XGBoost meets INLA: a two-stage spatio-temporal forecasting of wildfires in Portugal

Abstract: Wildfires pose a major threat to Portugal, with an average of over 115,000 hectares burned annually in the 45-year period of 1980-2024. Beyond a high number of ignitions, the country has experienced devastating mega-fires, such as those in 2017. Accurate forecasting of wildfire occurrence and burned areas is therefore essential for effective firefighting resource allocation and emergency preparedness. In this study, we present a novel two-stage ensemble approach that extends the widely used latent Gaussian modelling framework with the integrated nested Laplace approximation (INLA) for spatio-temporal wildfire forecasting. The first stage uses XGBoost, a gradient boosting model, to identify wildfire patterns from environmental covariates and historical fire records, producing one-month-ahead point forecasts for fire counts and burned area. These predictions are then incorporated as external covariates in a latent Gaussian model, which includes additional spatiotemporal random effects to produce the final probabilistic forecasts of monthly total fire counts and burned area at the council level. To effectively model both moderate and extreme wildfire events, we implement the extended generalised Pareto (eGP) likelihood (a sub-asymptotic distribution) within the INLA framework. We also develop and discuss penalised complexity priors (PC-priors) for the eGP parameters and provide a comprehensive comparison of the eGP likelihood against other commonly employed distributions in environmental modelling, such as the Gamma and Weibull distributions. The proposed framework addresses the challenge of accessing future environmental covariates, which are typically unavailable at prediction time, and demonstrates strong performance in one-month-ahead wildfire forecasting.