Does conserving roadless wildland increase wildfire activity in western US national forests?

National forests in the western United States are divided roughly in half between lands without roads managed for wilderness characteristics and lands with an extensive road system managed for multiple uses including resource extraction. We investigated the influence of these land use designations on fire ignitions, fire extent, and fire severity over the last three decades. Although roadless areas experienced fewer fire ignitions and are generally cooler, moister, and higher elevation landscapes less conducive to fire, wildfire extent was far greater in these areas than in roaded areas. An area equivalent to approximately one-third of roadless areas burned in the last three decades, while an area equivalent to less than one-fifth of roaded areas experienced fire. Most of the largest fires that have burned on national forest land in recent years began in roadless areas. Despite greater fire extent in roadless areas, there was no significant difference in fire severity between roadless areas and roaded areas after accounting for biophysical differences between these management regimes. Although fire patterns in roadless areas may pose challenges to land managers, the available evidence suggests that the greater extent of fire in roadless areas may confer resilience to these landscapes in the face of climate change.


Introduction
The number and size of wildfires in the western United States has increased dramatically over the last three decades, with significant social, economic, and ecological consequences (e.g. Calkin et al 2015, Sankey et al 2017, Bladon 2018. Understanding the drivers of fire activity is important to help society adapt to the new fire reality, particularly in the context of anthropogenic climate change (North et al 2015, Dunn et al 2020. Numerous studies have investigated the influence of fuel dynamics, topography, and climate change on fire in the Anthropocene (e.g. Parks et al 2014a, Keyser and Westerling 2017, Meddens et al 2018. This study examines the influence of public policy on wildfire, specifically the influence of different land use designations on fire ignitions, fire extent, and fire severity on national forest land managed by the US Forest Service (USFS).
The USFS manages 20% of forestland in the United States (1.5% of the global forestland total), most of which is found in the 11 western states (figure 1). For much of the agency's 115 year history, US Congressional legislation emphasized development of natural resources, e.g. the Multiple Use Sustained Yield Act and the National Forest Management Act (Clary 1986). In the last 60 years, protection of landscapes from development has received increasing attention from policy makers. In particular, the Wilderness Act of 1964 provided for the creation of wilderness areas to 'preserve natural conditions' . The Wilderness Act required the USFS to inventory roadless areas not designated as wilderness pending future Congressional action (Roth 1984, Booth 1991. Inventoried roadless areas not released for development in the 1970s and early 1980s served as an ad hoc administrative extension of the wilderness system until 2001, when the Roadless Area Conservation Rule ('Roadless Rule') generally prohibited road construction and commercial timber harvest in these areas (USDA (US Department of Agriculture) 1972, 1979, Loucks et al 2003, Voicu 2010; see also Wyoming Outdoor Coordinating Council v. Butz, 484 F.2d 1244(10th Cir.1973).
Decisions to protect roadless wildlands from development have thus created two strongly contrasting management regimes: an actively managed landscape with an extensive road system and a legacy of resource extraction, and a passively managed landscape with no roads and little or no history of resource extraction (figure 1, table 1). Actively managed roaded landscapes have been in the past and continue to be subject to a wide variety of anthropogenic modifications including but not limited to timber harvest, road construction and maintenance, and a wide variety of recreational developments. These activities have been largely absent within roadless areas. Passive management of roadless landscapes is largely a matter of decisions about suppression of natural disturbances like wildfire.
Examining the influence of different management regimes on wildfire is timely. Recent Congressional legislation such as the Healthy Forest Restoration Act of 2003, revisions to the National Forest Management Act planning regulations in 2012, and agency policy initiatives like the Collaborative Forest Landscape Restoration Program emphasize restoration of the natural fire regime to which different landscapes are adapted (Schultz et al 2012, Feldman andReimer 2019). Wildfire is a key disturbance process that shapes the structure, composition, and function of forests, and a better understanding of the influence of passive vs active management on fire patterns is critical for evaluating the ability of USFS managers to meet forest restoration objectives (Johnstone et al 2016, Thompson et al 2018, Schultz et al 2019. In this study we used USFS databases and satellite data to determine if there are important differences in (a) fire ignitions, (b) fire escape from initial suppression efforts, (c) fire extent, and (d) fire severity between passively managed wilderness and inventoried roadless areas ('roadless areas') and the roaded portion of national forest lands actively managed for multiple uses ('roaded areas').
In this paper we report results from a complete census of the extent and severity of wildfire based on satellite data encompassing all USFS roadless and roaded areas. However, simple summaries of ignitions, fire escape, fire extent, and fire severity between roadless and roaded areas may be an inadequate basis for evaluating the influence of passive vs active management because land use decisions that created different management regimes were not made at random. Many national forest wilderness and inventoried roadless areas were originally designated in large part because these lands are generally rugged, high-elevation landscapes that are difficult to develop (DeVelice andMartin 2001, Hjerpe et al 2017). Although roadless areas include extensive dry and fire-prone forests, at the broad spatial scales we consider, the history of wilderness and inventoried roadless area designations means that these lands are likely to be moister, cooler, and hence less conducive to fire ignition and fire spread (Agee 1996, O'Connor et al 2017.
We contextualized our comparison of fire activity across management regimes in several ways. First, we characterized differences in annual precipitation and maximum temperature between roadless and roaded areas. These interannual climate variables are good surrogates for differences in fuel abundance, site productivity, and vegetation type that exert a strong indirect control on the burning environment (Dwyer et al 2000, Westerling 2008, Krawchuk et al 2009. Second, we compared area burned between roadless and roaded areas within different western US ecoregions with distinctive climates and vegetation communities. Finally, we built statistical models that estimated the influence of management regime on fire probability and severity after accounting for temperature, precipitation, slope, elevation, vegetation type, and a wide range of other biophysical variables that potentially influence the fire environment.

Study area and data acquisition
We investigated the influence of management regime on fire ignitions, extent, and severity within national forests in the 11 conterminous western US states (WA, ID, MT, OR, CO, CA, NV, UT, CO, AZ, NM; figure 1). We limited our study to this geographic scope because the vast majority of national forest lands (82%) are found within these states and because the USFS has adopted lower size thresholds for protecting roadless areas and wilderness areas in the eastern states. Also, many eastern US national forests consist of formerly intensively managed private lands acquired by the federal government, which complicates inference about the effects of management regime on fire pattern. Most USFS lands in the west have remained in the public domain since Euro-American colonization (Bramwell 2012). We delineated national forest lands using USFS spatial data layers and divided them into two categories: roadless and roaded areas (figure 1, table 1). Roadless areas fall within the administrative boundaries of lands protected by the Wilderness Act or inventoried roadless areas protected by the Roadless Rule. Roaded areas include the remainder of the national forest lands managed for multiple use objectives.  We evaluated the influence of roadless management on fire ignitions and probability of those ignitions escaping initial suppression efforts using the Fire Occurrence Database (FOD) maintained by USFS fire management staff. We defined escaped fires as those ignitions that grew to be larger than 121 ha, a size at which the USFS generally expects wildfire to exceed initial attack capabilities (Fried andFried 1996, Calkin et al 2005). At the time of our analysis the FOD included information about fire ignition location and ultimate fire size for ∼2 million ignitions in the United States between 1992 and 2017 (Short 2017). Although other data sources, including the Monitoring Trends in Burn Severity (MTBS) program (see below), contain information about fires prior to 1992 we used the FOD because 90% of area burned in the study area occurred after 1991 and because the FOD contains detailed information about the spatial and temporal origin of fires (including fires that did not grow to the minimum size mapped by other data sources) essential for evaluating the influence of land use and biophysical variables on the fate of wildfire ignitions.
To evaluate the influence of management regime on fire extent we downloaded fire perimeter data from the MTBS program (www.mtbs.gov), which at the time of our analysis mapped all large (>405 ha) fires in the United States between 1984 and 2017 (Eidenshink et al 2007, Finco et al 2012. We added fire perimeter data compiled by the Geospatial Multi-Agency Coordination (GeoMAC; www.geomac.gov) to extend our fire extent data through the year 2018. We excluded GeoMAC fire perimeter data for fires less than 405 ha for consistency with MTBS fire perimeter data.
To evaluate fire severity we created seamless maps of the relative differenced normalized burn ratio (RdNBR) at 30 m resolution using Landsat satellite time-series within each MTBS fire perimeter (Kennedy et al 2018). RdNBR is an objective measure of fire-induced changes in vegetation (i.e. tree mortality) and soil reflectance that accounts for differences in vegetation across different regions (Miller and Thode 2007). We provide a detailed account of our fire severity mapping methods as supplementary materials (see supplementary materials S1 (available online at stacks.iop.org/ERL/16/084040/mmedia)).

Data analysis
We summarized the total number of ignitions documented in the FOD across both roadless and roaded portions of the western states national forest study area. Then we estimated the influence of biophysical variability and management regime on the probability of fire escape using generalized additive models (GAMs) with a binomial distribution implemented with the mgcv package in the R statistical environment (Wood 2004(Wood , 2006, R Core Team 2018). Many of the ignitions on national forest lands catalogued in the FOD that occurred between 1992 and 2017 were close together in time and space and may be temporally and/or spatially autocorrelated. We included individual national forest and year of fire as random effects in models, reasoning that variability in escape probability could be explained in part by the resources available to individual administrative units of the national forest system and the year in which ignitions occurred.
We summarized fire extent across roadless and roaded portions of the study area using the complete inventory of area burned within our western states study area provided by MTBS/GeoMAC data. To better understand variability in fire extent across national forest lands, we summarized total area burned in roadless and roaded areas within each of 23 different level III ecoregions delineated by the US Environmental Protection Agency located within the study area (Omernik 1987). We do not report results from 12 ecoregions that are composed of less than 2% national forest land or in which fire burned less than 1000 ha during the study period.
To account for environmental differences between roadless and roaded areas that potentially influenced fire extent, we analyzed 100 000 points randomly located across the study area. We constrained random location procedures so that points were at least 500 m apart, a distance that the literature suggests avoids spatial autocorrelation in fire effects that could bias coefficient estimates (van Mantgem and Schwilk 2009, Lanorte et al 2013, Kane et al 2015. We examined autocorrelation function plots and performed Moran's I tests to ensure that there was no significant autocorrelation among points. We estimated the influence of management regime (roadless or roaded) and a wide variety of biophysical variables (table 2) on the probability that these points would fall within a 1984-2018 fire perimeter using a GAM with a binomial distribution. We graphed the probability of fire along 30 year average precipitation and maximum temperature gradients because these climate variables are key indirect controls of natural fire regimes and good surrogates for environmental variability at the broad spatial scales we considered (Westerling 2008, Guyette et al 2012, Liu et al 2013, Whitman et al 2015. To evaluate differences in fire severity attributable to management regime, we summarized mean RdNBR for the roadless and roaded portions of fire perimeters between 1985 and 2017, leveraging pre-and post-fire Landsat imagery spanning 1984-2018. Like our summaries of fire extent, our summary of fire severity represents a comprehensive census of fire across the study area. To account for environmental differences between roadless and roaded areas that potentially influence fire severity, we modeled the influence of biophysical variability and management regime on RdNBR with a GAM using the subset of the 100 000 random points (20% of points) that fell within 1985-2017 MTBS fire perimeters.
Our goal with statistical models was to determine if management regime had a significant influence on fire escape, fire extent, and fire severity given environmental variables that influence fire. All continuous variables were modeled as smoothed functions in order to capture non-linear relationships between the fire response and environmental gradients. All categorical variables, including management regime, were modeled as parametric variables. We tested a variety of different vegetation classifications in wide use as predictor variables in models (table 2) and used the vegetation classification that explained the most variability in the response. We report the influence of management regime to be significant when confidence intervals (alpha = 0.05) for the effect of roadless and roaded management on the fire response examined do not overlap.
A number of the continuous environmental variables that we tested as explanatory variables for fire escape, fire extent, and fire severity responses were correlated. For instance, maximum temperature was strongly correlated with elevation and latitude. We evaluated full models that included all biophysical variables and parsimonious models created using forwards and backwards model selection procedures that maximized deviance explained and minimized Akaike Information Criterion while eliminating terms with concurvity indexes that exceeded 0.60. Previous research suggests that this threshold provides a conservative approach to eliminating variables that describe similar environmental gradients (Johnston et al 2019). Because there was little difference in the estimated effect of management regime between full and parsimonious models, we report results from parsimonious models to improve interpretability. Model specifications and a detailed summary of our statistical methods are

Roadless and wilderness designations are associated with fewer ignitions but more escaped fires
Of the 194 151 total ignitions within the western US national forest study area currently documented in the FOD, 3902 (2% of the total) escaped initial suppression efforts and grew to be >121 ha. The majority of ignitions (74% of the total) occurred within the roaded portion of the national forest landscape, but an almost equal number of fire ignitions (1925) escaped initial control efforts within roadless areas as escaped within roaded areas (1977). Of fire ignitions in roadless areas, 4% escaped control compared to 1.4% of ignitions in roaded areas. GAM models indicated that management regime exerted a significant influence on probability of fire escape even after accounting for topographic ruggedness, slope steepness, vegetation type and other factors that influence the ability of managers to suppress fires. Escaped fires that began in roadless areas grew to be one-third larger than escaped fires that began in roaded areas (table 3). Of the 20 largest fires on national forest land from 1984 to 2018, more than three-quarters began within roadless areas, which account for slightly less than half of the total area of western US national forests (table 4).

Roadless and wilderness designations are associated with greater fire extent
Roadless areas were associated with a far greater extent of fire relative to roaded areas. Between 1984 and 2018, an area equivalent to 30% of the total area of roadless lands experienced fire, whereas an area equivalent to 18% of roaded areas experienced fire (table 1). This difference is striking because roadless areas within the western states national forest study area contained a larger proportion of cooler and moister environments than roaded areas (figure 2). Management regime exerted a strong influence on the probability of fire even after accounting for vegetation type and biophysical variables. The general shape of the fire probability response while holding other vegetation and biophysical variables constant was similar across climate gradients in roadless and roaded areas (figure 3). In both roadless and roaded landscapes, probability of fire was lowest in very cool and moist settings where climate is less conducive to fire and in very hot and dry settings where fuel is sparse. Probability of fire was highest in areas with precipitation sufficient for fuel to accumulate and dry conditions suitable for fire spread (Krawchuk et al 2009, Pausas and Ribeiro 2013, Whitman et al 2015. However, the probability of fire was significantly higher in roadless areas across these gradients. There was little overlap in confidence intervals for estimates of probability of fire in roadless and roaded areas except within temperature and moisture regimes where fire was relatively rare (figure 3). Although roadless areas were associated with a substantially greater extent of fire than roaded areas, there was enormous variability in the extent of fire within different ecoregions (figure 4). Some ecoregions with extensive national forest land such as the Northern Rockies and Wasatch and Uinta Mountains have experienced relatively little fire, with fire evenly distributed between roadless and roaded areas. Other ecoregions with extensive national forest land such as the Idaho Batholith and Arizona/New Mexico Mountains have experienced extensive fire and a notably greater extent of fire in roadless areas. Several ecoregions stand out for the dramatic extent of fire in roadless areas, particularly when including multiple overlapping fires (i.e. reburn). An area equivalent to 86% of the large (1090 342 ha) national forest area managed as roadless wilderness in the Klamath Mountains/California High North Coast Range ecoregion burned between 1984 and 2018, while only 40% of the roaded portion of this ecoregion experienced fire during the same period. Similarly, an area equivalent to 100% of the relatively small (280 782 ha) roadless national forest land in the Central California Foothills and Coastal Mountains ecoregion burned during the same time period compared to 67% of the remaining roaded lands. Approximately 15% of total fire extent across the western states study area between 1984 and 2018 occurred within the roadless portion of the Idaho Batholith ecoregion. We provide a detailed summary of fire extent and severity in different ecoregions as supplementary materials (supplementary materials table S3.1).

Roadless and wilderness designations do not influence fire severity
Management regime had no influence on fire severity in any of the statistical models we tested. The difference in mean RdNBR between roadless and Table 4. Largest fires in the western states national forest study area over the period 1984-2018 in terms of national forest land burned (statistics exclude private and other public land burned). State is the state where largest area burned occurred. Fires without fire origin had multiple origins found within USFS roaded and roadless areas, or an origin on other public or private ownership. The mean RdNBR of the roaded portion of the Lake Creek Fire (RdNBR = 6) was excluded from the table and calculations of mean because of the very small area of roaded land that was burned in that fire. None of the grand mean statistics reported in the lower row are weighted by fire extent.
RdNBR weighted for fire extent was 481 and 394 for roadless and roaded portions of these fire perimeters, respectively. Data are from the FOD and MTBS/GeoMAC (see text).  Figure 2. Area of roadless and roaded western states national forest lands along two axes: annual precipitation and maximum temperature. Denser areas correspond to a greater proportion of the landscape located within that precipitation and temperature range. Vertical and horizonal lines represent mean annual precipitation and mean maximum temperature of roadless and roaded areas. Roadless areas are generally cooler and moister and hence less conducive to fire than roaded areas. Figure 3. Probability of fire at randomly located points across western US national forests along temperature and precipitation gradients in roadless and roaded areas. Thick line shows estimated probability of fire and transparent areas around the line show confidence intervals (alpha = 0.05) for the estimate. Probability of fire in roadless areas is higher across the maximum summer temperature gradient and peaks within cooler landscape settings than in roaded areas. Probability of fire is the same or higher in roadless areas across the annual precipitation gradient. roaded areas was negligible given the wide range of RdNBR values in our dataset ( figure 5, table 1). Vegetation type and biophysical variables that influence the distribution and abundance of vegetation including elevation, annual precipitation, and maximum summer temperature explained as much as 17% of the variance in the severity response compared to approximately 1% of the variance explained by management regime. Confidence intervals for the estimated effect of management regime on RdNBR overlapped zero in all models we tested that included both vegetation and biophysical variables. There were no consistent differences in fire severity between roadless and roaded areas at the scale of individual ecoregions (see supplementary materials table S3.1).

The influence of management regime and biophysical variability on fire
The greater incidence of fire ignitions in roaded areas but the greater extent of fire in roadless areas that are generally less conducive to fire is most likely related to three factors. First, proximity to developed areas is associated with more ignitions (Balch et al 2017). Second, a lack of road access limits deployment of the full range of suppression tools available to managers to control fires in roadless areas when they are small (Dunn et al 2017). Third, large remote landscapes allow managers flexibility to allow fires to burn when these fires are compatible with resource management objectives or when suppression . These trends, in conjunction with spiraling fire suppression costs, may ultimately result in divergent fire regimes, with fires in roaded areas influenced primarily by the exigencies of suppression efforts while fire extent and severity in roadless areas become increasingly selforganized (Riley et al 2018).
Although differences in fire escape and fire extent are strongly associated with management regime, our statistical analyses demonstrate that the primary drivers of fire severity across the national forest landscape are differences in the fire environment and not land use designations per se. The small difference observed in fire severity between roadless and roaded areas (figure 5, tables 1 and S3.1) likely reflects differences in the abundance of different tree species between these management regimes. Trees growing in roadless sites with generally greater moisture availability and lower temperatures are less fire tolerant and more susceptible to mortality from fire than species found in drier, lower-elevation landscape settings (Dunn and Bailey 2016, Johnston et al 2019, Stevens et al 2020. Even without accounting for differences in mortality attributable to differences in species composition and the fire environment, the difference in fire severity between roadless and roaded areas is minor. For example, one study that calibrated Landsat-derived RdNBR with plot-based fire severity data in the Pacific Northwest found that the mean fire RdNBR values for roadless vs roaded areas across our study corresponded to 58% and 52% mortality of overstory trees from fire (Reilly et al 2017).

Management and policy implications
Limiting smoke exposure to vulnerable populations and reducing risk to water supplies, habitat, and human infrastructure from uncontrolled 'mega-fires' are important goals of policy makers (McKenzie et al 2014, Vaillant andReinhardt 2017). Mechanical fuel treatments are a commonly used tool to accomplish these objectives (Stephens et al 2012), but more than half of all fires, including most of the largest fires in western US national forest lands, burn primarily in roadless areas where mechanical treatments are generally prohibited (tables 1 and 4). The extent of fire in landscapes where management options are limited drives home the need to adapt to fire rather than overcome fire in the American West (Dunn et al 2020).
Feedbacks between biophysical and policy systems may compound divergent trends in fire extent between roadless and roaded areas. The USFS is placing an increased emphasis on firefighter safety, which will result in fewer fire engagements in remote, complex terrain that are associated with firefighter casualties and which are more prevalent in roadless areas (Page et al 2019). Additionally, the greater extent of fire in roadless areas means that new fires are more likely to occur within previous fire perimeters. Suppressing these new ignitions may be extremely difficult and unsafe because of extensive patches of dead trees and difficult medical evacuation, potentially leading to a greater number of fires escaping initial suppression efforts in the future (Dunn et al 2019).
The disproportionate extent of fire in roadless areas may present challenges for managers depending on their objectives. For instance, roadless areas in coastal southern California are associated with large and costly fires that pose significant risk to adjacent densely populated communities (Keeley et al 2009). In other regions, repeated fire in dry mixed-conifer forests may result in new vegetation states with fundamentally different habitat potential (Coppoletta et al 2016, Serra-Diaz et al 2018, McCord et al 2020. Managers may need to adjust conservation strategies for sensitive species that rely on closed canopy forests within areas that are dominated by roadless areas given the extent of fire associated with this management regime (Spies et al 2006). For instance, in the Klamath ecoregion, a network of late-successional reserves designed to protect and restore habitat for the northern spotted owl (Strix occidentalis caurina) and other old-growth associates is anchored within roadless areas , Spies et al 2019. The extent of fire in these areas may require managers to conserve additional complex older forest within the roaded portion of the landscape or to develop a dynamic reserve system in which reserves are reallocated based on disturbance-mediated changes to habitat .
Despite these challenges, in many areas throughout the western states, fire associated with management for roadless and wilderness characteristics has the potential to confer resilience in the context of climate change. Fire suppression has sharply reduced wildfire activity on national forest lands over the last 100-150 years, and most western forests are in a 'fire deficit' relative to the natural fire regime (

Conclusions
Decisions to conserve the wild and roadless characteristics of almost half of western US national forest lands has resulted in a landscape-scale experiment that contrasts the effects of passive and active approaches to accomplishing management objectives. A broad range of management tools, including mechanical thinning, prescribed fire, and wildland fire will continue to be employed to accomplish objectives on the roaded landscape (North et al 2012). Absent significant policy changes, wildland fire will continue to be the primary tool available to accomplish fuel management objectives within roadless areas (North et al 2015). Managers and scientists should collaborate to quantify the degree to which objectives are being accomplished within different management regimes across time and space.

Data availability statement
The data that support the findings of this study are openly available at the following URL/DOI: www.mtbs.gov. risk science facilitates adaption of fire prone social ecological systems to the new fire reality Environ. Res