencies, microphysical characteristics, and cloud structures.