Prediction of regional wildfire activity with a probabilistic Bayesian framework

O_LIModelling wildfire activity is crucial for informing science-based risk management and understanding fire-prone ecosystem functioning worldwide. Models also help to disentangle the relative roles of different factors, to understand wildfire predictability or to provide insights into specific events. C_LIO_LIHere, we develop a two-component Bayesian hierarchically-structured probabilistic model of daily fire activity, which are modelled as the outcome of a marked point process in which individual fires are the points (occurrence component) and the fire sizes are the marks (size component). The space-time Poisson model for occurrence is adjusted to gridded fire counts using the integrated nested Laplace approximation (INLA) combined with the Stochastic Partial Differential Equation (SPDE) approach. The size model is based on piecewise-estimated Pareto and Generalized-Pareto distributions, also adjusted with INLA. The Fire Weather Index (FWI) and Forest Area are the main explanatory variables. Seasonal and spatial residuals as well as a post-2003 effect are included to improve the consistency of the relationship between climate and fire occurrence, in accordance with parsimonious criteria. C_LIO_LIA set of 1000 simulations of the posterior model of fire activity is evaluated at various temporal and spatial scales in Mediterranean France. The number of escaped fires ([≥]1ha) across the region can be coarsely reproduced at the daily scale, and is more accurately predicted on a weekly basis or longer. The regional weekly total number of larger fires (10 to 100 ha) can be predicted as well, but the accuracy decays with size, as the model uncertainty increases with event rareness. Local predictions of fire numbers or burnt areas likewise require a longer aggregation period to maintain model accuracy. C_LIO_LIRegarding the year 2003 -which was characterized by an extreme burnt area in France associated with a heat wave-, the estimation of the number of escaped fires was consistent with observations, but the model systematically underrepresents larger fires and burnt areas, which suggests that the FWI does not consistently rate the danger of large fire occurrence during heat waves. C_LIO_LIOur study sheds new light on the stochastic processes underlying fire hazard, and is promising for predicting and projecting future fire hazard in the context of climate change. C_LI