An A Priori Framework for Training-Source Selection in Deep-Learning Climate Downscaling

Abstract: Deep learning-based climate downscaling requires paired low-resolution and high-resolution training data, yet the low-resolution training source is rarely treated as an explicit experimental variable. This source generally falls into one of two categories: perfect prognosis (PP), in which the high-resolution target is coarsened to build its own training input, or model output statistics (MOS), in which a genuinely independent product, such as reanalysis or forecast output, serves as the input instead. This choice is not arbitrary: no training input can hold temporal correspondence with a future, unobserved climate, so every source necessarily sacrifices either temporal correspondence or distributional realism. However, no study treats this choice as an experimental variable in the observational downscaling setting, and although several independent studies document downstream symptoms of weak temporal correspondence, such as shared-scale mismatches, mass-conservation failures, and shifts in what a network learns to depend on spatially, none isolates temporal correspondence as a standalone, pre-training-computable quantity. This is the central gap this study addresses.

Therefore, in this study, we treat the low-resolution training source, its regridding operator, and temporal correspondence as explicit experimental variables for a fixed DL4DS U-Net backbone, comparing a coarsened Un-Restricted Mesoscale Analysis (URMA) source against real numerical weather prediction (NWP) and reanalysis products, including the High-Resolution Rapid Refresh (HRRR), the ICON Dual-resolution Reanalysis for Emulators, Applications and Monitoring (ICON-DREAM), and the fifth-generation ECMWF Reanalysis (ERA5), for wind speed, temperature, and precipitation. Each source is regridded with three operators, conservative, bilinear, and cubic, to test whether interpolation choice independently affects downstream skill beyond the source itself. We compute Wasserstein distance, a power-spectral-density-based effective-resolution proxy, and temporal correspondence for each source before any model is trained, then regress these pre-training diagnostics against post-training performance on historical hold-out and on projections from the Ouranos dynamically downscaled ensemble. We use New Mexico as an illustrative domain, chosen for its climatology and the severity of its climate-driven risk: extreme-precipitation increases here are projected to peak in autumn rather than winter, unlike most of the contiguous United States, and the region's water resources depend in part on a river ranked among the most stressed in the country.

We hypothesize that the optimal source is variable-dependent: distributional similarity should govern smooth variables such as wind speed and temperature, favoring finer-native-resolution reanalysis and NWP sources, while temporal correspondence should govern intermittent variables such as precipitation, favoring the coarsened-observation source despite its coarser native resolution. We anticipate this effect is sharper here than in humid-domain precedent, because monsoon-driven convection is more spatiotemporally discontinuous than synoptic-scale winter precipitation, amplifying the timing penalty paid by free-running sources. This study provides the first systematic, a priori, data-only protocol for training-source selection in deep-learning statistical downscaling, directly informing how water-management, agricultural-extension, and infrastructure-planning pipelines should preprocess coarse Earth-system-model output before downscaling it for the arid Southwest.