Bayesian deep learning for convective initiation nowcasting uncertainty estimation

Fan, D., Gagne, D. J., Greybush, S. J., Clothiaux, E. E., Schreck, J. S., et al. (2026). Bayesian deep learning for convective initiation nowcasting uncertainty estimation. Artificial Intelligence for the Earth Systems, doi:https://doi.org/10.1175/aies-d-25-0064.1

Title Bayesian deep learning for convective initiation nowcasting uncertainty estimation
Genre Article
Author(s) D. Fan, David John Gagne, S. J. Greybush, E. E. Clothiaux, John S. Schreck, C. Shen
Abstract This study evaluated the probability and uncertainty forecasts of five recently proposed Bayesian deep learning methods relative to a deterministic residual neural network (ResNet) baseline for 0–1-h convective initiation (CI) nowcasting using Geostationary Operational Environmental Satellite ( GOES-16 ) satellite infrared observations. Uncertainty was assessed by how well probabilistic forecasts were calibrated and how well uncertainty separated forecasts with large and small errors. Three Bayesian deep learning methods produced probabilistic forecasts that were comparable to those of the deterministic ResNet. Among them, the initial-weights ensemble + Monte Carlo (MC) dropout, a collection of deterministic ResNets with different initial weights to start training and dropout activated during inference, produced the most well-calibrated probability forecasts and the most reliable uncertainty estimates. The Bayesian ResNet ensemble performed worse than the deterministic ResNet at certain forecast times, likely due to the challenge of optimizing a larger number of parameters. To address this issue, the Bayesian–Model Priors with Empirical Bayes using Deep neural network (MOPED) ResNet ensemble was adopted, which constrained the hypothesis search near the deterministic ResNet solution and achieved forecast skill comparable to that of the deterministic ResNet. All Bayesian methods demonstrated well-calibrated uncertainty and effectively separated cases with large and small errors. In generalization tests, the initial-weights ensemble + MC dropout demonstrated better forecast skills than the Bayesian-MOPED ensemble and the deterministic ResNet on selected CI events in clear-sky regions but showed weaker generalization over broad clear-sky regions. In anvil cloud regions, all Bayesian methods produced skillful forecasts for the selected CI events but demonstrated poor generalization over the non-CI anvil regions.
Publication Title Artificial Intelligence for the Earth Systems
Publication Date Mar 1, 2026
Publisher's Version of Record https://doi.org/10.1175/aies-d-25-0064.1
OpenSky Citable URL https://n2t.net/ark:/85065/d7862n1r
OpenSky Listing View on OpenSky
CISL Affiliations MILES

< Back