Leveraging spatial patterns in precipitation forecasts using deep learning to support risk-averse flood management

Short-term forecasts of heavy precipitation are critical to regional flood
control operations, particularly in the Western U.S. where atmospheric
rivers can be predicted reliably days in advance. However, spatial error
in these forecasts may reduce their utility for risk-averse system
operations, where false negatives could be especially costly. Here we
investigate whether deep learning methods can leverage spatial patterns in
precipitation forecasts to (1) improve the skill of predicting the
occurrence of heavy precipitation events in a target region at lead times
from 1-14 days, and (2) balance the tradeoff between the rate of false
negatives and false positives (misses and false alarms) by modifying the
discrimination threshold of the classifiers. This approach is demonstrated
for the Sacramento River Basin, California, using the Global Ensemble
Forecast System (GEFS) v2 precipitation fields as input to convolutional
neural network (CNN) models. Results show that the deep learning models do
not significantly improve the overall skill (F1 score) relative to the
bias-corrected ensemble mean GEFS forecast. However, the models often
correct missed predictions from GEFS by compensating for spatial error at
longer lead times. Additionally, the deep learning models provide the
ability to adjust the rate of false negatives based on a desired level of
risk aversion, with the tradeoff of increasing the false positive rate.
Finally, analysis of the network activations (saliency) indicates spatial
patterns consistent with physical understanding of atmospheric river
events in this region, lending additional confidence in the ability of the
method to support flood management.