Landscape-scale fuel treatment and wild ﬁ re impacts on carbon stocks and ﬁ re hazard in California spotted owl habitat

. Forest managers are challenged with meeting numerous demands that often include wildlife habitat and carbon (C) sequestration. We used a probabilistic framework of wild ﬁ re occurrence to (1) esti­ mate the potential for fuel treatments to reduce ﬁ re risk and hazard across the landscape and within pro­ tected California spotted owl ( Strix occidentalis occidentalis ) habitat and (2) evaluate the consequences of treatments with respect to terrestrial C stocks and burning emissions. Silvicultural and prescribed ﬁ re treat­ ments were simulated on 20% of a northern Sierra Nevada landscape in three treatment scenarios that var­ ied in the land area eligible for treatment. Treatment prescriptions varied with topography, vegetation characteristics, and ownership. We then simulated many wild ﬁ res in the treated and untreated landscapes. Additional simulations allowed us to consider the in ﬂ uence of wild ﬁ re size on estimated emissions. Treat­ ments constrained to the land area outside of spotted owl activity centers reduced the probability of burn­ ing and potential ﬁ re intensity within owl habitat and across the landscape relative to no-treatment scenarios. Allowing treatment of the activity centers achieved even greater ﬁ re hazard reductions within the activity centers. Treatments also reduced estimated wild ﬁ re emissions of C by 45 – 61%. However, emis­ sions from prescribed burning exceeded simulated reductions in wild ﬁ re emissions. Consequently, all treatment scenarios resulted in higher C emissions than the no-treatment scenarios. Further, for wild ﬁ res of moderate size (714 – 2133 ha), the treatment scenarios reduced the C contained in live tree biomass fol­ lowing simulated wild ﬁ re. When large wild ﬁ res (8070 – 10,757 ha) were simulated, however, the treatment scenario retained more live tree C than the no-treatment scenario. Our approach, which estimated terres­ trial C immediately following wild ﬁ re, did not account for long-term C dynamics, such as emissions asso­ ciated with post-wild ﬁ re decay, C sequestration by future forest growth, or longer-term C sequestration in structural wood products. While simulated landscape fuel treatments in the present study reduced the risk of uncharacteristically severe wild ﬁ re across the landscape and within protected habitat, the C costs of treatment generally exceeded the C bene ﬁ ts.


INTRODUCTION
Forest managers in fire-prone ecosystems seek to balance a complex set of sometimes competing objectives that include providing wildlife habitat, avoiding catastrophic disturbance, and support ing local economies. In recent years, maintaining and increasing the capacity of forests to store car bon (C) has been added to these considerations due to concern over the effects of rising atmo spheric greenhouse gas concentrations on the earth's climate. In dry forests across much of the western United States, meeting these objectives is complicated by the increasing area and severity of wildfires occurring in concert with climate change (McKenzie et al. 2004, Stephens 2005, Westerling et al. 2006, Miller et al. 2009. A high-visibility example of competing objec tives in forest management is spotted owl (Strix occidentalis occidentalis) conservation in California. The northern (S. occidentalis caurina) and Mexican (S. occidentalis lucida) spotted owl subspecies have been listed as Threatened under the Endangered Species Act. Management directives for the Cali fornia subspecies focus on conserving nesting and roosting habitat by identifying protected activity centers (PACs): sites that include 121 ha (300 ac) of the best-quality habitat near known nest sites (Verner et al. 1992). Given the multi-storied, dense canopy forest characteristics of nesting and roost ing sites, the potential vulnerability of PACs to high-severity fire is a challenge to owl conserva tion (Collins et al. 2010, Stephens et al. 2016b. While low-to moderate-severity wildfire within nesting and roosting habitat may not negatively impact owls in the short term (Bond et al. 2002), longer-term effects of high-severity wildfire can include significant habitat loss due to direct and indirect tree mortality (Gaines et al. 1997, Jones et al. 2016, Stephens et al. 2016b). However, due to uncertainty concerning the effects of fuels reduc tion activities, management options for reducing wildfire hazard within PACs are restricted to light prescribed burning, although some thinning is permitted in the wildland-urban interface (USDA Forest Service 2004).
There is concern that such constraints on man agement activities limit the effectiveness of landscape-scale treatments intended to reduce the threat of uncharacteristically severe wildfire (Col lins et al. 2010, Tempel et al. 2015. Fire modeling studies have shown that treating a portion of the landscape can alter simulated fire behavior within and outside of treated areas and that strategically locating fuel treatments across the landscape has the potential to maximize treatment benefits while minimizing area treated , Schmidt et al. 2008. Restrictions on fuel treat ment location and severity limit real-world appli cation of treatment optimization methods. Even so, there may be significant opportunity for active management outside of high-quality owl habitat on fire-prone landscapes , Prather et al. 2008, Gaines et al. 2010. Given their demonstrated ability to alter wild fire behavior and effects (Martinson and Omi 2002, Pollet and Omi 2002, Ritchie et al. 2007, Fulm e et al. 2012, fuel treatments that address accumulated fuels and reduce stand density (e.g., prescribed burning, forest thinning, mastication) are com monly applied in dry western forests where wild fires were once frequent. It is less certain how treatments influence C stocks, and how to maxi mize C storage in frequent-fire systems. In the absence of disturbance, untreated forests may sequester the most C (Hurteau and North 2009, Stephens et al. 2009. How ever, high-severity wildfires can rapidly convert C sinks to sources, and burned forests may continue to be C sources for decades (Dore et al. 2008(Dore et al. , 2012. Treatments can reduce wildfire emissions (Finkral and Evans 2008, Hurteau and North 2009, North et al. 2009a, Reinhardt and Holsinger 2010, Wiedinmyer and Hurteau 2010, North and Hur teau 2011 and may retain more live tree C post fire (Hurteau and North 2009, North and Hurteau 2011, Stephens et al. 2012). Yet fuel treatments are associated with significant C emissions, releasing C to the atmosphere during harvest operations, burning, and/or biomass transport, and the C cost of treating forest fuels may exceed its C benefits (Campbell et al. 2011, Campbell andAger 2013). The circumstances under which treatments might lead to a net gain in C have not yet been resolved.
Recently, as a result of concern over the C costs of fossil fuel use and the threat of wildfire, inter est in harvesting historically low-value woody biomass has increased (Evans and Finkral 2009). Utilizing forest biomass for energy pro duction can help to reduce the cost of fuel treat ments, support local economies, offset fossil fuel use, and reduce the C and smoke emissions associated with fuel treatments (Reinhardt et al. 2008). Concerns remain over the sustainability of biomass removals, funding, and the availability of markets (Evans and Finkral 2009).
The focus of our research was to (1) evaluate whether withholding some land area from treat ment influences potential wildfire hazard across the landscape and within California spotted owl habitat, (2) estimate the short-term C conse quences of treatments, and (3) quantify the bio mass harvested in treatments. We simulated fuels reduction treatments and wildfire in a northern Sierra Nevada study area that encompassed 61 spotted owl PACs. In order to evaluate the C bal ance of the treatment scenarios, we quantified the C contained in the forest biomass harvested in each treatment scenario, the C emitted during prescribed fire and wildfires, and the C remaining within onsite pools. We confined our analysis to the immediate changes in C stocks and emissions, but recognize that a full accounting of treatment effects would also include long-term C dynamics (e.g., Dore et al. 2008, Malmsheimer et al. 2011.

Study area
The study area was defined by a long term demographic study site for the California spotted owl (S. occidentalis occidentalis). The 55,398-ha area contains 61 owl PACs. The study area is located ~20 km west of Lake Tahoe in the northern Sierra Nevada, with elevation ranging from 300 to 2400 m. The climate is Mediter ranean, with warm, dry summers and cool, wet winters. Vegetation at lower elevations in the study area is montane mixed-conifer forest. The forest type is dominated by ponderosa pine (Pinus ponderosa Dougl.), Douglas-fir (Pseudo tsuga menziesii var. menziesii (Mirb.), sugar pine (Pinus lambertiana Dougl.), incense-cedar (Caloce drus decurrens [Torr.] Florin.), white fir (Abies concolor (Gord. and Glend.)), Franco), and California black oak (Quercus kelloggii Newb.). California red fir (Abies magnifica var. magnifica Andr. Murray) has a stronger presence above ~2000 m (Barbour and Minnich 2000), but the red fir forest type is present on only ~5% of our core study area.
One-third of the study area is privately held in a generally checkerboard pattern of ownership (Fig. 1). The remaining 37,120 ha is managed by the Tahoe and Eldorado National Forests. Young forests dominate private land in the study area due to historical and active logging, while inter mediate and mature forests are relatively abun dant on public land (Laymon 1988, Bias andGutim errez 1992).

Vegetation and fuels data
The vegetation classification map developed in Chatfield (2005) forms the basis of our study area. Using aerial photographs combined with field accuracy assessment, Chatfield (2005) digitized eight land cover classes consistent with the California Wildlife Habitat Relationships (CWHR; Mayer and Laudenslayer 1988) system. A descrip tion of the cover classes is provided in Table 1. From the resulting cover class map, we delin eated polygons to represent stands of similar vegetation composition and structure (n = 4470) based on aerial photographs and topography (Fig. 2).
Stands were populated with vegetation data collected in 2007 in 382 sampling plots located within 10 km of the study area's northern bound ary, based on the assumption that the characteris tics of the plots are representative of the study area. These vegetation data included tree species, heights, diameters, and crown ratios. See Collins et al. (2011) for a detailed description of data col lection. To populate stands in the core study area with plot data, we first assigned a Chatfield cover class to each sampling plot based on species com position, canopy cover, and tree diameter distri bution. We then used a Most Similar Neighbor procedure (Crookston et al. 2002) to select five nearest neighbor plots for each stand using the Random Forest method with the R package yaim pute (version 1.0-22; Crookston and Finley 2008). Variables used in identifying nearest neighbors were topographic relative moisture index, east ness, northness, slope, and elevation. Stands were populated with data only from plots belonging to the same cover class. In order to increase variabil ity in stand conditions, three of the five plots ini tially selected to represent each stand were chosen randomly to contribute data to the stand. Each plot contributed data to an average of 35.5 stands (range: 1-437).
The method in which surface fuels are repre sented for fire modeling has important implica tions for findings related to expected fire behavior and effects. Fuel models are representa tions of fuelbed properties such as the distribu tion of fuel between particle size classes, heat content, and dead fuel moisture of extinction for use in the Rothermel (1972) surface fire spread model. As representations, fuel models artifi cially constrain the variation in surface fuel conditions. In order to represent a range of pre treatment fuel conditions for fire modeling, we overrode fuel model assignments made by the Fire and Fuels Extension to the Forest Vegetation Simulator (FVS-FFE, Dixon 2002, Reinhardt andCrookston 2003) and selected two fuel models for each stand. Fuel models representing the low end of the range were assigned following the selection logic of Collins et al. (2011); high-end models were selected to amplify surface fire behavior relative to the low-end models (App endix S1: Table S1; Collins et al. 2013). This approach to assigning fuel models to stands has been demonstrated to result in modeled fire behavior that is more consistent with observed fire effects than default fuel model assignments (Collins et al. 2013). An alternative approach could be to use the Landfire surface fuel model layer (e.g., Scott et al. 2016). However, we opted to tie fuel model assignments to the specific forest structural characteristics for each stand (Lydersen et al. 2015) as represented by the imputed plots rather than the remotely sensed dominant vege tation characteristics captured by Landfire.
Study area data were processed in the western Sierra variant of FVS to obtain the data layers required for fire behavior modeling. Due to the potential for spurious fire modeling results near study area edges, we obtained additional canopy fuel and surface fuel data layers from Landfire (www.landfire.gov) for an area adjacent to the study area boundary defined by a 10-km mini mum bounding rectangle (Fig. 2). The reason for using Landfire data for the buffer area was that  (Chatfield 2005) within the core study area, stand polygons, and 10-km minimum bounding rectangle for fire spread modeling. See Table 1 for description of classes.
we did not have a vegetation map with a similar classification scheme and level of detail outside of our core study area (Fig. 2). We merged study area and Landfire data layers to build 90 9 90 m resolution landscape files for fire behavior mod eling in Randig, described below. This allowed us to include wildfires originating outside of the study area in our analysis.

Wildfire, fuel treatments, and carbon loss modeling
We used ArcFuels (Ager et al. 2006) to stream line fuel treatment planning and analysis of effects. ArcFuels is a library of ArcGIS macros that facilitates communication among the array of models and other programs commonly used in fuel treatment planning at the landscape scale (vegetation growth and yield simulators, fire behavior models, ArcGIS, and desktop software). Our process, depicted in Fig. 3, involved: 1. fire behavior modeling in Randig (Finney 2006) to identify stands with high fire hazard; 2. prioritizing stands for treatment using the Landscape Treatment Designer (LTD) (Ager et al. 2012); 3. modeling fuel treatments in FVS-FFE; 4. fire behavior modeling for the post-treatment and untreated landscapes; and 5. developing C loss functions from simulated burning with FVS-FFE.
Conditional burn probability and flame length.-Wildfire growth simulations were performed in Randig, a command-line version of FlamMap (Finney 2006). Randig uses the minimum travel time algorithm (Finney 2002) to simulate fire growth during discrete burn periods under con stant weather conditions. Simulating many burn periods with Randig generates a burn probability surface for the study landscape. Simulations were conducted at 90-m resolution for computa tional efficiency. We simulated 80,000 randomly located ignitions with a 5-h burn period for all scenarios, including no treatment. The burn per iod was selected to produce fire sizes that approximated area burned in spread events of historical large wildfires near the study area. Large daily spread events in previous wildfires in the northern Sierra Nevada have burned >2000 ha (Dailey et al. 2008, Safford 2008); average fire sizes from our simulations ranged from 715 to 2133 ha. (The exceptional growth observed in the 2014 King Fire is addressed in a subsequent subsection.) The combination of igni tion number and burn period was sufficient to ensure that 99% of pixels in burnable fuel types experienced fire at least once (average: 64-1891 fires).
Randig outputs were used both in prioritizing stands for treatment and in evaluating the effects of treatment. We performed one Randig run for each fuel model range (low and high) within each scenario (no treatment, S1, S2, and S3) using land scape files representing the year immediately fol lowing treatment, 2009. Simulations were also completed for the 2007 pre-treatment landscape for use in treatment prioritization, for a total of 10 modeling runs.
To evaluate the effect of treatments on fire risk and fire hazard, we assessed changes in condi tional burn probability (CBP) and conditional flame length (CFL) between the treatment scenar ios and the untreated landscape based on wildfire simulations. It is important to note that the burn probabilities estimated in this study are not empir ical estimates of the likelihood of wildfire occur rence (e.g., Preisler et al. 2004, Brillinger et al. 2006, Parisien et al. 2012). Rather, we use CBP, the likelihood that a pixel will burn given a single ignition in the study area, and assuming the simu lation conditions described. From the simulation of many fires, Randig calculates a pixel-level dis tribution of flame lengths (FL) in twenty 0.5-m classes between 0.5 and 10 m. Conditional flame length, the probability-weighted FL given that a fire occurs , was calculated by combining burn probability estimates with FL dis tributions summarized at the stand level: ( ) where BP is CBP, BP i is the probability of burning at the ith FL class, and F i is the midpoint FL of the ith FL class.
To estimate the effect of treatment on fire risk and hazard, we first computed average pixel-level BP and CFL for treated and untreated stands in each scenario. Then, we calculated average BP and CFL for the same stands within the no-treatment landscape. The effect of each treatment scenario was estimated as the proportional change in each fire metric between the untreated and treated landscapes.
We obtained weather and fuel moisture inputs for wildfire modeling from the Bald Mountain and Hell Hole remote automated weather stations (RAWS), based on recommendations from local USDA Forest Service fire and fuel managers. We used 95th percentile weather conditions from the 1 June to 30 September period . This period represents the typical fire season for the study area, encompassing 85% of wildfires and 93% of the area burned within a 161-km (100-mi) radius of the study area between 1984 and 2012 (Monitoring Trends in Burn Severity database, Eidenshink et al. 2007).
Weather and fuel moisture inputs for wildfire simulations are provided in Appendix S1: Table S2. These conditions are similar to those occurring during recent large wildfires in and near the study area (e.g., 2001Star Fire, 2008 American River Complex, 2013 American Fire). In addition to using Randig to model fire spread and intensity, we used FVS-FFE to project effects of prescribed fires and wildfires (described below). Wind inputs varied somewhat between fire models: FVS-FFE requires only a single wind speed, while multiple wind scenarios were applied in Randig fire simulations. Wind speeds, azimuths, and relative proportions for Randig simulations followed Collins et al. (2011).
Spatial optimization of fuel treatments.-Stands were selected for treatment based on modeled pre-treatment wildfire hazard and stand structure using the LTD, which allows multiple objectives to be combined in the spatial prioritization of fuel treatments. Three treatment scenarios varied in the land designations eligible for treatment: Scenario 1: Public land, excluding spotted owl habitat Scenario 2: Public land, including spotted owl habitat Scenario 3: All lands: public and private ownerships Objectives were consistent across treatment sce narios, but differed in the land area available for treatment. For all LTD runs, we directed the model to maximize a total score that comprised numeric stand structure and fire hazard rankings (Appendix S1: Table S3). The stand structure rank ing (0, 1, 2) was based on cover class category: Cover classes most conducive to thinning were ranked highest. Fire hazard ranking (0, 2, 3) was assigned according to stand-level CFL as calcu lated from FL probability files generated in Randig simulations for the 2007 pre-treatment landscape.
To isolate the effect of varying land designa tions in the area available for treatment, total area treated was held constant between scenarios (20% of the core study area). In order to exclude small, spatially isolated treatment areas that would be impractical from a management stand point, we required a minimum treatment area of 12.1 ha (30 ac). To achieve this, the treatment pri oritization process was iterative. In each step, we eliminated all stands selected by LTD for treat ment that were not contiguous with a ≥12.1-ha treatment area. The rationale for this is based on the cost associated with re-locating equipment necessary to implement mechanical and/or fire treatments (D. Errington, personal communication, El Dorado National Forest). We then calculated the treatment area remaining. This process was repeated until total treatment area summed to the target area (~11,080 ha).
We simulated fuel treatments using FVS-FFE. Stands selected for treatment were assigned one of 13 treatment prescriptions depending on topog raphy, vegetation cover class, ownership, and overlap with owl PACs (Appendix S1: Table S4).
In an effort to promote landscape-scale hetero geneity, basal area targets for commercial thinning on public land varied with topography (aspect and slope position: canyon/drainage bottom, midslope, and ridge) (North et al. 2009b, North 2012. All thinning treatments were simulated as thin from-below harvests, and thinning within owl PACs was limited to hand thinning. We assumed that trees ≥25.4 cm (10 in) dbh would be har vested for wood products (FVS VOLUME key word) and that the biomass contained in smaller trees and in the tops and branches of larger trees would be utilized as feedstocks for bioenergy con version. Therefore, all thinning (except hand thin ning) treatments were simulated as whole tree harvests (FVS keyword YARDLOSS). Treatments preferentially retained fire-resistant species, with relative retention preference as follows: black oak>ponderosa pine>sugar pine>Douglas-fir>in cense-cedar>red fire>white fir.
Prescribed fires were simulated in the year fol lowing thinning (2009). Broadcast burning was applied except within owl PACs, on private land, and on steep slopes (>35%), where follow-up burning was limited to pile burning. To capture a more realistic range of post-treatment surface fuel conditions, stands selected for treatment were randomly assigned to one of three posttreatment fuel models for each fuel model range: TL1 (181), TL3 (183), or TL5 (185) (low range); TL3 (183), TL5 (185), or SB1 (201) (high range) (Scott and Burgan 2005). Weather conditions for prescribed fire modeling were based on recom mendations from a local fire management spe cialist (B. Ebert, personal communication).
Biomass and carbon effects of treatment.-Simu lated treatment prescriptions varied according to site characteristics such as topography and land ownership (Appendix S1: Table S4). We tracked the C emitted from burning, removed during har vesting, and contained in live and dead above ground biomass with FVS-FFE carbon reports Crookston 2003, Hoover andRebain 2008). FVS converts biomass to units of C using a multiplier of 0.5 for all live and dead C pools (Penman et al. 2003) except duff and litter pools, for which a multiplier of 0.37 is applied (Smith and Heath 2002). Stand C is partitioned into a number of pools including aboveground live tree, standing dead tree, herb and shrub, litter and duff, woody surface fuel, and belowground live and dead tree root C; we limited our analysis to aboveground pools of C. FVS-FFE also reports the C emitted during burning and that con tained in harvested biomass (Rebain et al. 2009). Treatment effects were assessed by comparing expected aboveground biomass C and emissions between the treated and untreated landscapes.
We developed C loss functions for each FVS treelist by simulating burning with FVS-FFE at a range of FLs (SIMFIRE and FLAMEADJ key words) .
The FL values supplied to FLAMEADJ were the 20 midpoints of the 0.5-m FL classes (0.5-10 m) found in Randig FL probability output files. As noted by Ager et al. (2010) and Cathcart et al. (2010), it is not currently possible to precisely match fire behaviors between Randig and FVS. The FLs reported in Randig outputs are the total of surface fire and, if initiated, crown fire. In contrast, the FLs supplied to FVS-FFE via the FLAMEADJ keyword are treated as surface fire FLs, and when FLAMEADJ is parameterized with only a predefined FL, the model does not use the input FL in crown fire simulations. To estimate fire effects in FVS-FFE, we parameterized FLA MEADJ with percent crowning (PC) and scorch height in addition to FL. Scorch height and critical FL for crown fire initiation (FLCRIT) were based on Van Wagner (1977). We estimated PC using a downward concave function where PC = 32% when flame length = FLCRIT and PC = 100% when FL is ≥30% of stand top height (the average height of the 40 largest trees by diameter) A. Ager, personal communication).
The derived C loss functions were combined with the probabilistic estimates of surface fire behavior produced in Randig simulations to esti mate the "expected C" emitted in wildfire or con tained in biomass. We estimated expected C emissions and post-fire biomass C for each pixel as follows: where E[C] j is the expected wildfire emissions of C from pixel j, or biomass C in pixel j, in mass per unit area; BP ij is the probability of burning at the ith FL class for pixel j; and C ij is the C emit ted from pixel j, or the biomass C remaining in pixel j post-wildfire, given burning at the ith FL class. Expected C emissions and biomass C were summed across all pixels in the core study area to obtain total expected wildfire emissions and expected terrestrial C for each treatment scenario.
In order to compare our modeling results to other analyses that reported wildfire emissions on a per area basis, we used a different method to esti mate C emissions per area burned. Because wild fires burned both the core and buffer areas of our study landscape while emissions were estimated only in the core area, we used conditional expected wildfire emissions to approximate the emissions from a wildfire burning entirely within the core study area. Conditional expected emissions are those produced for an area given that the area is burned. Conditional emissions were estimated for each pixel as follows: where C[WC] j is the C emitted by wildfire from pixel j in mass per unit area; BP j is the probability that pixel j is burned; BP ij is the probability of burning at the ith FL class, and WC ij is the C emitted from pixel j when burned at the ith FL class. Conditional expected emissions were averaged across all pixels to obtain an estimate of wildfire emissions per area burned.
Large fire revision. -Wildfire modeling was cali brated to produce fire sizes that approximated area burned in spread events of historical large wildfires near the study area. However, during the course of the study, a very large fire encoun tered our study area. The King Fire began on 13 September 2014 in El Dorado County and burned 39,545 ha-more than an order of magni tude greater than our modeled wildfires, includ ing >25% of the study area. Given the potential for very large wildfires in this region demon strated by the King Fire, we completed addi tional wildfire modeling to estimate the C effects of treatment given the occurrence of a very large fire.
Randig modeling was repeated for the notreatment and S3 scenarios using the high fuel model range and a revised burn period, number of simulated ignitions, wind speed, and wind directions. Burn period was increased from 5 to 12 h; number of ignitions was reduced by half to 40,000. Wind directions and relative probabilities (Appendix S1: Table S5) were those recorded at Hell Hole RAWS between 04:00 and 19:00 hours on 17 September, the day of the King Fire's lar gest spread event. We used the probable 1-min maximum wind speed as calculated from the maximum gust recorded on that day: 33 km/h (20.5 mph), based on maximum gust of 54.7 km/h (34 mph) (Crosby and Chandler 1966). These settings produced average fire sizes of NT = 10,757 ha (no-treatment scenario) and 8070 ha (S3). Average fire size was limited by the size of our buffered study area: Longer burn periods resulted in an increasing number of simulated wildfires that burned to the study area boundary. Table 2 provides a summary of the area trea ted in each scenario. Scenario 1 (S1) was the most restrictive with respect to the land area available for treatment, which more than dou bled between S1 and S3. Because treatment pre scriptions varied with land designation (public, owl PAC, private), and the designations avail able for treatment varied between scenarios, the relative proportions of thinning and burning methods also varied between scenarios. Com mercial and biomass thinning were applied most frequently in S3, which permitted treat ment of private land. Spotted owl activity cen ters composed 25% of the area treated in S2 vs. 10% in S3 and 0% in S1, the scenario in which PACs were not subject to treatment. As a result, the area treated with hand thinning in S2 was more than twice that in S1 and S3. Due to the Fig. 4. Low fuel model range treatment locations and difference in conditional burn probability (CBP) and con ditional flame length (CFL) (untreated-treated) for each treatment scenario. Negative values indicate an increase in CBP or CFL, while positive values represent a reduction. CBP is the likelihood that a pixel will burn given a single ignition on the landscape and assuming the simulation conditions described in Appendix S1: Table S1 and in the text. Conditional flame length is the probability-weighted flame length, given these same assumptions.

Treatment simulation
inclusion of PACs in S2 and both PACs and pri vate land in S3, the proportion of area treated with pile burning increased between S1 and S3, while broadcast burn area exhibited an opposite trend. Despite the variation in land designations available for treatment, the pattern of treat ment placement was generally similar between scenarios, with treatments concentrated in the central and eastern portions of the study area (Figs. 4, 5).

Landscape-scale burn probability and fire hazard
Conditional burn probability.-The pixel-to-pixel change in CBP between the untreated scenario and each treatment scenario is mapped in Figs. 4 (LO FM) and 5 (HI FM). Treatment reduced land scape burn probability by approximately 50% (Table 3), from 0.0124 (NT) to 0.0062 (S1), 0.0059 (S2), and 0.0055 (S3). Within treatment units, aver age CBP fell by 69-76% to 0.0033-0.0035; outside of treated stands, CBP fell to 0.0060-0.0069. Some increases in CBP were also observed, particularly for the low fuel model range (Fig. 4).
The influence of treatment on owl PAC likeli hood of burning was similar to that observed for stands in general. For treated PACs, average CBP fell by ~70% relative to no treatment for the same stands. Although PACs were not eligible for treatment in S1, all treatment scenarios had a large impact on estimated PAC CBP. Average PAC CBP was reduced from 0.013 to 0.0063 in S1, 0.0049 in S2, and 0.0054 in S3, a 49-64% decrease relative to PACs in the no-treatment landscape (Table 3).
Fire hazard.-Treatments reduced average land scape CFL by ~1 m, from 3.6 m (NT) to 2.5-2.7 m. Pixel-level CFL was reduced by a maximum of 8.0 m (LO FM) and 9.0 m (HI FM). Increases in CFL were also observed, however, particularly near the study area's western and southwestern boundaries where treatments were least concentrated (Figs. 4, 5). Maximum pixellevel CFL increases were 2.5 m (LO FM) and 3.1 m (HI FM).
Because fire hazard was used in prioritizing stands for treatment, the estimated pre-treatment CFL in stands selected for treatment (4.3-5.1 m) was greater than in stands not selected (3.2-3.3 m). After treatment, average CFL within trea ted stands fell to 1.3 (S1 and S2) and 1.7 m (S3). CFL in untreated stands was also reduced as a result of the influence of treatments on fire spread and intensity. CFL fell by 0.5-0.8 m (9-16%) relative to CFL in the same stands within the no-treatment landscape (Table 4).
Although spotted owl PACs were not treated in S1, relative to PACs in the NT landscape, PAC CFL was reduced by 10% (to 3.2 m) in S1. Treating PACs had a much larger impact on potential fire intensity, however. Average trea ted PAC CFL fell to 1.3 and 1.4 m in S2 and S3, respectively.

Carbon consequences of landscape fuel treatments
Prior to treatment, aboveground landscape carbon totaled 147.05 tonnes/ha, on average. Treatments removed 14% of pre-treatment C from treated stands, or 23.74 tonnes/ha, totaling 81,772-119,103 tonnes of C in harvested biomass and merchantable material (Tables 5 and 6).
Both the treatment scenarios and the choice of fuel models were important influences on esti mated C emissions from burning. As the least restrictive treatment scenario in terms of treat ment location and the only scenario to include treatment of private land, where broadcast burn ing was precluded as a treatment option, the S3 treatment scenario was associated with the low est wildfire and prescribed burning emissions (Tables 5 and 6). For each treatment scenario, expected wildfire emissions increased by more than an order of magnitude between the low and high fuel model ranges. This difference was the result of increasing fire intensity as well as wild fire size. Average wildfire size nearly doubled between fuel model ranges in the treatment sce narios and tripled in the no-treatment scenario (Fig. 6). For a given treatment scenario, wildfire emissions on a per hectare basis were approxi mately two tonnes greater for the high fuel  Notes: LF indicates the large fire scenarios. Expected C is that remaining in the core study area following treatment, if appli cable, and a random ignition and wildfire in the larger buffered study area, as estimated from the simulation of many wildfires. Live C is that contained in live aboveground herb, shrub, and tree biomass; dead C is the C contained in litter, duff, woody sur face fuel, and aboveground portions of tree snags. NT, no treatment. model range than for the low range. In contrast to the large influence of fuel model choice on wildfire emissions, the effect of fuel model range on prescribed fire emissions was minimal, with only a 1% increase in emissions between the low and high fuel model ranges for a given treatment scenario.
Although treatment significantly reduced wild fire emissions, combined emissions from pre scribed burning on 20% of the landscape and wildfire exceeded wildfire emissions in the notreatment scenarios (Tables 5 and 6). Relative to the no-treatment scenarios, treatment reduced estimated wildfire emissions by approximately 54% (low fuel model range), 59% (high fuel model range), and 45% (large fire scenarios). Yet pre scribed burning was a far more significant source of emissions than were wildfires of moderate size, with emissions from treatment exceeding wildfire emissions by 111,259-177,344 tonnes. Even for the large wildfire simulations, where landscape treatments nearly halved estimated wildfire emis sions, the combined carbon emissions from pre scribed burning and wildfire in the treatment scenario surpassed wildfire emissions in the notreatment scenario by 45% (Table 6, Fig. 7).
The total quantity of aboveground C expected to remain on the landscape following treatment and a randomly ignited wildfire was greatest for the no-treatment scenarios (Tables 5 and 6). For modeled wildfires of moderate size, treatment reduced both the live and dead C pools relative to the no-treatment scenarios, and terrestrial C in the no-treatment scenarios was 4-5% greater (323,316 -434,960 tonnes) than in any of the treatment scenarios (Tables 5 and 6). In comparison, under large wildfire conditions, the treatment scenario retained slightly more live biomass C:~15,000 ton nes, or 0.3% more than the no-treatment scenario. However, treatment also reduced necromass C by 288,000 tonnes (12%), resulting in a 3% overall decrease in onsite biomass C relative to an untre ated landscape (Table 6).
The proportional changes in aboveground biomass C pools between the treatment and notreatment scenarios are summarized in Table 7. For all treatment scenarios, the consumption of duff, litter, and downed woody fuels with pre scribed burning contributed to a net reduction in these C pools relative to the untreated landscape. Conversely, treatments protected more C in the live understory (herb and shrub) pool-the result of reduced wildfire size and intensity in the treat ment scenarios. Treatments in the moderate wild fire scenarios reduced live tree biomass C in comparison with no-treatment levels (Fig. 8). Table 6. Expected biomass carbon, expected wildfire C emissions, and C harvested and emitted in fuel treat ments for NT and treatment (S1, S2, S3) scenarios using the high fuel model range in fire modeling. Notes: LF indicates the large fire scenarios. Expected terrestrial C is that remaining in the core study area following treat ment, if applicable, and a random ignition and wildfire in the larger buffered study area, as estimated from the simulation of many wildfires. Live C is that contained in live aboveground herb, shrub, and tree biomass; dead C is the C contained in litter, duff, woody surface fuel, and aboveground portions of tree snags. NT, no treatment.
Notably, in the large modeled wildfire scenarios (NT-LF and S3-LF), treatments resulted in a 400,000-tonne increase in landscape-level live tree C over the no-treatment scenario.

Fuel treatments in protected habitat
Because there are often competing objectives between managing forests for resilience to fire and drought and protecting owl habitat, we assessed potential fire occurrence and hazard based on treatment scenarios that included and omitted treatment of PACs. Conducting fuels management outside of occupied owl habitat has been suggested as a means of reducing fire risk within occupied sites (Jenness et al. 2004, Tempel et al. 2015.  reported that fuel treatments on 20% of a western Oregon landscape reduced the probability of northern spotted owl nesting and roosting habitat loss by 44%, even though that habitat type was not treated. As in , we observed modifications in fire intensity and burn probability within owl habitat even when it was left untreated. In the S1 treatment scenario, in which owl activity centers were not eligible for treatment, the effect of treat ing other stands reduced both fire hazard (by 9-12% for CFL, or approximately 0.4 m) and CBP (by ~45%) within PACs. It is difficult to assess the significance of this proportional reduction in CBP given that the absolute differences in probabilities were relatively modest (Figs. 3, 4).  noted that allowing treatment of owl habi tat would have significantly reduced estimated habitat loss in their study. In this study, it is expected that habitat quality may be reduced within treated PACs in the short term through the removal of small-diameter trees (i.e., lower vertical structural heterogeneity). However, two structural attributes for suitable spotted owl nest ing habitat identified by Tempel et al. (2015), high canopy cover and large tree density, would not be altered. While we did not directly estimate habitat loss, we did observe much larger reductions in fire hazard within PACs that were treated as mea sured by CFL (71-75% reduction, equivalent to 2-3 m). It should be noted that the effect of wide spread treatments within PACs on spotted owl nesting and foraging behavior is unknown.
One concern with designating some land area unavailable for treatment is that it may limit the potential for treatments to alter fire behavior across the landscape (e.g., Finney 2001). Including all stands in the potential treatment pool allows the highest-priority stands, with respect to simulated effects on landscape-level fire behavior. The true effect of increasing the land area available for treat ment may be partially obscured by the varying frequency of treatment prescriptions between sce narios. For example, the hand thinning treatments applied within PACs would be expected to have a milder effect on potential wildfire behavior than more severe prescriptions, and hand thinning was twice as common in S2 as in the other scenarios.

Terrestrial carbon and burning emissions
Landscape treatments reduced wildfire emis sions by reducing the emissions produced per area burned by wildfire as well as average wildfire size. On average, wildfires in the treated landscapes released 19.3-21.6 tonnes C/ha, while wildfires in the untreated landscapes released 23.4-25.4 tonnes C/ha. Modeled wildfires decreased in size by 7% (low fuel model range), 36% (high fuel model range), and 25% (large fire scenario) relative to untreated landscapes (Fig. 6). Since the burn per iod for simulated wildfires was held constant between scenarios, this reduction in average wild fire size is the result of reduced spread rates derived from fuel treatments.
Despite the influence of treatments on wildfire intensity, size, and expected emissions, treatmentrelated emissions exceeded the avoided wildfire emissions conferred by treatment. Prescribed burning in our study, a combination of broadcast Fig. 6. Wildfire size relative frequency distributions from wildfire simulations. Bar color represents notreatment (NT) and treatment scenarios (S1-S3).
fire spread, to be treated, which would be expected to achieve the greatest modification of landscape fire behavior and effects. In the present study, although the land area potentially available for treatment more than doubled between S1 and S3, landscape-level effects of treatment on modeled fire risk and hazard were fairly similar (compared with the no-treatment landscape, all-stand CBP fell by 47% and 54% in S1 and S3, while CFL fell by 24% and 30%). Dow et al. (2016) also found that incorporating modest restrictions on treatment Fig. 7. Carbon emissions (tonnes) from wildfire and area availability (24% of the landscape unavail-prescribed burning. X-axis labels indicate no-treatment able) had minimal consequences for modeled fire (NT) and treatment scenarios (S1-S3); subscripts size and hazard. The modest changes in estimated denote fuel model ranges used in fire modeling fire metrics we observed may also be due to simi-(L: low, H: high). Large fire scenarios, which were larity in the general pattern of treatment placement modeled with the high fuel model range only, are indi between scenarios, which probably led to similar cated by LF. Notes: For example, a value of -0.10 represents a 10% dec line in biomass C from the no-treatment scenario. Treatment and no-treatment values were calculated as the average of low and high fuel model range values, except in the case of the large fire (LF) scenarios, which were modeled for the high fuel model range only. Expected C is that remaining after a random ignition and wildfire in the buffered study area as estimated from simulating 80,000 ignitions (LF: 40,000 ignitions).
C pool categories are those reported in Forest Vegetation Simulator Carbon Reports. Standing dead: aboveground por tion of standing dead trees, Down dead wood: woody surface fuels, Forest floor: litter and duff, Herb/shrub: herbs and shrubs, Live tree: aboveground portion of live trees. and pile burning, released 11.1-16.3 tonnes C/ha. For comparison, studies conducted in comparable forest types have estimated prescribed fire emis sions of 12.7 tonnes C/ha (warm, dry ponderosa pine habitat types; Reinhardt and Holsinger 2010) Fig. 8. Expected carbon contained in aboveground live and dead tree biomass. Expected C is that remain ing in the core study area following treatment (if appli cable) and a single random ignition within the larger buffered study area. X-axis labels indicate notreatment (NT) and treatment scenarios (S1-S3); sub scripts denote fuel model ranges used in fire modeling (L: low, H: high). Large fire scenarios, which were modeled for the high fuel model range only, are indi cated by LF. and 14.8 tonnes C/ha (an old-growth mixedconifer reserve in the southern Sierra Nevada; North et al. 2009a). Relative to the approximately 158,000 tonnes C emitted in prescribed burning, avoided wildfire emissions, at 1186-19,551 tonnes for wildfires of moderate size, were small. A simi lar study in southern Oregon with average mod eled wildfires of 2350 and 3500 ha (treatment and no-treatment scenarios, respectively) found that treatments reduced expected wildfire emissions by 6157 tonnes of C .
Surface fuels, represented with surface fuel models in commonly used modeling software, are the most influential inputs determining pre dicted fire behavior (Hall and Burke 2006). Fire behavior, fire sizes, and emissions in this study varied according to fuel model assignment, high lighting the importance of selecting the appropri ate fuel model to represent fuel conditions (see Collins et al. 2013). We show a 12-to 14-fold change in wildfire emissions due solely to the choice of fuel models (Tables 5 and 6). Indeed, the range of fuel models used in recent studies investigating fuel treatments and simulated fire behavior in mixed-conifer forests is noteworthy. Incorporating a range of fuel models into analy ses such that outcome variability can be reported facilitates comparison of effects across studies.
Our estimates of the aboveground C benefits of treatment under the moderate wildfire scenar ios, with average fire sizes of ≤2133 ha, are likely conservative. The effect of modeled wildfire size on the C consequences of fuel treatment was con siderable, emphasizing the importance of this variable in studies of the climate benefits of treat ment. Avoided wildfire emissions resulting from treatment increased to 61,276 tonnes C when large wildfires (8070-10,757 ha) were simulated. The treatment scenario, given large wildfires, also protected a greater portion of live tree C. If the ~40,000-ha King Fire is representative of the magnitude of future wildfires in the region, C accounting should improve with respect to treat ment favorability. Similarly, if multiple wildfires were to encounter the study area within the effective lifespan of treatments, the C gains asso ciated with avoided emissions in the treatment scenarios would increase.
Our approach to estimating the C conse quences of fuel treatments has a number of limita tions. A full accounting of treatment effects would project through time the consequences of both treatment and wildfire. Our analysis is static, incorporating only the short-term C costs and benefits of treatment. Simulating wildfire in the year immediately following treatment maximizes the apparent benefits of treatment. Over time, as surface fuels accumulate and vegetation regener ates, maintenance would be required to retain the effectiveness of treatments (Martinson and Omi 2013), increasing the C costs of reduced fire hazard. In addition, the C contained in fire-killed biomass will ultimately be emitted to the atmo sphere, although biomass decay could be delayed through conversion to long-lived wood products such as building materials (Malmsheimer et al. 2011). It is also important to note that our analysis did not include stochastic wildfire occurrence. Estimates of burn probability in the present study are not estimates of the likelihood of wildfire occurrence based on historical fire sizes and fre quency (e.g., Preisler et al. 2004, Mercer and Prestemon 2005, Brillinger et al. 2006), but rather are conditional on a single randomly ignited wildfire within the buffered study area.

CONCLUSIONS
Our findings generally support those of Camp bell et al. (2011), who concluded from an analysis of fire-prone western forests that the C costs of treatments are likely to outweigh their benefits under current depressed fire frequencies. In a more recent paper, Campbell and Ager (2013) con cluded that "none of the fuel treatment simulation scenarios resulted in increased system carbon," primarily from the low incidences of treated areas being burned by wildfire. However, our interpre tation of these findings differs from those dis cussed in Campbell et al. (2011) and Campbell and Ager (2013), especially in light of recent and projected future trends in fire activity (Westerling et al. 2011, Miller and Safford 2012, Dennison et al. 2014. The current divergence of increasing surface air temperatures and low fire activity is unlikely to be sustained, further suggesting greater future fire activity (Marlon et al. 2012). If increased fire activity is realized, then the likeli hood of a given area being burned in a wildfire increases. This differs from the simple increase in stand-level fire frequency modeled by Campbell et al. (2011) because increases in fire likelihood are not necessarily associated with corresponding decreases in fire severity, as assumed by Campbell et al. (2011). Increased fire likelihood could very well lead to positive feedbacks in fire severity, and ultimately to vegetation type conversion (Coppo letta et al. 2016)-effects that would have signifi cant implications for carbon storage.
Due to the significant emissions associated with treatment and the low likelihood that wild fire will encounter a given treatment area, forest management that is narrowly focused on C accounting alone would favor the no-treatment scenarios. Landscape treatments protected more C in live tree biomass only when large wildfires were simulated. While treatment favorability improved with large wildfire simulation, the notreatment scenario still produced fewer emissions than the treatment scenario. Given the potential for large wildfire in the region as demonstrated by the 2014 King Fire, and the increasing fre quency of large wildfires and area burned in Cali fornia expected from climate modeling studies (Lenihan et al. 2008, Westerling et al. 2011), we suggest that future studies of fuel treatment-wild fire-C relationships should incorporate the poten tial for large wildfires at a frequency greater than those observed over the last 20-30 yr. Others have argued that treatments to increase forest resilience should be a stand-alone, top land management priority independent of other ecosystem values such as carbon sequestration and fire hazard reduction (Stephens et al. 2016a).
We also note that the potential benefits of fuels management are not restricted to avoided wild fire emissions. Here, we show that landscape fuel treatments can alter fire hazard across the land scape both within and outside of treated stands, and have the potential to affect the likelihood of burning and fire intensity within protected Cali fornia spotted owl habitat. Underscoring the risk to sensitive habitat, the 2014 King Fire encoun tered 31 PACs within our study area, leading to the greatest single-year decline in habitat occu pancy recorded over a 23-yr study period (Jones et al. 2016). Modest simulated treatments within activity centers significantly reduced potential fire intensity relative to both the no-treatment landscape and a treatment scenario that did not permit treatment within PACs, indicating that active management may be desirable to protect habitat in the long term (Roloff et al. 2012).
However, treatments conducted outside of owl habitat also reduced wildfire hazard.

ACKNOWLEDGMENTS
We are grateful to a number of individuals for shar ing their time and fuel treatment and fire modeling expertise: Nicole Vaillant (ArcFuels), Alan Ager and Andrew McMahan (carbon estimation), and Coeli Hoover (harvest simulations in FVS-FFE). Brian Ebert and Dana Walsh of the Eldorado National Forest pro vided invaluable advice on weather inputs for fire modeling and on designing fuel treatment prescrip tions. Ken Somers, Operations Forester on the Blodgett Forest Research Station, gave advice concerning tim ber management on private land. We also thank two anonymous reviewers, whose comments greatly improved this article. The California Energy Commis sion provided funding for this project.