*4.2. Climate Change and Very-Large Fire Occurrence*

For both RCP scenarios and nearly all divisions, complex changes to wildfire activity are predicted that will result in an overall increase in the frequency of VLFs, which is largely consistent with many other projections (Flannigan et al., 2009 [28], Barbero et al., 2015 [21] ). While overall increases in the frequency of these events are predicted using robust methods, the exact nature of these changes remain unclear. It is not certain, for instance, if the range of fire sizes will remain largely static in the future and only frequency of exceptionally large events will increase; or if the size distribution will shift, so that burn areas exceed historic records. These distinctions are important because the relative costs of these two competing possibilities are likely to vary across decision makers. The Mediterranean division was somewhat of an exception to the overall reported increases in VLF activity. Westerling et al. (2011 [26]) project either no change or modest increases in LF activity in much of lowland California, and large increases in mid- and high-elevation locations, which at first seems inconsistent with the predicted decrease in VLF activity, although there are a few explanations. Firstly, by considering a larger number of climate models and predictive models, the range of results in this analysis will be inherently more variable, and marginal results seen in other studies could emerge as significant when these structural uncertainties are incorporated. Secondly, as shown in this study, the environmental drivers of large and conditional VLF probabilities can vary, and differences regarding the definition of LF can result in variability amongst methodologies (Slocum et al., 2010 [60]). Thirdly, differences between the covariates considered and model structure are likely to alter the predictions across analyses. For instance, anthropogenic and vegetation effects on wildfire activity were omitted in this study, but are known to be an important influence of wildfire activity in California (Syphard et al., 2007 [62]) and elsewhere (Syphard et al., 2017 [63]).

The months in which VLF activity was historically highest may not necessarily apply in the second half of the 21st century, and noticeable changes in intra-annual patterns, usually increases, of VLF activity were predicted in most scenarios and regions. Some regions, like Temperature Desert Regime Mountains and Marine Regime Mountains Redwood Forests, are predicted to have increases in VLF frequency only during a limited portion of the year, while others, like the Subtropical division, are predicted to have a relatively uniform increase in VLF frequency throughout the year. Given that simultaneous increases in VLF probabilities are anticipated in multiple independent regions, it is likely that VLF activity will change in ways that will increase resource strain. Indeed, the results of this study suggest that, depending on the emission scenario, between 12–13 regions will have future VLF frequencies that exceed the historical record, and that intra-annual increases in VLF occurrence are often predicted during the same time of year in spatially distinct regions.

In addition to changes to intra-annual patterns of overall VLF frequency, it is important to acknowledge that the overall increases in VLF frequency are the product of two processes: changes in LF and conditional VLF probabilities. Any increase in VLF frequency is then the result of one of three scenarios: an increase in both probabilities, and increase in LF frequency only, or an increase in the frequency that LFs become VLFs. These specific changes in model components may be of particular relevance to firefighting, public health professionals, and other decision-makers who will—due to differences in the impacts of the events—react to no-fire, LF, and VLF months differently and require guidance regarding the characteristics of the novel future wildfire regimes. Reducing the uncertainty as to which emission scenario the future will resemble should also be a priority for decision-makers and researchers, as the predicted changes tend to be more exaggerated under the RCP 8.5 scenario compared to the RCP 4.5, which should influence adaptation and mitigation efforts of future wildfire impacts.

### *4.3. Caveats and Future Work*

While the simultaneous acknowledgement of structural uncertainties in the climate models and PETs represents an interesting approach, there are still a number of uncertainties that were not addressed in this climate impact analysis. The limited availability of reliable and consistently recorded (e.g., satellite-based) measurements of wildfire activity (Taylor et al., 2013 [64]) and the inherent rarity of VLF events remain significant obstacles to validating predictions and estimating underlying model structures. The validation results should be considered as the current state of knowledge regarding the ensemble's predictive ability, and may change when more data becomes available in the future. If inter-annual variability in wildfire activity is high, then the validation results used in this study may be based on particularly predictable or unpredictable fire years, and therefore not be representative of the actual performance. Longer duration datasets would be preferred, and thirty year climatologies are often considered ideal (Arguez and Vose 2011 [65]), but the entire range of available burn area data only extends 33 years and it is unlikely that longer time scale meteorological associations with VLF activity will be accurately captured with the relative brevity of data (Westerling and Swetnam 2003 [66], Marlon et al., 2012 [67]). Moreover, if recent increases in VLF activity are indicative of a sudden a shift into overall wildfire patterns unlike what has been observed in the past, then forecasting future activities based on historical relationships could be inadequate. For instance, the two events occurring in the Hot Continental Regime Mountains in 2016 were quite unusual in historical terms, as only four VLF months were reported from 1984–2005, and only one VLF was reported from 2006–2015.

Data limitations may also be qualitative, and many of the remaining important structural uncertainties are due to unconsidered covariates, like vegetation changes, suppression effort, and population growth, that were not modeled due to data inavailability, practical considerations, and challenges related to predicting these quantities in the future. While the PETs used in this study produced a diverse suite of predictive models and are known to be highly unstable (Wang et al., 2016 [50]), there are many other lingering sources of structural uncertainty that could still be incorporated. For instance, generalized linear models could be used instead, which take a number of mathematical structures depending on the choice of link and response functions (Clyde 2003 [68]). Similarly, various data transformations could be used to generate competing models of the wildfire activities. Alternative models could be constructed that condense the two model components into VLF occurrence probabilities only, so that the event space of each month is purely binary. Instead of biogeographical classification of regions, the Continental United States could be partitioned using administrative or other boundaries to generate VLF predictions relevant to specific stakeholders. Hence, clearly a broad variety of other structural uncertainties still exist that could potentially influence predictions of future VLF frequency in the second half of the century.

It is important to understand that the VLF probabilities do not inform us as to what will actually happen, but rather communicates the degree of uncertainty about future outcomes conditional on carbon emission scenarios. For this reason, some tolerance to deviations between observed and expected VLF frequencies should be considered, as should the fact that the predictions were based on modeled climate data as opposed to direct observations. Still, in many regions, the ensemble performance was relatively adequate and the simulated distribution of fire counts covered the observations. Moreover, when deviations occurred, they tended to underestimate the future VLF counts. Hence, the overall claim that VLF counts will increase in the future under climate change is supported by the results of this study, as well as through the work of others (Stavros et al., 2014 [56], Barbero et al., 2015 [21]). Stochastic uncertainty will be critical when explicitly linking changes to VLF occurrence to human activities and for assessing the future levels of VLF simultaneity (Tedim et al., 2018 [69]) and is a factor that would be well addressed using the methods described here, but is beyond the scope of this paper. The inherent stochasticity of the PET construction process suggests that repeated applications of this methodology in the future may yield slight variations to the results presented here.

Interestingly, a standard factor analysis revealed that more than 86 percent of the variability in predicted probabilities could be attributed to variance amongst the PETs rather than variance amongst the climate models, and while the PETs are an inherently unstable choice of predictive model, this suggests that structural uncertainties should receive the attention of climate impact researchers in much the same way that the choice of climate model does. Further exploration of these structural uncertainties in climate impact analyses cannot be recommended enough in future analyses, as they inform not only of future impacts, but the reliability of these predictions, which can influence decision-maker behaviors in a variety of ways (Weber and Johnson 2009 [70]).
