Physically Explainable Deep Learning for Convection Initiation Nowcasts

Convection initiation (CI) nowcasting remains a challenging problem for both numerical weather prediction models and existing nowcasting algorithms. In this study, object-based probabilistic deep learning models are developed to predict CI based on multichannel infrared GOES-R satellite observations. The data comes from patches surrounding potential CI objects identified in Multi-Radar Multi-Sensor Doppler weather radar products by an objective radar-based approach from June and July 2020 and June 2021. The deep learning models significantly outperform the classical logistic model at lead times up to 1 hour. To explain the reasoning behind the complex structure, two model explanation methods, Shapley additive explanations and gradient times input, are used. The explanations through these methods reveal that ice generation is important behind upcoming CI events, while clear-sky environments are favorable for no-CI events. To further explore how trustworthy these results are, the uncertainty of the results is quantified and decomposed into aleatoric uncertainty and epistemic uncertainty. Results reveal that the uncertainties are high for incorrect predictions, and low for correct predictions. The dependence of model performance on these uncertainties, as well as predictability as a function of lead time, is further explored.