Spatially Informed Convolutional Neural Networks for Predicting Winter Orographic Precipitation Gradients of the Western CONUS: The Python TensorFlow-GPU OPG-CNN Models
Description
Convolutional neural networks (CNN) were developed to predict winter orographic precipitation gradients (OPGs; Bohne et al. 2020) across the western Contiguous United States (CONUS) from the European Center for Medium-Range Weather Forecasting (ECMWF) Reanalysis, Version 5 (ERA5; Hersbach et al. 2023). OPGs were formulated as a linear regression of elevation change on topographic facets and accumulated precipitation from the Global Historical Climatology Network (GHCN)-Daily, Version 3.22 (Menne et al. 2012). Spatial evaluation of western CONUS OPGs found that absolute correlation decreased exponentially with distance, while mean absolute differences increased and plateaued with distance. With this finding, custom combination loss functions were used to penalize the CNN by the mean squared error of OPG values and the mean squared error of expected spatial relationships of OPGs. The relative importance of each component in the loss function was explored through up-weighting the spatial relationship (alpha) and down-weighting the OPG value error (1-alpha) over many CNN trainings. The following folders contain the Python Tensorflow-GPU developed OPG-CNN models with varying alpha weights (Chollet et al. 2015). Additionally, this repository contains an OPG-CNN control case trained only on the mean squared error of OPG values, the OPG data set, and example atmospheric data from ERA5. An example loading each of the models is provided in example_load_opg-cnn_mdl.py and the custom loss functions are listed within custom_combination_loss_functions.py for quick reference. Support for this project was provided by the Wilkes Climate Center Seed Grant (Award: 003921), the Utah Division of Water Resources’ Great Salt Lake Basin Integrated Plan, and the Cooperative Institute for Research to Operations in Hydrology (CIROH) with funding from the NOAA Cooperative Institute Program (Award: NA22NWS4320003). Undergraduate research support was provided by the National Science Foundation Research Experiences for Undergraduates Program (Award: 2244272) through the University of Utah Research Experience in Alpine Meteorology (REALM) program, and the Undergraduate Research Opportunities Program (UROP) at the University of Utah. The support and resources from the Center for High Performance Computing at the University of Utah are gratefully acknowledged.
Files
Institutions
- University of UtahUtah, Salt Lake City