Contemporary patterns of ﬁ re extent and severity in forests of the Paci ﬁ c Northwest, USA (1985 – 2010)

. Fire is an important disturbance in many forest landscapes, but there is heightened concern regarding recent wild ﬁ re activity in western North America. Several regional-scale studies focus on high-severity ﬁ re, but a comprehensive examination at all levels of burn severity (i


INTRODUCTION
Fire plays an essential role in the structure and function of many forested ecosystems in western North America (Agee 1993, Sugihara et al. 2006, Baker 2009). Despite the historical role of fire, increases in fire activity since the 1980s (Westerling et al. 2006, Littell et al. 2009, Dennison et al. 2014) have had profound socio-economic impacts , Fischer et al. 2016, and there is much concern regarding the ecological effects of contemporary patterns of burn severity (i.e., tree mortality). Temporal trends and spatial patterns of high-severity fire have been the focus of several regional-scale studies (e.g., Miller et al. 2009, Dillon et al. 2011, Cansler and McKenzie 2014, Harvey et al. 2016a). However, low-and moderate-severity fires have received comparatively little attention (but see Mallek et al. 2013, Kolden et al. 2015 despite the potential ecological benefits (i.e., fuels reduction, restoration of forest structure) associated with them in forests where fire was frequent historically (North et al. 2012). A comprehensive analysis of trends and spatial patterns at all levels of burn severity-and particularly how they vary among biophysical settings (e.g., vegetation zones) differing in environment, climate, and historical fire regimes-is needed to better inform our understanding of the ecological effects of recent fires.
Current knowledge of historical fire regimes and landscape dynamics in the western United States comes primarily from retrospective studies of fire history (e.g., Table 1). These studies provide guidance for management (e.g., Cissell et al. 1999, Swetnam et al. 1999 and are the foundation for historical fire regime characterizations based on fire frequency and idealized landscape patterns of burn severity (e.g., Heinselman 1981, Kilgore 1981, Agee 1993, 1998 Fig. 1). Lowseverity fire regimes are typically found in environments where conditions conducive to burning occur in most years, and frequent fire limits the accumulation of surface and ladder fuels. Small patches (<1 ha) of high-severity fire do occur, but most of the area burned is comprised of large patches of low-severity fire (Agee 1998). Mixedseverity fire regimes are typically found in environments of intermediate moisture and are characterized by a mosaic of patches at all levels of severity (Agee 1993, Perry et al. 2011. Large patches of high-severity fire (≥100 ha) may occur (Agee 1998, Tepley andVeblen 2015), but most of the area burned is made up of patches of lowand moderate-severity fires depending on topdown climatic drivers and bottom-up fuels and topographic controls (Perry et al. 2011). Highseverity fire regimes are typically found in cold and wet environments where ignition or conditions (e.g., drought) conducive to burning occur infrequently (Gedalof et al. 2005, Krawchuk andMoritz 2011). Large patches of high-severity fire drive landscape dynamics and comprise most of the area burned.
A long period of fire exclusion and increasingly warm and dry conditions have sparked concerns that contemporary patterns of burn severity are uncharacteristic of historical landscape dynamics in some forests (see Noss et al. 2006). Fire exclusion may have exerted little influence on patterns of burn severity in historically high-severity regimes, where the period of 20thcentury fire exclusion is shorter than historical fire return intervals , Schoennagel et al. 2004, Noss et al. 2006). On the contrary, exclusion of frequent fire in historically low-and mixed-severity regimes is hypothesized to have increased vulnerability to high-severity fire (Hessburg et al. 2005, Perry et al. 2011). There is increasing evidence that some forest types where fire was historically frequent are burning with greater proportions of high-severity fire (Mallek et al. 2013, O'Connor et al. 2014) and larger patches of high-severity fire (Miller et al. 2012). However, the balance between characteristic and uncharacteristic fire activity is poorly understood due to a lack of information on spatial patterns (e.g., relative proportions and patch size distributions) at all levels of burn severity and how these compare with idealized characterizations of historical landscape dynamics.
Despite recent increases in fire activity, there is growing recognition that current fire frequency and extent are well below historical estimates (Stephens et al. 2007) and that many regions may still be accruing an overall "fire deficit" (Marlon et al. 2012, Mallek et al. 2013, Parks et al. 2015. However, prior studies on fire deficit are limited to coarse regional scales that either do not distinguish among individual vegetation zones or do not assess spatial patterns of burn severity. Given the potential variability in vegetation, environment, and historical fire regimes at regional scales (Guyette et al. 2012), knowledge of how temporal and spatial patterns of burn severity vary among vegetation zones is needed to provide important context for assessing the ecological effects of recent fires. This knowledge can inform strategic use of managed wildfire (e.g., North et al. 2015 to promote restoration of fire as an ecological process and help develop mitigation and adaptation strategies that promote resilience for the future (e.g., Deser et al. 2012).
In this study, we assess temporal trends in fire extent and burn severity across the major forested vegetation zones of the Pacific Northwest from 1985 to 2010. This biophysically diverse region includes a variety of vegetation zones (Fig. 2) that differ in historical fire frequency by over an order of magnitude (Table 1) and are representative of historical fire regimes across much of the western United States (Guyette et al. 2012). A period of fire exclusion during the early and mid-20th century altered composition and structure in dry forests that historically experienced low-and mixed-severity fire regimes (Hessburg et al. 2005, Hagmann et al. 2013, 2014, Merschel et al. 2014), but wildfires have increased in frequency and extent since the mid-1950s (Littell et al. 2009). Dillon et al. (2011) found no temporal trends in fire extent from 1984 to 2006 for the eastern or western portion of the region, but little is known about how trends and spatial patterns of burn severity vary among vegetation zones at sub-regional scales.
We use a novel approach to map burn severity by integrating field data from a systematic regional forest inventory with a spatially and (  Fig. 1. Conceptual diagram characterizing the cumulative proportions of low-, moderate-, and highseverity fires in three major fire regime types. Inset panels represent idealized landscape dynamics associated with each regime based on proportions and size class distributions of patches at each of the three levels of severity. Adapted from Agee (1993Agee ( , 1998 temporally normalized 26-year Landsat time series. This approach provides a direct measure of burn severity as an ecological process (i.e., tree mortality; Morgan et al. 2014) by relating spectral change [relativized difference in the normalized burn index (RdNBR), Miller and Thode (2007)] to tree mortality measured from pre-and post-fire plots (i.e., percentage of basal area change). Our specific objectives were to (1) quantify temporal trends in fire extent and spatial patterns (i.e., proportions and patch size class distributions) at all levels of burn severity in relation to drought and annual fire extent; (2) compare these trends and patterns of burn severity among the major forested vegetation zones; (3) evaluate contemporary fire activity in relation to expectations based on previous studies of historical fire regimes and landscape dynamics in the region.

Study region
Our study region encompasses approximately 26 million ha of forest across Oregon and Washington and includes a variety of potential vegetation zones (Fig. 2) distributed along strong climatic and topographic gradients (Franklin and Dyrness 1973). Potential vegetation zones represent broad biophysical environments and geographic ranges distinguished by the tree species that would dominate in later-successional stages in the absence of disturbance (Pfister and Arno 1980). We used a map of potential vegetation zones for Oregon and Washington that was created following Henderson et al. (2011) using species distributions from existing forest inventory data in relation to dominant climatic and topographic gradients (available online: www.ecoshare.info/category/gis-data-vegzones). The major wet vegetation zones include warm western hemlock (Tsuga heterophylla), cool silver fir (Abies amabilis), and cold mountain hemlock (Tsuga mertensiana). Major dry vegetation zones include warm ponderosa pine (Pinus ponderosa), cool Douglas-fir (Psuedotsuga menziesii), cool grand fir/ white fir (Abies grandis/Abies concolor), and cold subalpine. The subalpine zone includes subalpine fir (Abies lasiocarpa) and Engelmann spruce (Picea engelmannii), as well as subalpine parklands dominated by whitebark pine (Pinus albicaulis). Other minor forested vegetation zones include Jeffrey pine (Pinus jeffreyi), lodgepole pine (Pinus contorta), Sitka spruce (Picea sitchensis), Shasta red fir (Abies magnifica), and tanoak (Lithocarpus densiflorus). These latter zones are included in our analysis of regional patterns, but are not analyzed individually as they comprise a small proportion of the total region.
Fire was an important part of the historical disturbance regime across the entire region (Agee 1993), but its role varied geographically and temporally with periods of regional synchrony related to drought (Weisberg and Swanson 2003, Hessl et al. 2004, Wright and Agee 2004, Gedalof et al. 2005, Heyerdahl et al. 2008. Historical fire was infrequent in wet vegetation zones (Table 1). These vegetation zones are traditionally characterized as having a high-severity fire regime, except for the drier portions of the western hemlock vegetation zone in the southern portion of the Cascade Range, which experienced a mixed-severity regime (Agee 1993). Dry vegetation zones (Table 1) represent a broader range of historical fire regimes. These include low-severity in the ponderosa pine, mixedseverity in the Douglas-fir and grand fir/white fir vegetation zones (a.k.a. mixed conifer), and highseverity in the subalpine vegetation zone.

Development of burn severity maps
We created burn severity maps for all fires in Oregon and Washington from 1985 to 2010 with a mosaicked set of Landsat TM images from 1984 to 2011. Images were acquired from mid-July to late August and processed using the LandTrendr algorithm. LandTrendr is a temporal and spatial normalization process that uses a segmentation algorithm to smooth spectral reflectance of individual pixels (Kennedy et al. 2010). The process uses relative radiometric normalization and cloud screening rules to create mosaics of multiple images per year. Multiple images may be used in individual years to improve spatial coverage where pixel-level data are missing due to cloud cover or the Landsat 7 scan line failure. Consistency of seasonality trumps absence of clouds as the highest priority during image selection. The algorithm removes noise related to variability in illumination, atmospheric conditions, geometric resolution, and phenology. Thus, common post-processing practices to facilitate comparisons among multiple fires (e.g., the dNBR offset ; Miller and Thode 2007) are accounted for by the LandTrendr algorithm (see Kennedy et al. 2010 for more details).
Following pre-processing and scene selection, the normalized burn ratio (NBR) was calculated for each pixel where: The LandTrendr segmentation algorithm was then applied to the time series of NBR values at each pixel to create annual NBR grids. Annual NBR grids were rescaled to 0.81-ha grid cells by calculating the mean of a 3 9 3 neighborhood to approximate the footprint of the field plots (described below). We then used the annual NBR grids to calculate the relativized difference in the normalized burn ratio (RdNBR) following Miller and Thode (2007). We used NBR pre-fire (year À 1) and NBR post-fire (year + 1) to avoid situations where imagery was taken either before or after the time of burning. We clipped annual RdNBR layers with the perimeters of all fires ≥400 ha occurring between 1985 and 2010 from the Monitoring Trends in Burn Severity Program (available online: http://mtbs.gov; Eidenshenk et al. 2007). Manual inspection of burn severity maps revealed issues in nine fires associated with the Landsat 7 scan line corrector failure in 2003, so we selected new pre-or post-fire imagery and recalculated RdNBR for these fires.

Burn severity thresholds
To classify our RdNBR burn severity maps in terms of tree mortality, we quantified changes in live basal area from 304 plots with pre-and postfire field observations from the Current Vegetation Survey (CVS) inventory of USDA Forest Service Region 6. The CVS is a systematic regional inventory across all the forest lands administered ❖ www.esajournals.org 6 March 2017 ❖ Volume 8(3) ❖ Article e01695 by the USDA Forest Service in Oregon and Washington with the first measurement from 1992 to 1997 and the second from 1997 to 2007 (Max et al. 1996). Plots include four variable radius subplots within a circular 1-ha plot and are located on a 2.5 by 2.5 km grid, with the exception of federally designated wilderness areas where plots are located every 5 km. Plots were not necessarily visited immediately pre-and post-fire, so we only used plots with no evidence of a disturbance (except the fire of interest) after the initial plot measurement based on field notes and visual inspection of LandTrendr spectral trajectories. We calculated the percent of basal area mortality based on re-measurements of tagged live trees present at the initial sampling. We used regression analysis to assess the relationship between RdNBR and the percent of basal area mortality (RdBA). We evaluated multiple non-linear model forms used in other studies (e.g., Miller and Thode 2007, Cansler and McKenzie 2012) based on model significance, coefficient of determination (r 2 ), and visual inspection of model fit and residuals during preliminary analysis. We selected a second-order polynomial with the following form: We then classified annual RdNBR layers into severity classes with the following thresholds based on the percent of basal area mortality: low (<25%), moderate (25-75%), and high (>75%). Only five of the plots (1.6%) within fire polygons lacked reference to fire occurrence in field notes during the re-measurement interval and were actually unburned. We chose not to separate unburned from low-severity given the small sample size for validation.

Severity patch size class distributions
We used Fragstats (McGarigal et al. 2012) to calculate patch size distributions for each year of our study period. This was done for low-, moderate-, and high-severity patches for individual vegetation zones as well as all vegetation zones combined. Patches were delineated using a four-cell rule where individual patches were based on adjacency such that cells touching diagonally were not considered part of the same patch. We calculated the Gini coefficient for size class distributions of all patches of low, moderate, and high severity using the package "reldist" in R (R Development Core Team 2014). The Gini coefficient has been applied in previous studies of spatial patterns of burn severity (Cansler and McKenzie 2014) as a metric of inequality in patch size class distributions. Values closer to one have a greater proportion of total class area made up of larger patches, whereas values closer to zero have a greater proportion of total class area made up of small patches.

Drought data
We acquired gridded 4-km data for the Palmer Drought Severity Index (PDSI) (Palmer 1965) from the University of Idaho (available online: http://me tdata.northwestknowledge.net) that were created following Abatzoglou (2013). PDSI is a measure of meteorological drought based on precipitationand temperature-driven evapotranspiration (Littell et al. 2016) and represents climatic deviation from normal. Negative values indicate drier-thanaverage conditions, while positive values indicate wetter-than-average conditions. Past work has shown that PDSI promotes regional synchrony in historical fire activity across the Pacific Northwest (e.g., Hessl et al. 2004, Heyerdahl et al. 2008. For each year, we calculated a seasonal average of PDSI for the forested portion of the region and individual vegetation zones for January-March (winter), April-June (spring), July-September (summer), and October-December (fall).

Temporal trends in fire extent and fire rotation
We tested for temporal trends in total fire extent, and extent of low-, moderate-, and highseverity fire, using multiple regression, first for all vegetation zones combined and second for individual vegetation zones in a single global model. In all analyses, we began by evaluating temporal autocorrelation using autocorrelation function (ACF) plots. Because no temporal autocorrelation was observed, we proceeded with tests of normally distributed residuals and homoscedasticity. Since neither was met, we used a generalized linear model (GLM) on the count of hectares burned for each year. We observed overdispersion when using a Poisson distribution, so we tested for correlations using a negative binomial distribution ( trends were exclusively related to drought or additional temporal factors in both analyses. In the second analysis, we used vegetation zone as a categorical variable in the global model and tested all two-way interactions among these variables for statistical significance. We calculated total fire extent as well as the extent of low-, moderate-, and high-severity fires for the entire region and individual vegetation zones. We used this information to calculate fire rotation, or the time required to burn an area equal to the area of interest (Agee 1993). We estimated fire rotation by dividing the length of study interval by the proportion of the study area of interest burned during the interval. Fire rotation has been used to compare historical and contemporary fire extent (Stephens et al. 2007, Miller et al. 2012) and may be a better descriptor of landscape fire frequency than fire return interval (Agee 1993). We recognize the inherent limitations of the fire rotation metric which may be sensitive to large fire years over a relatively short study period (Miller et al. 2012). Nevertheless, fire rotation estimates provide us with a means to evaluate contemporary and historical fire extent under the assumption that fire history studies are representative of individual vegetation zones at broader scales.

Temporal trends in burn severity proportions and patch size distributions
We assessed temporal trends in the proportions and patch size distributions (Gini coefficients) at low, moderate, and high severity in relation to PDSI and total fire extent with multiple regression analysis. We first conducted these analyses for all vegetation zones combined for a regional estimate, and second in a global model using vegetation zone as a categorical variable. We conducted the same analyses for low, moderate, and high severity. Potential covariates included year, seasonal averages of PDSI with a lag of up to 1 yr, vegetation zone (except when analyzing combined regional estimates), and all two-way interactions. Because severity responses exhibited strong relationships with fire extent, we evaluated total fire extent in one model and year and PDSI in another model to isolate the effects of these variables. We log-transformed total fire extent to normalize these highly variable estimates. ACF plots indicated no temporal autocorrelation in any of the variables we tested. We used beta regression for these analyses because both response variables (severity proportion and Gini coefficient) were bounded by zero and one, and homoscedasticity is typically violated without complicated transformations (Ferrari and Cribaru-Neto 2004). We excluded years with no fire since proportions and Gini coefficients were contingent on some amount of fire. We used AIC values to determine the final model and included only covariates with a statistically significant effect at a ≤ 0.05. We report the final models only and derived all figures from these models, including intercept-only models when the relationship with a covariate of interest was not significant. We isolated the relationship with year and PDSI for figures by holding all other covariates at their mean values. These statistical models are reported on the scale of the logit link function used in beta regression.
Finally, we quantified the relationships between maximum patch size of low-, moderate-, and high-severity fires for all vegetation zones combined and then for each of the major vegetation zones. We excluded years with no fire since the maximum patch size was contingent on some amount of fire occurring. For temporal trends, we again evaluated temporal autocorrelation using ACF plots and did not observe temporal autocorrelation. We used a GLM on the maximum patch size count of hectares for each year. We observed overdispersion when using a Poisson distribution, so we tested for changes using a negative binomial distribution (Venables and Ripley 2002; R package for glm.nb). Potential covariates included year, seasonal averages of PDSI with a lag of up to 1 yr, vegetation zone (except when analyzing regional estimates), and all two-way interactions. Because maximum patch size exhibited strong relationships with fire extent, we evaluated total fire extent in one model and year and PDSI in another model to isolate effects of these variables. We isolated the relationship with year and PDSI for figures by holding all other covariates at their means and report model results on the log scale used in these GLMs. We also tested for the relationship between annual fire extent and maximum patch size using multiple linear regression. We log-transformed total annual fire extent and maximum patch size to normalize residuals and meet assumptions of homoscedasticity. We tested for interactions across vegetation zones, but did not observe significant associations.

Temporal trends in fire extent
Small but significant increases in annual fire extent were associated with drought in most vegetation zones (Table 3, Fig. 4). Although total fire extent across the region did not increase significantly, this was likely an artifact of reduced statistical power due to the small sample size (i.e., 26 yr) when all vegetation zones were combined.
Additionally, we found that in all cases, fire extent was negatively related to drought, with more area burned during warm, dry climatic conditions indicated by negative values of summer season PDSI (Table 3). Despite evidence of increasing fire extent over time, increases were slight compared to the area of individual vegetation zones and the total extent of the region. More importantly, fire extent was highly episodic, exhibiting high interannual variability associated with periods characterized by large fire events toward the end of the study period (e.g., 2002-2003Fig. 4). In particular, increases in fire extent in wet vegetation zones were driven largely by a few years of relatively extensive fire and several years in the early part of the time series with little or no fire.

Temporal trends in burn severity proportions and patch size distributions
Unlike fire extent, the proportion burned at any level of burn severity did not increase significantly during the study period for the region combined or any vegetation zone (Table 4, Fig. 5).   Notes: Field data come from 304 pre-and post-fire plots from the Current Vegetation Survey (see Methods). Overall accuracy was 70.1% (j = 0.48, P < 0.0001).
The average proportion for each vegetation zone did vary by forest type, with larger proportions of high-severity fire burning in wet vegetation zones relative to dry vegetation zones as expected given their respective historical fire regimes. For the region as a whole, greater proportions of moderate-and high-severity fires occurred during years with greater drought conditions (more negative PDSI), but this relationship varied among vegetation zones, as indicated by statistically significant interaction terms (Table 4, Fig. 5). For example, subalpine forests had larger proportions of high-severity fire (smaller proportions of low-severity fire) during wetter years, suggesting drought was not a consistent predictor of these proportions for all vegetation zones. Generally, the most significant relationship was between the severity proportions and annual fire extent both within vegetation zones and across the region. Years that exhibited high fire extent also had the greatest proportion of high-severity fire. For the combined regional estimates, these increases were reflected with a decrease in low-severity fire (Fig. 5).
Severity patch size distributions (Gini coefficients) did not vary significantly over time for any severity class when all vegetation zones were combined at the regional scale (Table 5, Fig. 6). However, the average Gini coefficient did vary by vegetation zone for all severity classes showing the benefits of describing these trends at this scale. In addition, Gini coefficients were highly correlated with annual fire extent; as total fire extent increased, so did the Gini coefficient for all severity classes. In terms of drought, the regional-scale Gini coefficient for high-severity fire exhibited the only statistically significant relationship with PDSI. Specifically, higher Gini coefficients were associated with increasingly droughty conditions. Variation among vegetation zones was evident, with the most rapid rate of increase observed for high-severity fire in the silver fir and mountain hemlock vegetation zones.
Maximum patch size for high-severity fire increased over time for all vegetation zones combined as well as for each individual major vegetation zones (Table 6, Fig. 7). However, low-and moderate-severity patches exhibited more variable responses among vegetation zones. Specifically, the maximum low-severity patch size decreased over time for all vegetation zones except ponderosa pine, grand fir/white fir, and mountain hemlock, and maximum moderate-severity patch size declined over time only for the subalpine and western hemlock vegetation zones. The maximum patch size for all severity classes increased with increasingly droughty conditions, with substantial variation in maximum size across all vegetation zones (Table 6, Fig. 7). The mountain hemlock,

Cumulative patterns of fire extent and burn severity
Total fire extent over the study period varied by vegetation zone, with dry vegetation zones experiencing greater extent and shorter fire rotations than wet vegetation zones (Table 7). The cold, dry subalpine vegetation zone experienced the greatest proportion of total extent burned (~15%) and had the shortest fire rotation (176 yr), which fell within historical estimates (109-275 yr; Table 1). Rotations in ponderosa pine (380 yr), Douglas-fir (265 yr), and grand fir/ white fir (277 yr) were longer than historical estimates (11-64 yr). Combined, these vegetation zones made up most of the total fire extent in the region, but each experienced fire across <10% of its individual total extent. Fire cumulatively affected <3% of any wet vegetation zone where current rotations exceeded 1000 yr and were far longer than most historical estimates (<600 yr).
The cumulative percent burned by low-and high-severity fire varied among vegetation zones, but the percent of moderate-severity fire was relatively similar across all vegetation zones (Fig. 8). Ponderosa pine and western hemlock had the lowest percentages of high-severity and greatest percentage of low-severity fire, whereas subalpine, mountain hemlock, and silver fir had the greatest percentages of high-severity and the lowest percentages of low-severity fire. Distributions of burn severity were similar in the Douglas-fir and grand fir/white fir zones, where most of the extent burned at low or moderate severity (Fig. 8).
Patch size class distributions for low-severity fire in dry vegetation zones indicated a greater proportion of fire extent in larger patches (Gini coefficients ≥0.8) than in wet vegetation zones (Fig. 9). In addition, the majority of low-severity patches (≥50%) in dry vegetation zones occurred in patches ≥100 ha, whereas most low-severity Notes: Douglas-fir is the reference group for estimates by vegetation zone. Parameter estimates with P ≤ 0.05 are indicated with Ã . Variables that were not included in the final models are indicated by n.a. Vegetation zones are abbreviated: WH, western hemlock; SF, silver fir; MH, mountain hemlock; SA, subalpine; GF, grand fir/white fir; and PP, ponderosa pine. patches in wet vegetation zones were ≤100 ha. Gini coefficients for moderate-severity patch size distributions were similar among most vegetation zones, but were highest in grand fir/white fir and Douglas-fir. Gini coefficients for highseverity patches were highest in subalpine and mountain hemlock (0.89-0.92) and lowest in western hemlock (0.69), with similar values for the remaining vegetation zones (0.82-0.85). In all vegetation zones, most of the extent burned at high severity occurred in patches ≥100 ha.

DISCUSSION
Our results depict a 26-year period characterized by slight increases in fire extent and patterns of burn severity that differed among vegetation zones and varied over time with drought. Cumulatively, fire affected <3% of wet vegetation zones and most dry vegetation zones experienced less fire than expectations from fire history studies. The relatively large proportion of high-severity fire (43-48%) occurring primarily in large patches ≥100 ha was consistent with expectations in vegetation zones with historically high-severity regimes, yet low and moderate severity still comprised~50% of the total extent burned. Low (45-54%)-and moderate-severity (24-36%) fires were prevalent in vegetation zones that historically experienced low-and mixed-severity regimes, but the proportion of high severity (23-26%), almost half of which occurred in large patches (≥100 ha), was much greater than expectations from most fire history studies in the region. This comparison of historical and recent fire activity is challenged by the relatively short period of our analysis and a lack of historical studies that address patterns of burn severity at correspondingly large spatial scales, but provides valuable new information to scientific and policy discussions about contemporary patterns of burn severity (e.g., Schoennagel et al. 2004, Noss et al. 2006. Our results support concerns about large patches of high-severity fire in some dry vegetation zones. However, they also suggest that spatial patterns of burn severity across much of the extent burned are generally consistent with current understanding of historical landscape dynamics in the region.

Recent trends of fire extent and severity
Although we found that total fire extent did not increase significantly across the entire region (Dillon et al. 2011), our analysis at sub-regional scales indicated increases in individual vegetation zones after controlling for inter-annual variability in drought (Fig. 4). Greater fire extent during drought is consistent with contemporary (Collins et al. 2006, Abatzoglou andKolden 2013) and historical studies that document regional synchrony in fire activity across most Notes: Douglas-fir is the reference group for estimates by vegetation zone. Parameter estimates with P ≤ 0.05 are indicated with Ã . Variables that were not included in the final models are indicated by n.a. Vegetation zones are abbreviated: WH, western hemlock; SF, silver fir; MH, mountain hemlock; SA, subalpine; GF, grand fir/white fir; and PP, ponderosa pine. vegetation zones in the region (Weisberg and Swanson 2003, Hessl et al. 2004, Wright and Agee 2004, Gedalof et al. 2005, Heyerdahl et al. 2008. Regardless, observed increases in fire extent were small compared to the total extent of most vegetation zones, and fire was less common than it would have been under the historical regimes in most dry vegetation zones. Our results are consistent with multiple studies that document a fire deficit in forested regions in the western United States (Stephens et al. 2007, Marlon et al. 2012, Mallek et al. 2013. Relationships between the Gini coefficient for size class distributions of patches of low (top), moderate (middle), and high severity (bottom) with year (left), Palmer Drought Severity Index (PDSI) (center), and total annual fire extent (right) for the major forested vegetation zones of the Pacific Northwest, USA, from 1985 to 2010. The Gini coefficient is a measure of equality in the size class distributions where values closer to 1 represent increasing dominance by larger patches. PDSI values are for July, August, and September with increasing drought conditions represented at lower values of PDSI. Vegetation zones are abbreviated: WH, western hemlock; SF, silver fir; MH, mountain hemlock; SA, subalpine; GF, grand fir/white fir; DF, Douglas-fir; PP, ponderosa pine; and All, all combined. et al. 2015), but also provide much needed context on how extent and severity vary among vegetation zones at sub-regional scales. Fire rotations (265-380 yr) were much longer than historical estimates in dry vegetation zones where historic fire regimes are characterized as low severity (ponderosa pine) and mixed severity (Douglasfir and grand fir/white fir). While high-severity fire affected a small proportion of these vegetation zones, the extent of low-and moderate-severity fires is insufficient to address the current fire deficit or restore fire as an ecological process at a meaningful scale. Fire rotation in the subalpine vegetation zone (176 yr) was within historical estimates (109-275 yr), but these estimates should be interpreted with caution due to the relatively small extent of the zone and disproportionate impact that large fire years can have on fire rotation (Miller et al. 2012). Long fire rotations (>800 yr) in wet vegetation zones (western hemlock, silver fir, and mountain hemlock) also suggest a fire deficit compared to the late 1800s, but are consistent with a period of limited fire from the 1600s to the 1800s (Weisberg and Swanson 2003).
Similar to a recent study across the western United States (Picotte et al. 2016), we found very little evidence of increases in the proportion of extent burned at any level of severity over time. However, our results indicate an increasing amount of high-severity fire occurring in large patches over time that was partially associated with drought and increasing annual extent burned, particularly at the regional scale. The relatively weaker and more variable relationships at the scale of vegetation zones may reflect the limitations of PDSI in vegetation zones where most precipitation falls as snow (Littell et al. 2016) and could also be related to a stronger role of topography and weather during individual events (Dillon et al. 2011). Although fires are not necessarily getting more severe with time per se (i.e., proportionally), it is evident that drought years with a greater fire extent have unique ecological effects including greater proportions and larger patches of high-severity fire (Lutz et al. 2009, Miller et al. 2012, Cansler and McKenzie 2014. At the same time, drought years also had the largest patches of low-and moderate-severity Table 6. Regression results and parameter estimates for the effects of a) time (year) and Palmer Drought Severity Index (PDSI), and b) total burned extent (LnTotal) on the maximum patch size of low-, moderate-, and highseverity fire from 1985 to 2010 in the Pacific Northwest, USA. Notes: Douglas-fir is the reference group for estimates by vegetation zone. Parameter estimates with P ≤ 0.05 are indicated with Ã . Variables that were not included in the final models are indicated by n.a. Vegetation zones are abbreviated: WH, western hemlock; SF, silver fir; MH, mountain hemlock; SA, subalpine; GF, grand fir/white fir; and PP, ponderosa pine.
fire, highlighting the potential ecological tradeoffs that occur during years with the greatest extent burned.

Spatial patterns of burn severity and contemporary landscape dynamics
Patterns of fire severity were highly variable among geographic locations at smaller spatial scales ( Fig. 10) but became more characteristic and differed among individual vegetation zones when examined cumulatively across the study period (Agee 1998). We found the greatest proportions of high-severity fire in colder and wetter vegetation zones (except for western hemlock) and the greatest proportions of low-and moderate-severity fires in warm and cool-dry vegetation zones. However, there was relatively little variability in the proportion of moderate-severity fire, which theoretically peaks and makes up the greatest proportion in mixed-severity fire regimes ( Fig. 1; Agee 1993). The occurrence of patches at all levels of burn severity in most size classes underscores the limitations of generalizing observations from a single fire or landscape-scale study to dynamics across an entire vegetation zone or region. Likewise, defining expectations from regional patterns to an individual landscape may be equally problematic, as multiple factors (e.g., fuels, weather, topography; Dillon et al. 2011) may contribute to substantial geographic variability in patterns of severity among fire events.
The warm, dry ponderosa pine vegetation zone appears to be the vegetation zone most departed from historical patterns of burn severity in the region. Low-severity (54%) and moderateseverity (24%) fires were generally consistent with a historically low-severity fire regime across the majority of the extent that burned. However, the 23% of the vegetation zone affected by highseverity fire is much greater than would be expected from a number of studies documenting frequent fire during the pre-settlement period ( Table 1). The percent of high-severity fire is also greater than estimates from some parts of the region during the late 1800s and early 1900s (12-16.5%; Hessburg et al. 2007, Baker 2012, Williams and Baker 2012, especially given substantial concern that reconstructions based on tree density and size (Baker 2012, Williams andBaker 2012) overestimate the amount of highseverity fire (Hagmann et al. 2013, Ful e et al. 2014). In addition, estimates derived from aerial photographs in Hessburg et al. (2007) include patches of early seral vegetation (e.g., shrublands, meadows), making these results difficult to compare with ours, which are based on tree mortality. Anecdotal observations of "extensive" patches of high-severity fire do exist from central Oregon during the early 20th century (Weaver 1959(Weaver , 1961. However, references to patches ≥100 ha, which comprised over half of the recent high-severity fire in this zone (Fig. 10c), are limited to a single~400-ha stand of even-aged ponderosa pine (Bork 1985). Mixed-severity fire regimes in the grand fir/ white fir and Douglas-fir vegetation zones are potentially the most complex yet least understood in the Pacific Northwest (Agee 1998, Halofsky et al. 2011, Perry et al. 2011. Low (~45%)-and moderate-severity (~30%) fires together comprised the majority of the fire extent in these vegetation zones, with most occurring in large patches (≥100 ha; Fig. 10b, d). Estimates of recent highseverity fire (25%) are greater than would be expected from a number of studies documenting frequent fire during the pre-settlement period (Table 1) but are similar to estimates of 23-24% in some portions of the region during the late 1800s and early 1900s (Hessburg et al. 2007, Baker 2012. These studies, however, have challenges or limitations when comparing with our results. Very large patches of high-severity fire (>1000 ha) observed during individual events (Fig. 10b, e, g) are uncharacteristic of historical fire regimes in these vegetation zones as evidence of high-severity patches during the pre-settlement era is limited to intermediate spatial scales (10-100 ha; Wright and Agee 2004). There is some evidence of the rare occurrence of extremely large (>10,000 ha) patches of high-severity fire in the wetter and cooler portions of these vegetation zones after Euro-American settlement (Leiberg 1903, Hessburg et al. 2007). However, these events are likely anomalous, as Leiberg (1903) explicitly refers to his observation as "the most thorough and complete sweep of a standing forest by fire that I have ever seen" (p. 274).
Greater proportions and large patches of highseverity fire in cold and wet vegetation zones are consistent with the role of high-severity fire in environments where ignition or conditions (e.g., drought) conducive to burning occur infrequently (Gedalof et al. 2005, Krawchuk andMoritz 2011). Similarly, very large patches of high-severity fire ( Fig. 10a, 8. Proportion of total fire extent burned at low (<25%), moderate (25-75%), and high severity (>75%) from 1985 to 2010 in the major forested vegetation zones in the Pacific Northwest, USA. Severity is based on percentage of change in basal area. Vegetation zones are abbreviated: WH, western hemlock; SF, silver fir; MH, mountain hemlock; SA, subalpine; GF/WF, grand fir/white fir; DF, Douglas-fir; and PP, ponderosa pine.  fires still account for about half of the extent burned in vegetation zones traditionally considered high severity (silver fir, mountain hemlock, subalpine). The higher prevalence of low-and moderate-severity fires in western hemlock (81%) was unique among wet vegetation zones (Fig. 8). However, most of the extent burned in this vegetation zone occurred in the southern portion of the western Cascades of Oregon (Fig. 10f), where the fire regime is characterized as mixed severity (Agee 1993, Weisberg 2004, Tepley et al. 2013. Challenges to comparing contemporary and historical fire dynamics Despite a growing network of tree-ring and fire history studies (Falk et al. 2011), our understanding of historical dynamics associated with fire remains incomplete in many regions (Conedera et al. 2009). One of the greatest challenges facing a comparison such as ours lies in uncertainty surrounding the role of anomalous high-severity events in vegetation zones that historically experienced low-and mixed-severity fire regimes. In these vegetation zones, most historical evidence of large patches (≥100 ha) of high-severity fire is limited to the mid-to late 19th and early 20th centuries. Given the known influence of human settlement during this period (e.g., Pyne 1982), these high-severity patches are not necessarily indicative of pre-settlement dynamics and may obscure comparisons and desired reference conditions for management (e.g., Cissell et al. 1999, Swetnam et al. 1999. Recent studies show promise for detecting large patches of high severity (Tepley and Veblen 2015) prior to settlement, but there is a need to further expand the spatial coverage of landscape-scale studies to better understand historical geographic variation within and among vegetation zones.
In addition to discrepancies in the spatial and temporal scales of observation, comparisons between contemporary and historical fire activity are further challenged by fundamentally opposing views of fire as an ecological process. Remotesensing studies typically view burn severity as a process associated with the mortality of live vegetation. These studies have high overall classification accuracy, but poor resolution on live residual structure and composition, especially for low and moderate severity. On the other hand, most fire history studies are based on residual structure and fire scars (Beaty and Taylor 2001) or cohort establishment (e.g., Heyerdahl et al. 2001, O'Connor et al. 2014, Tepley and Veblen 2015, Yocom-Kent et al. 2015 and view burn severity as a process related to survival or stand initiation. There is a need to reconcile these two approaches by increasing the ecological resolution of remotesensing studies and development of new methods (e.g., Tepley and Veblen 2015) for estimating patterns of burn severity at large spatial scales (e.g., relative proportions, patch sizes).

Implications and management considerations
Despite a long period of fire exclusion and the warmest decade (1998)(1999)(2000)(2001)(2002)(2003)(2004)(2005)(2006)(2007)) the region has experienced in the last century (Abatzoglou et al. 2014), our results suggest that much of the fire activity during our study period has potentially provided beneficial ecological effects in many forested vegetation zones. There is increasing evidence in the Pacific Northwest that low-and moderate-severity fires may be restoring some aspects of historical structure (Reilly and Spies 2015) and increasing within-stand heterogeneity , but additional fieldbased studies are needed to better assess correspondence between fire effects and management goals. High-severity fire provides habitat for some avian species (e.g., Hutto 2008) and may enhance biodiversity through the creation of structurally diverse, early seral conditions with biological legacies (i.e., high snag densities, large remnant trees), which are currently rare in the region (Reilly and Spies 2015). However, large patches of high-severity fire have uncharacteristic effects in dry vegetation zones that historically experienced low-and mixed-severity fire regimes, and losses of dense, old-growth habitats to high-severity fire have negatively affected species such as the northern spotted owl (Strix occidentalis caurina; Davis et al. 2016).
Future projections indicate increasing fire extent across the Pacific Northwest throughout the 21st century , Krawchuk et al. 2009, Littell et al. 2010, Rogers et al. 2011, Liu et al. 2013, Wimberly and Liu 2014. Our results suggest that increasing fire extent will likely be associated with larger patches of highseverity fire under the current management paradigm of aggressive suppression, where a small number of ignitions during extreme weather conditions escape initial attack and account for 98% of the total extent burned (Calkin et al. 2005). The risk of large patches of high-severity fire may be mitigated through landscape-scale strategies to increase heterogeneous fuel conditions, including reducing fire suppression under moderate weather conditions (Hessburg et al. 2016). Large high-severity patches may require long-distance dispersal events (e.g., Lesser and Jackson 2013) and be slow to regenerate in drier ponderosa pine-dominated forests (Dodson and Root 2013). These patches also may be at risk of transitioning to alternative, stable states such as shrub fields (e.g., Lauvaux et al. 2016) if re-burned at high severity. Despite this risk, more fire activity in the future could be ecologically beneficial given current deficits and the prevalence of low-and moderate-severity fire effects in recent fires across the dry vegetation zones. Studies from dry forests in other regions suggest that low-and moderateseverity fires have the potential to reduce the occurrence of high-severity fire in subsequent re-burns where fuels have increased (Parks et al. 2014, Coppoletta et al. 2016, Harvey et al. 2016b and that increases in fire frequency may potentially lead to decreased burn severity in the future (Parks et al. 2016).
The use of managed fire has been proposed as a means to expand the influence and benefit of existing and future fuels treatments and restoration efforts (North et al. 2015). A new large-fire management paradigm deemphasizing fire suppression and promoting variable response strategies to fire ignition while still protecting highly valued resources is necessary . Although drought is considered when making fire management decisions (Kolden and Brown 2010), information on the relationship between drought and burn severity has been lacking (Littell et al. 2016). Our results suggest that allowing wildfires to burn during cool, wet conditions may reduce the proportion and size of patches of high severity at broad regional scales. Differences between wet and dry vegetation zones are less clear, however, and more work is needed to better understand interactions with bottom-up drivers of burn severity at finer spatial scales (e.g., fuels, topography; e.g., Dillon et al. 2011).

CONCLUSIONS
Our study supports concerns about recent fire activity in some dry vegetation zones, but it also suggests that patterns of burn severity across much of the extent burned are generally consistent with current understanding of historical landscape dynamics in the region. Although the proportions and spatial patterns of burn severity vary tremendously with scale and among geographic areas, our results agree with field-based studies that most recent fire activity in the Pacific Northwest has been low and moderate severity and that high-severity fire has affected a relatively small proportion of most vegetation zones (Reilly and Spies 2015, Whittier and Gray 2016. While fires are not necessarily getting more severe with time per se (i.e., proportionally), it is evident that drought years with greater fire extent have unique ecological effects, including greater proportions and especially larger patches of high-severity fire. In wet forests where high-severity fire has been rare in recent decades, more high-severity fire could be ecologically beneficial by increasing the amount of structurally diverse, early seral habitats. In dry forests where fire extent was greater but still generally below what occurred historically, much recent fire activity has likely been ecologically beneficial despite the occurrence of large high-severity patches with little historical precedent. Our results support the use of managed wildfire (sensu North et al. 2015) to promote biodiversity and restore an important ecological process. Our study also highlights the need to recognize the trade-offs associated with the positive and negative ecological effects of large fire years. Integrating these trade-offs into forest policies and strategies intended to adapt to projected increases in fire activity will be essential for biodiversity conservation and the continued provisioning of ecosystem services during the 21st century. data, Zhiqiang Yang for assistance with PDSI data, and Dave Bell, Andrew Merschel, and Meg Krawchuk for comments that greatly improved the final version of manuscript.