Here, we propose a new approach leveraging an ML method based on Support Vector Machines (SVM) that has been used to generate a dataset of fire progression data for many past wildfire events. The method can integrate information from satellite fire and ground detections, infrared airborne fire perimeters, and time and location of ignitions to create a continuous fire progression. Then, the estimated fire progression is used to derive the rate of spread at each location using a numerical approximation of the Eikonal equation and to drive constrained WRF-SFIRE simulations to obtain multiple weather features that can be used to build a relationship to the rate of spread. The features obtained can be divided into static (elevation, slope, fuel type), wind, temperature, solar radiation, moisture, precipitation, pressure, and planetary boundary layer height. They are interpolated spatially and temporally to the rate of spread locations and times into a final tabular dataset.
After that, exploratory data analysis and feature selection are performed to find the best features driving the RoS. Using the best combination of features, different ML models are tuned, trained, and compared to select the best-performing one using different metrics. The final ML-based RoS model shows promising results and is integrated into WRF-SFIRE. The integration is done by storing the model in a data file and loading it at the initialization stage in WRF-SFIRE. Then, the prediction is performed inside WRF-SFIRE at every pixel and time step, similar to the other included RoS models. Simulations using the new ML-based RoS model are compared to other RoS models in the coupled atmosphere-fire model, WRF-SFIRE.

