*4.1. Important Predictors of Very-Large Fires*

Wildfire events are associated with a number of factors (Flannigan et al., 2009 [28]) that may vary in space (Stavros et al., 2014 [56], Barbero et al., 2015 [57], Arpaci et al., 2013 [58], Flannigan et al., 2006 [59]), and may reveal themselves only under certain conditions (Slocum et al., 2010 [60], Krueger et al., 2015 [61]); it should not then be unexpected that model variability can often be high. Attempts to identify any single factor as most closely associated with VLFs are frustrated by the complex behavior of wildfires, competition among models, data limitations, and diversity of performance criterion. Despite the ubiquity of structural and other uncertainties, the relative importance of various coarse scale meteorological factors to specific wildfire activities could be gauged by observing the frequency with which they were utilized to make predictions. In some cases, a meteorological variable could, with high confidence, be readily identified as important to predicting VLFs in a particular region. In the Temperate Desert division, seasonality was frequently utilized in PETs for both wildfire probabilities, and was also often identified as the optimal splitting criterion. More typically, however, some level of structural uncertainty was present and identifying a best predictor was not always as obvious. In the Subtropical division, LF forest, temperature and precipitation based variables were identified as the optimal splitting criterion with nearly equal frequency. In the Mediterranean Regime Mountains division, seasonality was frequently the optimal splitting criterion in the LF forest, but it was much less common in the VLF forest. Moreover, in the Mediterranean division, wind-based metrics were frequently utilized in LF forests in the Mediterranean forest, but not as a first-split in the PETs. Model variability could be particularly high in the LF forests in the Eastern Continental United States. Precipitation based variables were overwhelmingly preferred in extreme southern Florida and in the Appalachians, but wind based variables were preferred in the Hot Continental division; Temperature was slightly preferred in the Warm Continental division, and as already mentioned, the Subtropical region showed no strong preference with regard splitting criterion. Although some regions showed preferences for certain weather variables, model variability was fairly high in the VLF forests.

Although these structural uncertainties are sometimes obstacles to identifying important meteorological relationships with VLFs, they are also critical to understanding the true level of confidence we have in observed correlations and safeguard against overconfident conclusions. While clearly notable levels of model variability could be encountered across multiple factors, robust patterns and trends could still be inferred. For instance, we note that, in most of the West, with the exception of the Great Plains and the Tropical/Subtropical portions of the Southwest, temperature based metrics were often the best predictor of LFs and were commonly used in LF forests. In the remaining Western areas, temperature metrics were less useful and instead precipitation metrics were selected as the optimal splitting criterion. This apparent preference for precipitation based metrics over temperature based ones in these regions may be related to the characteristics of fuel-limited versus climate-limited fire regimes (Meyn et al., 2007 [32]), or due to a relative inability of seasonal temperature fluctuations to match wildfire activity compared to precipitation. The relative popularity of wind-based variables in very-large forests compared to very-large forests is also interesting, as wind

has been reported to have variable influence on wildfire activity depending on fire size and geographic location (Slocum et al., 2010 [60]).
