Learning-based prediction of wildfire spread with real-time rate of spread measurement

Abstract

A learning-based wildfire spread model was developed in this study to predict short-term wildfire spread. Real-time rate of spread (RoS) measurement was first conducted by calculating normal movements of fire fronts. Subsequently, machine learning was employed to correlate the local RoS and environmental parameters and predict the RoS in the unburnt area. After that, a narrow-band level-set method was utilized to simulate the evolution of fire front. RoS measurement, machine learning, and level-set method were individually verified with numerically generated fire fronts, and applied in a real scale shrubland fire scenario. Results show that the proposed learning-based method is capable of predicting short-term fire spread without employing an empirical RoS model, which is beneficial for modeling spreading of a real wildfire.