A Probabilistic U-Net Approach to Downscaling Climate Simulations

TL;DR

This paper tackles the computational bottleneck of obtaining high-resolution climate projections by applying a probabilistic U-Net to downscale data from $16\times$ coarser resolutions, enabling uncertainty quantification for precipitation and temperature fields. It couples a deterministic U-Net backbone with a variational latent space and evaluates four training objectives to balance extreme-event fidelity and fine-scale variability. The study finds that MS-SSIM-based losses excel at reproducing extremes under certain settings, while afCRPS better captures spatial variability across scales, with none universally dominating across metrics. The work highlights the practical value of probabilistic downscaling for climate impact analyses, suggesting that a combination of objectives may yield robust representations of both extremes and fine-scale structures.

Abstract

Climate models are limited by heavy computational costs, often producing outputs at coarse spatial resolutions, while many climate change impact studies require finer scales. Statistical downscaling bridges this gap, and we adapt the probabilistic U-Net for this task, combining a deterministic U-Net backbone with a variational latent space to capture aleatoric uncertainty. We evaluate four training objectives, afCRPS and WMSE-MS-SSIM with three settings for downscaling precipitation and temperature from $16\times$ coarser resolution. Our main finding is that WMSE-MS-SSIM performs well for extremes under certain settings, whereas afCRPS better captures spatial variability across scales.

Figures (7)

• Figure 1: Precipitation return levels for four training objective variants at two grid cells.
• Figure 2: Ground truth versus four training objectives for precipitation: log-frequency histograms (left) and power spectral density (right).
• Figure 3: Probabilistic U-Net architecture for statistical downscaling, showing the prior and posterior networks, the U-Net backbone, and the latent variable fusion during (a) training and (b) inference.
• Figure 4: From left to right: the coarse-resolution input, three sampled high-resolution realizations from the model (out of an arbitrarily large ensemble), and the ground-truth high-resolution field. This figure illustrates how the probabilistic U-Net generates diverse yet physically consistent realizations. While the large-scale precipitation pattern is reproduced across all predictions, variability appears in regions of higher intensity, reflecting the model’s stochastic sampling of fine-scale structures. This ensemble spread is precisely what enables the model to represent uncertainty in extremes that a deterministic baseline would smooth out.
• Figure 5: Minimum (left panel) and maximum (right panel) temperature return levels for four training objective variants at two grid cells.