DEUCE: A Neural Network for Probabilistic Precipitation Nowcasting with Aleatoric and Epistemic Uncertainties

Precipitation nowcasting (i.e. forecasting locally at 0-6h lead times) can help with reducing the economic and humanitarian costs incurred by the consequences of heavy precipitation. The usual approach is to predict future radar echoes and to convert them to precipitation estimates, as NWP is ill-suited for this problem. The past decade has seen lots of advances in precipitation nowcasting techniques, in particular in the rise of deep learning models as a popular alternative to classical methods. However, the highly useful probabilistic approach has yet only been narrowly explored with deep learning.

We propose the DEUCE (Deep Ensemble-based Uncertainty Combining radar Echo nowcasting) neural network model for probabilistic precipitation nowcasting. DEUCE models the epistemic (knowledge-related and reducible) uncertainty of the prediction with an underlying U-Net architecture by placing probability distributions on the weights, which are learned through variational inference. The uncertainty from sampling the weights is combined with the heteroscedastic aleatoric (input-dependent and irreducible) uncertainty, modeled as the variance of the Gaussian likelihood of the prediction, provided by a separate decoder branch. The combined uncertainty is used to produce an ensemble of predictions.

DEUCE is trained and verified using Finnish Meteorological Institute radar data. It produces both skillful and reliable probabilistic nowcasts ranging to 60 minutes, showing the highest area under the ROC curve, as well as competitive Expected Calibration Error and CRPS scores against STEPS and LINDA-P baseline models. The relative importance of the uncertainties is also explored, showing the aleatoric part to be more substantial and to exhibit a clearer dependence on the input variables than the epistemic part.