Towards a Data-Driven Simulation of Wildfire Spread: a Data Assimilation Methodology for Parameter Calibration

Despite recent progress in wildfire spread modeling, operational models are far from being predictive. Due to the fire complexity and computational requirements, they cannot integrate all the multi-scale multi-physics processes involved in a wildfire. Instead, they are mainly based on a parameterization of the Rate Of Spread (ROS) in terms of input data that characterize the vegetation, the wind conditions and the terrain topography. The models rely on parameters that are mostly fitted using laboratory-scale data; they have therefore a limited domain of applications and are subject to significant uncertainties. A promising approach to overcome the difficulties found in wildfire spread simulations is to integrate fire modeling and fire sensing technologies: recent progress made in remote sensing technology provides new ways to monitor the fire front position, which can then be incorporated into a data assimilation system. The purpose of this study is to provide a proof-of-concept that modeling uncertainties can be reduced using a data assimilation methodology. We use a classical data assimilation algorithm (Best Linear Unbiased Estimator) with synthetically-generated measurements to estimate parameters of a simplified fire spread model. For parameter calibration, the observation operator is non-linear as it includes the model integration; its Jacobian is calculated from a numerical approximation since it is not available analytically. In this study, the fire spread model mimics the real evolution of a wildfire at a regional scale (e.g. over square-kilometer areas). The propagation of the front was simulated in a two-dimensional domain, using a Level-Set method in which the local ROS is the main physical quantity. A parameterization of the ROS was developed as a function of a reduced number of dominant factors characterizing vegetation heterogeneities. The local ROS was adjusted using a 2-parameter correction, these parameters representing the vegetation layer depth and the moisture content. The data assimilation prototype was applied with two types of observations: field observations of the burnt/unburnt status at fixed spatial locations; or direct tracking of the position of the fire front. Both assimilations provided an improved estimate of the ROS, resulting in front positions closer to the true trajectory. However, tests showed that front observations are more powerful than field observations. More representative of the fire physics, the front observations have a higher contribution to the assimilation and they are able to better handle the model non-linearities, even for a high perturbation in the control parameters. As a conclusion, it was shown that integrating airborne-like observations of front positions into a fire spread model is a promising approach to obtain more accurate predictions of wildfire propagation. A validation study using experimental data is the focus of ongoing research.