Integrating physical drivers for wildfire hazard modelling in support of disaster risk reduction: A case study in Castile and León, Spain

Wildfires represent a global threat, particularly in southwestern Europe and the Mediterranean basin, a climate change hotspot where rising temperatures and drought exacerbate risk. The present work addresses the urgent need for updated, spatially explicit hazard models, focusing on Castile and León, Spain, a strongly affected region within southwestern Europe that experienced an unprecedented fire season in 2025. The objective was to develop and evaluate a robust and globally transferrable wildfire hazard assessment framework at a daily, 1-km scale. This involved integrating biophysical and anthropogenic data from 2007 to 2022 with an external evaluation using data from August 2025. The rigorous validation within this complex, high-risk socioecological context ensured the fundamental capacity of the models for broader application. Two machine learning models were used: Convolutional Long Short-Term Memory (ConvLSTM) and Extreme Gradient Boosting (XGBoost). The results demonstrate the robustness of both models for predicting the ignition hazard. Interpretability analyses consistently identified thermal and hydric forcings as the main determinants of danger. Both models had strong predictive performance with complementary strengths, i.e., ConvLSTM excelled in restricting false alarms, and XGBoost maximized the detection of actual events. The integration of both architectures offers a solid foundation for developing hybrid decision-support systems. This very transferable modeling framework will empower managers and policymakers to optimize the allocation of firefighting resources and develop dynamic early warning systems, thereby facilitating a shift towards adaptive and sustainable risk management strategies in southwestern Europe and informing management practices in other fire-prone regions worldwide.