Understanding Predictability of Extreme SEUS Precipitation Using Explainable Machine Learning

We previously used XAI to investigate the predictability of the sign of aggregated daily South-East US (SEUS) precipitation anomalies associated with large-scale atmosphere and ocean predictors. We developed an accurate and reliable convolutional neural network (CNN) with gridded fields of sea surface temperature (SST), outgoing longwave radiation (OLR), zonal wind, and geopotential height as predictors. The high reliability of the CNN allowed us to identify forecasts of opportunity or times of higher predictability. We then used XAI to provide insights into the sources of this higher predictability learned by the CNN. Specifically, we identified synoptic-scale circulation patterns for winter and summer that are sources of predictability for daily SEUS precipitation. These circulation patterns were linked to ENSO during winter and a connection between North Atlantic SSTs and the North Atlantic Subtropical High during summer. These results lead to questions about whether the pattern and amplitude of large-scale atmospheric circulation that contributes to forecasts of opportunity for the sign of precipitation anomalies is the same or different for precipitation extremes in this region.

In this presentation we show the results of investigating sources of predictability for extreme precipitation in the SEUS. We use similar CNN and XAI methodologies designed to predict whether a wet or dry extreme will occur in the SEUS during a given week. Temporal and spatial aggregation are used to mitigate the potential for sample size issues when designing and training neural networks for predicting extremes. This allows us to capture events that occur frequently enough within a region to train a data-driven model but are locally extreme at a given gridpoint. The percentile thresholds vary by region, season, and wet vs. dry extreme. Preliminary results for dry extremes of <3 %-tile precipitation and >95 %-tile precipitation indicate a different circulation pattern associated with predicting forecasts of opportunity for dry vs. wet extremes and for extremes vs. sign of anomalies.