Fire spread predictions: Sweeping uncertainty under the rug.

Predicting fire spread and behavior correctly is crucial to minimize the dramatic consequences of wildfires. However, our capability of accurately predicting fire spread is still very limited, undermining the utility of such simulations to support decision-making. Improving fire spread predictions for fire management purposes, by using higher quality input data or enhanced models, can be expensive, unfeasible or even impossible. Fire managers would benefit from fast and inexpensive ways of improving their decision-making. In the present work, we focus on i) understanding if fire spread predictions can be improved through model parameter calibration based on information collected from a set of large historical wildfires in Portugal; and ii) understanding to what extent decreasing parametric uncertainty can counterbalance the impact of input data uncertainty. Our results obtained with the Fire Area Simulator (FARSITE) modeling system show that fire spread predictions can be continuously improved by 'learning' from past wildfires. The uncertainty contained in the major input variables (wind speed and direction, ignition location and fuel models) can be 'swept under the rug' through the use of more appropriate parameter sets. The proposed framework has a large potential to improve future fire spread predictions, increasing their reliability and usefulness to support fire management and decision making processes, thus potentially reducing the negative impacts of wildfires.

Deciphering the impact of uncertainty on the accuracy of large wildfire spread simulations.

Predicting wildfire spread is a challenging task fraught with uncertainties. 'Perfect' predictions are unfeasible since uncertainties will always be present. Improving fire spread predictions is important to reduce its negative environmental impacts. Here, we propose to understand, characterize, and quantify the impact of uncertainty in the accuracy of fire spread predictions for very large wildfires. We frame this work from the perspective of the major problems commonly faced by fire model users, namely the necessity of accounting for uncertainty in input data to produce reliable and useful fire spread predictions. Uncertainty in input variables was propagated throughout the modeling framework and its impact was evaluated by estimating the spatial discrepancy between simulated and satellite-observed fire progression data, for eight very large wildfires in Portugal. Results showed that uncertainties in wind speed and direction, fuel model assignment and typology, location and timing of ignitions, had a major impact on prediction accuracy. We argue that uncertainties in these variables should be integrated in future fire spread simulation approaches, and provide the necessary data for any fire model user to do so.