A Machine Learning Model for Predicting Wildland Surface Fire Spread According to Rothemel’s Equations ^†

Our ability to predict wildland surface fire behaviour is of great significance for planning and optimising risk-mitigation and fire-suppression strategies. This applies especially for Euro-Mediterranean countries, where an intensified fire-related risk is expected in the decades to come due to both climatic and anthropogenic factors [1,2,3]. In the midst of a broad variety of existing fire spread models, Rothermel’s quasi-empirical mathematical model is one of the most extensively employed, even though the impact of the drivers of fire spread on the model evaluation and uncertainty is not fully understood [4,5,6].

This study aims to contribute to the knowledge required to improve the applicability of Rothermel’s model for operational purposes. A synthetic dataset of the drivers of fire spread has been generated by randomly sampling realistic combinations of parameters in a pilot area, namely Sardinia, Italy, and computing the rate of fire spread for each of those combinations by means of Rothermel’s equations. A machine learning model has been trained and validated on the synthetic dataset to perform a feature importance analysis aimed at enhancing the comprehension of the interdependence of the drivers of fire spread and understanding their impact on each model prediction.

The outcomes suggest that the horizontal wind fluxes and the characterisation of the existing vegetation in terms of fuel models produce an individual average impact on the evaluation of the rate of spread that is more than twice as high with respect to any other driver of fire spread. Moreover, the results demonstrate that the developed model operates analogously to an optimised implementation of Rothermel’s model with the advantage of a reduction of up to 90% of the computing time, thus making it particularly useful for operational applications.