Abstract:
Satellite passive microwave (PMW) radiances provide a wealth of convective and microphysical information in data sparse regions like over the open ocean when available, but there is insufficient temporal and spatial PMW data coverage of individual convective phenomena resulting from the limitations of low-Earth orbits. In contrast, geostationary satellites can provide great spatial and temporal data coverage of individual convective phenomena within their domain, but their data in the visible and infrared spectrum provides less information about the convective structure below cloud tops. Therefore, this study seeks to investigate the possibility blending the benefits of both types of satellites by utilizing Bayesian Convolutional Neural Networks (CNNs) to produce synthetic “geostationary” PMW radiances from geostationary infrared brightness temperatures. Since the benefit of Bayesian CNN artificial intelligence/machine learning (AI/ML) models over other AI/ML architectures is the ability to quantify uncertainty, specific emphasis is placed on exploring how the choice of Bayesian CNN architecture impacts the skill and utility of the model. Results are presented in two parts, where Part 1 investigates how the choice of “Flipout,” “Reparameterization,” or “Monte Carlo Dropout” Bayesian architecture impacts model skill, and Part 2 extends the results of Part 1 towards the development of a new Bayesian architecture that decomposes predicted uncertainty into its “aleatoric” and “epistemic” components. Part 2 will then also explore the practical utility of this uncertainty decomposition, which has implications for active machine learning, training dataset optimization, data assimilation, and real-time convection identification as explored through Hurricane Teddy (2020).