Title : From spread to suppression: A computational wildfire model with machine learning guided suppressant optimization
Abstract:
As climate change and urbanization continue to intensify, an estimated 115 million individuals face heightened risks from wildfires. Existing operational models such as Rothermel’s, however, are constrained to rate-of-spread calculations and cannot reproduce full wildfire propagation patterns or burn area. To overcome these challenges, this project introduces a three-dimensional cellular automaton (CA) wildfire model that simulates fire spread through terrain interaction. This framework includes three main components: (1) landscape generation from satellite imagery using k-means clustering and elevation data; (2) a probabilistic CA that models heat radiation transfer, terrain effects, and ember transport; and (3) a Bayesian optimization model that identifies high risk zones to strategically place fire suppressant using a Gaussian Process (GP) surrogate to minimize computationally expensive simulations. Applied to the 2020 Los Angeles Bobcat Fire, the model closely reproduces mid-stage and final burn areas, achieving 93.2% and 89.9% accuracy respectively, with additional metrics such as precision, recall, F1-score, and IoU demonstrating further predictive strength. Then, Bayesian optimization was used to pinpoint suppressant placements, effectively reducing burn area 26.6% without exhaustive search. Overall, this integrated approach offers a powerful tool for anticipating wildfire behavior and optimizing real-time response strategies as fire risks continue to rise.
