Derivation of a Bayesian fire spread model using large-scale wildfire observations

Models that predict wildfire rate of spread (ROS) play an important role in decision-making during firefighting operations, including fire crew placement and timing of community evacuations. Here, we use a large set of remotely sensed wildfire observations, and explanatory data (focusing on weather), to demonstrate a Bayesian probabilistic ROS modelling approach. Our approach has two major advantages: (1) Using actual wildfire observations, instead of controlled fire observations, makes models developed well-suited to wildfire prediction; (2) Bayesian modelling accounts for the complex nature of wildfire spread by explicitly considering uncertainty in the data to produce probabilistic ROS predictions. We show that highly informative probabilistic predictions can be made from a simple Bayesian model containing wind speed, relative humidity and soil moisture. We provide current operational context to our work by calculating predictions from widely used deterministic ROS models in Australia.

Automated summary

This study demonstrates a Bayesian probabilistic ROS modeling approach using actual wildfire observations and explanatory data, providing highly informative probabilistic predictions. Bayesian modeling, by explicitly considering uncertainty in the data, is suitable for capturing the complexity of wildfire spread.