A Fast Fire Rate of Spread Model Leveraging Machine Learning

Previous studies have tested machine learning (ML) methods for enhancing wildfire science and management. However, none of these approaches attempted to build the ML rate of spread model directly based on the observational data. They either focused on optimizing the existing Rothermel model (not addressing its fundamental limitations) or produced new fire locations without building any universal relationship between the rate of spread and input parameters. Here, we present a novel approach leveraging the machine learning support vector machines (SVM) method to process historical observational fire data, and to provide a structured training data set suitable for training of random forest (RF) machine learning model. The method leverages the continuous fire progression reconstructed by the SVM model based on the satellite fire detections, ground detections, fire perimeters, and origin times and locations. It encodes historical fire progression as the fire arrival map and reconstructs the observed rate of spread using a numerical approximation of the Eikonal equation. Then, simulations of a coupled atmosphere-fire model, WRF-SFIRE, with previously reconstructed fire progression are conducted to reconstruct weather variables affecting the rate of spread. After that, these weather variables, plus several static input parameters (fuel properties, topography, etc.), are used to train a supervised machine-learning (RF) model to predict the rate of spread. The described ML model was successfully applied to the training and testing datasets, and the simulated rate of spread showed good performance when compared to the observed rate of fire spread. Additionally, the ML model enabled the assessment of the relevance of the input variables, which indicated the friction velocity, air temperature, ground moisture flux, and boundary layer height as the critical predictors of the fire rate of spread.