Switzerland, like many countries, is becoming increasingly vulnerable to extreme rainfall as a result of climate change. Heavy precipitation events are occurring more frequently and are expected to intensify further as the climate warms, according to the Federal Office for Meteorology and Climatology (MeteoSwiss)1. Combined with the Alps’ steep terrain, these changes are likely to exacerbate the occurrence of landslides and flooding.
Yet climate models still struggle to reliably capture extreme precipitation events. Most global and regional models operate at coarse resolution of roughly 100 and 15 kilometres, respectively. This limits their ability to capture localised extremes, particularly in complex terrains such as the Alps.
With more extreme events there is a need for accurate local-scale projections at resolutions of a few kilometres. Typically, communes and towns want to know in granular detail who and what could be affected by future flooding or landslides. Improved projections are crucial if local authorities, insurers and businesses are to manage and underwrite risk, plan and adapt infrastructure, invest in flood defences, or create preparedness schemes.
The super-resolving approach
Super-resolving climate projections is one answer. This is a rapidly evolving scientific field focused on transforming low-resolution global weather and climate model outputs into high-resolution, localised forecasts.
The Expertise Center for Climate Extremes (ECCE) at the University of Lausanne, which brings together specialists from HEC Lausanne and the Faculty of Geosciences and Environment (FGSE), conducts interdisciplinary research in this area. “We have developed a novel approach to super-resolving climate variables, aimed at inferring the local statistical distribution of extremes.” explains Erwan Koch, Director of ECCE.
When applying their super-resolution framework to extreme rainfall in Switzerland, Koch and his team combine machine learning with extreme-value theory, specifically the Generalised Extreme-Value distribution.2 They assess how models trained under current conditions generalise to future climates by leveraging present-day data alongside future climate simulations.
To this end, they introduce the concept of the “robustness gap”, defined as the difference in performance under future climate conditions between models trained on present-day data and those trained directly on future climate data. This metric captures how reliable a model remains when applied under future climate conditions. The overarching objective is to reduce this gap and improve the model over time.
“We have applied our methodology to extreme precipitation in Switzerland. We use easily interpretable statistical models, enabling us to identify and analyse the drivers of the robustness gap — how can this gap be explained?” asks Koch.
This research from ECCE advances this field by offering a new interpretable methodology, enabling more accurate projections at much finer spatial resolution. “Our model and methodology are highly relevant in practice, providing decision-makers with detailed local insights into future extreme events, along with a clear understanding of uncertainty and reliability. They are also easy to deploy, as they are far less computationally demanding than many AI-based approaches.” explains Koch.
He concludes: “Such local insights are essential for accurately assessing risks and guiding climate change adaptation strategies.”
Reference:
1. More frequent and more intense heavy precipitation, Federal Office of Meteorology and Climatology, MeteoSwiss, 2025
2. Investigating the robustness of extreme precipitation super-resolution across climates, Largeau L., Beucler T., Leutwyler D., Mariethoz G., Chavez-Demoulin V., Koch E., Weather and Climate Extremes, Volume 52, June 2026
