Projected carbon stocks in the conterminous USA with land use and variable fire regimes

Introduction

There has been much interest in quantifying terrestrial carbon fluxes to evaluate their capacity to offset fossil fuel emissions. Despite the efforts of many research teams using a bottom-up approach based on observations from either flux towers (e.g., Turner et al., 2013) or inventories (e.g., Pan et al., 2011) and mechanistic models (e.g., Huntzinger et al., 2012) or a top-down approach based on inversion models (e.g., Schuh et al., 2013) and atmospheric sampling (e.g., Deng & Chen, 2011), the magnitude of these fluxes remain uncertain. At the end of the 20th century, terrestrial ecosystems have been shown to act as net carbon sinks likely a result of a combination of increased atmospheric CO₂ concentrations and warmer temperatures as well as recovery from land use (e.g., Nemani et al., 2003). In the USA, carbon gains can be associated with a combination of fire suppression, regrowth from timber harvests and agricultural abandonment, reduction in harvests and increased nitrogen deposition (Houghton & Hackler, 2000a and Houghton et al., 2000b). However, droughts have also caused declines in productivity and increased mortality and general warming has triggered pest outbreaks and facilitated extensive fires that have contributed to a decline in global productivity and carbon stocks (e.g., Westerling et al., 2006; Allen et al., 2010; Zhao & Running, 2010). Understanding the variability of the carbon budget requires a good understanding of the interactions between climate, plant succession, disturbance regimes and human land use legacies (e.g., Loudermilk et al., 2013).

Land use practices have directly affected the carbon cycle by increasing anthropogenic carbon emissions and harvesting large fractions of ecosystem production, but also indirectly by changing land surface albedo and modifying the planet's radiative balance while transforming its hydrological cycle (Foley et al., 2005). As society looks for options to reduce anthropogenic emissions to mitigate climate change effects, the role of management in maintaining carbon stocks and enhancing carbon sequestration potential has become critical. Agriculture reduces soil carbon storage by increasing soil carbon decomposition through increased aeration (Arora & Boer, 2010). Management decisions, such as thinning forests to reduce fire risk or choosing to use land for crops vs. pasture, become choices between higher and lower emissions, lower and higher carbon sequestration options. To evaluate the repercussions of such decisions, simulation models are useful tools to test hypotheses and quantify the results.

Dynamic global vegetation models (DGVMs) were developed to evaluate the impacts of climate change on a variety of ecosystems globally (e.g., Cramer et al., 2001; Kucharik et al., 2006; Sitch et al., 2008) and are now often embedded in earth system models to evaluate the potential biological feedbacks to the atmosphere. They generally include algorithms representing biogeochemical processes as well as biogeography rules based on climate and biomass thresholds, and some disturbance mechanisms that cause shifts in potential vegetation (e.g., Woodward et al., 1995; Moorcroft, 2006; Prentice et al., 2011; Rogers et al., 2011). To provide results that are applicable anywhere around the globe, modelers have simplified their representation of vegetation and described land cover with broad plant functional types and physiognomic groups, rather than specifying individual species assemblages or vegetation communities. Most DGVMs provide results at a spatial grain that is coarser than that at which land use decisions are made (Hurtt et al., 1998) but that are well suited to provide estimates of nationwide carbon stocks and fluxes and the various ecosystems’ capacity to mitigate anthropogenic emissions.

This study was designed to complement earlier results from a national carbon sequestration assessment, the LandCarbon project (http://www.usgs.gov/climate_landuse/land_carbon/), conducted by the US Geological Survey (USGS) in response to section 712 of the U.S. Energy Independence and Security Act of 2007 (US Government Printing Office, 2007). The assessment examined carbon storage, carbon fluxes and other GHG fluxes (methane and nitrous oxide) in all major terrestrial and aquatic ecosystems in two time periods: baseline (2001–2005) and future (from baseline to 2050). Three regional assessments (Zhu et al., 2011; Zhu & Reed, 2012, 2014) were generated using three biogeochemical models including Century version 4.0 (Metherell et al., 1993; Parton et al., 1994), the Erosion Deposition Carbon Model (Liu et al., 2003), and the Land Greenhouse-Gas Accounting Tool biogeochemical models implemented in the General Ensemble Biogeochemical Modeling System (Liu, 2009; Liu et al., 2012). Our contribution to this project had three objectives: (i) Because initial results from the LandCarbon project simulated the response of actual vegetation, our simulation results could provide comparable estimates of the response of potential natural vegetation to climate change and its effect on ecosystem carbon balance; (ii) Because our model includes a dynamic fire model, it could help estimate the amount of carbon lost to wildfires and the effect of fire suppression on carbon stocks; (iii) by simulating even a simplified representation of land use with the same model, we could evaluate the magnitude of the effect of land use scenarios associated with each emission scenario on the US carbon budget. The MC2 model, the C++ version of the original MC1 DGVM (Bachelet et al., 2001a,b), was run for 9 climate futures (3 GCMs-CSIRO, CGCM3 and MIROC and 3 emission scenarios – A1B, A2 and B1) following the LandCarbon methodology described by Zhu et al. (2010) to provide estimates of plant and soil carbon pools, carbon fluxes such as primary production, decomposition, fire emissions for potential natural vegetation, as well as possible effects of fire suppression and a simple land use scheme on the US carbon balance. The C++ version of the DGVM was designed to be a faster more efficient version of the MC1 model (Creutzburg et al., 2015), better suited to run on a supercomputer (NASA NEX), which was needed to run on such large domain (12 million grid cells) at fine scale. Results provide spatial and temporal estimates of the carbon sequestration capacity and of the magnitude of carbon fluxes emitted by US potential natural ecosystems under different future climates and disturbance regimes and include a simple glimpse at how human land use might affect C sequestration capacity under each of those futures.

Materials and methods

The DGVM MC (version) 2, the C++ version of MC (version) 1 (Bachelet et al., 2000), simulates vegetation type, plant growth and associated biogeochemical cycles, as well as their response to natural wildfires. MC1 was used extensively at a variety of spatial scales and domains (e.g., Daly et al., 2000; Bachelet et al., 2003; Lenihan et al., 2008; Gonzalez et al., 2010; Rogers et al., 2011; King et al., 2013; Creutzburg et al., 2015; Yospin et al., 2015) and has been included in several climate change assessments (Intergovernmental Panel on Climate Change, 2007; Karl et al., 2009; Melillo et al., 2014). MC2 consists of a more efficient version of MC1, better suited to run on a supercomputer to simulate large domains at fine scales. The model structure and algorithms of MC1 were conserved. The model simulates the dynamics of life-forms rather than species, including evergreen and deciduous needleleaf and broadleaf woody life-forms (trees and shrubs), C3 and C4 herbaceous life-forms (grasses, forbs and sedges) as they respond to both climate change and increasing atmospheric CO₂. It simulates competition between woody and herbaceous life-forms for light and available soil water (Bachelet et al., 2001a,b). The model simulates net primary production and monthly values of live and dead plant compartment pools from which it derives the size of the various fuel types. Using litter characteristics, it simulates decomposition, soil respiration and nutrient release over time. Nitrogen inputs are calculated as in the original CENTURY model (Parton et al., 1994) to include release from organic matter turnover, wet/dry deposition and biological nitrogen fixation. The model also estimates the height, trunk diameter and bark thickness of average-sized trees from live aboveground woody biomass using life-form-specific allometric relationships. Based on this information and climate indices, it calculates fire occurrence and effects including mortality (without consumption) and fire emissions (Lenihan et al., 2008). It modifies carbon stocks after a fire accordingly, thereby allowing for the simulation of a postfire ‘succession’. Code updates have been documented with versions maintained on a stable server (https://sites.google.com/site/mc1dgvmusers/home/mc2). Further details on the model can be found in the Appendix S1.

Model inputs include soil characteristics (mineral depth, texture, bulk density), annual atmospheric CO₂ concentrations and monthly climate (minimum and maximum temperature, precipitation and vapor pressure). MC2, like MC1, is run in four distinct successive phases: first, an equilibrium phase where potential vegetation cover is generated using average historical climate (1895–1924) and fire return intervals (FRI) are fixed. Carbon pools are initialized for each potential vegetation type until the resistant soil carbon reaches stability (less than 1% change per year), which may last up to 3000 years. During the second or spinup phase, the model uses detrended historical climate data (adjusted to the 1895–1924 mean to avoid an abrupt transition with the historical phase) iteratively for 600 years to readjust the carbon pools and vegetation types in response to dynamic fire simulations. This phase ends when the carbon exchange stabilizes and the net biological production (net ecosystem production – fire losses) nears zero. The third or historical transient phase is run with transient historical climate data from Daly et al. (2008) starting in 1895. The final transient phase starts in 2010 and ends in 2100 using climate futures.

To be consistent with the methods of the LandCarbon project (Zhu et al., 2010), we used CMIP3 climate futures (Fig. 1) from three general circulation models (GCMs) that spanned much of the range in temperature increases associated with the IPCC SRES greenhouse gas emission scenarios (Nakicenovic et al., 2000): CSIRO Mk3 (Gordon et al., 2010), CGCM3 (Boer, 1995) and MIROC 3.2 medres (Hasumi & Emori, 2004) (henceforth CSIRO, CGCM3 and MIROC). Projections for the three emission scenarios A2, A1B and B1 were downscaled using the delta or anomaly method of Fowler et al. (2007). For each climate variable and each future month, anomalies (differences for temperature, ratios for precipitation and vapor pressure) between future and mean monthly historical (1971–2000) GCM-simulated values were calculated for each GCM grid cell over the conterminous USA. Anomalies were downscaled to the 30 arc-second grid using a Gaussian interpolation filter and applied to the monthly historical PRISM baseline (1971–2000) available at the same spatial scale. CO₂ concentrations matching each of the emission scenarios were used to run the vegetation model.

Fire suppression was imposed for some model runs (with and without land use) using thresholds of fire line intensity, energy release component and rate of spread (900/60/100 = defaults), above which fires could occur (Rogers et al., 2011). Below these thresholds, fires were not allowed to burn assuming complete effective suppression. In the model, suppression starts in 1950 since in reality it only became truly effective following the advent of vehicles particularly air tankers, equipment and roads several years after the end of World War II (Pyne, 1982; Veblen et al., 2003) causing an order of magnitude drop in fires (Dombeck, 2001).

To maintain consistency with the work of our collaborators in the LandCarbon project (Zhu et al., 2010), we used three time series of land use each starting in 1992 and diverging in 2006 with each emission scenario (Fig. 2). Based on Intergovernmental Panel on Climate Change (2000) definitions, ‘land use refers to the total of arrangements, activities and inputs undertaken in a certain land cover type (a set of human actions). The term land use is also used in the sense of the social and economic purposes for which land is managed (e.g., grazing, timber extraction and conservation)’. However, in this study, we did not refer to fire suppression as land use and only used the term land use when we imposed a specific land cover type. For the land use runs, the model was first run with fire suppression alone starting in 1950 and together with land use starting in 1992 (Sleeter et al., 2012; Sohl et al., 2013, 2014) such that agricultural lands were harvested annually and managed forests were harvested periodically. More specifically with regard to forest lands, when land use was imposed in 1992, 30% of all forests was transformed into agricultural land harvested annually, 3% of all forests was lost to mining and/or urban development, and 2% of all forests became harvested at regular intervals driven by the imposed rotation information. The model assumed natural regeneration after harvest, and closed-canopy forest would be reached only if climate and soil conditions would allow it to happen. Because this was the first time the model was run with land use, we kept management rules very simple (Table 1). Prescribed categories of land use are described in detail by Sleeter et al. (2012). We aggregated the 3 categories of managed forests (mechanically disturbed national, mechanically disturbed other public and mechanically disturbed private) because we did not have access to past, current or future land ownership information. We also aggregated the developed and mining categories and assumed all growth was eliminated. We did not represent grazing lands (pastures and rangelands) but aggregated hay pastures with agriculture. The categories of natural vegetation were simulated as potential vegetation types determined by the biogeography rules.

The model was run on the Pleiades NASA supercomputer. At 30 arc-sec spatial grain over the conterminous USA, the model was run on about 12 million grid cells using 1.6 Terabytes of input data. Output files include ~30 model variables, which corresponds to about 2.2 Terabytes and 6000 CPU hours for each of the 3 sets of runs, from 1895 to 2100, without fire suppression, with fire suppression and with land use. To reduce the size of the climate future input files (1.3 Terabytes × 3 climate models × 3 emission scenarios) and the I/O bottleneck that slowed the run time, downscaling was performed ‘on the fly’.

We subsampled the 30 arc-second (~800 m) dataset to a resolution of 5 arc minutes using a stride of 10 (every 10th row and every 10th column). This subsampled dataset allowed multiple runs for calibration purposes and ensured reasonable time for processing results and generating model runs to further test and evaluate trends in C cycle dynamics. When results of the full-resolution (30 arc sec) simulations were compared with the subsampled results, regional summaries were virtually identical. Consequently, they were used to generate summary statistics.

As in the past for MC1 calibration (e.g., Bachelet et al., 2001a,b), Küchler's potential vegetation map of the United States (Küchler, 1975) was used to test the MC2 simulation of vegetation cover during the historical period. After calibration to best match it, 47% of the grid cells had a vegetation type that matched Küchler's description and 33% of the grid cells were qualified as minimal mismatch (in either density or the type of vegetation, e.g., a woodland instead of a forest, or a C4 grassland instead of a C3 grassland). Only 9% of the grid cells were qualified as a total mismatch with Küchler's types and the other 11% were qualified as mediocre match (temperate shrubland vs. temperate forest for example). Simulated fire rotation period was tested against Leenhouts’ (1998) potential FRI estimates, which were designed to specifically match Küchler's potential vegetation types. With the best calibration efforts, the normalized root-mean-square error remains at 0.35. We also tested model against simulated aboveground biomass results with National Biomass and Carbon Dataset data (Kellndorfer et al., 2012), and with the best calibration efforts, the normalized root-mean-square error remains at 0.11.

As a validation exercise after calibration, simulated historical results were compared with a variety of published values to evaluate general model skill. Model results generally agree with the range of NPP, NEP and NBP published values (Fig. S1). Because our model includes the effect of fire on C dynamics, it is difficult to compare individual years in terms of carbon fluxes for which there are large variations even among published values. We generally use averages over several years to dampen the possible fluctuations in climate variability and disturbance (fire emissions) and report variation over time by reporting the standard deviation around the mean for C fluxes. We also compared our model results to matching records from EPA (2013) for forest area, forest carbon stocks and fire emissions over the conterminous USA (data not shown here). As expected, the simulated emissions of naturally occurring wildfires exceeded observations; however, uncertainty also looms large on the national estimates of fire emissions used as a benchmark. Our simple scheme to include land use lowers our estimates of forest area, so they compare better with observations, but the model underestimates forest C stocks. The cause of this discrepancy is the overly simplified representation of harvests in the model, removing more than usually occurs in timberlands. We are working to improve our treatment of land use in the next phase of this project to more realistically represent regional variations in harvests.

Results

For potential vegetation, U.S. carbon stocks, soil and litter as well as live vegetation pools, are projected to increase almost linearly throughout the 21st century, following a similar pattern to what was observed during the 20th century, particularly after fire suppression is imposed in 1950 (Fig. 3). The trend is similar after 2050 without fire suppression. When land use is artificially imposed in 1992, soil carbon levels drop because harvest of crops and forests reduce the inputs to litter and soil carbon reservoirs while decomposition increases under warmer conditions in areas where moisture is not limiting. After ~70 years, the soil carbon trace stabilizes except under the A2 scenario with losses continuing to be simulated as managed lands area continue to increase. Vegetation carbon increases without land use over the 21st century especially with fire suppression. Without suppression, fire losses early in the 21st century delay the increase by about by 2–3 decades. Only under the mildest scenario (B1) does vegetation continue to increase its carbon sink capacity when land use is imposed. Under the warm scenario (A2), both soil and vegetation carbon stocks decrease below historical levels (Fig. 3). From 2041 through 2060, the average ecosystem carbon stored (live biomass, soil and dead biomass) varies between 155 and 162 Pg C across the three emission scenarios, and between 122 and 126 Pg C with imposed land use. Carbon losses due to fire for the same period vary between 131–173 Tg and 314–409 Tg yr⁻¹ across climate scenarios with and without land use, respectively.

Spatially, the effect of fire suppression clearly enhances carbon sequestration in Western states and on the eastern edge of the Great Plains where woody life-form expansion increases both above- and belowground carbon pools (Fig. 4). Land use imposes significant losses to soil carbon throughout the more urbanized and agricultural east coast and Midwest regions as well as along the West Coast. The Willamette Valley, the Great Lakes area and the southeast are 3 regions where carbon losses are simulated across all scenarios. Drought conditions and fires contribute to carbon losses in these areas.

The number of fire occurrences in Western states is much larger than in the rest of the country but is not always cause for vegetation instability (Fig. 5). It allows the maintenance of both western shrublands and central grasslands by reducing woody life-form establishment and tree growth (King et al., 2013). Natural variability in the climate records causes some instability at ecotones between eastern forest types visible as linear features. Southwest deserts oscillate between semi-arid grasslands and desert vegetation in the model as rainy periods allow grass expansion (often during spring), while extended drought conditions only allow the survival of scarce deep-rooted shrubs. These shifts are caused entirely by internal biomass thresholds and reflect rainfall variability and the rapid response of annual vegetation rather than discernable changes in land cover type. Climate thresholds in MC2 take into account 15 years of records, so these vegetation shifts do not correspond to year-to-year variability but to wet vs. dry decades. Wet decades allow higher levels of seasonal grass biomass that biogeography rules interpret as a shift to grass-dominated systems defined in this case as ‘grasslands’ even though grass biomass remains low. The Arizona monsoon dynamics is key to these shifts.

When land use and fire suppression are imposed together, the instability of eastern ecotones disappears as most of the area is now either managed or transformed into agricultural land (Fig. 6). Mountainous regions continue to be prone to wildfires, especially along the Sierra Nevada, but the number of small fires is reduced dramatically (e.g., across the Great Plains). Instability remains high in the southeast where high temperature and drought conditions stress vegetation growth and increase fire risk.

Because forest ecosystem services are so important to the US economy (e.g., timber) and to local populations (e.g., water quality, wildlife habitat, microclimate decoupled from regional warming), we summarized our projections for forested areas in the conterminous USA (Fig. 7). While under most (all except one) scenarios forest area increases, forest carbon density decreases except under CGCM3 and CSIRO A1B without land use. With land use, forest carbon density decreases because timber extraction removes carbon for the forest areas and reduces organic matter inputs to the soil. Fire suppression causes the greatest amount of forest expansion overall, this is particularly true in areas where grasslands are usually maintained by low-intensity frequent fires such as at the ecotone between grasslands and eastern deciduous forests but also in the intermountain West where woody expansion is currently being observed. With the Canadian model projections (CGCM3), carbon stocks increase with suppression, but the greatest magnitude of increase occurs under CSIRO A1B. With the Japanese (MIROC) climate projections, carbon stocks only increase under fire suppression, regardless of the emission scenario, because wildfires are reducing stocks when no land use is simulated while land use prevents woody encroachment.

Discussion

There are been few modeling studies simulating the response of natural ecosystems to the combined interactions of climate change, rising atmospheric CO₂, natural disturbances such as fire, and land use change (e.g., Arora & Boer, 2005; Bachelet et al., 2008; Liu et al., 2011; Prentice et al., 2011). This is partially due to the uncertainty inherent to the input data, the scarcity of available calibration datasets, but also our limited understanding of these interactions.

Because MC2 includes a dynamic fire model that is very responsive to fuel load and fuel moisture, it is very sensitive to differences in both seasonality and magnitude of rainfall projected by the climate models. When production increases during wet periods, fuels build up in dry areas where fuels are usually limiting, increasing fire risk when drier conditions follow. Extended drought periods reduce production and thus fuels in the long term but in the short term increase the amount of dead fuels as well as dry all the fuels increasing the probability that fire occurrence thresholds are exceeded. The uncertainty associated with precipitation projections has been discussed extensively in the literature (e.g., Meehl et al., 2007), but it is important to reiterate here how much precipitation affects simulated impacts particularly fire occurrence and effects and thus how much its variability contributes to the uncertainty of our results. This influence is only partially lessened with prescribed fire suppression as fire occurrence thresholds can be exceeded during drought conditions and cause simulated wildfires to escape.

On the same subject of water, model results are greatly affected by the quality of the soil characteristics used to estimate soil water availability at various depths. Precipitation can be intercepted by foliage, run off on the soil surface, penetrate and get distributed through the soil profile before being considered as deep soil water recharge or stormflow. Soil depth and texture affect the amount of water available for uptake in each of the soil layers, and their accuracy affects MC2 skill to represent land cover, primary production and fire effects. In the model, while frequent fires reduce woody encroachment, drought conditions favor woody encroachment by allowing deeply rooted woody plants to outcompete shallow rooted herbaceous where surface soil layers are dry and deep soil water is available. We found several areas, especially in complex terrain, where NATSGO soil depths were not representative and caused inaccurate simulation of ecosystem processes (e.g., Conklin, 2009; Peterman et al., 2014).

Woody encroachment is also affected by the effect of increasing CO₂ concentrations on plant physiology. The effect of CO₂ on mature forests and on the various tree species that were not studied in FACE experiments, in areas where nutrient limitation may reduce the CO₂ fertilization effect, remains a gap in our knowledge (e.g., Asshoff et al., 2006). As a consequence, we used a moderate effect of CO₂ (from the original CENTURY code) that provided some mitigation effect in the form of enhanced water use efficiency to woody life-form under drought conditions. To estimate its effect on shrub expansion, we ran the model without CO₂ effect (results not shown here) and without fire (results not shown here) over the historical period and found that while the size of the carbon pools was affected, the CO₂ effect did not affect woody encroachment across western landscapes as much as fire did. However, it may have caused the underestimation of tree mortality through drought stress causing carbon starvation or embolism as well as tree vulnerability to pest outbreaks (McDowell et al., 2013), neither of which is simulated by MC2.

The MC2 fire model assumes that fire occurrence is not limited by ignition. Such assumption stems from the original intent for the model (Lenihan et al., 1998) to simulate fire only during extreme weather and fuel conditions (i.e., severe drought and low coarse fuel moisture). Large and severe fires occurring under these conditions account for a very large fraction of the annual area burned historically (Strauss et al., 1989). High-severity fires, such as the 1988 Yellowstone fire and the 2002 Hayman fire, happened in response to extreme climate signals, which will become more dominant in a warmer future (Running, 2006; Seager et al., 2007). However, under future conditions, it is possible that natural ignition sources might be limiting, but there is little information on the occurrence of lightning strikes under future climate change scenarios (Price & Rind, 1994). On the other hand, it is probable that with increases in population, human ignition sources will increase (Bowman et al., 2011).

We also have assumed that fire suppression was 100% effective below the set thresholds but we know that in areas of high population density, fire risk increases, so under land use our assumption may need to be revised particularly in high-density areas such as California or in other wildland–urban interface areas (Syphard et al., 2009). As we improve our treatment of management practices in future model runs with land use, we are also planning to simulate ignitions in the fire model taking into account human infrastructures such as roads and powerlines that could affect the location and timing of fire occurrence.

This brings up the validity of our land use scheme as: since we did not use any land management before 1992, the legacy of land use is not taken into account when we report changes in C stocks. To maintain consistency with the work of our collaborators in the LandCarbon project, we had to use the exact same methods and thus use the land use data provided by the USGS, which only started in 1992. During the project, the MC2 code was for the first time modified to allow the reading of a new input file that corresponded to the gridded land use data prescribing the timing of harvests but not their magnitude. We think we generally overestimated harvests, which means we simulated the rapid loss of live carbon pools without compensatory increase in litter pools. Whether complete harvest removal affected the carbon sequestration potential and the soil buildup to a greater extent than the decomposition fluxes would have over the longer term remains to be tested. The quantity of slash remaining after a timber harvest certainly affects the size of the fuel classes and the fire risk. Regional and land ownership differences will require new sets of data used as inputs before we can be confident we are simulating reasonable management practices and can evaluate their effect on carbon fluxes. We know we used a simplistic treatment of harvest, but we are currently trying to improve this aspect of the model in the second phase of the project. For this first phase, we can only report the immediate impact of simulated complete harvests and associated carbon losses.

Another limitation in our approach is our lack of age structure in the simulation of forest recovery after disturbance, although the model simulates a pseudo-succession, especially after a stand-replacing fire, by simulating the transition between a low-biomass grass-dominated ecosystem with recovering small woody plants towards a full canopy restored forest. To take into account age distribution and the different growth rates between young and older plants would require simulating subgrid cell heterogeneity and population dynamics, requiring an entirely new model structure. While we are well aware of the importance of forest age structure to accurately estimate carbon stocks, it is beyond the scope of this project to address this issue properly and furthermore, as Williams et al. (2012) noted, ‘the paucity of observed net ecosystem productivity and biomass chronosequences limits our ability to evaluate modeled responses’ as well as our ability to accurately parameterize et calibrate the models.

Our ultimate goal was to improve our understanding of how carbon dynamics were affected by combinations of changes in climate, fire regimes and vegetation cover across the conterminous USA characterized by a variety of topography and climate, soil types, potential vegetation types and human land uses. We simulated the effects of 9 different climate futures on carbon pools and fluxes, fire occurrence and effects, as well as vegetation types and concluded that despite large effects of climate on natural processes, human actions such as fire suppression and land use have the greatest impacts of the magnitude of the carbon pools and the direction of carbon fluxes.

Much of the carbon gained in the first half of the century is due to the simulated increase in woody cover in the central and western USA. Shrub expansion has been widely documented in the Great Plains (e.g., Heisler et al., 2003), but land use is limiting its extent. Forests and woodlands are simulated in the interior West replacing shrublands and grasslands but because the model does not simulate either browsing by ungulates or invasive species affecting the fire regime, it overestimates the woodification of the Western states. We also assumed and probably overestimated the effectiveness of fire suppression in the last 20 years in Western states. To increase carbon capture and reduce increasing fire risks, managers will be faced with the dilemma of choosing best practices that optimize both wood production but reduced fuel loads. Another challenge will be to decide whether resources should be spent to try and eradicate invasives or to develop management strategies that include them future landscapes. In a future with increased fire risk, ecosystems that optimize carbon sequestration belowground, out of reach from fires, should be protected. Woody species adapted to warm and dry conditions with physiological traits of fire resistance should be considered for future plantations and restoration of forest ecosystems affected by insect outbreaks and fires.