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

Cool-season precipitation plays a significant role in the complex terrain of the Western contiguous United States (CONUS), impacting water resources, recreation, and public safety. However, the coarse grid spacing of current operational modeling systems cannot sufficiently resolve fine-scaled precipitation patterns in this region. In this study, convolutional neural networks (CNNs) were utilized to downscale winter precipitation in mountainous terrain based on observed orographic precipitation gradients (OPGs), allowing for more fine-scaled prediction. The observed OPGs were formulated from the Global Historical Climatology Network-Daily dataset as linear regressions onto faceted terrain, similar to mountain faces. These daily wintertime OPGs were found to vary due to synoptic conditions, suggesting the potential for OPGs to be predicted based on synoptic-scale atmospheric patterns with a CNN.

Previous work in our group tested a CNN in the Northern Rockies region of the CONUS accounted for 34% of OPG variance with a mean absolute error (MAE) of about 2.9 mm (1000 m)-1; however, facets with fewer OPG observations had higher MAE. To improve results, we speculated the strong spatial relationship between facet OPGs could be leveraged to improve facets with fewer observations within the loss function. This research aims to expand the application of the CNN tested in the Northern Rockies to predict OPGs for the entire Western CONUS and to improve model performance via hyperparameter tuning and custom loss functions.

With the goal of reducing errors in facets with fewer OPG observations, two separate custom loss functions were tested that leveraged a strong spatial relationship found between facets with long observational records. One loss function was based on the spatial correlation of facet OPGs while the other was based on the absolute difference in OPGs between facets. By penalizing deviations from an expected spatial relationship between facets with at least a 75% length of record, facets with fewer observations could benefit from neighboring facets with more observations. The correlation of the CNN’s predicted facet OPGs by distance followed a similar exponential decay to that of the relationship based solely on facets with a 75% length of record but had an overestimation of magnitude. The mean absolute difference in the CNN’s predicted facet OPGs increased logarithmically with distance, similarly to the relationship based on facets with a 75% length of record, but with an underestimation of magnitude. Thus, when these relationships were leveraged in custom loss functions, the resulting predicted OPGs contained random heterogeneity, which was effective at decreasing (increasing) correlation (absolute difference), but ineffective at improving OPG prediction. Ultimately, the CNN for the entire Western CONUS only accounted for 15% of OPG variance with a MAE of 4.69 mm (1000 m)-1, possibly indicating that the entire Western CONUS could be too large and diverse of a region for a single CNN to be effective.