Statistical Prediction of Wildfires in the U.S.: A Baseline Study

  

Brian Bonnlander

 

Although the fire science community has long assumed a connection between weather conditions and the probability of a wildfire occurring, this study appears to be the first to assess the extent to which statistical models can predict fire occurrence from weather measurements or fire hazard models. Prediction data are derived from two large, previously independent databases maintained by the U.S. Forest Service. Techniques we explored include logistic regression, classification trees, and three different neural network approaches. Results show that using weather data alone can predict fires in a binary choice problem (with chance = 50% accuracy) with a 66.4% accuracy for large, naturally occurring fires, and 64% accuracy for large fires created by humans or natural events. Predictive ability drops when the dataset includes both large and small fires, and it also drops when inputs are derived from fire hazard models instead of direct weather observations. There is little variation in predictive performance among the various classification techniques, which suggests that prediction using ground-based data sources is inherently difficult; time dependence in the data appears to contribute to the difficulty of the problem.