Machine Learning-Based Wildfire Susceptibility Modeling for Catalonia

Climate change is amplifying wildfire risk in many regions worldwide due to a variety of factors, such as the increasing frequency and intensity of heatwaves and droughts. To support mitigation strategies, accurate and timely prediction of wildfire susceptibility is essential.  

We present a wildfire susceptibility prediction model based on the Extreme Gradient Boosting (XGBoost) algorithm, designed to generate daily regional-scale susceptibility maps throughout the wildfire season. The model is implemented over Catalonia and trained using a diverse set of inputs, including population density, distance to the electric network, terrain elevation, the Normalized Difference Vegetation Index (NDVI), land cover classifications, and historical Fire Weather Index (FWI) and burned-area records. The training dataset covers the period 2007–2022 and includes 65 documented wildfire events, three of which correspond to large-scale fires affecting extensive areas, while the remaining events were of smaller magnitude. Model training focuses on the fire season (April–September), and performance is evaluated through external validation using data from 2023–2024. To ensure robust and generalizable predictions, we applied an extensive hyperparameter optimization procedure combined with a 5‑fold cross‑validation strategy, enabling the development of an optimized model and the creation of a consistent historical fire susceptibility dataset. 

Evaluation of the model predictions for the 2020–2024 period using the quadratic weighted Kappa metric shows moderate to strong agreement with the official fire danger maps produced by the regional forest fire prevention service across most of Catalonia. Reduced skill is observed in southern Lleida and in high‑elevation sectors of the northern Pyrenees, where additional analysis will be required to better understand the sources of these regional discrepancies and guide future model improvements. Importantly, the developed model consistently outperforms fire‑danger assessments based solely on the Fire Weather Index. For a comparable recall level (0.6), it achieves twice the precision, demonstrating substantially higher predictive skill in identifying areas at risk of ignition.

This model is currently under development within the MedEWSa project, funded by the EU Horizon Europe Programme (grant agreement No 101121192) and represents a step toward operational tools for wildfire risk management and climate adaptation in Mediterranean environments.  

Keywords: wildfire susceptibility, machine learning, XGBoost, fire danger prediction