Generalizable Spatial Downscaling of 2-m Temperature with Multiple Data Sources - Insights from the Recent Advances in Deep Learning

Abstract:

Statistical downscaling (SD) bridges the gap between low resolution (LR) numerical model fields and high-resolution (HR) meteorological inputs that are requested by regional impact studies. In this research, we provide a novel deep-learning-based approach for the downscaling of gridded daily mean 2-m temperature (TMEAN) in western North America, with the focus of generalization abilities on unseen spatial domains and numerical models. Our method collaborates the Cycle-consistent Generative Adversarial Network (Cycle-GAN) that transforms raw numerical model mesh TMEAN, typically with 25-100 km grid spacings, into a 0.25-degree regular latitude-longitude lattice, and an encoder-decoder CNN, the UNet, to downscale the transformed 0.25-degree LR TMEAN and produces an HR version with 4-km grid spacings. Our method is trained by the Parameter-Elevation Regressions on Independent Slopes Model (PRISM) from 2015 to 2018 and evaluated by the Global Historical Climate Network Daily (GHCN-daily) observations from 2018 to 2020. The evaluation result shows that our method can perform high quality downscaling for LR inputs obtained from reanalysis products, numerical weather prediction, and global climate model runs, with up to 20% mean absolute error reduction compared to a classic SD downscaling baseline. Our method is a good option for the downscaling of TMEAN in areas with insufficient observation data.