Predicting future extreme wildfire events in the western US

Predictive mapping of extreme wildfire disasters presents a daunting challenge to the field of risk and hazard science.  Wildfire risk to developed areas is the cumulative product of complex interacting factors including spatiotemporal patterns of human and natural ignitions, fire propagation over long distances through heterogeneous fuel types and terrain, extreme weather events, variation in building susceptibility, and suppression capacity.  Specific combinations of these factors ultimately create the scenario for wildland urban interface (WUI) disasters such as the extreme Camp fire event that destroyed the town of Paradise, California  and resulted in 82 fatalities.  Similar disasters have been observed elsewhere, notably the Portugal fires in 2017 that resulted in 61 fatalities, and the Black Saturday disaster in SE Australia  in 2009 resulting in 173 fatalities.

Future disasters of similar or greater magnitude are inevitable given predicted climate change but remain highly uncertain in terms of location and timing. As with other natural disasters, simulation models are one of the primary tools to map risk and design prevention strategies.  However, prior risk assessment methods have focused on estimation of mean values rather than prediction of extreme events that are increasingly becoming a reality in many parts of the globe. 

Using the western US as a study area, we synthesized newer wildfire simulation and building location data (54 million fires, 25 million building locations) and compared the outputs to several sources of observed building exposure data. The simulations used synchronized weather among spatial simulation subunits, thereby providing estimates of extreme fire seasons, fire events within them, and exceedance probabilities at multiple scales.  The study is the first to explore large scale extreme wildfire events in terms of both uncertainty and magnitude, providing a broad foundation of methods to advance wildfire disaster prediction.

We found that:

  • Annual area burned was accurately predicted by simulations but building exposure was substantially overestimated, although small historical sample size might have influenced the comparison. 
  • We identified extreme fire seasons in the simulation record (10,000 fire years) that exceeded historical fire seasons by 278% in terms of area burned, and 1,255% in terms of buildings exposed, under contemporary climate. 
  • The worst observed fire season in terms of area burned was  ~8 million hectares in a year that caused 1782 fires and exposed 76,210 buildings.
  • The largest number of buildings exposed in a single simulated year was 495,843, when 3,964,265 hectares burned in 436 fires.
  • The worst single fire in terms of exposure was reported on VOX News. This simulated fire exposed 106,378 buildings in California in a fire that would have burned 470,578 hectares.
  • The worst historical year in terms of buildings exposed was 39,505 buildings in 2018. The actual number of buildings lost to the Camp fire in 2018 was 19,558.
  • Predicted building exposure was optimized in specific regions along gradients of building density and burnable fuels.  Optimal conditions for building exposure was observed at a fuel density of 0.7 (proportion of the land that has burnable vegetation) per square kilometer and a building density between 1400 and 1500 buildings per square kilometer.

Key findings:

  • Simulation modelling predicts wildfire disasters well beyond observed in the western US,
  • Predicted building exposure was maximized in specific regions along gradients of building density and burnable fuels.  Prior studies have examined this problem of maximum exposure using WUI classes and thus our approach provides a quantitative model of exposure rather than predicting exposure for a relatively small number of discrete classes. 
  • Extreme fire events were added to the Fireshed Registry, an interactive geospatial data portal that provides  access to data describing past, present and future trends regarding wildfire exposure to communities and forest and fuel management. The Registry employs a nested spatial framework that organizes landscape variation in wildfire risk to developed areas into containers or “firesheds” and displays these data on a background of maps on management and disturbance including past and predicted wildfire events and their potential impacts.


We investigated the prediction of extreme wildfire events using fire simulation methods developed at the Fire Lab. These are plausible but rare scenarios in the simulation library that describe yet to be observed fire disasters.  Climate change studies suggest that extreme wildfires will become more frequent in the future.  The study was motivated in part by the tragedy in Paradise, California as well as similar events in Australia and Portugal.  We also compared predicted building exposure with historical data and prior empirical studies to understand potential bias in exposure estimates from simulation.  Lastly, we discussed the idea that assessment of plausible extreme events might be a more effective risk communication tool than green-yellow-red risk maps that display long-term averages.

Additional references:

Academic Times: Wildfire risk maps aren’t capturing the threat of extreme events and seasons

VOX News: This is a worst-possible wildfire scenario for Southern California


Select Publications & Products

Ager, A. A., M. A. Day, F. J. Alcasena, C. R. Evers, K. C. Short, and I. C. Grenfell. 2021. Predicting Paradise: Modeling future wildfire disasters in the western US. Science of the Total Environment 784: 147057. doi: 10.1016/j.scitotenv.2021.147057