Extreme wildfires are more becoming more frequent and getting larger, but predicting extreme events at management-relevant scales remains difficult. Here we integrate a 30 year wildfire occurrence record with meteorological and housing data to explain and predict wildfire extremes using a spatiotemporal Bayesian model that allows the drivers of fire dynamics to have nonlinear, spatially-varying effects.
Results/Conclusions
Across the United States, extreme events can be explained primarily by increased fire frequency, rather than dramatic changes in expected wildfire burn area. Dryness, air temperature, precipitation, and human housing density regulate wildfire risk, but these effects strongly depend on location and ecological context. A zero-inflated negative binomial model for fire occurrence achieves 98.9% interval coverage for the number of fires over 1000 acres in a withheld data set over a five-year prediction time horizon. Combining this count model with a log-normal model for the area burned by each fire, we demonstrate that observed extreme wildfires can be anticipated at a spatiotemporal resolution that is useful for wildfire management and disaster planning. We conclude that recent wildfire extremes need not be surprising, and that the chance of much larger million-acre wildfires is alarmingly high.