Understanding Wildfire Patterns in Italy Through Bayesian Marked Point Process Modeling

Wildfires exhibit pronounced spatial and temporal heterogeneity, and the processes governing their occurrence and burned-area severity may respond differently to climatic, environmental, and topographic conditions. We propose a Bayesian marked spatio-temporal point process framework for analyzing wildfire occurrence and burned-area severity while allowing distinct associations with potential drivers. The study is based on 20,154 large wildfires recorded in Italy during 2007–2023. Wildfire occurrence is modeled through a log-Gaussian Cox process, whereas burned-area severity is modeled conditionally on the observed events through a Gaussian model for log-burned area. Both components include covariate effects, nonlinear elevation effects, temporal effects, and latent spatial Gaussian fields. The latter are represented as Matérn Gaussian fields and approximated using the SPDE approach, while Bayesian inference is performed with INLA through the inlabru package. The results highlight distinct but complementary patterns for occurrence and severity: NDVI is associated with both components, whereas latitude and Fire Weather Index are mainly associated with wildfire occurrence.