We investigate the impact of climate change on wildfire activity and carbonaceous aerosol concentrations in the western United States. We regress observed area burned onto observed meteorological fields and fire indices from the Canadian Fire Weather Index system and find that May–October mean temperature and fuel moisture explain 24–57% of the variance in annual area burned in this region. Applying meteorological fields calculated by a general circulation model (GCM) to our regression model, we show that increases in temperature cause annual mean area burned in the western United States to increase by 54% by the 2050s relative to the present day. Changes in area burned are ecosystem dependent, with the forests of the Pacific Northwest and Rocky Mountains experiencing the greatest increases of 78 and 175%, respectively. Increased area burned results in near doubling of wildfire carbonaceous aerosol emissions by midcentury. Using a chemical transport model driven by meteorology from the same GCM, we calculate that climate change will increase summertime organic carbon (OC) aerosol concentrations over the western United States by 40% and elemental carbon (EC) concentrations by 20% from 2000 to 2050. Most of this increase (75% for OC and 95% for EC) is caused by larger wildfire emissions with the rest caused by changes in meteorology and for OC by increased monoterpene emissions in a warmer climate. Such an increase in carbonaceous aerosol would have important consequences for western U.S. air quality and visibility.
 Many of the above studies predict changes in fire indices, but estimates of emissions from fires require predictions of area burned. Flannigan et al.  investigated relationships between climate and the areas of fires in Canada. Stepwise linear regression was used to derive the best predictors of area burned, chosen from meteorological variables (surface temperature, rainfall, wind speed, and relative humidity) and calculated values of forest fuel moisture from the Canadian Fire Weather Index (FWI) System. Temperature and fuel moisture explained between 36 and 64% of the variance in monthly area burned depending on the ecosystem. The Canadian and Hadley Centre GCMs were used to predict increases in area burned of 74–118% under a 3 × CO2 scenario. RCMs have also been used to study wildfire area burned in limited regions of Canada. For the boreal forests of Alberta, Tymstra et al.  used an RCM to predict a 13% increase in area burned in a 2 × CO2 scenario and a 30% increase in a 3 × CO2 scenario. Most of these studies did not account for any future changes in ignition sources. Price and Rind  used empirical lightning and fire models along with the Goddard Institute for Space Studies (GISS) GCM to predict that more intense convection under a 2 × CO2 scenario leads to increased lightning and a 78% increase in area burned in the United States.
 Despite these efforts to predict the effect of future climate on wildfires, there have not been studies of the impact of these future wildfires on air quality. In this paper we predict how wildfires in the western United States will respond to changes in climate between the present day and 2050 and evaluate the impacts on aerosol air quality (see Figure 1). We apply the technique of Flannigan et al.  to the western United States, building regressions between observed wildfire area burned [Westerling et al., 2003] and observed climate. Projections of future climate, calculated by the GISS GCM, are used to predict changes in wildfire area burned. We use the GEOS-Chem chemical transport model (CTM) driven by meteorology from the GISS model to quantify the impact of changing wildfire on carbonaceous aerosol concentrations.
2. Predicting Wildfire Emissions for 2000–2050
 Here we describe our prediction of future wildfire emissions of carbonaceous aerosol in the western United States, defined as the domain 31°–49°N, 125°–100°W (from the Pacific Coast to eastern Colorado and from the Mexican border to the Canadian border).
2.1. Area Burned Predictions
 We extend the approach of Flannigan et al.  to the western United States, building regressions of observed area burned with surface meteorological data and output from the FWI model. Observed area burned was taken from the database of Westerling et al. . They used reports from various agencies in the United States that provided the area burned on federal land, and the start and end date of individual fires, from 1980 to 2000. This database has been extended to 2004. Westerling et al.  assumed that the fires burn entirely in the month during which they started (end dates are often unreliable), and the areas were aggregated on a grid of 1° × 1°. Because the gridded database used only the start date of each fire, it may not accurately reflect the seasonal dependence of each fire season. In addition, the wildfire time series is relatively short, and if there are only a few extreme events, it is difficult to fit with the least squares approach used here. For these reasons we chose to predict annual area burned.
 Area burned was binned according to the ecological stratification of Bailey et al. . This system defines 18 ecoprovinces in the western United States. These ecosystems were further aggregated to produce six ecoregions with similar vegetation and climate, as shown in Figure 2. We tested our regressions with the original 18 ecosystems, but found we could better fit area burned for the larger ecoregions, as did Flannigan et al.  for Canada. This is probably caused by meteorological factors that influence area burned operating at synoptic scales, and larger spatial units providing some statistical smoothing of noisy data for area burned. Littell et al.  find strong relationships (R2 of between 0.31 and 0.87) at the original ecoprovince level, potentially because they also include the Palmer Drought Severity Index and antecedent meteorological variables within their regressions, and because they first extracted the principal components of the times series of meteorological data before fitting the area burned data. The coarse spatial resolution of the global climate model limits the scale at which we can make future projections of climate and so area burned. For this reason we use the six aggregated ecoregions (Pacific Northwest, California Coastal Shrub, Desert Southwest, Nevada Mountains/Semidesert, Rocky Mountains Forest, and Eastern Rocky Mountains/Great Plains) for the rest of this work.
 We obtained from the USDA Forest Service data for four meteorological variables important to wildfire frequency and required as input to the FWI model: daily 12.00 local standard time temperature, relative humidity, wind speed, and 24-h accumulated rainfall (available at http://fam.nwcg.gov/fam-web/weatherfirecd/). Meteorological stations were selected if they reported data for at least two thirds of the 1980–2004 time period and if the altitude of the station was within 500 m of the mean altitude of all stations within that ecoregion. Table 1 shows the number of stations selected for each ecoregion. Temperature and relative humidity values at each station were adjusted to the mean elevation of the stations. For temperature the adjustment is based on the U.S. Standard Atmosphere lapse rate of −6.5 K/km. Relative humidity was then recalculated using the adjusted temperature. For 1980–2004 we calculated a daily value for each meteorological variable and each ecoregion as an average across the selected meteorological stations for each ecoregion.
Table 1. Wildfire Area Burned Regressions for Aggregated Ecoregions in the Western United Statesa
Description of Bailey ecosystems appears in Figure 1. For each ecoregion the number of meteorological stations and the mean elevation of the stations above sea level is shown.
Predictors are chosen from maximum and mean daily May through October values of meteorological variables and components of the Canadian Fire Weather Index System (see Figure 3): T (°C), Temperature; DC, Drought Code; FFMC, Fine Fuel Moisture Code; FWI, Fire Weather Index; Rain, accumulated 24-h rainfall (mm); BUI, Build-up Index; DSR, Daily Severity Rating.
= −3.1 × 106 + 9.4 × 104 T + 1.7 × 103 DC
Californian Coastal Shrub
= −3.6 × 106 + 3.4 × 104 Tmax + 2.6 × 104 FFMC
= −1.64 × 106 + 6.3 × 104 T + 296 DCmax
= 3.9 × 105 + 1.2 × 104 FWImax−1.4 × 106 Rain
Rocky Mountains Forest
= −6.55 × 106 + 3.2 × 105 T +5.3 × 103 BUImax
Eastern Rocky Mountains/Great Plains
= −3.6 × 105 + 3.4 × 104 DSR
 The calculated daily values of the four meteorological variables were used as input to the Canadian FWI System [Van Wagner, 1987]. The model calculates daily fuel moisture codes and fire severity indices using these four variables to track changes in forest fuel moisture. A schematic of this model is shown in Figure 3. The fuel moisture codes describe the moisture content of three distinct fuel layers in the forest floor. The Fine Fuel Moisture Code (FFMC) represents surface fuel litter and changes rapidly with short-term changes in atmospheric moisture (time lag of 2/3 day). The Duff Moisture Code (DMC) represents loosely compacted organic layers (time lag of 15 days) and the Drought Code (DC) represents deep layers of compacted fuel and reacts to seasonal droughts (time lag of 52 days). The fire severity ratings combine information from the fuel moisture codes to give an indication of the fire danger or rate of fire spread. The Build-up Index (BUI) combines DMC and DC and is an indication of the availability of fuel for consumption. The potential rate of spread of a fire is calculated by combining wind speed and FFMC to give the Initial Spread Index (ISI). The ISI and BUI are combined to give the Fire Weather Index (FWI) which is commonly used as a general index of fire danger. The Daily Severity Rating (DSR) is an exponential function of the FWI and gives an indication of the difficulty of fire control.
 Linear forward stepwise regression was used for each of the six ecoregions with annual area burned as the predictand. For predictors we used the maximum and mean of the daily May to October values of temperature, relative humidity, and wind speed and the seven output fields from the FWI model (described in section 2.1). In addition we used May to October mean daily rainfall and total May to October rainfall. This gives 22 potential predictors. We used the same test for significance as Flannigan et al. ; terms were accepted only if they met a significance level (p value) of 0.15. The predictor with the highest correlation coefficient was added to the regression first. Predictors were then added in the order that maximized the correlation coefficient, until the correlation coefficient did not increase by a preselected amount (typically 2%), or until a predictor was selected that resulted in a nonphysical (counterintuitive) relationship between area burned and fuel moisture. In general, two predictors were selected for each ecoregion. We tested our method using both area burned and natural logarithm of area burned. We found similar correlation coefficients with the two predictands. For the rest of this work, prediction of linear area burned was used.
Figure 4 shows a comparison of observed and predicted annual area burned in the Pacific Northwest and Rocky Mountains Forest ecoregions. The observations show large interannual variability in area burned, with a range of 7500 ha/a to 440,000 ha/a in the Pacific Northwest and 4800 ha/a to 1.45 million ha/a in the Rocky Mountains Forest. In the Pacific Northwest the regression explains 52% of this interannual variability. The chosen predictors for this region are mean drought code and mean temperature, which correlate with area burned with R2 of 46 and 43%, respectively. The greatest observed area burned occurred in 1987 and 2002 coincident with high DC and T. Predicted area burned is also maximum during these 2 years but is underestimated by about 40%. In the Rocky Mountains Forest the regression explains 47% of the variability in annual area burned. The best predictors for this region are mean temperature and maximum Build-up Index, which correlate with area burned with R2 of 42 and 40%, respectively. The regression underpredicts the largest fire year in 1988 by about 60%, but other large fires years (e.g., 2000 and 2001) are well predicted.
Table 1 shows the best predictors and the explained variance for the six ecoregions in the western United States. The regressions explain 24–57% (mean 45%) of variance in annual area burned. Our relationships are weaker than recent work by Littell et al.  who found 33–87% (mean 64%) for the reasons discussed earlier in this section. The explained variance in our work is generally greater in forest dominated ecosystems (48–52%) than in shrub and grass dominated ecosystems (24–49%). The lower explained variance in these ecosystems is likely due to the importance of the previous year's climate for fire activity in these areas [Westerling et al., 2002; Westerling and Bryant, 2008; Littell et al., 2009] which we do not take into account with our method. However, as we show in section 2.3, these shrub and grass dominated ecosystems have limited impact on regional particulate air quality because of low fuel loads.
 Best predictors of area burned are ecosystem dependent but generally include temperature and fuel moisture codes (FFMC or DC). Temperature is the most commonly chosen predictor in the western United States, as has been found previously in Canada [Flannigan et al., 2005]. High temperatures are associated with clear skies, persistent stagnation, and dry fuel: conditions that favor wildfire occurrence.
2.2. Simulation of Future Area Burned
 To calculate future area burned, we archived daily mean temperature, relative humidity, and local noon wind speed as well as 24-h accumulated rainfall from the GISS simulation for 2000–2050. We used the “q flux” version of the GISS GCM 3 [Rind et al., 2007], which has a horizontal resolution of 4° × 5° and 23 vertical sigma levels between the surface and 0.002 hPa. In the q flux version, ocean heat transport fluxes are kept fixed while ocean temperatures and ocean ice respond to changes in climate. Observed concentrations of well-mixed greenhouse gases, ozone, and aerosols were used for the model spin-up between 1950 and 2000, starting from a climate equilibrium [Hansen et al., 2002]. For 2001 to 2055 we used concentrations of well-mixed greenhouse gases from the IPCC SRES A1B scenario, with CO2 calculated using the Bern-CC model [Houghton, 2001]. Under this scenario, CO2 mixing ratios reach 522 ppm by 2050. We assumed no changes in ozone or aerosols from 2001 to 2055 for the purpose of calculating climate change. This model predicts global mean July temperatures to increase by 1.8°C from 2000 to 2050.
 Because the GISS GCMs tend to have a warm continental bias [Schmidt et al., 2006], we scaled temperature as well as the other fields to match observations by multiplying the GISS output by the ratio of mean observed to mean GISS values in each model grid square for May to October of 1990–2000. The adjusted GISS meteorology was used as input for the FWI model to calculate daily fuel moisture parameters. The regressions developed in section 2.1 were then applied to GISS and FWI output to predict ecosystem specific annual area burned.
Figure 5 shows the simulated changes between 2000 and 2050 in mean May through October noon values of the four meteorological variables over the western United States; we compare the means for 1996–2005 and for 2046–2055. Temperatures increase across the western United States by 1–3°C, with the largest increases in the Pacific Northwest and Nevada Mountains/Semidesert ecoregions. The projected change in temperature over the western United States is large, 1.5–1.8 times the standard deviation in May–October mean temperature, even though 2049 has a mean temperature below the 1996–2006 mean. Precipitation and relative humidity increase by 7% and less than 2%, respectively, across the western United States with the greatest increases in the Eastern Rocky Mountains/Great Plains and Desert Southwest ecoregions. Mean wind speeds are projected to decrease slightly throughout the western United States. These climate projections lie within the range reported by the Intergovernmental Panel on Climate Change [Seager et al., 2007; Christensen et al., 2007], giving us confidence that the climate projections from the GISS model are robust.
 We applied the adjusted GCM meteorology to our area burned regression model to predict annual ecoregion area burned from 1996 to 2055. Table 2 shows the average observed area burned for 1980 to 2004 and average calculated area burned for a 10-year period in the present day (1996 to 2005) and in the future (2046 to 2055) for each of the six ecoregions. Average area burned in the present day is calculated with a normalized mean bias (NMB = 100% × (Pi − Oi)/Oi, where Pi is the predicted area burned during year i and Oi is the observed area burned during year i and N is the number of years in the analysis) of +10% to −25% depending on the region. Figure 5e shows simulated annual area burned for 1996–2005, and Figure 5f shows the ratio of area burned in 2046–2055 to that in 1996–2005.
Table 2. Annual Mean Observed (1980–2004) and Simulated Area Burned ± 1σ by Ecoregion in the Western United Statesa
The standardized departure is the absolute change [F-PD] divided by the standard deviation [σ(1996–2055)] in predicted area burned.
The trend in predicted area burned (1996–2055) is fitted using linear regression and the slope of the best fit line is reported.
1.08 ± 0.39
1.08 ± 0.70
1.92 ± 0.79
1440 ± 470
Californian Coastal Shrub
0.59 ± 0.17
0.60 ± 0.43
0.84 ± 0.31
320 ± 310
0.74 ± 0.14
0.81 ± 0.22
1.16 ± 0.18
690 ± 160
2.84 ± 0.71
3.04 ± 1.03
3.14 ± 0.67
−140 ± 590
Rocky Mountains Forest
2.07 ± 0.71
1.53 ± 1.52
4.19 ± 1.76
5010 ± 1220
Eastern Rocky Mountains/Great Plains
0.78 ± 0.26
0.55 ± 0.85
0.50 ± 0.67
−400 ± 540
 Total area burned across the western United States is projected to increase by 54% for 2046–2055 relative to 1996–2005. This projected increase is significant (student's t test p = 0.03). Area burned is predicted to increase in all regions except the Eastern Rocky Mountains/Great Plains where the change is not significant. Statistically significant (p < 0.05) increases in area burned are projected for the Rocky Mountains Forest (78%), Pacific Northwest Forest (175%), and Desert Southwest (43%) ecoregions (see Table 2). In these ecoregions, area burned by midcentury is predicted to increase by more than one standard deviation. We used our regression equations (Table 1) along with the predicted changes in meteorological and FWI parameters to quantify the contributions of the different predictors to the change in predicted area burned (see Table 3). Simulated increases in temperature (Figure 5) are responsible for more than 80% of the predicted increase in area burned in these ecoregions.
Table 3. Present-Day (1996–2005) and Future (2046–2055) May–October Values of Predictors Simulated Using the GISS GCM and FWI System and Used to Calculate Area Burneda
Standardized departure is the future minus present-day simulated means divided by the standard deviation for each predictor.
Student's t test calculated from the difference between the present-day and future simulated means.
The percentage contribution to the change in area burned is calculated for each predictor using the regressions in Table 1 and the change in predictor reported here.
Californian Coastal Shrub
Rocky Mountains Forest
 For most of the west, temperature plays the main role in driving future changes in area burned. However, the small (but insignificant) reduction in area burned in the Great Plains/Eastern Rocky Mountains is due to increased precipitation simulated by the GCM. In the Nevada Mountains/Semidesert and to a lesser extent in the Desert Southwest ecoregions the impact of increasing temperature is partly offset by increasing precipitation; there is no significant change in area burned in the former region and a 43% increase in the Desert Southwest ecoregion.
Figure 6 shows the interannual variability in predicted area burned and in the predictors used to calculate area burned in the Pacific Northwest and Rocky Mountains Forest ecoregions. Interannual variability in predicted area burned is similar to that in observed area burned. We fit the predicted trend in area burned for each ecoregion during the period 1980 to 2004 using linear regression and found that the Pacific Northwest, Desert Southwest, and Rocky Mountains Forest ecoregions have significant positive trends, as shown in Table 2.
2.3. Production of Wildfire Emissions
 To calculate emissions from wildfires, we took the following steps. We first converted annual area burned to monthly area burned by using the average observed seasonal variability of wildfire in each ecoregion in 1980–2004. We assumed that the seasonality of wildfire remains the same in the future and that the emissions of wildfires are constant across a month. Calculated ecoregion area burned was mapped onto a 1° × 1° grid using the observed area burned data to constrain the typical spatial extent of fires within each ecoregion. Figure 7 shows the fraction of 1° × 1° grid squares that contain 70% of observed annual area burned in any 1 year in each of the six ecoregions. We repeated this procedure for 60, 80, and 90% of area burned, but changing this had little impact on our results. For all ecoregions, 70% of area burned in a particular year occurs in 5–25% of the ecoregion. To match this observed behavior, we place 70% of projected area burned in 10% of 1° × 1° grid squares in each ecoregion. We locate these grid squares randomly within each ecoregion. The remaining 30% of area burned was averaged across the remaining 90% of grid squares within the ecoregion. To check for bias introduced by the random placement of wildfires within each ecoregion, we created a 100-member ensemble of simulations of yearly biomass consumption, each with a different randomly chosen set of wildfire locations. Our results showed only a 1% one-sigma variation in total biomass consumption over the simulation period.
 Fuel consumption was calculated from a detailed database of fuel loading (i.e., available fuel) for the United States developed by the U.S. Forest Service, the Fuel Characteristic Classification System (FCCS) (http://www.fs.fed.us/pnw/fera/fccs/), and the fraction of fuel loading that is burned as a function of fire severity for different categories of fuel. We assume that fires burn with 25% high, 25% medium, and 25% low severity, and that 25% of the area is unburned, on the basis of an analysis of the largest wildfires in 2002 in the lower 48 states [Randall, 2004]. The FCCS gives the fuel loadings for many fuel beds with resolution of 1 km × 1 km [McKenzie et al., 2007; Ottmar et al., 2007; Riccardi et al., 2007]. The loading for each fuel bed is given for various size categories of fuels that correspond to time-lag fuel categories commonly used in fire behavior modeling in the United States. Fuel consumption was calculated from fuel loadings by adopting the fraction of fuel that is burned in low-severity, medium-severity, and high-severity fires in six classes of fuel given in Table 4. The fractions in Table 4 are used by the U.S. Geological Survey to derive the severity of a burn using field observations of fractional consumption after the fire (http://frames.nbii.gov/ffi/docs/Composite_Burn_Index.pdf). We matched the FCCS fuel categories to the fuel classes of the Composite Burn Index in Table 4 as follows: (1) litter/light fuels: FCCS categories of litter and woody fuels <7.6 cm in diameter, (2) medium fuel: FCCS woody fuels 7.6–20.3 cm in diameter, and (3) heavy fuel: FCCS woody fuel >20.3 cm in diameter. Fuel consumption was calculated for each 1 km2 grid, weighted by severity, and then an average formed for each 1° × 1° grid, the resolution of our area burned product. Average fuel consumption is shown in Figure 8a.
Table 4. Fractional Fuel Consumption as a Function of Fire Severity
Depth of duff was converted to fuel loading in western forest ecosystems using a density of 1.63 kg/m2/cm depth [van Wagtendonk et al., 1998].
 Emissions of carbonaceous aerosol from wildfires were calculated using the predicted 1° × 1° wildfire area burned maps combined with the 1° × 1° fuel consumption product and emission factors from Andreae and Merlet . We assume that fuel loadings and fire severity do not change between present day and 2050, so that the emissions of carbonaceous aerosol per unit area burned do not change over the simulation period. That is, the simulated change in wildfire emissions is driven solely by changes to wildfire area burned. A database of monthly biomass consumption on the 1° × 1° grid is available on request.
 We find that wildfires in the Pacific Northwest and Rocky Mountains Forest ecoregions dominate present-day biomass consumption by fires in the western United States, accounting for 30 and 43%, respectively, of the total consumption for 1980–2004 using the observed area burned from Westerling et al.  with updates (see Table 5). Wildfire in these two ecoregions will therefore have the largest potential impact on regional air quality. The Nevada Mountains/Semidesert ecoregion has the greatest area burned, 35% of the total for 1980–2004 (Table 2), but accounts for only 7% of the total biomass consumption because of low fuel loads. Similarly, the Californian Coastal Shrub and Desert Southwest with 16% of area burned account for only 6% of total biomass consumption. Figures 8b and 8c contrast the different spatial distributions of area burned and fuel consumption in the western United States.
Table 5. Annual Mean Dry Biomass Consumption by Wildfire in the Western United States
Student's t test p value calculated from the difference between the present-day and future simulated means.
California Coastal Shrub
Rocky Mountains Forest
Eastern Rocky Mountains/Great Plains
Western U.S. Total
 The calculated annual mean dry biomass consumption based on observed area burned is 14.2 Tg for 1980–2004. Predicted mean dry biomass consumption in the western United States, averaged over the 100-member ensemble of simulations, increases from ∼13.8 Tg/a for 1996–2005 to ∼26.4 Tg/a for 2046–2055, an increase of ∼90% (see Table 5). This increase is statistically significant (p < 0.01). The linear trend in biomass consumption is 0.23 ± 0.07 Tg/yr. Figure 9 shows the trend in predicted dry biomass consumption by wildfire in the western United States for 1996–2055 plotted as the standardized departure from the mean for 1996–2005 (Standardized departurei = [Pi− (1996–2005)]/standard deviation(P1996–2005), where Pi is the biomass consumption in year i). A low fire year in 2046–2055 (except for 2049) is about one standard deviation above the 1996–2005 mean and is equivalent to a high fire year during 1996–2005.
 The projected increase in future area burned may result in a decrease in the fire return interval, and thus a decrease in the average stand age and in the mean fuel load, with implications for our projected future emissions from wildfires. A maximum reduction in fuel burned can be estimated by assuming that there is no recovery of vegetation after a wildfire; that is, the fuel loading is zero if the same area reburns, with random placement of fires within each ecozone each year. When we make this assumption our projected biomass consumption in the western United States in 2046–2055 is 18% less than when we assume constant fuel consumption. This is obviously an unrealistic scenario but gives a plausible estimate of the maximum reduction in fuel consumption caused by an increase in area burned over 50 years. For the rest of this work we assume that fuel loads do not vary over the simulation period.
3. Simulations of Atmospheric EC and OC
 We use our simulated wildfire emissions along with a global chemistry model to calculate the changes in aerosol air quality over the western United States.
 Meteorological output from the GISS GCM was archived with 6-h resolution (3 h for surface quantities and mixing depths) and used as input to the CTM. We used GEOS-Chem model version v7.04.01 (see http://www-as.harvard.edu/chemistry/trop/geos) with the same spatial resolution as the GCM, 4° × 5°.
 The model treats EC and OC with a hydrophobic and hydrophilic fraction for each (giving four advected tracers). Combustion sources emit hydrophobic aerosol which become hydrophilic with an e-folding time of 1.2 days [Cooke et al., 1999; Chin et al., 2002]. We assumed that 80% of EC and 50% of OC emitted from primary sources are hydrophobic [Cooke et al., 1999; Chin et al., 2002; Chung and Seinfeld, 2002]. Anthropogenic emissions of OC over the United States are from Cooke et al.  with the correction factor from Park et al. . Biofuel OC emissions are from Yevich and Logan  and from Park et al.  for the United States. In the western United States, we used fire emissions calculated as described in section 2.3, using one member of the ensemble of simulations with random placement of fires within each ecoregion. Outside of the western United States we used climatological biomass burning emissions derived from Lobert et al.  with seasonality from Duncan et al. . We assumed that wildfire emissions are constant across each month. Biomass burning emissions were emitted into the boundary layer. The model includes secondary organic aerosol (SOA) formation from biogenic terpenes. Emissions of monoterpenes were calculated using Guenther et al.  and vary according to temperature and solar radiation. We did not account for the effects of changing CO2 concentrations on vegetation [e.g., Constable et al., 1999] or changing land cover [Sanderson et al., 2003] on biogenic monoterpene emissions. We assumed a 10% carbon yield of hydrophilic OC from terpenes [Chin et al., 2002]. We do not account for the formation of SOA from isoprene which over the western United Stated may contribute a similar amount of SOA to that from monoterpenes [Liao et al., 2007]. A global evaluation of GEOS-Chem EC and OC is given by Chin et al. , and a more detailed evaluation over the United States can be found in the work of Park et al. .
3.2. Impact of Future Wildfires on Carbonaceous Aerosol Concentrations
 The short lifetime of EC and OC results in concentrations over the western United States being dominated by emissions from the United States. In the work of Spracklen et al.  we showed that interannual variability in western United States wildfire emissions controls much of the observed interannual variability in summertime atmospheric OC concentrations. Park et al.  showed that trans-Pacific transport from natural and anthropogenic Asian sources contributes only 2% of the United States OC burden. Here we make a first prediction of the impacts of climate change on future carbonaceous aerosol concentrations in the western United States resulting from a change in the area of western United States fires, assuming that wildfires outside the western United States remain constant. To isolate the impacts of changes in fires resulting from changes in climate, we maintained anthropogenic emissions of EC and OC from fossil fuel and biofuel sources at their present-day values.
 We performed two 5-year simulations for the present day (1996–2000) and for the mid-21st century (2046–2050). Each model run was initialized with a 1-year spin-up. Figure 10 shows simulated summertime OC and EC concentrations in the western United States for these two time periods. Summertime mean concentrations of OC over the western United States increase from 1.4 μg m−3 to 2.1 μg m−3 over 50 years (an increase of 40%) whereas EC increases from 0.18 μg m−3 to 0.21 μg m−3 (18%). The smaller fractional increase in EC concentrations is a result of EC in the western United States being more dominated by fossil fuel emissions than is OC [Spracklen et al., 2007]. The maximum increase in carbonaceous aerosol concentrations occurs over the northwest United States (Oregon, Washington, Idaho, western Montana, and northern California, United States) where absolute OC increases 1–4 μg m−3(15–70%) and EC aerosol increases 0.05–0.2 μg m−3 (10–70%). These large increases in carbonaceous aerosol are caused by the large increase in area burned simulated for the Pacific Northwest and Rocky Mountains Forest ecoregions (Table 2). The smaller increases in carbonaceous aerosol in the southwest (southern California, Arizona, and New Mexico) reflect the smaller predicted increases in wildfires in these areas.
 This projected change in aerosol concentrations is caused by a combination of different effects including changes in wildfire emissions, atmospheric transport, wet and dry aerosol deposition, and biogenic emissions of monoterpenes which changes the amount of secondary organic aerosol (SOA). To evaluate the contribution of these different effects on future aerosol concentrations, we perform two additional simulations: one with future wildfires driven by climate change but where we maintain atmospheric transport, deposition, and monoterpene emissions at present-day values, and one with future atmospheric transport, deposition, and monoterpene emissions but where we maintain wildfires at present-day values. Figure 11 summarizes the results for the western United States (31–49°N, 125–100°W) for these scenarios and for the standard present-day and future simulations described in section 3.2. Interannual variability in weather is important as can be seen in the calculated aerosol concentrations: a cold future summer (year 4) shows up clearly as a year with low regionally averaged concentrations of both OC and EC aerosol. As we saw above, a combination of changing atmospheric transport, deposition, wildfires, and monoterpene-derived SOA results in western United States EC concentrations (2046–2050) increasing by 18% and OC concentrations increasing by 40% relative to 1996–2000. In the model simulations where we increase wildfire emissions but keep atmospheric transport, aerosol deposition, and monoterpene emissions constant (future fires and PD climate) we calculate that EC concentrations increase by 17% and OC concentrations increase by 30%. Increased wildfire emissions caused by changing climate therefore dominate our projected future change in carbonaceous aerosols, resulting in 75% of the change in OC and 95% of the change in EC concentrations.
 Simulated EC concentrations with future atmospheric transport, deposition, and monoterpene emissions but with present-day wildfire emissions (PD fires and future climate) increase by only 3%. Using the same GISS model, Wu et al.  calculated a 5% decrease in afternoon mixing depths over the northwest in the future climate, which would increase EC concentrations. However, this effect is offset by increasing precipitation in this region which will increase aerosol wet deposition and reduce EC concentrations. Concentrations of OC in the future climate, but with present-day wildfires, are 14% greater than present day. Most of this change (80%) is caused by increasing temperature driving increased monoterpene emissions and SOA formation, with the remainder caused by changing atmospheric transport and aerosol deposition. The temperatures projected by the GISS model combined with the Guenther et al.  biogenic emissions algorithm predict a 20% increase in monoterpene emissions in the United States by 2050. Biogenic emissions of isoprene may also change in the future resulting in an additional change to SOA that is not considered here. However, the projected future change in isoprene emissions is currently uncertain [Heald et al., 2009].
4. Discussion and Conclusions
 We have quantified for the first time the effect of changing wildfire activity in a warming climate on carbonaceous aerosol concentrations in the western United States in future decades. We used stepwise linear regression to derive relationships between observed meteorology and observed wildfire area burned for 1980–2004. Our regressions are ecosystem dependent, with temperature and fuel moisture explaining 24–57% of the variance in annual area burned. Our focus is on the prediction of wildfire in forest dominated ecosystems that contribute most substantially to carbonaceous aerosol emissions because of their greater fuel loads. Our approach works well for these ecosystems where the meteorology of the particular fire season has the dominant control on fire, but less well for shrub and grass dominated ecosystems where fuel loads, and hence wildfire, is strongly influenced by the previous year's precipitation [Littell et al., 2009].
 Following the IPCC A1B greenhouse gas scenario, the GISS GCM predicts a 2 K increase in summertime temperature and a ∼7% increase in summertime precipitation by midcentury in the western United States. These predicted changes in climate increase projected area burned in the western United States in 2046–2055 by 54% relative to 1996–2005. Predicted changes to area burned vary regionally, from no change to an increase of 175%, because of regional changes in simulated climate combined with varying ecosystem response to a changing climate. The largest increases in area burned are projected for the Pacific Northwest (78%) and Rocky Mountains Forest (175%) ecoregions where wildfire appears to depend most strongly on temperature. This is consistent with the change in wildfire activity observed in the western United States over the past few decades; sixty percent of the observed increase in large (>400 ha) forest fire area burned that occurred between 1970–1986 and 1987–2003 was located in the northern Rockies, and 18% in the Sierra Nevada, Cascades, and coast ranges of Oregon and California [Westerling et al., 2006]. Our method projects little change in area burned by 2050 for the Nevada Mountains/Semidesert and Eastern Rocky Mountains/Great Plains ecoregions because simulated increases in precipitation compensate for increases in temperature in these regions.
 We use the GEOS-Chem CTM, driven by meteorology from the GISS GCM, to predict changes in carbonaceous aerosol concentrations over the western United States. We predict that mean summertime OC concentrations in 2046–2050 increase by 40% (from 1.5 to 2.1 μg m−3), and EC concentrations increase by 18% (from 0.18 to 0.22 μg m−3) relative to 1996–2000. The largest projected increases are in the northwest United States, colocated with the greatest increases in wildfire. Most of the increase in carbonaceous aerosol concentrations (95% for EC and 75% for OC) is caused by increases in wildfire emissions, which are predicted to increase by 90%. Changes in meteorology that occur between present day and midcentury contribute about 5–10% of the predicted change in carbonaceous aerosol concentrations. For OC, the remainder of the change (20%) is caused by the predicted increase in monoterpene emissions, due to rising temperature, and the resulting increase in secondary organic aerosol formation which has been previously predicted [Heald et al., 2008].
 In common with previous studies of future wildfires, there are limitations in this study. We did not account for changes to vegetation, ignition, the length of fire season, or human activity on future area burned. Future wildfires may be modified by changes in the distribution of vegetation caused either by direct anthropogenic land-use change or by climate change. In this paper, we considered a 50-year timescale over which vegetation will not change substantially, unless it is driven by wildfire [McKenzie et al., 2004] or pest outbreaks [Logan et al., 2003; Hicke et al., 2006]. Changes to wildfire due to changes in lightning frequency [Price and Rind, 1994] and change to anthropogenic ignition [Wotton et al., 2003] are also not considered here. A longer wildfire season is possible in a future climate, and an earlier start to the wildfire season has already been observed in the western United States [Westerling et al., 2006]. We predict annual area burned and do not explore possible changes in the length of the fire season. Consequently, we focused on the air quality impact of wildfires in summer (June–August). This is the period of greatest wildfire activity and also the period when air quality degradation from wildfires is at its worst. However, longer wildfire seasons in the future may extend the negative impacts on air quality further into the spring and autumn.
 We assume that fuel consumption per unit area burned does not change over our simulation period. In reality, both fuel loads and fire severity may change, altering future fuel consumption. Increased wildfire area burned may drive reduced fire return intervals, vegetation stand age, and fuel consumption per unit area burned [Fellows and Goulden, 2008]. We estimated the maximum reduction in fuel consumption this effect can have by assuming no recovery of vegetation after a wildfire, such that areas that reburn over the period 1996–2055 result in no fuel consumption. With this assumption we calculate that biomass consumption in the western United States in 2046–2055 is 18% less than when we assume constant fuel consumption. Changes in vegetation caused by climate change, anthropogenic activity, or pest outbreaks will further affect fuel loads. Fuel consumption is also impacted by potential changes in fire severity. Climate change (increased temperature and wind speeds and reduced fuel moisture) may drive increases in fire severity, and hence increases in biomass consumption per unit area. A recent study has made the first projections of future wildfire emissions accounting for both changes in area burned and fire severity. Amiro et al.  find that changes in wildfire biomass consumption from forest floor fuels in the Canadian boreal forest region are dominated by changes in area burned which are projected to increase by 93% under a 3 × CO2 scenario. Under the same scenario, changes in fire severity are projected to increase biomass consumption by 0–18%, depending on the ecosystem. The impact of changing climate on fire severity in the western United States, where the forest types differ from those studied by Amiro et al. , is currently not known. These combined changes in fuel loads and fire severity will likely have important impacts on future biomass consumption, either amplifying or suppressing the projected change in wildfire emissions calculated in this work.
 The vertical extent of wildfire emissions is a further uncertainty in our projected future aerosol concentrations. There is evidence that some fraction of wildfire emissions are injected above the boundary layer [Leung et al., 2007; Mazzoni et al., 2007; Kahn et al., 2007]. How this injection height will vary in the future with potential changes in fire severity and changes in atmospheric stability is also uncertain. In this study we have assumed wildfire emissions are injected into the boundary layer, and we do not consider any future change.
 This study has explored the role of future climate change on wildfire in the western United States. In this work we have evaluated the impacts of future climate change on wildfire area burned and carbonaceous aerosol concentrations over the western United States. We have not been able to account for the full set of variables and processes which are likely to control future wildfire emissions. In particular, future studies need to account for potential changes in vegetation, fuel loading (within the same vegetation type), fire severity, and fire ignition under a changing climate. In addition, it is important to explore the potential uncertainty associated with climate model projections, anthropogenic greenhouse gas emission scenarios, and fire prediction tools. Our predicted increase in wildfire has potential implications for ecology, carbon balance, land management, and fire suppression in western forests. In this work we focussed on the impacts of changing wildfire on carbonaceous aerosol concentrations. We predict summertime mean OC concentrations across the western United States to increase by 40% and EC concentrations to increase by 20% by midcentury relative to present day. This increase will have negative impacts on atmospheric visibility and human health. Carbonaceous aerosol currently accounts for 40% of fine aerosol mass in the western United States [Malm et al., 2004]. Assuming other aerosol components remain unchanged, carbonaceous aerosol will account for ∼50% of fine aerosol mass by midcentury. Future work is required to extend this study to the boreal forests of North America and Siberia, expand the impact to other atmospheric species such as ozone, and to study the potential climate feedbacks of increased wildfire [Randerson et al., 2006].
 This work was funded by the National Aeronautics and Space Administration (NASA-MAP grant NNG06GB48G) and STAR Research Assistance agreement RD-83227501-0 awarded by the U.S. Environmental Protection Agency (EPA). This publication has not been formally reviewed by the EPA. The views expressed in this document are solely those of the authors, and the EPA does not endorse any products or commercial services mentioned in this publication.