Uncertainty Estimation of Bayesian Deep Learning for Convective Initiation Nowcasting

This study evaluates the probability and uncertainty forecasts of five Bayesian deep learning methods for 0-1 h convective initiation (CI) nowcasting using GOES-16 satellite infrared observations. Uncertainty is evaluated by how well probabilistic forecasts are calibrated and how well uncertainty separates forecasts with large and small errors. Of all the methods, the initial-weights ensemble + Monte Carlo (MC) method, an ensemble of deterministic residual neural network (ResNet) models with different initial weights to start training and MC dropouts activated during inference, produced more skillful and calibrated forecasts than the other Bayesian methods at 10- to 60-min lead times. Across all lead times, this method achieved 8-20% Brier skill score increases compared to the deterministic ResNet. The superior performance of the initial-weights ensemble + MC method is attributable to its enhanced forecast capabilities on low uncertainty cases, likely benefiting from a more extensive exploration in the hypothesis space. The Bayesian ResNet, a popular Bayesian method, performs worse than the deterministic ResNet models at lead times longer than 10 minutes. The MOdel Priors with Empirical Bayes using Deep neural network (MOPED) method enhances the skill of the Bayesian ResNet by initializing the priors with the pretrained deterministic ResNet. The Bayesian-MOPED ensemble method shows improved CI forecast skill over both the deterministic and Bayesian ensemble methods. The discard test shows the initial-weights ensemble + MC method effectively separates forecasts with large and small errors. Through case studies, the initial-weights ensemble + MC method demonstrates better forecast skill than the Bayesian-MOPED and deterministic ResNet on selected CI events in clear-sky regions. However, the initial-weights ensemble + MC method exhibits poorer generalization in clear-sky and anvil cloud regions without occurrence of CI compared to the deterministic and Bayesian-MOPED methods.