Largeau et al. (2026) Investigating the robustness of extreme precipitation super-resolution across climates

Identification

• Journal: Weather and Climate Extremes
• Year: 2026
• Date: 2026-04-01
• Authors: Louise Largeau, Tom Beucler, David Leutwyler, Grégoire Mariethoz, Valerie Chavez-Demoulin, Erwan Koch
• DOI: 10.1016/j.wace.2026.100885

Research Groups

• Expertise Center for Climate Extremes, University of Lausanne, Switzerland
• Faculty of Geosciences and Environment, University of Lausanne, Switzerland
• Faculty of Business and Economics (HEC), University of Lausanne, Switzerland
• Federal Office of Meteorology and Climatology MeteoSwiss, Switzerland

Short Summary

This study introduces a novel framework for super-resolving the Generalized Extreme Value (GEV) distribution parameters of hourly precipitation extremes using Vector Generalized Additive Models (VGAMs) and Vector Generalized Linear Models (VGLMs). It quantifies model robustness to climate change via a "robustness gap" and identifies limits to super-resolution factors based on spatial correlations.

Objective

• To develop and evaluate a framework for super-resolving the parameters of the Generalized Extreme Value (GEV) distribution for hourly precipitation extremes from coarse-resolution data and topography. A secondary objective is to quantify the robustness of these super-resolution models to climate change using a "robustness gap" metric and identify limits to super-resolution factors.

Study Configuration

• Spatial Scale: Switzerland, using a fine resolution of 2.2 km and coarse resolutions of 13.2 km (6x), 26.4 km (12x), and 52.8 km (24x) obtained via mean pooling.
• Temporal Scale: 11 European summer (June, July, August) seasons from historical (1999–2009) and projected (2079–2089) climates, with hourly precipitation data and monthly maxima.

Methodology and Data

• Models used:
  • Statistical models: Vector Generalized Additive Models (VGAMs) and Vector Generalized Linear Models (VGLMs) to parameterize the Generalized Extreme Value (GEV) distribution.
  • Climate model: COSMO regional weather and climate model for pseudo-reality simulations.
• Data sources:
  • Hourly precipitation output from COSMO regional climate model simulations at 2.2 km resolution (Hentgen et al., 2019).
  • Digital elevation model (DEM) for elevation statistics (mean and standard deviation of elevation).
  • Pseudo-global warming (PGW) method for future climate scenarios, using CMIP5 runs of the MPI-ESM-LR GCM for climate deltas.

Main Results

• VGAMs effectively super-resolve GEV distribution parameters for summer hourly precipitation extremes, outperforming VGLMs and coarse-resolution baselines in the reference climate (e.g., mean Cramér–von Mises error of 3.61 for VGAM vs. 3.67 for VGLM on 13.2 km test set).
• Coarse-resolution GEV parameters (location and scale) are the dominant predictors for super-resolution at a 6x factor, accounting for 81% of the model's explanatory power.
• The predictive advantage of coarse precipitation fields over elevation statistics diminishes beyond a super-resolution factor corresponding to a block size of approximately 30.8 km, where the spatial correlation of precipitation drops below its cross-correlation with elevation.
• Models trained on present-day data show significant "robustness gaps" at high quantile levels (above the 90th percentile) when applied to future climate data, indicating limitations in generalizing to climate change.
• This degradation in robustness at high quantiles is primarily driven by changes in the scale and shape parameters of the GEV distribution, while at lower quantiles, the location parameter dominates the robustness gap.
• Normalizing the robustness gap reveals that large absolute gaps at high quantiles partly reflect the intrinsic difficulty of predicting very rare extremes, rather than solely a lack of model transferability.

Contributions

• Introduction of the novel concept of super-resolving distributions of weather/climate extremes, focusing on distribution parameters rather than raw fields.
• Development of an interpretable "robustness gap" metric to diagnose and explain the generalization capabilities of empirical downscaling models under climate change, particularly for quantiles.
• Identification of an upper bound on super-resolution factors based on spatial auto- and cross-correlation functions, beyond which coarse-resolution target variables lose predictive value relative to auxiliary covariates.
• Provision of a model-agnostic diagnostic framework, applicable to any variable governed by parametric distributions, for understanding the factors driving generalization and robustness failures in non-stationary systems.

Funding

• Swiss National Science Foundation (SNSF) under Grant No. 10001754 ("RobustSR" project).
• Expertise Center for Climate Extremes (ECCE) at UNIL.
• Swiss National Supercomputing Centre (CSCS) under project ID pr144 for supercomputer time (Piz Daint).

Citation

@article{Largeau2026Investigating,
  author = {Largeau, Louise and Beucler, Tom and Leutwyler, David and Mariethoz, Grégoire and Chavez-Demoulin, Valerie and Koch, Erwan},
  title = {Investigating the robustness of extreme precipitation super-resolution across climates},
  journal = {Weather and Climate Extremes},
  year = {2026},
  doi = {10.1016/j.wace.2026.100885},
  url = {https://doi.org/10.1016/j.wace.2026.100885}
}