Application of particle filters to regional-scale wildfire spread

This paper demonstrates the capability of particle filters for sequentially improving the simulation and forecast of wildfire propagation as new fire front observations become available. Particle filters, also called Sequential Monte Carlo (SMC) methods, fit into the domain of inverse modeling procedures, where measurements are incorporated (assimilated) into a computational model so as to formulate some feedback information on the uncertain model state variables and/or parameters, through representations of their probability density functions (PDF). Based on a simple sampling importance distribution and resampling techniques, particle filters combine Monte Carlo samplings with sequential Bayesian filtering problems. This study compares the performance of the Sampling Importance Resampling (SIR) and of the Auxiliary Sampling Importance Resampling (ASIR) filters for the sequential estimation of a progress variable and of vegetation parameters of the Rate Of fire Spread (ROS) model, which are all treated as state variables. They are applied to a real-world case corresponding to a reduced-scale controlled grassland fire experiment for validation; results indicate that both the SIR and the ASIR filters are able to accurately track the observed fire fronts, with a moderate computational cost. Particle filters show, therefore, their good ability to predict the propagation of controlled fires and to significantly increase fire simulation accuracy. While still at an early stage of development, this data-driven strategy is quite promising for regional-scale wildfire spread forecasting.