3.1. Historical Comparisons
 During the historical period (hereafter 1971–2000 means), the study domain averaged 8.4°C and 1182 mm mean annual precipitation (MAP), with a clear distinction between cool wet winters and warm-hot dry summers. Simulated NPP peaks in June, decreases mid-summer due to water stress, and again in the fall due to temperature limitations (Figure 2).
Figure 2. Mean monthly historical (1971–2000) temperature, precipitation, water stress, and net primary production (NPP). Water stress is defined as 100*(1 – ws/PET), where ws = mean monthly available soil water and PET = potential evapotranspiration.
Download figure to PowerPoint
 MC1 results compared favorably with observations during the historical period, although there are a few sources of disagreement. First, MC1 vegetation distribution fails to capture the mixed open forests of the Willamette Valley and the pattern of grasslands and shrublands on the exterior edges of the Columbia Plateau (Figure 3). However, Native Americans and European settlers greatly modified the fire regime in the Willamette Valley [Whitlock and Knox, 2002], thus allowing the establishment and maintenance of mixed oak forests and woodlands. Additionally, the Columbia Plateau is particularly vulnerable to both grass [Keane et al., 2008] and woodland [Belsky, 1996] encroachment, promoted by late-twentieth-century climate trends, fire suppression, and early-twentieth-century grazing. These issues complicate the assignment of potential vegetation communities in the two regions.
 Second, while MC1 captures the broad spatial patterns of burn areas (Table 1), it often misses the magnitude and timing of individual fire years (Figure 4). Although climate exerts a dominant control on fires at large spatial and temporal scales [Flannigan et al., 2000], fine-scale patterns of ignition, weather, and suppression render temporal patterns of burn area more difficult to match on an annual basis. Simulated large (un-suppressed) fires account for 95.2% of the overall simulated burned area.
Table 1. Historical Burn Areas From Westerling et al.  and MC1 Simulated With Fire Suppression
|Region||Observations (1980–2004) (% area yr–1)||MC1 (1980–2004) (% area yr–1)|
 Third, MC1 overestimates combustion factors (fraction of biomass combusted by fire) in the Biscuit Fire and may therefore overestimate carbon losses from fires throughout the domain and time series. High severity fires typically account for ∼20% of burned area in the PNW (G. Meigs, manuscript in review), and observations made by Campbell et al. , disaggregated by carbon pools to match MC1's structure, show low, medium, and high severity combustion factors from the Biscuit Fire to be 0.14, 0.17, and 0.23, respectively. MC1 simulates a mean of 0.26.
 Finally, MC1 overestimates carbon stocks in Western Forests (Figure 5b). However, many areas within Western Forests, particularly the Coast Range, the Willamette Valley, and to some extent the western Cascades, were subject to heavy human influence during the twentieth century, including land conversion, logging, and urbanization. These activities all acted to decrease total ecosystem carbon [Smithwick et al., 2002]. Comparisons with old-growth forests yielded mixed results depending on the data source (Figure 5c), which may in part result from differing methods of plot selection.
Figure 5. Comparisons of carbon pools and fluxes between MC1 and observations: (a) FIA plots across Oregon from Hudiburg et al.  (5,093 plots, 90.4% on public lands, with stand ages of 212 ± 134 years), (b) aboveground live forest carbon from Blackard et al. , (c) old-growth plots and MC1 run without fire (Smithwick data [Smithwick et al., 2002] contain 37 plots on public lands with stand ages of 429 ± 257 years; EPA data [Hudiburg et al., 2009] contain 8 plots on public lands with stand ages of 417 ± 215 years; and FIA data [Hudiburg et al., 2009] contain 1,607 plots, 98.2% on public lands, with stand ages of 332 ± 123 years), (d) domain-wide monthly net primary production (NPP) from the MODIS Aqua product [Running et al., 2004] (with a 24-day moving average filter), and (e) NPP derived from flux measurements at the Metolius Intermediate Pine site tower [Law et al., 2003] (44.4523° lat, −121.5574° lon).
Download figure to PowerPoint
3.2. Future Projections
 Future climate projections (2070–2099 means) display both similarities and differences in the seasonality and magnitudes of changes. Temperatures rise ubiquitously (Figure 6a) with larger increases in summer than winter (Figure 6c). Consistent with the findings of Mote and Salathé , increased precipitation generally occurs in winter and decreased precipitation occurs in summer months (Figure 6d). Comparatively speaking, the CSIRO climate projection is cool and wet (+2.6°C and +176 mm MAP), MIROC is hot and wet (+4.2°C and +82 mm MAP), and Hadley is hot and dry (+4.2°C and −78 mm MAP). Calculating the interannual variability of precipitation relative to a scenario's linear trend, the future CRISO projection displayed more (167 mm yr−1), MIROC displayed similar (149 mm yr−1), and Hadley displayed less (124 mm yr−1) variability than the historical time series (149 mm yr−1). Results were similar when variability was calculated relative to a scenario's previous year's precipitation (CSIRO = 184 mm yr−1, MICOC = 149 mm yr−1, Hadley = 125 mm yr−1, and Historical = 151 mm yr−1).
Figure 6. Projected annual (a) temperature and (b) precipitation, and monthly changes in (c) temperature, (d) precipitation, (e) water stress, and (f) net primary production (NPP). Shaded areas in (a) and (b) indicate the time periods used for analysis in (c)–(f) (1971–2000 and 2070–2099).
Download figure to PowerPoint
 Mote and Salathé  compared 21 GCMs used in the IPCC AR4 [Alley et al., 2007] to CRU version 2.02 climate data [Mitchell et al., 2004] over the PNW and found that each of our three selected GCMs showed its own strengths and weaknesses. Notably, Hadley produced one of the lowest precipitation biases (both annually and seasonally), yet was relatively poor in representing the spatial distribution of meteorological fields and had a near-zero twentieth-century temporal temperature trend. MIROC displayed one of the lowest temperature biases, yet one of the highest precipitation biases. CSIRO ranked highest of all models for its twentieth-century temperature trend, yet had the highest positive temperature bias. Otherwise, the three GCMs selected here performed close to average for all other metrics considered.
 With MIROC and Hadley, climate projections lengthen the growing season and amplify the already strong simulated seasonal climatic cycles, thereby increasing NPP during the rainy season and decreasing summer NPP by exacerbating summer drought. Under CSIRO's milder conditions, NPP increases and water stress decreases year-round (Figures 6e and 6f). As a general trend, the relative seasonal amplitude of simulated plant functions and stresses are amplified by future climate projections, thereby increasing summer drought stress and susceptibility to fires, but also increasing productivity during the rest of the year. These ecosystem responses have been observed in the tree ring record [Villalba et al., 1994] and predicted under projected future climates for lodgepole pines in other parts of the American West [Smithwick et al., 2009]. It should be noted that MC1 does not contain a radiation constraint on NPP. The model may therefore overestimate increased productivity due to warmer temperatures in non-summer months, causing the build-up of fuel loads and depletion of soil water, and hence overestimate responses to summer drought and susceptibility to fire.
 Simulated fires increase under all climate projections across the domain (Table 2). Although these increases appear late in the twenty-first century under CSIRO and MIROC, Hadley's hot and dry conditions cause large fires early to mid-twenty-first century (Figure 7), primarily in Western Forests. Because of woody encroachment in the Columbia Plateau and larger and more frequent fires in Western and Eastern Forests under all three scenarios, burn severities (kg C m−2 burned) steadily increase across the domain throughout the twenty-first century (Figure 7) and result in large increases in biomass consumption (Table 2). More frequent forest fires also generally decrease the interannual variability in burn severities. These intensifications of PNW fire regimes are caused by three main factors in the model: (1) increased frequency and intensity of droughts in mesic regions, (2) elevated fuel loads in xeric regions, and (3) higher interannual variability of precipitation, particularly in CSIRO. When a three-year running average filter was applied to CSIRO precipitation by month, which preserves seasonality and long-term means but dampens variability, burn area decreased by 88% compared to historical. It should be noted that while precipitation variability is important for the dynamic fire model, overall drying trends can be the most important factor in some cases, such as under the Hadley scenario, where interannual precipitation variability decreases but burn area increases dramatically due to more frequent summer droughts in Western Forests.
Figure 7. Changes in burn area, burn severity, and ecosystem carbon with fire suppression, as well as ecosystem carbon with full fire and no fire. Shaded areas cover the historical period. Fire suppression was initiated in 1940.
Download figure to PowerPoint
Table 2. Historical (1971–2000 Means) and Future (2070–2099 Means) Changes to Fire Regimes and Carbon Stocks by Ecoregion
|All Domain||Burn areaa||0.326||+76.3||+95.0||+310.1|
|Western Forests||Burn area||0.143||+161.5||+159.6||+1177.4|
|Eastern Forests||Burn area||0.638||+110.6||+141.0||+133.4|
|Columbia Plateau||Burn area||0.362||−28.5||−11.2||+28.4|
 The simulated twenty-first-century PNW carbon budget is a balance between biomass losses from intensified summer drought and fire, and biomass gains from higher rainy season NPP due to increased precipitation, longer growing seasons, and/or CO2 fertilization. The domain gains 1.1 Pg C under CSIRO and 0.9 Pg C under MIROC. Thresholds of summer drought are surpassed such that 1.2 Pg C are lost under Hadley due to large and frequent fires in Western Forests. To put this in context, 1 Pg C is approximately of our current global annual fossil fuel emissions [Le Quéré et al., 2009] and 23 times the size of Oregon and Washington's current combined annual emissions [Oregon Department of Energy, 2010; Waterman-Hoey and Nothstein, 2006].
 Domain-wide impacts have distinct regional differences. Western Forests, typically considered stable with long fire return intervals, proved to be the most vulnerable of the three regions in MC1. These mesic maritime forests are largely unable to benefit from increased winter precipitation because, as has been observed [e.g., Harr, 1977], soils are already saturated in winter. Instead, the region suffers from more intense summer droughts and incurs the greatest relative increases in fires (Table 2). Large fires are simulated in years with summer droughts substantially worse than those during the historical period. These droughts occur more often under CSIRO and MIROC after 2070 and are mainly limited to the southwest part of the domain, but they occur much more frequently under Hadley throughout the twenty-first century and cause fires throughout most of the Western Forests region. Under this latter projection, burn area and biomass consumption increase by an order of magnitude and the region loses nearly a quarter of its large ecosystem carbon stocks. Western Forests are also subject to relatively large-scale forest type conversions, with expansions of subtropical mixed forests under MIROC and temperate coniferous forests under Hadley, and losses of subalpine forests under all three projections (Figure 3). Taken together, simulations of Western Forests under Hadley conditions resemble the climate, vegetation, and fire regimes of the late Holocene Thermal Maximum [Whitlock et al., 2003]. In comparison, both the Columbia Plateau and Eastern Forests gain carbon in all three scenarios despite intensified fire regimes because of longer growing seasons, greater synchrony between spring growth and precipitation, and woody encroachment in the case of the Columbia Plateau (Table 2, Figure 8). The highest spatial agreements between projections occur in the cases of increased fires and ecosystem carbon in the eastern domain, and vegetation shifts in Western Forests (Figure 8).
Figure 8. Number of future scenarios that agree on a change from the historical baseline. Changes in the positive and negative direction are given by positive and negative numbers for carbon and burn area. Changes of less than 5% (carbon) and 10% (burn area) from historical on a gridcell basis were deemed insignificant.
Download figure to PowerPoint
3.3. Sensitivity Analyses
 A number of full-domain fire sensitivity analyses were conducted in order to assess the influence of fire and fire suppression on the carbon balance. MC1 was first run with fire suppression turned off (full fire) and second with all fires turned off (no fire). As expected, fire suppression always produces less burn area and biomass consumption than full fire. However, when compared to results for historical periods with the same fire rules, fire suppression results in greater relative and absolute increases in burn area and biomass consumption than does full fire, under all scenarios (Table 3). This suggests an intensification of future PNW fire regimes due to suppression because (1) simulated (and observed) historical fire suppression is causing elevated fuel loads in semi-arid ecosystems and (2) current fire suppression (as assumed in the model) will not be as effective against intense future fires. Paradoxically, because absolute biomass consumption is less and ecosystems continue to gain carbon after the initiation of simulated fire suppression, suppression results in greater carbon gains (or smaller losses) than full fire simulations (Figure 7). Nonetheless, suppression is unable to curtail the large carbon losses under Hadley's hot and dry climate. Because the domain is a carbon sink under Hadley when fire is turned off, carbon losses are due entirely to large conflagrations in high-biomass forests. A third sensitivity analysis was conducted wherein the suppression thresholds of fireline intensity and rate of spread were raised until future burn area matched historical levels. To reach that goal, fire suppression needed to be effective on fires that were 30% more intense under CSIRO, 41% more intense under MIROC, and 287% more intense under Hadley climate projections. Finally, MC1 was run with the previous fire suppression rule that reduced all burn areas by 87.5%. While historical burn area is similar to those simulated with the new rule (both rules were calibrated to historical data in the U.S.), this simple reduction in burn area results in distinctly different impacts under climate change: ecosystem carbon increases under all scenarios (7%–15%) and burn area changes only slightly (±25%).
Table 3. Burn Area and Biomass Consumed Simulated With Full Fire and Fire Suppressiona
|Scenario||Burn Area (% area yr−1)||Biomass Consumed (g C m−2 yr−1)|
|full fire||fire suppression||full fire||fire suppression|
|CSIROc||1.92 (−1.1)||0.57 (+76.3)||31.5 (+30.1)||20.4 (+127.9)|
|MIROCc||2.01 (+3.9)||0.63 (+95.0)||33.7 (+39.5)||23.7 (+165.3)|
|Hadleyc||2.75 (+41.6)||1.34 (+310.1)||61.5 (+154.3)||51.7 (+477.6)|
 These fire sensitivity analyses suggest that fire will exert a dominant control over future PNW biomass stocks and that the simulation of fire impacts is strongly influenced by how fire suppression is modeled. More importantly in terms of management, the model suggests that current suppression efforts may become less effective against more intense future fires.
 Two Monte Carlo parameter sensitivity analyses were also conducted. The first, in which 100 points were run through 100 random selections of parameters related to fundamental model processes, revealed that our model conclusions are relatively conservative (Figure 9). Because the number of augmented parameters was high (30), the output spread is considerable. However, when the 100 points were aggregated together to represent a sample of the full domain, the median historical ecosystem carbon and burn area values were close to those from the original run. Additionally, in almost every case, the future Monte Carlo simulations result in less ecosystem carbon and more future burn area than our original run. The exception is MIROC burn area, where the median change is slightly less than value from the original run (+78% versus +95%). This suggests that uncertainty in our input parameters leads to more unfavorable projections for the PNW (i.e., more fires and less carbon sequestration).
Figure 9. Probability density functions of (a) historical and (b) future changes to ecosystem carbon, and (c) historical and (d) future changes to burn area from a Monte Carlo sensitivity analysis. One hundred points were simulated with 100 different choices of 30 parameters, chosen from a ±20% Latin hypercube. Points were aggregated to a single value for each of the 100 runs. Stars represent median values from the Monte Carlo runs, and dots represent values from the original full domain run.
Download figure to PowerPoint
 A second parameter sensitivity analysis was conducted by varying only thresholds for fire suppression. With the exception of simulations using the CSIRO climate, results from this analysis are much more tightly constrained than those from the previous (see above). Compared to the original full domain run, slightly less area burns during the historical period (mean of 0.26% versus 0.33%) while slightly more area burns in the future (+350% versus +310% changes under Hadley and +127% versus +95% changes under MIROC). The effects on ecosystem carbon are negligible: changes are within 1% of the original run. Under CSIRO, however, much more area burns than does in the original run (+352% versus +76.3%) and the domain gains substantially less carbon (+8.4% versus +12.2%). This again suggests that uncertainty in our fire suppression thresholds results in exacerbated future projections.
 These results come with numerous sources of uncertainty, some of which may be quantified using Monte Carlo sensitivity analyses such as those above, and others that are more overreaching. The primary control on twenty-first-century climate change will be the trajectory of anthropogenic CO2 emissions, which depends on political and social decisions and are thus highly unpredictable. Additionally, as seen in this study, individual GCMs display their own biases and unique projections for the PNW. Our downscaling method cannot correct for these GCM-simulated biases in annual means and seasonal, interannual, and/or interdecadal climate variability. The method can also not account for local biosphere-atmosphere feedbacks such as increased warming over regions that lose snowpack. Moreover, the monthly time step in MC1 may miss physiologically important daily changes such as differential warming between day and night. This version of MC1 assumes no nitrogen limitation and may therefore miss carbon-nitrogen feedbacks associated with warming and changing fire regimes. Because MC1 does not take radiation effects on NPP into account, it may overestimate future increases in productivity and vulnerability to fire by overestimating growth responses to temperature and related water use. MC1 does not include inter-cell communication, which at smaller scales is important for hydrology, erosion and sedimentation, and fire spread. While some vegetation changes are driven by competition and incorporate biogeochemical processes, others rely strictly on physiologically based climatic indices. The rates of vegetation change, and influences of adaptation, may therefore not be accurately captured.
 MC1 does not account for many direct impacts on the landscape. Changes to agriculture area, farming practices, harvested land, and/or rotation ages may affect carbon stocks in ways unaccounted for in this study. Our fire suppression rule is based on physical metrics of fire intensity, but it includes no information on human population densities, forest fragmentation, fire fighting policies and budgets, or ignition sources. While lightning is the primary ignition source in the PNW [Rorig and Ferguson, 1999], changing storm patterns, arson rates, or population expansion could influence future fires. MC1 does not simulate insects and pathogens, yet mortality from pest outbreaks has been increasing [van Mantgem et al., 2009] and tends to increase fire vulnerability in low elevation dry forests. Finally, the model does not account for herbivory, which could greatly reduce post-fire forest regeneration. Many of these issues suggest model results should be considered relatively conservative.