Towards predictive data-driven simulations of wildfire spread – Part I: Reduced-cost Ensemble Kalman Filter based on a Polynomial Chaos surrogate model for parameter estimation

Abstract

Abstract. This paper is the first part in a series of two articles and presents a data-driven wildfire simulator for forecasting wildfire spread scenarios, at a reduced computational cost that is consistent with operational systems. The prototype simulator features the following components: an Eulerian front propagation solver FIREFLY that adopts a regionalscale modeling viewpoint, treats wildfires as surface propagating fronts, and uses a description of the local rate of fire spread (ROS) as a function of environmental conditions based on Rothermel's model; a series of airborne-like observations of the fire front positions; and a data assimilation (DA) algorithm based on an ensemble Kalman filter (EnKF) for parameter estimation. This stochastic algorithm partly accounts for the nonlinearities between the input parameters of the semi-empirical ROS model and the fire front position, and is sequentially applied to provide a spatially uniform correction to wind and biomass fuel parameters as observations become available. A wildfire spread simulator combined with an ensemble-based DA algorithm is therefore a promising approach to reduce uncertainties in the forecast position of the fire front and to introduce a paradigm-shift in the wildfire emergency response. In order to reduce the computational cost of the EnKF algorithm, a surrogate model based on a polynomial chaos (PC) expansion is used in place of the forward model FIREFLY in the resulting hybrid PCEnKF algorithm. The performance of EnKF and PC-EnKF is assessed on synthetically generated simple configurations of fire spread to provide valuable information and insight on the benefits of the PC-EnKF approach, as well as on a controlled grassland fire experiment. The results indicate that the proposed PC-EnKF algorithm features similar performance to the standard EnKF algorithm, but at a much reduced computational cost. In particular, the re-analysis and forecast skills of DA strongly relate to the spatial and temporal variability of the errors in the ROS model parameters.

A comparative study of parameter estimation and state estimation approaches in data-driven wildfire spread modeling

“…Thus, the EnKF algorithm is able to handle model non-linearities. In the PE approach, the control vector includes the n input parameters of the Rothermel-based ROS model that are subject to uncertainties and to which FIREFLY is sensitive (Rochoux et al 2014a). In the SE approach, the control vector is made of the coordinates of the Nfr front markers along the simulated fire front such that n = 2Nfr (Rochoux et al 2014b).…”

“…Thus, both PE and SE approaches are able to retrieve an accurate estimation of the fire front location, even though the prior information is subject to high levels of uncertainties and the terrain topography is complex. Note that as for the real case presented in Figure 7, the SE approach is more effective at retrieving the topology of the front at the analysis time than the PE approach; the latter provides valuable information on the input parameters to forecast the front at future lead-times (see Rochoux et al 2014aRochoux et al , 2014b for further details). …”

“…The key idea is that, when used alone, neither measurements nor computer models can provide a reliable and complete description of the real state of the physical system. The present study is an extension of our previous work in Rochoux et al (2013Rochoux et al ( , 2014aRochoux et al ( , 2014b and Rochoux (2014), in which a prototype data-driven wildfire spread simulator capable of forecasting the wildfire spread behavior was developed. The prototype simulator features the following main components: a level-set-based fire propagation solver (Rehm and McDermott 2009) that treats wildfires as propagating fronts and uses a description of the local rate of fire spread (ROS) as a function of vegetation, topographical and meteorological conditions based on the model due to Rothermel (1972); a series of observations of the fire front location; and a DA algorithm based on the ensemble Kalman filter (EnKF) to account for some of the non-linearities in the mapping between control space and observation space.…”

“…The prototype simulator features the following main components: a level-set-based fire propagation solver (Rehm and McDermott 2009) that treats wildfires as propagating fronts and uses a description of the local rate of fire spread (ROS) as a function of vegetation, topographical and meteorological conditions based on the model due to Rothermel (1972); a series of observations of the fire front location; and a DA algorithm based on the ensemble Kalman filter (EnKF) to account for some of the non-linearities in the mapping between control space and observation space. The DA algorithm also features a choice between a parameter estimation (PE) approach in which the control variables are the input parameters of the ROS model (Rochoux et al 2014a) and a state estimation (SE) approach in which the control variables are the spatial coordinates of the fire front (Rochoux et al 2014b). This study presents the extension of the FIREFLY-based datadriven wildfire spread simulator to problems with complex terrain topography; the formulation of the PE-/SE-based DA strategies is not affected by the increasing complexity of the wildfire spread model.…”

Short-term fire front spread prediction using inverse modelling and airborne infrared images

“…Their work showed promising results while raising some concerns about spurious fire corrections and the computing time required. Following their idea, Rochoux et al (2014aRochoux et al ( , 2014bRochoux et al ( , 2015 explored a data-driven wildfire simulator based on parameter and state estimation that assimilates fire front positions and corrects the wildfire forecast by means of a level set model based on Rothermel's. They explored a parameter and state estimation strategy with stochastically based estimation of the error covariance matrices.…”