Conference Agenda

Wildfires have become an increasingly recurrent hazard in central Chile, producing severe environmental and socioeconomic impacts and motivating probabilistic, spatially explicit tools for prevention and operational planning. Focusing on the Maule Region, this study estimates wildfire occurrence probability in high-risk zones using a spatiotemporal modeling approach that captures both local environmental conditions and their temporal evolution. We compile and harmonize a gridded dataset combining wildfire occurrence records with dynamic drivers (e.g., meteorological variables and vegetation-related proxies) and static descriptors (e.g., topography), and we formulate risk estimation as a sequence-to-probability task at the grid-cell level. We develop a Transformer-based spatiotemporal architecture that learns nonlinear dependencies across time and space and outputs calibrated probabilities of wildfire occurrence for multiple forecast windows. Performance is assessed on held-out periods using discrimination and calibration analyses, and the predicted probabilities are converted into risk-zonation maps to support consistent operational thresholding. Results show that the proposed model produces coherent probability fields that concentrate risk in historically fire-prone areas under conducive conditions, while delivering stable probabilistic outputs suitable for decision-making. The resulting risk maps provide actionable guidance for prioritizing surveillance and allocating prevention and response resources in the Maule Region, and the framework is readily transferable to other regions where wildfire risk exhibits strong spatiotemporal variability.