Prediction of regional wildfire activity in the probabilistic Bayesian framework of Firelihood

Modeling wildfire activity is crucial for informing science-based risk management and understanding the spatiotemporal dynamics of fire-prone ecosystems worldwide. Models help disentangle relative influences different factors, understand predictability, provide insights into specific events. Here, we develop Firelihood, a two-component, Bayesian, hierarchically structured, probabilistic model daily fire activity, which modeled as outcome marked point process: individual fires are points (occurrence component), sizes marks (size component). The space-time Poisson occurrence adjusted to gridded counts using integrated nested Laplace approximation (INLA) combined with stochastic partial differential equation (SPDE) approach. size based on piecewise-estimated Pareto generalized distributions, INLA. Fire Weather Index (FWI) forest area main explanatory variables. Temporal spatial residuals included improve consistency relationship between weather occurrence. posterior distribution Bayesian provided 1,000 replications that were compared observations at various temporal scales in Mediterranean France. number larger than 1 ha across region was coarsely reproduced scale, more accurately predicted weekly basis or longer. regional total (10–100 ha) well, but accuracy degraded size, uncertainty increased event rareness. Local predictions numbers burned areas also required longer aggregation period maintain accuracy. estimation consistent during extreme season 2003 unprecedented heat wave, systematically underrepresented large areas, suggests FWI does not consistently rate actual danger waves. Firelihood enabled novel analysis stochasticity underlying hazard, offers variety applications, including hazard projections context climate change.