Methodical framework of the PRE4IMPACT-AT project: Exploiting explainable machine learning for impact-based early warning and trend analysis in Austria

Recent extreme precipitation events across Europe, including those in autumn 2024, underscore the need to strengthen proactive disaster risk reduction through improved impact-based early warning. In Austria, precipitation-related hazards such as landslides, flash floods and hailstorms repeatedly result in considerable impacts on people, infrastructure and economic assets. These challenges are expected to intensify under ongoing climate and environmental change. In response, European national meteorological and hydrological services are increasingly pursuing a paradigm shift in their warning strategies, from traditional weather warnings towards impact-based warnings (IbW). IbW focus on the consequences of weather events (“what the weather will do”) rather than solely on meteorological conditions (“what the weather will be”). However, data-driven and applicable approaches to predict precipitation-induced impacts at the national scale remain limited.

The PRE4IMPACT-AT project is part of the Austrian Climate Research Programme (ACRP) and addresses this gap by developing explainable and user-oriented impact-based predictive models for precipitation-related hazards in Austria. This contribution presents the overall project concept and the methodological framework, exemplified through a recent transferable and generalizable approach (Steger et al., 2025; https://doi.org/10.5194s/egusphere-2025-4940). PRE4IMPACT-AT focuses on processes whose impacts that typically occur in temporal and spatial proximity to precipitation events, namely landslides, flash floods and hailstorms.

Adopting a risk-oriented perspective, PRE4IMPACT-AT first conceptualizes impacts as the outcome of interacting atmospheric drivers, biophysical and geomorphological preconditions, and socioeconomic exposure and vulnerability. These relationships are formalized using an impact-chain framework, which supports the systematic identification and prioritization of key impact drivers for each hazard type. In subsequent steps, the selected drivers are parameterized and harmonized using a wide range of national datasets, including meteorological and geo-environmental information, as well as socioeconomic data. Model training relies on available national and international damage databases (landslides, flash floods) and agricultural insurance loss data (hail). Based on these datasets, explainable machine learning is applied to derive spatiotemporal predictive rules linking static and dynamic drivers to observed impacts. The resulting models aim to characterize typical impact conditions, with a strong emphasis on interpretability to enhance transparency and allow plausibility checks. The models are evaluated in hindcast and nowcast settings to assess their suitability for short-term impact-based warning applications. In addition, long-term analyses, synthesizing large numbers of hindcasts, are used to identify trends in critical conditions and emerging patterns. Finally, individual hazard-specific models are combined to provide a multi-hazard impact perspective. A core element of PRE4IMPACT-AT is continuous user engagement through iterative evaluation workshops with stakeholders who hold warning mandates. Overall, the project contributes to advancing impact-based forecasting, early warning and climate impact assessment by providing Austria with a transparent and operationally relevant foundation, while offering transferable insights for national services facing similar challenges across Europe. 

This project is funded by the Climate and Energy Fund in the course of the Austrian Climate Research Programme (ACRP) and the FFG (www.ffg.at). The FFG is the central national funding agency and strengthens Austria’s innovative capacity.