A Machine Learning Approach to Forecasting California Wildfire Risk

Abstract:

Large wildfires in the western US are becoming more frequent and intense, but predicting their occurrence and intensity remains a challenge due to their spatio-temporal complexity and inter-dependence on coupled climatic and human factors. There is an increasing opportunity leverage remote-sensing data products to understand wildfire risk and to develop data-driven models for fire occurrence, spread and dependence on climate change. In this work we develop a random forest model for the occurrence of large wildfires given antecedent meteorological and vegetation parameters, trained on a remote-sensing dataset of observed historical wildfires from 2002-2016 and use it to predict wildfire risk in California. We show that the model predictions are consistent with historical large fire occurrences, outperforming existing statistical measures. We also demonstrate model interpretability by decomposing model weights of the individual physical drivers of fire risk. The results suggest both the potential for use in early warning and detection and also for understanding changes in fire risk in a changing climate.