Modeling Wildfire Ignition Distribution and Making Prediction of Human-caused Wildfire

This paper proposes a further exploration of machine learning algorithms within the context of modelling the spatial distribution patterns of the human-caused wildfires over a Southern California landscape. In this research, the wildfire distribution problem is defined as a Binary Classification task conducted on a cellular lattice overlay the study area. A fifteen-year historical wildfire occurrence data as target variable was used, along with eight independent variables derived from anthropogenic factors such as distance measure to road-network and Wildland-Urban Interface. Meteorological factors such as temperature and humidity have also been used in the model training process. Both of the two machine learning algorithms, the Conditional Inference Tree and the Random Forest methods, combining with the Synthetic Minority Over-Sampling Technique, demonstrate a significant improvement over traditional method. And the predicted result shows that the location with high proximity with WUI and road tend to be more vulnerable towards wildfire incidence.