A supranational machine learning approach to assess wildfire losses under climate change in Southeastern Europe

Abstract Wildfires pose significant threats to both ecosystems and socio-economic assets, particularly under changing climate conditions. This paper presents a comprehensive, data-driven approach for wildfire risk assessment that integrates machine learning modeling, fuel-type classification, and economic loss estimation. First, a Random forest-based susceptibility model is developed to identify the probability of fire spread, using diverse predictors such as climate indices, topographic parameters, and vegetation continuity. The model is calibrated at a pan-European scale and then tailored at national levels to capture interannual variability. Next, wildfire hazard is determined by combining susceptibility output with aggregated vegetation classes to approximate potential fire intensity. This susceptibility-hazard combination highlights the range of wildfire behaviors, from low-intensity surface fires in grasslands to high-intensity crown fires in coniferous forests. To account for potential economic impacts, the approach incorporates exposure data—buildings, forests, and critical infrastructure, along with vulnerability tables that translate hazard intensity into monetary losses. Average Annual Loss is then derived by pairing hazard-based probabilities with these exposure and vulnerability layers. The results comprehensively quantify potential financial impacts. By merging detailed ML susceptibility modeling with explicit hazard and loss calculations, this study offers a scalable, robust tool for policymakers and land managers seeking to mitigate wildfire risks in an era of intensifying climate extremes.