A Probabilistic U-Net Approach to Downscaling Climate Simulations

2025-11-05

AI Summary

Coarse-resolution climate model outputs inadequately resolve localized climate features required for regional impact assessments. To address this, we propose a probabilistic statistical downscaling method based on a U-Net architecture that integrates a deterministic U-Net backbone with a variational latent space, explicitly modeling stochastic uncertainty in precipitation and temperature fields. Our key innovation is a multi-objective loss function combining adaptive Continuous Ranked Probability Score (afCRPS) and Weighted Mean Squared Error–Multi-Scale Structural Similarity (WMSE-MS-SSIM): afCRPS improves representation of cross-scale spatial variability, while WMSE-MS-SSIM significantly enhances fidelity in reproducing extreme events. Experiments demonstrate that our method outperforms existing baselines in physical consistency, probabilistic calibration, and extreme-value fidelity. It establishes a new paradigm for high-resolution, high-fidelity climate downscaling with improved reliability for impact-oriented applications.

Technology Category

• Machine Learning: Calibration & Uncertainty Quantification
• Reasoning under Uncertainty: Uncertainty Representations
• Multiagent Systems: Multiagent Systems under Uncertainty

Application Category

• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking
• Graph Algorithms and Modeling for the Web: Representation, reconstruction, and subgraph or motif discovery in Web-related graphs
• Web Mining and Content Analysis: Large pretrained models with web data

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 $16 imes$ 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.

Problem

• Downscaling coarse climate simulations to finer resolutions
• Capturing aleatoric uncertainty in precipitation and temperature data
• Evaluating training objectives for extreme weather and spatial variability

Innovation

• U-Net backbone with variational latent space
• Training objectives include afCRPS and WMSE-MS-SSIM
• Downscaling climate data from 16x coarser resolution