Evaluation of a Convolutional Neural Network to Predict Wintertime Orographic Precipitation Gradients of the Western CONUS

Despite remarkable advances in forecasting precipitation for regions of complex terrain, many current operational modeling systems lack adequate grid spacing to resolve fine-scaled precipitation patterns. As a result, statistical or dynamical downscaling is often used to produce higher-resolution precipitation forecasts. However, statistical downscaling can overlook non-stationary atmospheric relationships and often relies on a monthly or annual climatology, whereas dynamical downscaling features physically-based calculations but is computationally expensive. This study aims to downscale precipitation by predicting orographic precipitation gradients (OPG) over the Northern Rockies of the western continental United States (CONUS) using a convolutional neural network (CNN).

Based on prior work by Bohne et al. (2020), we divided the western CONUS terrain into facets based on regional terrain orientation. Linear regressions were then used to quantify daily OPGs within each facet from observed precipitation of the Global Historical Climatology Network-Daily dataset from 1979 to 2018. These daily OPGs frequently vary from the climatological means due to the effects of large-scale circulation patterns, suggesting the potential for a machine learning algorithm to predict OPG based on these large-scale patterns.

Our CNN is trained using the ECMWF ERA5 Reanalysis and daily OPG from all suitable facets in the Northern Rockies. Current training results for the Northern Rockies region show the CNN model accounting for 50% of the OPG variance with a mean absolute error of about 2.5 mm of water equivalent per km of elevation. Composite Grad-CAM analysis of regional high OPG events indicates the CNN is focusing on appropriate areas for OPG prediction. In contrast, composite Grad-CAM analysis of southerly high OPG events indicates that the CNN focuses on seemingly unrelated regions for OPG prediction. The development of a combination CNN-OPG model enables insights into the relationships between large-scale circulations and western CONUS OPG and enables novel approaches for prediction.