Accounting for extremes in modeling the size and likelihood of large fires in the United States

Wildfires pose growing threats to ecosystems, infrastructure, and communities, yet fire size distributions are often modeled without sufficient attention to rare, high-impact events. Most statistical approaches emphasize the central body of the distribution, which can obscure the behavior of extreme wildfires (i.e., tail events) that have the most significant impact. Here, we introduce a framework that explicitly distinguishes between the statistical characteristics of body and tail fire size distributions. We analyzed 30,331 large fire perimeters from the Monitoring Trends in Burn Severity (MTBS) data set (1984–2024), disaggregated across 105 Level III ecoregions and 10 Geographic Area Coordination Centers (GACCs) across the United States. For each region, we fit three candidate distributions (Pareto Type II, lognormal, Weibull) to both the full data set and an exceedance-based tail sample. Results show that models calibrated only to the body can substantially misestimate the exceedance probability (by up to three orders of magnitude) and burned area (by hundreds of thousands of acres) for large fires. For the body of the distribution, the lognormal and Pareto Type II distributions generally outperform the Weibull. For the tail, the picture is more nuanced: the Pareto Type II and lognormal perform comparably at the ecoregion level, while the Weibull, despite being the weakest body model, provides the best tail fit at the GACC scale. These findings highlight the need to model body and tail behavior separately and show that optimal distributions vary across scales, underscoring the importance of region-specific tail models for suppression budgeting, fuel treatment, and resilience planning.