Novel method for a posteriori uncertainty quantification in wildland fire spread simulation

Abstract

Simulation is used to predict wildland fire spread in real-time. Nevertheless, the large uncertainties in these simulations must be quantified in order to provide better information to fire managers. Ensemble forecasts are usually applied for this purpose, with an input parameter distribution that is defined based on expert knowledge. We propose a novel approach to generate calibrated ensembles whose input distribution is defined by a posterior PDF with a pseudo-likelihood function that involves the Wasserstein distance between simulated and observed burned surfaces of several fire cases. Due to the high dimension and the computational requirements of the pseudo-likelihood function, a Gaussian process emulator is built to obtain a sample of the calibrated input distribution with a MCMC algorithm in about one day of computation on 8 computing cores. The calibrated ensembles lead to better overall accuracy than the uncalibrated ensembles. The a posteriori probability distribution of the inputs favors lower values of rate of spread and lower uncertainty on wind direction. This strongly limits overprediction, while keeping the ability of the ensemble to cover the observed burned area.

Emulation of wildland fire spread simulation using deep learning

“…Some intervals follow those of a previous study that focused on uncertainty quantification (see notably Table 1 in Allaire, Mallet, & Filippi, 2021).…”

“…For this case, some reference inputs are defined from weather predictions and a presumed ignition point is identified, as explained in (Allaire, Filippi, & Mallet, 2020). Then, an ensemble of perturbed simulations is generated, where the inputs presented in Table follow a calibrated distribution that was obtained in a previous study (Allaire et al, 2021) with β = 1/2. It should be noted that the resulting ensemble of burned surface areas in the present study is not the same as in (Allaire et al, 2021) because supplementary inputs were variable in the previous study (such as perturbations in the times of fire start and fire end, which could make the simulated fire duration different from one hour).…”

Characteristics and Mechanism of Fire Spread between Full-Scale Wooden Houses from Internal Fires

“…Due to China's long history and ethnic diversity, ancient wooden buildings with regional characteristics are widespread throughout the country [1][2][3]. These buildings have not only residential value but also are integrated into the local folklore, culture, history, economy, and natural scenery and have both historical and touristic value [4][5][6].…”

From the probability density function of the rate of spread to that of the corresponding burned area

“…As a consequence of this complexity, the evaluation of fire spreading predictions relies on scoring methods (Filippi 2013, Filippi 2014. Because of such unpredictability, in analogy with the weather forecast, ensemble forecasting has been formulated for wild fires, see, e.g., (Finney 2011, Allaire 2021, together with the application of data assimilation procedure, see, e.g., (Mandel 2008, Rochoux 2014, Rochoux 2015. The idea of ensemble forecasting is based on the concept of stochastic dynamics that describes the evolution of the probability of certain observables.…”

“…Uncertainties in prediction by fire simulators are statistically related to the simulators' dependence on the required parameters, such that sensitivity analysis is considered for improving the reliability of predictions (Trucchia 2019, Asensio 2020) and also on the regional-scale weather prediction. Thus, such uncertainties are covered by probabilistic predictions in terms of ensemble forecasts through a proper distribution of input parameters (Allaire 2018, Allaire 2020, Allaire 2021.…”