Scalable & explainable ML for wildfire risk modeling in Southern Europe: A case-study in Portugal (Papers Track) Spotlight
Ophélie Meuriot (Denmark Technical University); Jorge Soto Martin (Denmark Technical University); Beichen Zhang (Lawrence Berkeley National Laboratory); Francisco Camara Pereira (Denmark Technical University); Martin Drews (Denmark Technical University)
Abstract
Wildfires constitute one of the most devastating natural disasters, with wide impacts across economic sectors and society in Europe. The study shows how data-driven models can be used to combine human, topographic, land cover and climate data to quantify wildfire risk. Four machine learning models are trained on a historical record of fires (from June to October between 2008 and 2023) in Southern Europe excluding Portugal. The Random Forest (RF) model (AUC = 0.91 and F1 = 0.85) is then run daily in Portugal (12 x 12 km grid) during October 2017, a month during which devastating wildfires were the cause of 51 casualties. The model is found to represent high-risk areas more accurately than the widely used Fire Weather Index (FWI). Key features influencing the high fire risk during the period are identified using explainable AI. This study provides a scalable and lightweight model, which can be used to support climate impact assessments by (1) quantifying high-risk areas and (2) identifying key drivers to inform adaptation strategies.