Projected Impact of Mid-21st Century Climate Change on Wildﬁre Hazard in a Major Urban Watershed outside Portland, Oregon USA

: Characterizing wildﬁre regimes where wildﬁres are uncommon is challenged by a lack of empirical information. Moreover, climate change is projected to lead to increasingly frequent wildﬁres and additional annual area burned in forests historically characterized by long ﬁre return intervals. Western Oregon and Washington, USA (westside) have experienced few large wildﬁres (ﬁres greater than 100 hectares) the past century and are characterized to infrequent large ﬁres with return intervals greater than 500 years. We evaluated impacts of climate change on wildﬁre hazard in a major urban watershed outside Portland, OR, USA. We simulated wildﬁre occurrence and ﬁre regime characteristics under contemporary conditions (1992–2015) and four mid-century (2040–2069) scenarios using Representative Concentration Pathway (RCP) 8.5. Simulated mid-century ﬁre seasons expanded in most scenarios, in some cases by nearly two months. In all scenarios, average ﬁre size and frequency projections increased signiﬁcantly. Fire regime characteristics under the hottest and driest mid-century scenarios illustrate novel disturbance regimes which could result in permanent changes to forest structure and composition and the provision of ecosystem services. Managers and planners can use the range of modeled outputs and simulation results to inform robust strategies for climate adaptation and risk mitigation.


Introduction
Moist forests historically characterized by infrequent wildfire are projected to experience significant increases in wildfire frequency before the end of the century as a result of climate change and anthropogenic activities [1][2][3]. Many of these forests are not adapted to rapid shifts in disturbance regimes and future wildfire could lead to irreversible changes in vegetation structure and composition [4][5][6][7]. The pace and magnitude of climate change and resulting shifts in disturbance regimes could lead to reduced forest resilience, which, in this setting, refers to a reduced capacity of forests to continue to provide ecosystem services on which communities rely [6,[8][9][10][11]. In the present, human communities rely on these forests for drinking water, economic value, carbon storage and sequestration, and cultural significance, but climate-wildfire adaptation strategies that would maintain resilience are underdeveloped [12][13][14][15]. variable and top-down regional climate controls [16,[48][49][50][51]. The relatively limited empirical fire record and the overwhelming success of fire suppression efforts make it difficult to define and characterize contemporary westside fire regimes [52,53]. Given the regional heterogeneity in fire regime characteristics within the westside and limited direct experience with wildfire, managers and planners benefit from modeled assessments of fire regimes and fire risk to inform wildfire response, landuse planning, and climate adaptation [54][55][56]. To date, most assessments of future westside fire regime characteristics have relied on relatively coarse scale dynamic global vegetation models that fail to capture local variation in fire regime characteristics [41,[57][58][59]. The LANDIS II model has also been used to forecast future changes in forest structure and composition, including as a result of wildfire [60]. However, Monte Carlo-based fire simulation models have become one of the most widely used techniques for conducting wildfire risk assessments at national, regional, and state levels [61,62]. Ref. [63] developed a novel technique for using the large fire simulator FSim, a Monte Carlo approach model, to project changes in future fire regimes over the Northern Rockies, but no similar studies have been conducted for the westside Cascade Range to the authors' knowledge.
The primary objective of this study was to characterize plausible mid−21st century changes in wildfire exposure to a municipal watershed in a way that captures locally specific spatial and temporal variation in fire regime characteristics. It is expected that such information will help managers and planners identify sources of risk and will inform development of risk mitigation and climate adaptation strategies. We selected the Clackamas River watershed (Clackamas Basin) east of Portland, Oregon because it is located in a region with spatially diverse fire regime characteristics and which has not experienced a very large fire (greater than~5000 ha) in more than a century [16]. Additionally, the Clackamas Basin is the second largest source of drinking water for the Portland metro area. In 2017, the Eagle Creek Fire threatened Portland's primary drinking water source the Bull Run watershed adjacent to the Clackamas Basin, and elevated the importance of evaluating future fire exposure in municipal watersheds. Using a range of plausible climate scenarios for 2040-2069 and modeling methods presented in [63], we assessed the range of projected changes in fire size, frequency and seasonality of large fire events. In addition, we evaluated projected changes in the spatial distribution and magnitude of wildfire exposure to the watershed using a quantified hazard metric described in [64].

Study Area
Clackamas County, Oregon is part of the Portland metropolitan area and encompasses the majority of the Clackamas Basin. Portland is the largest metro area in Oregon with 2.5-3 million residents, only slightly smaller than the Seattle metro area. The Clackamas Basin, 244,002 ha, provides drinking water to over 300,000 Portland area residents and is the second largest drinking water source for the Portland region. Land ownership and management in the Clackamas Basin is heavily segmented (Figure 1). Approximately two-thirds of the watershed, the upper or higher elevation portion, is managed by the Mt. Hood National Forest (MHNF) ( Table 1) division of the United States Forest Service (USFS). In the Clackamas Basin, the MHNF includes wilderness, motorized recreation areas, campsites and day-use facilities, and active timber management units. The middle portion of the watershed is comprised of a patchwork of land managed primarily for commercial timber value by the U.S. Bureau of Land Management (BLM) and a variety of private industrial timber landowners. The lower watershed transitions from very low-density rural development and agriculture to increasingly dense suburban and urban housing in the Portland metro area. Clackamas County's population is expected to grow by nearly 270,000 people in the next fifty years, placing increasing demand on land and water resources [65].  [66,67] compares the importance of forests to surface drinking water on a relative scale of 0-100. Forests are especially important to drinking water in westside landscapes (left), many of which have experienced few to no significant wildfires in the past 20 years [27]. Prior to 2020, the largest wildfire in recent history in the Clackamas Basin (right) was the 36 Pit fire in 2014 which burned just over 2200 ha on the Mt. Hood National Forest, but adjacent to private land [68].
The basin ranges from the crest of the Cascade Range at approximately 2200 m in elevation, to about 3 m in elevation at the confluence of the Clackamas and Willamette Rivers ( Figure 1). The Clackamas Basin is dominated by moist mixed-conifer forests largely characterized by western hemlock (Tsuga heterophylla) [69]. Potential vegetation types indicate that Clackamas Basin forest types are adapted to infrequent (>200 year fire return intervals) but stand replacing fires, as well as to moderately frequent fires  year return intervals) of mixed severity [16]. However, existing species assemblages and forest structure throughout much of the watershed are the result of human land management, predominantly the legacy of intensive forest management during the 20th century which favored relatively young, even-aged, and single species stands [70]. The vast majority of the watershed is coniferous forests (Table 1) including Douglas-fir (Pseudotsuga menziesii var. menziesii), western hemlock and Pacific silver fir (Abies amabilis) as well as other conifers in smaller extents. The watershed receives over 150-200 cm of precipitation on average annually, the vast majority of which falls between October and April each year [17,71]. The watershed has two main types of geology: the lower watershed is dominated by the older Western Cascade mountains (inactive with shallow subsurfaces), whereas the upper watershed ( Figure 1) encompasses the younger High Cascades (with active Volcanoes) and porous geology [72]. Summers are generally dry, and, as a result of groundwater storage, seasonal variations in drought are highly dependent on snowpack levels and snowmelt timing [32]. Although annual temperatures are moderate, late summer temperatures regularly exceed 32 • C during same period of time when water resources are most scarce. Ref. [16] characterize the Clackamas Basin as a 'moderate-frequency/mixed-severity' fire regime. The national LANDFIRE dataset indicates that fire return intervals greater than 200 years dominate the mid-and low-elevation portions of the watershed, approximately 55% of the watershed, while the upper watershed, where lightning strikes are more common and forests are drier, return intervals are between 35 and 200 years [73]. Despite thousands of recorded ignitions in the Clackamas Basin from 1992-2015, nearly 75% of which are caused by human-activities there have been only a handful of large fire events in recent history ( Figure 1) [74]. The fact that most ignitions do not result in a large fire reflects mostly the lack of fire conducive climate/weather conditions as well as the fact that most human ignitions are identified immediately and able to be quickly extinguished (i.e., campfires, equipment fires, and escaped brush piles). At the time of this writing, one of the largest wildfires on record is actively burning in the watershed. Favored by very dry and hot, fast-moving easterly winds that rapidly cured fuels, within days the Riverside Fire ignited on 7 September 2020 and was just one of several major fires during one of the largest wildfire events in western Oregon's recorded history. Within days the Riverside Fire grew to nearly 60,000 hectares; the cause is listed as unknown but is not suspected to be lightning. Prior to the Riverside Fire, the 36 Pit Fire burned just over 2000 hectares in the middle of the watershed in 2014, and the B & B Complex fires in the southeast of the watershed in 2003 covered over 30,000 hectares, some of which were located in the uppermost section of the watershed.

Wildfire Modeling
We used the large fire simulator, FSim, to model contemporary fire regime characteristics and wildfire exposure as well as the effect of climate change during 2040-2069 (mid-century) [75]. FSim uses a Monte Carlo approach to model the interaction of fuel, topography, weather and spatial ignition patterns on a daily time-step and simulate fire occurrence, growth and suppression in thousands of statistically plausible, unique fire seasons. For calibration and modeling purposes, large fires were defined as fires greater than or equal to 100 hectares. The particulars of FSim as well as probabilistic fire risk assessments have been well documented elsewhere [62,63,[75][76][77][78]. We simulated contemporary and future wildfire occurrence in two model domains in order to reasonably account for all fires that might ignite outside but spread into the Clackamas Basin, similar to recent experience with, for example, the B & B Complex fires ( Figure 2). The upper reaches of the Clackamas Basin abut the Cascade Range crest which divides two drastically different fire environments, but records show that fires can and do ignite on the east side of the mountains and spread west into the watershed (e.g., as what the Lionshead fire did on 7 September 2020). For each domain, east and west, we used unique fire histories,  Extent of FSim modeling. The model extent was divided into two fire occurrence areas (FOAs) in order to distinguish between very different fire environments on either side of the Cascade Range crest. Each FOA was associated with a unique weather station from which empirical observations were used in the baseline scenarios, and modified to reflect projected future climate conditions. The FOA defines the area within which FSim simulated ignitions and is significantly larger than the analysis area, the Clackamas Basin, in order to allow for fires that ignite outside the analysis area but eventually burn into it.
We used daily weather records (1992-2015) from two Remote Automated Weather Stations (RAWS) (Figure 2). RAWS data were processed in Fire Family Plus software to compute a distribution of daily Energy Release Component values based on Fuel Model G (ERC-G) [75,79]. ERC quantifies the amount of energy released at the flaming front of an active fire and is a proxy for fuel moisture and ERC-G was selected because [80] demonstrated its utility across a range vegetation types, including closed canopy conifer forests which dominate the Clackamas Basin as well as dry forests, and shrublands in the eastern FOA ( Figure 2) [63,75,81]. FSim uses ERC to determine on what days a fire may ignite, how many fires may ignite, and when fires do ignite FSim uses ERC values in conjunction with wind speed/direction and landscape features to simulate how a fire grows and is extinguished.
To simulate mid-century climate scenarios we adapted methods from [63]. We selected four global climate models (GCMs) from the Coupled Model Intercomparison Project phase 5 (CMIP 5). CMIP5 GCMs have been evaluated for use and applied across the PNW, including the study area [59,82]. Each model was downscaled from native to 4 km resolution using the Multivariate Adaptive Constructed Analogs (MACA) method which was specifically designed for wildfire applications [83]. Only Representative Concentration Pathway (RCP) 8.5 scenarios were considered based on guidance from local resource managers who expressed a primary concern for high emissions scenarios. Models were selected based on average annual changes in temperature and relative humidity over the Clackamas Basin between historical (1970-1999) and mid-century periods. Among the nineteen Figure 2. Extent of FSim modeling. The model extent was divided into two fire occurrence areas (FOAs) in order to distinguish between very different fire environments on either side of the Cascade Range crest. Each FOA was associated with a unique weather station from which empirical observations were used in the baseline scenarios, and modified to reflect projected future climate conditions. The FOA defines the area within which FSim simulated ignitions and is significantly larger than the analysis area, the Clackamas Basin, in order to allow for fires that ignite outside the analysis area but eventually burn into it.
We used daily weather records (1992-2015) from two Remote Automated Weather Stations (RAWS) ( Figure 2). RAWS data were processed in Fire Family Plus software to compute a distribution of daily Energy Release Component values based on Fuel Model G (ERC-G) [75,79]. ERC quantifies the amount of energy released at the flaming front of an active fire and is a proxy for fuel moisture and ERC-G was selected because [80] demonstrated its utility across a range vegetation types, including closed canopy conifer forests which dominate the Clackamas Basin as well as dry forests, and shrublands in the eastern FOA ( Figure 2) [63,75,81]. FSim uses ERC to determine on what days a fire may ignite, how many fires may ignite, and when fires do ignite FSim uses ERC values in conjunction with wind speed/direction and landscape features to simulate how a fire grows and is extinguished.
To simulate mid-century climate scenarios we adapted methods from [63]. We selected four global climate models (GCMs) from the Coupled Model Intercomparison Project phase 5 (CMIP 5). CMIP5 GCMs have been evaluated for use and applied across the PNW, including the study area [59,82]. Each model was downscaled from native to 4 km resolution using the Multivariate Adaptive Constructed Analogs (MACA) method which was specifically designed for wildfire applications [83]. Only Representative Concentration Pathway (RCP) 8.5 scenarios were considered based on guidance from local resource managers who expressed a primary concern for Fire 2020, 3, 70 7 of 24 high emissions scenarios. Models were selected based on average annual changes in temperature and relative humidity over the Clackamas Basin between historical (1970-1999) and mid-century periods. Among the nineteen available GCMs CNRM-CM5 (CNRM), HadGEM-ES365 (HadGEM), MIROC5, and MIROC-ESM-CHEM (MIROCCHEM) represent the outer boundaries of possible changes in mid-century temperature and relative humidity. All four GCMs project warmer average annual temperatures in mid-century, but MIROCCHEM and MIROC5 project increased average annual relative humidity where CNRM and HadGEM project a decrease in relative humidity. HadGEM is the warmest and driest GCM, CNRM is dry and slightly less warm than HadGEM, MIROCCHEM is warm and more humid, while MIROC 5 is the coolest and wettest GCM. Projected monthly average changes in temperature and relative humidity were used to augment observed RAWs data and create a distribution of ERC values that reflect plausible mid-century ERC values ( Figure 3). 1 Figure 3. Average monthly temperature and relative humidity departure from historical  predicted by the four GCMs selected to model future fire weather conditions. These four GCMs represent a full range of predicted climate futures and were downscaled over the study area.
We generated an ignition probability raster quantifying the relative likelihood of an ignition at any pixel in the Clackamas Basin using historical ignitions of fires greater than 100 hectares in the Fire Program Analysis, Fire Occurrence Dataset [74]. A raster stack of eight 90 m resolution LANDFIRE layers were used as inputs for each model domain to describe topography and fuels [84][85][86][87][88][89][90][91]. The same landscape files were used in baseline and mid-century scenarios.

Data Analysis
For each scenario, we merged simulated fire perimeters from the two domains (east and west of the Cascade Range crest) and removed the overburn. Overburn occurred when fires from the same simulation year burned the same pixel(s); the fire perimeter which ignited earlier in the year was maintained and overburnt pixels were removed from the subsequent fire. Under baseline conditions and in the least warm mid-century scenarios overburn was less than 5%, but in the hot and dry mid-century scenarios it was as high as 12%. After correcting for overburn, we generated new burn probability rasters for the Clackamas Basin that reflected the occurrence of fires from both domains east and west of the Cascade Range crest. All spatial and quantitative analysis was performed in R statistical software 3.5 [92].
Fire season timing and length were measured using average daily ERC values. Our treatment of ERC and fire season is similar to that in [63], which used the relationship between ERC and National Fire Danger Rating System classifications to measure fire season. The bottom threshold for "moderate fire danger" and an indication of active fire season, was defined by the 80th percentile ERC values [93]. The 80th percentile was derived from baseline climate data and was applied to all mid-century scenarios because the fuel moisture thresholds to fire occurrence are not expected to change. We also calculated the number of days on which daily average ERC values exceeded the 90th ("high fire danger") and 97th ("extreme fire danger") percentiles for all scenarios.
Using the overburn-corrected perimeters we calculated average annual number of fires, average fire size, and average annual area burned to evaluate the impact of changes in temperature and relative humidity associated with potential future climate change effects. Fire size and annual area burned distributions from all four mid-century scenarios were compared to baseline using a Mann-Whitney test with an alpha of 0.05. Fire size distributions were compared between the four mid-century scenarios using a Kruskal-Wallis test followed by Dunn's Multiple Pairwise Comparison using the Benjamani-Hochberg adjustment method with an alpha of 0.05. In addition to annual metrics, we included two unique fire characteristic analyses. First, we sub-sampled 300 thirty-year samples of fire perimeters for each climate scenario to evaluate the variation in annual area burned during each thirty-year climate period. FSim uses a Monte Carlo simulation approach in which each iteration is assigned a year and represents a statistically plausible annual fire season and, because the weather module is dynamic and ignitions are stochastic, each iteration is unique. [94] presented a method for subsampling the 10,000 iterations to create statistically plausible multidecadal scenarios. Similar to [94], each iteration was assigned a "fire year" and within each thirty-year sample, we assumed that the same pixel would not burn more than once every five years and removed overlapping portions of fire perimeters when they were within five years of each other. A multidecadal analysis is particularly useful in low frequency fire regimes in contrast to metrics like annual averages which may reveal trends over time, but have little managerial application.
Second, modeling studies elsewhere have demonstrated the value in assessing trends over time in the frequency and magnitude of extremely large fires [95,96]. In this study, extremely large fires are fires that exceed 99th percentile fire size under baseline climate conditions (9861 ha). In addition to evaluating climate change impacts on average fire size, we used a generalized Pareto distribution model to evaluate trends in the size and return intervals of extremely large (99th percentile) fires. Generalized Pareto models have been used to evaluate probabilities in the tails of highly skewed distributions in which, like wildfire, extremely rare events are highly impactful [97]. Extreme event analysis was performed in R using the 'extRemes' package [92,98].
Finally, [64] introduced a metric that quantifies integrated wildfire hazard as the product of burn probability and mean fireline intensity at each pixel in the study area. This method has the advantage of being quantitative and easily comparable across space and time, aiding managers to identify areas that are at highest risk of burning at high intensity. Evaluating the potential intensity of future fires is particularly important where risk to drinking water is concerned [99][100][101]. We calculated integrated hazard for all burnable pixels for each climate scenario and determined the change in hazard that results from all four GCMs. All spatial analyses are presented using quantiles of the modeled data distribution.

Fire Season
Three of four future scenarios project that the fire season as measured by ERC will expand by mid-century ( Figure 4). Under baseline conditions 80th percentile ERC value was 29, 90th percentile Fire 2020, 3, 70 9 of 24 was 39 and 97th percentile was 51. There were on average 70 days above 80th percentile ERC in the baseline scenario. In Figure 4, MIROC5, the least warm and wettest of the GCMs, shows an eight day increase in average fire season length but it is not significantly different than baseline (Mann-Whitney, p = 0.289). The remaining three GCMs project significantly higher ERC values between May and November. MIROCCHEM and CNRM both projected 99 days above the 80th percentile (Mann-Whitney, p < 0.0001), and HadGEM projected 102 days on average each year during which ERC values will be above the 80th percentile (Mann-Whitney, p < 0.0001). Under HadGEM climate conditions ERC values exceeded the 80th percentile earlier in the season but peaked in August and September similar to baseline conditions. In contrast MIROCCHEM and CNRM models illustrated the potential for a similar start to fire season compared to today, but the peak might not occur until late September or October. Seasonal changes in ERC from Figure 4 were reflected in the seasonal pattern and magnitude of simulated ignitions where HadGEM simulated large numbers of fires in July and August, while MIROCCHEM and CNRM simulated more fires late in the summer and early fall ( Figure 5).
Fire 2020, 3, x FOR PEER REVIEW 9 of 24 will be above the 80th percentile (Mann-Whitney, p < 0.0001). Under HadGEM climate conditions ERC values exceeded the 80th percentile earlier in the season but peaked in August and September similar to baseline conditions. In contrast MIROCCHEM and CNRM models illustrated the potential for a similar start to fire season compared to today, but the peak might not occur until late September or October. Seasonal changes in ERC from Figure 4 were reflected in the seasonal pattern and magnitude of simulated ignitions where HadGEM simulated large numbers of fires in July and August, while MIROCCHEM and CNRM simulated more fires late in the summer and early fall ( Figure 5). Under baseline conditions, average daily ERC values never exceeded 90th percentile ERC, the high fire danger threshold. Daily ERC values are highly variable from one year to the next and have exceeded 90th percentile thresholds, but average daily baseline ERC values do not. In contrast, HadGEM, CNRM, and MIROCCHEM projected more severe fire weather with 33−69 days on average exceeding 90th percentile ERC (Mann-Whitney, p < 0.0001) (Figure 4). No scenarios projected average daily ERC values above the 97th percentile extreme fire danger threshold, but HadGEM projected average ERC will be close to the 97th percentile throughout August and September. Under baseline conditions, average daily ERC values never exceeded 90th percentile ERC, the high fire danger threshold. Daily ERC values are highly variable from one year to the next and have exceeded 90th percentile thresholds, but average daily baseline ERC values do not. In contrast, HadGEM, CNRM, and MIROCCHEM projected more severe fire weather with 33−69 days on average exceeding 90th percentile ERC (Mann-Whitney, p < 0.0001) (Figure 4). No scenarios projected average daily ERC values above the 97th percentile extreme fire danger threshold, but HadGEM projected average ERC will be close to the 97th percentile throughout August and September.

Large Fire Size and Frequency
In response to longer fire seasons and longer continuous stretches of fire-conducive weather within each season, the average annual number and size of simulated large fires increased under each of the climate scenarios compared to baseline (Table 2). Increased large fire occurrence in mid-century simulations mirrors seasonal ERC patterns ( Figure 4). Under all mid-century climate conditions there were more projected opportunities for ignitions to occur, and higher potential for ignitions to turn into large fires. The size and probability of extremely large wildfire events also increased under midcentury conditions. The largest fire simulated under baseline conditions was 37,509 ha. The largest fire size increased significantly (Mann-Whitney, p < 0.001) to 42 Accordingly, changes in fire size and frequency generally resulted in increased annual area burned. Excluding fire-free years, approximately 1200 ha burned annually on average under baseline conditions (Table 2). Compared to baseline, projected annual average area burned increased significantly (p < 0.001) by 50% under MIROC5, 95% under MIROCCHEM, nearly 150% under CNRM, and 540% under HadGEM. Proejcted annual area burned was significantly different among all scenarios, not just between mid-century and baseline scenarios (Dunn's test, p < 0.01). Under baseline conditions during years in which fires occurred, less than 0.5% of the watershed burned, but the average extent of the watershed burned on average increased to 1% under CNRM and over 3% under HadGEM. However, the range of projected annual area burned values in Table 2 illustrates that large regions of the watershed could be impacted in a given year under any of the scenarios, including baseline, in which the largest annual area burned was 56,656 ha, or 23% of the watershed.
When we sub-sampled the 10,000 iterations in each scenario and created thirty-year samples we learned that increasingly more of the watershed will be exposed to wildfire during the mid-century. Under baseline conditions approximately 8% of the watershed is projected to experience wildfire in a thirty-year period. From 2040-2069, a projected 13% of the watershed would be affected under the

Large Fire Size and Frequency
In response to longer fire seasons and longer continuous stretches of fire-conducive weather within each season, the average annual number and size of simulated large fires increased under each of the climate scenarios compared to baseline (Table 2). Increased large fire occurrence in mid-century simulations mirrors seasonal ERC patterns ( Figure 4). Under all mid-century climate conditions there were more projected opportunities for ignitions to occur, and higher potential for ignitions to turn into large fires. The size and probability of extremely large wildfire events also increased under mid-century conditions. The largest fire simulated under baseline conditions was 37,509 ha. The largest fire size increased significantly (Mann-Whitney, p < 0.001) to 42 Accordingly, changes in fire size and frequency generally resulted in increased annual area burned. Excluding fire-free years, approximately 1200 ha burned annually on average under baseline conditions (Table 2). Compared to baseline, projected annual average area burned increased significantly (p < 0.001) by 50% under MIROC5, 95% under MIROCCHEM, nearly 150% under CNRM, and 540% under HadGEM. Proejcted annual area burned was significantly different among all scenarios, not just between mid-century and baseline scenarios (Dunn's test, p < 0.01). Under baseline conditions during years in which fires occurred, less than 0.5% of the watershed burned, but the average extent of the watershed burned on average increased to 1% under CNRM and over 3% under HadGEM. However, the range of projected annual area burned values in Table 2 illustrates that large regions of the watershed could be impacted in a given year under any of the scenarios, including baseline, in which the largest annual area burned was 56,656 ha, or 23% of the watershed. Table 2. Simulated mean fire size of all fires and the mean annual number of large fires, fires greater than 100 ha. Range and standard error are shown below mean values. Conditional area burned is the size of the study area effected in those years that fires were simulated-not all years experienced fire. 30-year area burned is the estimated cumulative extent of wildfire between 2040-2069. Mid-century fire size and frequency distributions were compared to baseline distributions using Dunn's tests and in all instances, mid-century distributions were significantly different from baseline (all p < 0.0001). When we sub-sampled the 10,000 iterations in each scenario and created thirty-year samples we learned that increasingly more of the watershed will be exposed to wildfire during the mid-century. Under baseline conditions approximately 8% of the watershed is projected to experience wildfire in a thirty-year period. From 2040-2069, a projected 13% of the watershed would be affected under the MIROC5 scenario, 18% under MIROCCHEM, 23% under CNRM, and 56% of watershed under HadGEM.

Annual
Extremely large wildfires are fires exceeding the 99th percentile of modeled fire sizes in each scenario. Compared to baseline conditions, the number and size of extremely large fires is projected to increase under all mid-century climate scenarios (Table 3). Under baseline conditions the average extremely large fire was 15,789 ha (sd = 5808). This average increased significantly under MIROC5 (mean = 19,917 ha, sd = 6193), under MIROCCHEM (mean = 22,104 ha, sd = 6628), under CNRM (mean = 23,254 ha, sd = 7392), and under HadGEM (mean = 30,746 ha, sd = 7334). A generalized Pareto model illustrated that the threshold for extremely large fires is projected to increase by more than a factor of two (Table 3). Likewise, fires with a 0.01 probability of occurrence are projected to be increasingly larger (Table 3). Annualized metrics are useful for illustrating the effect of climate change on fire characteristics, but results demonstrate that fire occurrence will continue to vary intra-and inter-annually. In all future scenarios there were iterations during the 10,000 simulations during which no fires were simulated because ERC never exceeded 80th percentile values, or because simulated ignitions occurred but were located in unburnable pixels (i.e., water or urbanized areas). We can interpret iterations without fire as "fire free years". Under baseline conditions, 43% of iterations included no fires. However, climate change is projected to decrease the number of fire free years in mid-century. 36% of iterations were fire free under MIROC5, 27% under MIROCCHEM, 23% under CNRM, and 18% under HadGEM.

Wildfire Hazard
Wildfire hazard under baseline conditions is concentrated at higher elevations in the upper watershed where both burn probability and mean fireline intensity are highest ( Figure 6). Increases in wildfire hazard under all mid-century climate scenarios were driven by increased burn probability (Figure 7). Increased fire size and more frequent wildfires resulted in 226%, 357%, and 476% of average watershed-wide increases, compared to baseline, in burn probability in MIROC5, MIROCCHEM, and CNRM, respectively. However, under the hot and dry conditions of HadGEM, average burn probability increased by 1663% (Table 4). Despite large changes in burn probability the actual annual likelihood of wildfire remains less than 1% across the entire watershed in all mid-century scenarios, except for HadGEM under which the watershed-wide average probability of wildfire in any given year exceeds 2% ( Table 4). The distribution of mid-century burn probabilities across the watershed is similar to the spatial pattern in baseline conditions but under hotter and drier conditions, the lower and upper watershed became increasingly fire prone (Figure 8).    Increases in future wildfire hazard roughly follow the same spatial pattern as under baseline conditions with increasing hazard under hotter and drier conditions (Figure 9). In baseline conditions, approximately 30,000 ha (12%) are classified as very high hazard (Figure 9), but under mid-century conditions, very high hazard classifications increased to 37,000 ha (15%), 42,000 ha (17%), 45,000 ha (18%), and 62,000 ha (26%) under MIROC5, MIROCCHEM, CNRM, and HadGEM, respectively. Under CNRM and HadGEM, the lower watershed adjacent to the Portland metro area is exposed to increasing wildfire hazard (Figure 9).  Mean fireline intensity did not change as a result of changes in temperature and relative humidity (Table 3 and Figure 7). Mean fireline intensity in the watershed varies by approximately one order of magnitude from the lower watershed where the average value is about 300 kW/m to the upper watershed where the average value is about 3000 kW/m. Compared to baseline, none of the mid-century mean fireline intensities were significantly different (Mann-Whitney, p > 0.05). HadGEM is the warmest and driest future scenario, followed by CNRM. Both MIROCCHEM and MIROC5 project small increases in summer relative humidity and are less hot than HadGEM and CNRM but are still warmer than baseline conditions. Non-burnable pixels indicate water, rock/ice, high density urban areas, and agriculture. Results are classified using a quantile method.
Increases in future wildfire hazard roughly follow the same spatial pattern as under baseline conditions with increasing hazard under hotter and drier conditions (Figure 9). In baseline conditions, approximately 30,000 ha (12%) are classified as very high hazard (Figure 9), but under mid-century conditions, very high hazard classifications increased to 37,000 ha (15%), 42,000 ha (17%), 45,000 ha (18%), and 62,000 ha (26%) under MIROC5, MIROCCHEM, CNRM, and HadGEM, respectively. Under CNRM and HadGEM, the lower watershed adjacent to the Portland metro area is exposed to increasing wildfire hazard (Figure 9). HadGEM is the warmest and driest future scenario, followed by CNRM. Both MIROCCHEM and MIROC5 project small increases in summer relative humidity and are less hot than HadGEM and CNRM, but are still warmer than baseline conditions. Non-burnable pixels indicate water, rock/ice, high density urban areas, and agriculture. Results are classified using a quantile method.

Discussion
The primary objective of this study has been to characterize plausible impacts of climate change on mid-century wildfire hazard in a major municipal watershed which has experienced few recent large fires, but which has a demonstrated potential for infrequent, stand-replacing fires. Our results show that while wildfire hazard in the Clackamas Basin could significantly increase in the near The four mid-century scenarios are based on projected climate scenarios. HadGEM is the warmest and driest future scenario, followed by CNRM. Both MIROCCHEM and MIROC5 project small increases in summer relative humidity and are less hot than HadGEM and CNRM, but are still warmer than baseline conditions. Non-burnable pixels indicate water, rock/ice, high density urban areas, and agriculture. Results are classified using a quantile method.

Discussion
The primary objective of this study has been to characterize plausible impacts of climate change on mid-century wildfire hazard in a major municipal watershed which has experienced few recent large fires, but which has a demonstrated potential for infrequent, stand-replacing fires. Our results show that while wildfire hazard in the Clackamas Basin could significantly increase in the near future, the magnitude of change is uncertain because of variability in projections of future climate.
Our analysis projects an average three week increase in fire season length between baseline and mid-century time periods. Our methods were based on [63] which projected only a 12 day increase in average fire season length in the Northern Rocky Mountains of the United States. There are many factors which could explain the difference including very different biophysical settings, but also the fact that [63] used climate data representing a moderate emissions scenario. We used RCP 8.5, a scenario in which little to no action is taken to curb global emissions and which has been shown to be consistent with actual emissions [102]. However, contrasting research shows that RCP 8.5 may be less accurate for mid-century projections compared to more moderate emissions pathways and so our results should be interpreted as being based on plausible, but high end climate change projections [103]. [59] demonstrated that RCP 4.5 scenarios produce more moderate impacts on fire regime characteristics across the PNW, including the westside, compared to RCP 8.5. Likewise, our projections of longer future fire seasons could be moderated based on changes in future precipitation amount and seasonal timing. For instance, in a study that included western Oregon and Washington, [59] noted that earlier autumn precipitation could offset changes in temperature which might otherwise lead to longer fire seasons, but there is a high degree of uncertainty around future precipitation patterns [34,39,82].
Based on our method, future wildfire hazard in the Clackamas Basin will be largely a function of changes in annual burn probability, which we project could increase between 66-730%. Our assessment used simulation methods very similar to the PNW Quantitative Wildfire Risk Assessment (QWRA) but baseline burn probability values from our study were sometimes higher than burn probability modeled in the QWRA [104]. In particular, in the upper Clackamas Basin we projected burn probabilities as much as 80% greater than the QWRA (0.007109 compared to 0.003931). In part, the difference could be explained by our inclusion of simulated fires that ignited east of the Cascade Range, outside the Clackamas Basin, and spread into the basin. For regional and national fire simulations conducted with FSim, like the QWRA, it is not standard practice to account for fires transmitted between fire occurrence areas because of associated computational demands [94]. However, the top of the Clackamas Basin along the crest of the Cascade Range is a relatively low saddle and forests east of the crest are prone to significantly more frequent fires which have in the past spilled over onto the westside (i.e., B & B Fire in 2003). Our inclusion of fires transmitted across the saddle from east to west demonstrates the value of conducting simulations at a relatively local scale. In our analysis, generally longer fire seasons and more severe fire weather facilitated increases in burn probability which are consistent with recent studies of future westside fire frequency and annual area burned. Estimates of mid-and late-21st century annual area burned in westside study areas range from increases of 76-500% and our results indicate a similar range of the 56-540% increase in the Clackamas Basin [41,58]. Under hotter and drier conditions, we project that mid-century annual burn probabilities (in excess of 1.5%) in the upper Clackamas Basin will be increasingly similar to contemporary burn probabilities in high-frequency fire regimes like those of central Oregon, which arguably represents a new disturbance regime that would plausibly lead, through flammability feedbacks, to changes in dominant forest types [4,11,[104][105][106].
Notably, our results showed no significant difference in future fireline intensity compared to baseline. Fireline intensity is a function of fuel type, fuel load, fuel moisture and rate of spread [107,108]. In FSim, fuel type and load are determined by fire behavior fuel models (FBFM) which were not changed between baseline and mid-century simulations, and the third, rate of spread, is strongly influenced by fuel structure, slope, and wind speed which also were not changed between baseline and mid-century simulations. The decision to not alter future FBFM and vegetation characteristics was based on research showing relatively stable westside forest structure and composition into mid-century [58,59]. Despite the fact that fireline intensity appears to be best explained by fuel structure and wind speed, we did anticipate that modeled changes in temperature and relative humidity would result in lower fuel moisture in the future which would in turn facilitate higher mean fireline intensity in the Clackamas Basin [109,110]. In one study of projected future fireline intensity across Canada, [110] showed a 20% increase in the occurrence of fireline intensities high enough to limit ground and aerial suppression under RCP 8.5 in the near future. However, future fireline intensities in [110] were spatially heterogeneous and strongly associated with particular FBFM, so we might not expect similar results in the Clackamas Basin. In contrast, [111], similar to this study, found little to no change in mid-century conditional flame lengths, a fire intensity metric derived from fireline intensity. Fireline intensity is an important metric for understanding fire effects on vegetation, soil, and ecological processes, and it also informs wildfire response strategies [112,113]. Yet, wildfire simulations have been shown to underestimate fire intensity during severe fire weather events and so future work should focus on evaluating the sensitivity of modeled fireline intensity to fuel structure, wind speed and fuel moisture in westside forest types, as well as evaluating the accuracy of simulated fireline intensity under severe weather [114].
In addition to increased wildfire hazard in the future, extremely large wildfires are increasingly plausible in the Clackamas Basin. In fact, at the time of this writing, current events are demonstrating the plausibility of extreme westside wildfires. On 7 September 2020 a strong east wind, low relative humidity, and record temperatures fueled western Oregon wildfires to burn in excess of 300,000 hectares. One of those fires, the Riverside Fire (~56,000 ha and 72% contained) has burned nearly 40,000 hectares in the Clackamas Basin, very similar to the largest single-fire area burned that we simulated under baseline climate conditions (37,509 ha). The Riverside Fire confirms that westside landscapes are vulnerable to infrequent, large fires and that simulated very large fire sizes are not unrealistic.
Our results indicate that very large fire sizes could increase significantly and that very large fires could be as much as eight times more frequent. This is consistent with [115] which projected an approximate fivefold increase in fires greater than 20,000 ha across the PNW under RCP 8.5. in mid-century. In a study of future fire size across Canada, [116] found that increasingly severe fire weather led to more days conducive to fire growth and a three to nine fold increase in the occurrence of extremely large wildfires. However, both [115] and [116] illustrate that projections of future fire sizes are closely linked to the emissions scenario, in which case our results ought be interpreted as plausible, but high-end projections of future fire sizes in the Clackamas Basin.
Our results demonstrate the importance of and provide an opportunity for climate adaptation and risk reduction planning in westside forests. The objective of this paper was to provide information that will be locally actionable for planners and managers in the Clackamas Basin, but we stop short of attempting to evaluate potential adaptation strategies. Our results illustrate a wide range of plausible future wildfire hazard, and, given that uncertainty, robust adaptation plans will be ones that maintain essential ecosystem services across the broadest range of future hazard scenarios [8,117,118]. Ref. [12] recommends continued, strategic fire suppression as a way to maintain ecosystem services in westside forests characterized by infrequent, stand replacing fires. However, our results project that fires could be burning more frequently under severe weather conditions which would limit the success of suppression and lead to higher costs [119,120]. Moreover, suppression in and of itself is not an effective land management strategy and could proliferate negative feedbacks and inhibit post-wildfire drought-adapted vegetation transitions [112,113]. A logical next step is for managers to assess likely consequences to specific resources and assets (i.e., timber value, water quality, communities, etc.) and then to use methods like those presented in [54,[121][122][123], to develop climate adaptation strategies that balance forest management, fire suppression, and community preparedness to achieve a range of risk reduction goals.

Conclusions
This study simulated a plausible range of changes to mid-century fire regimes in a historically low-to moderate-frequency fire regime in western Oregon, USA. Simulation models in relatively low-frequency fire regimes are useful risk planning tools given the lack of actual wildfires in recent history, and the concentration of resources and assets. Moreover, projecting future fire regime characteristics is important given that climate change is projected to create annual conditions more conducive to large fires.
Our results demonstrated that wildfire hazard will likely increase by mid-century as a result of larger, more frequent fires. Future wildfires are expected to grow larger as fire weather conducive to large fires becomes more common and longer lasting. Projected changes in temperature and relative humidity led to longer fire seasons and more severe fire weather in three of the four scenarios. Under the hottest and driest mid-century scenarios, annual burn probabilities were similar to those found in higher frequency fire regimes. All climate and baseline scenarios illustrate that extremely large, intense fires are plausible, and that they will become more plausible under hotter and drier climate scenarios. The range of plausible impacts to mid-century fire regimes should be cause for managers and planners to adopt robust approaches to climate adaptation and risk mitigation. The hottest and driest mid-century scenarios could lead to disturbance regime changes resulting in significant changes to forest structure and composition, and as a result could pose significant challenges to ecosystem service provision. Even under more moderate mid-century climate scenarios, infrequent and very large wildfires remain a potential threat to ecosystems and communities.
Our work provides actionable information for regional managers and planners to guide decisions around climate adaptation in land use management, wildfire management, forest management, water provision and other fields which would be affected by wildfire. Ongoing research aims to evaluate the relative importance of fuel structure, fuel moisture and wind speed on fire size and intensity in westside forests. Based on this study, future work is needed to develop and test risk mitigation strategies that might be effective given the uncertainty demonstrated in mid-century fire regime characteristics. Finally, our results do not specifically address the role of rare wind events, such as those that led to the catastrophic westside fires in 2020. Additional work evaluating the ability of FSim to capture these events would be beneficial and supplement our analysis.