High-severity fire: evaluating its key drivers and mapping its probability across western US forests

Wildland fire is a critical process in forests of the western United States (US). Variation in fire behavior, which is heavily influenced by fuel loading, terrain, weather, and vegetation type, leads to heterogeneity in fire severity across landscapes. The relative influence of these factors in driving fire severity, however, is poorly understood. Here, we explore the drivers of high-severity fire for forested ecoregions in the western US over the period 2002–2015. Fire severity was quantified using a satellite-inferred index of severity, the relativized burn ratio. For each ecoregion, we used boosted regression trees to model high-severity fire as a function of live fuel, topography, climate, and fire weather. We found that live fuel, on average, was the most important factor driving high-severity fire among ecoregions (average relative influence = 53.1%) and was the most important factor in 14 of 19 ecoregions. Fire weather was the second most important factor among ecoregions (average relative influence = 22.9%) and was the most important factor in five ecoregions. Climate (13.7%) and topography (10.3%) were less influential. We also predicted the probability of high-severity fire, were a fire to occur, using recent (2016) satellite imagery to characterize live fuel for a subset of ecoregions in which the model skill was deemed acceptable (n = 13). These ‘wall-to-wall’ gridded ecoregional maps provide relevant and up-to-date information for scientists and managers who are tasked with managing fuel and wildland fire. Lastly, we provide an example of the predicted likelihood of high-severity fire under moderate and extreme fire weather before and after fuel reduction treatments, thereby demonstrating how our framework and model predictions can potentially serve as a performance metric for land management agencies tasked with reducing hazardous fuel across large landscapes.


Introduction
Wildland fire is a critical natural disturbance and ecological process in many ecosystems around the globe, particularly in the forested regions of the western US (Agee 1993, Bond et al 2005. Fire affects a wide range of ecosystem components and processes such as post-fire successional trajectories, nutrient cycling, hydrology, and carbon dynamics (Turner 2010, Such concerns are heightened in forests with a legacy of past logging and fire exclusion, where significant shifts in ecosystem composition, structure, and function have triggered fuel conditions at greater risk for high-severity fire (Mallek et al 2013, Hessburg et al 2015.
Over the last decade, our understanding of factors that influence fire severity has improved, but the relative importance of these factors remains unclear. Topography, for example, is clearly an influential factor (Holden et al 2009, Dillon et al 2011, as is the amount and composition of live vegetation and dead fuel (Fang et al 2015, Harris andTaylor 2015). A limited number of studies also indicate that long-term climate (i.e. 30-year climate normals) is an important factor driving fire severity (Parks et al 2014c, Kane et al 2015a, although many suggest that climate likely has an indirect influence via its effect on productivity and dominant vegetation type (e.g. Miller and Urban 1999a, Pausas and Bradstock 2007, Krawchuk et al 2009. Surprisingly, empirical evidence is extremely varied pertaining to the importance of fire weather as a driver of high-severity fire; some studies have shown that its influence is marginal (Fang et al 2015, Birch et al 2015 whereas others have concluded it is a highly influential factor (Keyser andWesterling 2017, Lydersen et al 2017). Other factors that influence fuel, such as vegetation management activities (Thompson et al 2007, Prichard andKennedy 2014) and the presence of previous wildland fire (Parks et al 2014b, Stevens-Rumann et al 2016, have also been shown to influence fire severity. Research to date pertaining to the key drivers of high-severity fire has been either comprehensive in ecological scope but geographically limited, or geographically broad but lacking important environmental components. Dillon et al (2011) conducted perhaps the most comprehensive evaluation (in terms of geographic scope and number of fires) of the drivers of high-severity fire using data from three ecoregions in the northwestern US and three in the southwestern US (∼1500 total fires). Dillon et al (2011), however, did not evaluate some of the factors likely responsible for high-severity fire such as fuel, thereby making it difficult to interpret their findings from an ecological perspective. Keyser and Westerling (2017) also conducted a comprehensive evaluation of fire severity in the western US, but their unit of analysis was coarser-at the individual fire (i.e. fires were categorized as either 'high-severity' or 'other'). Conversely, most studies to date (and this study) evaluated pixels within individual fires as the unit of analysis, thereby preserving and analyzing intra-fire variability.
Some studies have evaluated a more inclusive set of environmental drivers but were often conducted at disparate temporal and spatial scales, ranging from those of individual fires (Thompson et al 2007, Harris andTaylor 2015) to landscapes with ∼50-100 fires (Fang et al 2015, Birch et al 2015, thereby making broader-scale generalizations challenging. Differences in methodology among these studies also complicate interpretation. An evaluation using consistent data and methods across the broad geographic range of forested landscapes of the western US will allow for an improved understanding of the most influential factors driving fire severity and will provide forest managers with highly relevant information for planning and mitigation purposes. In this study, we assessed a comprehensive suite of potential drivers of high-severity fire using a consistent, repeatable approach that was not only geographically extensive but also predictive in nature. We built a statistical model describing high-severity fire for each ecoregion in the contiguous western United States (hereafter western US) with the exception of those with insufficient data (e.g. Sonoran Desert was excluded; see Methods). We defined high-severity fire as those that are stand-replacing as inferred by the relativized burn ratio (RBR) (Parks et al 2014a), a gridded satellitebased fire severity metric. Our evaluation included explanatory variables representing live fuel, topography, climate, and fire weather. The models we developed have the potential to support fire and fuel management (cf Hessburg et al 2007) because several of the explanatory variables are dynamic (i.e. varying on daily to annual time scales), such as those representing live fuel and daily fire weather. Consequently, raster maps representing predictions of high-severity fire (cf Holden et al 2009) can be updated annually and under different weather scenarios to assess, for example, the potential for high-severity fire in an upcoming fire season. Such products may facilitate the development of more adaptive strategies for addressing the contemporary challenges of wildland fire management. Similarly, model predictions have the potential to monitor and quantify potential changes in the probability of high-severity fire resulting from management actions, such as fuel reduction treatments.
Our overarching objectives were three-fold. First, we aimed to identify the most influential factors driving high-severity fire for each ecoregion in the western US. Second, we designed a quantitative framework such that the model predictions for each ecoregion can be updated annually using recent (e.g. 2016) satellite imagery and implemented to evaluate the probability of high-severity fire (were a fire to occur) under a range of potential weather scenarios. Third, we incorporated the capability for model predictions to assess and monitor the effectiveness of fuel treatments in changing the probability of highseverity fire.

Methods
This is an abridged version-see appendix A available at stacks.iop.org/ERL/13/044037/mmedia for a detailed description of the Methods.

Data
We built a statistical model describing high-severity fire for each ecoregion in the western US (Olson and Dinerstein 2002) (figure 1). Fire severity was measured using the relativized burn ratio (RBR), a satellite measure of fire severity (resolution = 30 m) that differences pre-and post-fire Landsat data. We classified the RBR into binary categories representing high-severity (RBR ≥ 298) and other severity (RBR < 298) (Parks et al 2014a). High-severity fire can be considered standreplacing fire in the context of this study.
We evaluated 16 explanatory variables in the model for each ecoregion which can be categorized into four groups representing live fuel, topography, climate, and fire weather (table 1). The fuel group is comprised of three satellite vegetation indices: NDVI, NDMI, and EVI (table 1). These metrics implicitly incorporate management activities and disturbances such as fuel reduction treatments and wildland fire. Inclusion of 'static' fuel metrics such as vegetation type or cover (e.g. www.landfire.gov) (cf Birch et al 2015, Keyser and Westerling 2017) was not considered since such products are only updated periodically and are thus not sensitive to annual dynamics. The variables representing topography, climate, and fire weather are summarized in table 1; see appendix A for further details.

Sampling design and statistical models
We sampled individual 30 m pixels within fires that occurred from 2002-2015. Each ecoregion was modeled separately. We only sampled pixels identified as forest (i.e. forest, woodland, and savanna). We removed all pixels <100 m from the fire perimeter to reduce edge effects common at fire boundaries (Stevens-Rumann et al 2016). All analyses and predictions were conducted using the native resolution of the response variable (30 m). For each ecoregion, we used boosted regression trees (BRT) using the 'gbm' package in R to model high-severity fire (binary response) as a function of live fuel, topography, climate, and fire weather (table 1). A handful of ecoregions were not evaluated because they contained a low proportion of forest or did not have enough fire data (e.g. Sonoran Desert and North Cascades ecoregion, respectively) (appendix A).
In an effort to reduce overfitting and build the most parsimonious model for each ecoregion, we employed a cross-validated stepwise procedure in which specific variables were removed if they did not provide unique information that improved model fit. Models for each ecoregion were evaluated with five-fold cross validation that was spatially and temporally structured such that 20% of fires (as opposed to pixels) within an ecoregion were held out in each iteration. Specifically, we built a model for each ecoregion using the full suite of variables (table 1) and evaluated it with the area under curve (AUC) statistic derived from the receiver operating characteristic curve as measured with the 'verification' package in R. We then built an additional set of models for each ecoregion in which each explanatory variable was removed and calculated the  (2017) Tmax.day Maximum daily temperature. Raw value converted to per-pixel percentile.
Abatzoglou (2013) HM.ann Heat moisture for the year in which the fire occurred. HM is calculated as follows: (annual temperature + 10) / (annual precipitation/1000). Raw value converted to per-pixel z-score. AUC as previously described. If the cross-validated AUC increased when any given variable was removed from the model, it indicates that the model is overfit and that the variable does not provide any unique information. In these cases, the variable that resulted in the largest increase in AUC was permanently removed and the process was repeated until all variables resulted in a decreased AUC when removed from the model. As such, all variables in the final models provided unique information and ensured that our models were spatially and temporally transferable. Once the final model for each ecoregion was identified, the relative influence of variable groups was calculated using the AUC of a five-fold cross validation using a process that excluded all variables from a particular group. Specifically, we compared the fivefold cross validated AUC of the full model to models that iteratively excluded all variables representing live fuel, topography, climate, and fire weather. The specific equation was as follows: where AUC.full was the AUC of the full model, AUC.no.var was the AUC of the model excluding any particular variable group, and i represented one of the four variable groups.

Model implementation and map production
From the BRT models, we produced wall-to-wall raster maps (objective 2) depicting the probability of high-severity fire, if a fire were to occur, for each ecoregion in which the cross-validated AUC ≥0.70. For the fuel inputs (NDVI, NDMI, and EVI), satellite imagery from 2016 spanning the entirety of each ecoregion was obtained using Google Earth Engine (GEE; https://developers.google.com/earth-engine/). Consequently, these raster predictions represent fairly current fuel conditions across each ecoregion. Predictions theoretically range from zero to one and depict the probability of high-severity fire. We aimed to produce these severity predictions representing the average weather conditions under which fires burn. This is somewhat challenging, however, given that weather is spatially and temporally dynamic. Consequently, we produced 100 initial predictions and varied the weather for each of these predictions; all other inputs across each ecoregion (fuel from 2016, topography, and climate) were held static. To vary the weather, we randomly selected 100 records from our fire severity datasets. Each record represents one burned pixel with a unique combination of observed fire weather. We used the observed fire weather from each random record for each of the 100 initial predictions. We then averaged the 100 initial predictions over each 30 m pixel, resulting in one raster map depicting the probability of high-severity fire under average weather conditions in which fires burn. An important consideration here is that the severity predictions do not represent 'average weather conditions', but the 'average weather conditions under which fires burn'. That is, because fires often burn under more extreme fire weather, our predictions implicitly incorporate weather associated with high fire activity. This consideration also pertains to our mapped predictions under moderate and extreme fire weather, as described in the next paragraph. For those ecoregions in which the relative influence of fire weather ≥15%, we produced two additional raster maps, one depicting the probability of highseverity fire under conditions representing moderate weather and the other under conditions representing extreme weather. To do so, we calculated the 50th and 95th percentile for each pixel out of the 100 previously described initial predictions. While these maps represent the 50th and 95th percentile in predicted outcomes for each pixel, we use them to represent the outcomes of moderate and extreme fire weather, respectively. Neither says anything specific about the percentile of weather conditions under which they occurred, but they can be interpreted as resulting from moderate and extreme fire weather conditions.
To illustrate how our models can potentially be used to monitor changes in the probability of high-severity fire due to fuel treatments (objective 3), we made preand post-treatment predictions using the BRT model from the Arizona-New Mexico Mountains ecoregion. We obtained imagery representing the live fuel variables using GEE for the years 2007 (pre-treatment) and 2011 (post-treatment). Again, we produced two sets of predictions for each time period (pre-and posttreatment) representing moderate and extreme fire weather, as previously described.

Results
We incorporated data from over 2000 fires across all ecoregions to describe and explain the probability of high-severity fire (appendix B). On average, the BRT models performed moderately well for the 19 ecoregions for which we modeled (table 2). The average spatially and temporally independent cross-validated AUC statistic was 0.72 and ranged from 0.66 (Okanagan) to 0.81 (Colorado Plateau). Following Mason and Graham (2002), all five cross-validated models were statistically significant (p < 0.01) for each of the 19 ecoregions.
Although there was substantial variation across ecoregions (table 2), live fuel was the most important variable group, with an average relative influence of 53.1% among ecoregions; this ranged from 5.1% (California North Coast) to 99.0% (Utah-Wyoming Rockies). Fire weather was the second most influential variable group (22.9% average), ranging from 0% (California Central Coast and Utah -Wyoming Rockies) to 66.2% (California North Coast). Climate was the third most influential variable group (13.7% average) and topography the least influential (10.3% average) (table 2). The cross-validated variable selection approach reduced overfitting and produced parsimonious models (i.e. all variables provided unique information) (table 3).
Raster maps depicting the probability of highseverity fire were built for the 13 ecoregions in which the cross-validated AUC ≥0.70 (figure 2; appendix C). These gridded probabilities represent fuel conditions (i.e. as measured with Landsat imagery) in 2016 and   average weather conditions under which fires burn and show substantial spatial variability in the probability of high-severity fire. For ecoregions in which the relative influence of weather ≥15% (n = 6), we produced two additional raster maps depicting the probability of high-severity fire under conditions representing moderate and extreme fire weather (figure 3; appendix D). Maps of pre-and post-treatment predictions provide an example of how our models and approach can potentially be used to quantify and monitor changes in the probability of high-severity fire due to fuel treatments (figure 4). This example shows that, under conditions representing both moderate and extreme fire weather, there is an overall reduction in the probability of high-severity fire within treatment units.

Discussion
High-severity fire is often of high ecological and societal consequence, thereby motivating increasing attention and research towards better understanding its drivers and distribution ( Research to date has been either comprehensive in ecological scope but geographically limited or geographically broad but capturing only a subset of the key elements affecting fire severity. Our study expands upon these previous investigations of fire severity by (a) including a more complete suite of relevant explanatory variables, (b) evaluating fires over a large geographic extent (i.e. forests of the western contiguous US) at fine spatial resolution (30 m), and (c) including a high number of fires in our models (n = 2061 unique fires among all ecoregions; appendix B). Our results show that fuel is the most important driver of high-severity fire in forested regions of the western US, followed by fire weather, climate (i.e. 30 year normals), and topography. Our results are supported by the findings of past research but also contrast with several previous studies (see below) and provide important new insights regarding the drivers of high-severity fire. Our study is also a substantial step forward by providing a modelling framework that enables the prediction for high-severity fire while incorporating fuel and fire weather inputs. In particular, this framework involves the inclusion of fuel and fire weather inputs as dynamic variables (i.e. those that change over time) and gives us the ability to produce maps depicting the probability of highseverity fire, were a fire to occur, over entire ecoregions (e.g. figures 2 and 3). This framework also provides the means to evaluate changes in the probability of high-severity fire due to fuel treatments (e.g. figure 4).
Live fuel, as measured with Landsat vegetation indices, was on average the most important group of  . We found that live fuel was 2.3 times (on average) more influential than fire weather across the 19 ecoregions in the western US we analyzed (but see 'caveats' section). This finding is not trivial in terms of management efforts to reduce fire severity because land managers can control fuel via fuel treatments, prescribed fire, and managed wildland fire (formerly termed wildland fire use) but cannot control fire weather.
Our study found that fire weather was, on average, the second most important variable group driving highseverity fire. Previous studies have reported somewhat conflicting findings pertaining to the relative influence of fire weather in driving fire severity in the western US. Whereas some studies found weather to be moderately to highly influential (e.g. Collins et al 2007, Lydersen et al 2017, others found that the influence of weather was marginal to negligible (e.g. Kane et al 2015a, Harris and Taylor 2015). We posit that the variability we observed pertaining to the influence of fire weather among ecoregions could partly explain the divergent findings of previous studies: the relative influence of fire weather ranged from 0% to 66.2% and was the most important variable group in five out of the 19 ecoregions we analyzed. Although our results show that weather was less influential in driving highseverity fire than fuel, its influence was important in most ecoregions and should not be discounted in terms of managing fuel and fire. For example, the maps we generated (e.g. figure 3) clearly show that the probability of high-severity fire is reduced under moderate vs. extreme fire weather.
On average, climate ranked as the third most influential variable group. This contrasts with some previous studies. For example, Kane et al (2015a) found that climate was highly influential in driving fire severity in the Sierra Nevada. However, we suspect that climate was less important in our study because, over broad spatial and temporal extents, climate provides an indirect measure of fuel associated with inherent biophysical environments (Parks et al 2014c). More specifically, biomass amount is known to vary along climatic gradients (Meyn et al 2007, Krawchuk andMoritz 2011), which implies that climate can serve as an indirect surrogate for biomass. However, satellitederived vegetation indices such as those used in this study are a more direct measure of biomass (Zhao et al 2005). Consequently, when fuel and climate are both included as variables (as was done in our study), climate is ranked as less important. This said, climate was a non-negligible factor in most ecoregions. We suggest, as do others (Miller and Urban 1999b), that climate may indirectly measure factors that were not well accounted for by our variables. We believe that climate may correspond to dominant vegetation type, in that climate promotes particular physiognomic vegetation types and species that are more or less susceptible to fire (Parks et al 2018). For example, cooler and wetter climates are more likely to support species that are more susceptible to fire-induced mortality (e.g. Engelmann spruce), whereas warmer and drier climates are more likely to support species that can survive fire (e.g. ponderosa pine) (Lutz et al 2010).
Nearly every fire severity study to date has found that topography had a moderate to high influence on fire severity (e.g. Holden et al 2009, Dillon et al 2011, Fang et al 2015, Kane et al 2015a, Birch et al 2015, Estes et al 2017. Conversely, our study indicates that topography is on average the least important variable group. We posit that topography is an indirect measure of fuel, and that because we directly account for fuel (using satellite-derived vegetation indices), topography is deemed a relatively unimportant factor. It is worth noting that many of these previously mentioned studies do not incorporate any measure of fuel or vegetation into their analyses (Holden et al 2009, Dillon et al 2011, Kane et al 2015b, and consequently, the influence of topography may be unintentionally elevated. For example, even though Dillon et al (2011) found topography to be the strongest driver of fire severity across large regions of the western US, they clearly stated that topography was serving as a proxy for variation in fuel and bioclimatic variables (i.e. fuel moisture and temperature) which were not accounted for in their study. Since we capture such variability in live fuel using satellite-derived vegetation indices, the influence of topography on its own is greatly diminished.

Caveats
There are several difficulties associated with building statistical models describing fire severity across broad geographic regions. Our ability to characterize fuel, for example, was limited to satellite indices that generally characterize overstory vegetation and have limited capacity to measure live and dead surface and ladder fuels known to drive fire behavior and effects (Rothermel et al 1972, Scott andBurgan 2005). Our ability to adequately characterize fire weather was also limited. For example, we estimated the day at which any given pixel burned using MODIS fire detection data. These day-of-burning estimates are not without error (Parks 2014, Veraverbeke et al 2014; this error increases uncertainty and likely diminishes our ability to characterize the influence of weather. Also, the temporal and spatial resolution of currently available gridded weather datasets do not necessarily match the realized spatial and temporal weather variability associated with any given fire (Wagenbrenner et al 2016). Fires have even been known to generate their own weather (Potter 2012) and gridded weather datasets do not and will not likely be able to account for such phenomenon in the foreseeable future.
It is also worth noting that our data is potentially biased due to undersampling of non-extreme weather conditions, thereby limiting our ability to completely capture the full range of weather conditions conducive to fire and to completely characterize the influence of fire weather. Specifically, fire suppression success rates are higher under less-than-extreme weather conditions (Arienti et al 2006, Fernandes et al 2016, Beverly 2017, thereby reducing the amount of area burned under more moderate weather; this biases the weather associated with our fire severity data. For example, we found that the relative influence of weather was zero in two ecoregions; this result is more likely an artifact of biased data (and the other caveats mentioned in this section) than an unconditional representation of the relative influence of weather. Simply put, we potentially mischaracterized the relative influence of weather since we could not sample the full range of weather under which fire can burn. We also suggest that more data (i.e. fires) are needed in some ecoregions to better characterize the influence of fire weather and other variables (e.g. the coastal ecoregions in California; see appendix B).

Management implications
Managing for wildland fire has become incredibly complex as we face the nexus of increasingly large and intense wildfires linked to a warming climate and more frequent drought, landscapes with heavy fuel accumulations due to prolonged fire exclusion, and a rapid expansion of the wildland-urban interface (Littell et al 2016, Stephens et al 2016, Schoennagel et al 2017. Land management agencies have a daunting challenge to reduce risks from fire to communities and fire fighters while simultaneously restoring forests to more resilient conditions (www.forestsandrangelands.gov) (Barnett et al 2016). In response, land management agencies in the US established a long-term fuel reduction program in which millions of hectares have been treated since 2001 using a variety of methods such as mechanical thinning and prescribed burning (US Congress 2003). Various efforts are underway to assess how to best focus such fuel reduction activities given that land management agencies have limited resources. In particular, spatially explicit planning frameworks have offered an effective means to strategize locating treatments across landscapes (e.g. Ager et al 2016, Scott et al 2016. These planning frameworks are often built on spatial assessments of quantitative wildfire risk that incorporate the probability of wildfire occurrence across a range of simulated fire intensities, and the effects of fire on specific values at risk (e.g. natural resources, built assets) (Finney 2005, Scott et al 2013. We suggest that the modeling framework in this study could complement these efforts and allow predictions of high-severity fire to be integrated with fire occurrence and behavior predictions to provide managers with a more comprehensive set of riskanalysis information to target locations in wildfire mitigation planning. We also suggest that our models and the resulting predictions of high-severity fire could potentially serve as a performance metric for evaluating hazardous fuel treatments (see figure 4). For example, the US Forest Service often uses 'acres treated' as a performance measure, but this measure does not capture anything about whether treatment objectives have been met (USDA Forest Service 2016). Specifically, the primary objective of most hazardous fuel treatments is to reduce the intensity and resulting severity of potential wildland fires ( The protocols developed in this study offer a means to provide predictions that are objective, consistent, updateable, spatially detailed (30 m resolution), and spatially extensive as a measurable benchmark to characterize changes in the potential for high-severity fire. We acknowledge, however, that substantial financial resources would be necessary to implement our framework to monitor the potential for high-severity fire, but such a tool is essential if not timely for filling a key information gap in fire management on public lands in the US and elsewhere.

Conclusions
Fuel is on average the most influential factor driving high-severity fire in forests of the western US. Consequently, efforts to reduce fuel will likely reduce the potential for high-severity fire (Pollet and Omi 2002, Stephens et al 2009, Arkle et al 2012. Our results also indicate that fire weather has a substantial influence on fire severity and highlight that the probability of highseverity fire is reduced under conditions representing moderate vs. extreme fire weather. This finding, when considered with the fact that fire suppression is more effective under less-than-extreme fire weather (Arienti et al 2006, Beverly 2017, underscores that land management agencies may be paradoxically selecting for high-severity fire by aggressively suppressing fire (cf Calkin et al 2015). Simply put, aggressive fire suppression reduces the occurrence of low severity fire, thereby increasing fuel on the landscape and selecting for higher severity fire when the inevitable fire occurs. This has substantial ecological and social consequences, particularly for dry forests that historically experienced low-and mixed-severity fire. For example, fire-facilitated conversions from dry forest to nonforest vegetation (shrubland and grassland) are now evident, but it is important to note that such conversions appear to be triggered only by high-severity fire (not low severity) (Savage and Mast 2005, Coop et al 2016, Coppoletta et al 2016. Consequently, to limit the probability of high-severity fire, fire-facilitated conversions to non-forest, and altered successional trajectories in dry forests (Johnstone et al 2016, Walker et al 2018, land managers could consider, in addition to traditional fuel reduction treatments, expanding opportunities that allow wildland fires to burn under less-than-extreme weather conditions.

Data accessibility statement
All ecoregional fire severity predictions shown in appendices C and D are available for download in a georeferenced raster format through the Fire Research and Management Exchange System (FRAMES; www.frames.gov/NextGen-FireSeverity).

Acknowledgments
We acknowledge funding from the Joint Fire Science Program under project 15-1-3-20 and from the National Fire Plan through the Rocky Mountain Research Station. We thank two anonymous reviewers whose feedback substantially improved this manuscript. We also thank Diana Olson and Michael Tjoelker (University of Idaho) for facilitating the distribution of our mapped severity predictions through the Fire Research and Management Exchange System (FRAMES; www.frames.gov/NextGen-FireSeverity).