Exploring the monthly contribution of drivers on European summer wildfires with explainable artificial intelligence (XAI)

• Established a European-wide time series wildfire model using 18 indicators. • Meteorological and vegetation features are the most important indicators. • Summer and the winter-spring transition are key periods for fire prevention. • The high contribution of indicators in the Mediterranean emerges as early as spring. • Early monitoring of SHAP values for indicators can provide effective warnings. As climate change continues, wildfires are becoming more frequent, posing significant challenges to both society and ecosystems. The time-lagged effects of wildfire drivers make them difficult to quantify, and a comprehensive understanding of how preceding weather and vegetation conditions influence wildfire risk across Europe is still lacking. In this study, we focus on summer wildfires in European forests, shrubs, and herbaceous vegetation areas from 2014 to 2023. Using the long short-term memory (LSTM) method, we developed a reliable model with an Area Under the ROC Curve of 0.928, incorporating 18 indicators related to burned areas, meteorology, vegetation, topography, and human activity. We then used SHapley Additive exPlanations (SHAP), an Explainable Artificial Intelligence method, to interpret the model and obtain the contribution of drivers in the 11 months preceding wildfire events. The results indicate that the four main contributors are the condition indices of Land Surface Temperature, Solar Radiation, and Soil Moisture, along with NDVI. Wildfire severity in summertime is most strongly tied to current-season drivers, with the winter-to-spring transition as the next key period. Wildfire risk in the Mediterranean is the highest among biogeographic regions, with the relatively high contribution of drivers to August wildfires emerging as early as spring. For wildfires caused by the cumulative effects, early monitoring of SHAP values can provide effective warnings. Our study provides a new method for quantitative analysis of the time-lag effects of wildfire drivers, which will help to better understand the mechanisms of wildfire occurrence and enhance prevention and mitigation.