Predicting Wildfire Ember Hot-Spots on Gable Roofs via Deep Learning

Abstract

Ember accumulation on and around homes can lead to spot fires and home ignition. Post wildland fire assessments suggest that this mechanism is one of the leading causes of home destruction in wildland urban interface (WUI) fires. However, the process of ember deposition and accumulation on and around houses remains poorly understood. Herein, we develop a deep learning (DL) model to analyze data from a series of ember-related wind tunnel experiments for a range of wind conditions and roof slopes. The developed model is designed to identify building roof regions where embers will remain in contact with the rooftop. Our results show that the DL model is capable of accurately predicting the position and fraction of the roof on which embers remain in place as a function of the wind speed, wind direction, roof slope, and location on the windward and leeward faces of the rooftop. The DL model was augmented with explainable AI (XAI) measures to examine the extent of the influence of these parameters on the rooftop ember coverage and potential ignition.