CMStatistics 2020: Start Registration
View Submission - CMStatistics
Title: Spatiotemporal wildfire modeling through point processes with extreme marks Authors:  Jonathan Koh - EPFL (Switzerland) [presenting]
Thomas Opitz - BioSP, INRA (France)
Abstract: Accurate modeling of wildfires is essential to gain a better understanding of the mechanisms driving fire-prone ecosystems and to improve risk management. Here, we consider daily summer wildfire records in the French Mediterranean basin during the period 1995--2018. We jointly model the occurrence intensity and the wildfire sizes by combining extreme-value theory and point process tools within a Bayesian hierarchical modelling framework. In the occurrence component, the wildfire ignition locations and times are modelled as a spatiotemporal point pattern generated by a log-Gaussian Cox process. We use the concept of thinning a point process to model the points associated with extreme fires exceeding a high threshold of burnt area. For the size component, we consider the burnt areas as numerical marks for the points and define two subcomponents to model extreme and non-extreme fires, respectively. We capture non-linear relationships between important covariates, such as weather conditions and forest cover, and the different aspects of fire risk, by incorporating component-specific smooth functions that capture seasonal variation. To reveal common effects driving different aspects of wildfire activity, we share latent effects between different components and highlight how this improves interpretability, parsimony and prediction. To achieve tractable inference in our setting with millions of observations, we propose a stratified subsampling scheme that limits information loss.