Downscaling climate projections to 1 km with single-image super resolution

TL;DR

Helps address the gap between coarse-resolution climate projections and local decision-making needs by downscaling to 1 km using single-image super-resolution trained on observations. The authors compare EDSR, FNO, and SwinIR and evaluate downscaled outputs via weather-station climate indicators (TG, GDD, HDD, CDD) rather than pixel-wise errors. They show that SR-based downscaling improves indicator RMSE over the low-resolution REMO2015 baseline and bilinear/bicubic interpolation. The work also proposes a practical evaluation framework for downscaled projections in the absence of ground-truth high-resolution climate fields, highlighting directions for uncertainty quantification and extension to additional variables.

Abstract

High-resolution climate projections are essential for local decision-making. However, available climate projections have low spatial resolution (e.g. 12.5 km), which limits their usability. We address this limitation by leveraging single-image super-resolution models to statistically downscale climate projections to 1-km resolution. Since high-resolution climate projections are unavailable, we train models on a high-resolution observational gridded data set and apply them to low-resolution climate projections. We cannot evaluate downscaled climate projections with common metrics (e.g. pixel-wise root-mean-square error) because we lack ground-truth high-resolution climate projections. Therefore, we evaluate climate indicators computed at weather station locations. Experiments on daily mean temperature demonstrate that single-image super-resolution models can downscale climate projections without increasing the error of climate indicators compared to low-resolution climate projections.

Figures (2)

• Figure 1: Low-resolution (upper left) and downscaled climate projection for May 1, 2003
• Figure 2: ReKIS evaluation weather stations