Improving Wildfire Predictability via Machine Learning

Wildfires become high cost and high disruption events, and cannot be ignored for much longer. The fundamental objective for the work is to understand the wildfire phenomena better by using the state of the art technology in machine learning to harvest patterns from "Big Data" pools. The study employs publicly available reanalysis data, human infrastructure data, lightning patterns, climatic indices, and NASA's MODIS satellite images to capture continuously any changes that might affect wildfire risks and severity. A range of machine learning models are developed - CARTs, Random Forests, CNNs - to enable short term (1 month) and long-term (6 months) predictions, and to estimate year-to-year change in wildfire risks. Moreover, the study estimates predictor significance for different forecast time horizons. The findings are shown for the case study of North America at 10 km resolution.