Title: Predicting Severe Wildfire Occurrence in Canada

Abstract:

About 8000 wildfires occur in the protected area of Canada each year. Approximately 2% of these fires exceed 100+ ha in size, but account for most of the suppression costs and are the greatest threat to our communities. Although statistical approaches to fire occurrence prediction (FOP) have evolved over the past 40 years, FOP has not yet been implemented at a national scale in Canada. We develop a big data based statistical modeling approach applying Lasso-logistic regression and supervised machine learning methods (e.g. random forests) to a set of spatially gridded meteorological, topographic and demographic covariates to predict severe large wildfire occurrences in Canada one and two weeks ahead. Case control sampling was used to tackle the class-imbalance problem inherent to rare events problems. Both LASSO logistic and random forest models allowed for the inclusion and selection of a large number of covariates with useful predictive skill.  Here we present model performance results for the LASSO-logistic approach to spatio-temporally predict number of: i) human- and lightning-caused ignitions, and ii) large wildfires (100+ ha) in British Columbia. We anticipate that the implemented models will better facilitate agency preparedness as well as tactical decisions regarding resource allocation and sharing between fire management agencies in Canada.  However, predicting surges in ignitions following large lightning storms remains challenging and an area for future work.