An atmosphere-ocean-vegetation coupled model is used to quantify the biogeophysical feedback that emerges as vegetation adjusts dynamically to a quadrupling of atmospheric CO2. This feedback amplifies global warming by 13%. About half of it is due to climatically induced expansion of boreal forest into tundra, reinforced by reductions in snow and sea ice cover. The other half represents a global climatic effect of increased vegetative cover (an indirect consequence of plant physiological responses to CO2) in the semi-arid subtropics. Enhanced absorption of shortwave radiation in these regions produces a net surface warming, which the atmosphere communicates poleward. The greatest vegetation-induced warming is co-located with large, vulnerable carbon stores in the north. These lose carbon, so that in the long term, the biospheric response to CO2 and climate change becomes dominated by positive feedbacks that overwhelm the effect of CO2 fertilization on terrestrial carbon stocks.
 The present study investigates what may happen to land carbon storage in the longer term, when (a) vegetation patterns are allowed to equilibrate interactively with a high-CO2 climate, and (b) carbon stores are subsequently allowed to stabilize in the changed climate. The results indicate an important role for biogeophysical feedback mechanisms, involving climate- and CO2-induced changes in vegetation patterns, in determining both climate and the terrestrial carbon balance.
2. Experimental Setting
 The MIROC atmospheric general circulation model (AGCM) [Hasumi and Emori, 2004], as used in the IPCC Fourth Assessment Report [Meehl et al., 2007], was coupled synchronously to the Lund-Potsdam-Jena (LPJ) dynamic global vegetation model [Sitch et al., 2003]. Ten plant functional types (PFTs) are modelled. Photosynthesis, plant respiration, carbon allocation, plant establishment, growth, tissue turnover, mortality, competition among PFTs, disturbance (fire) regime and soil carbon dynamics are calculated. Monthly mean temperature, precipitation and fractional cloud cover required as input to LPJ are derived from MIROC as running means of the most recent 20 years of simulation. In each model year, the combination and proportion of PFTs in every grid cell is used to re-assign the grid cell in the land-surface classification system. This assignment updates the parameter values (vegetation height, root depth, leaf albedo and photosynthetic properties) in MATSIRO [Takata et al., 2003], the land-surface component of MIROC. Sea surface temperature and sea ice extent are predicted by a slab-ocean model with prescribed seasonal ocean heat transport. All components of the coupled model run at 2.8° × 2.8° resolution, providing reasonable representations of large-scale vegetation and climate patterns. Simulated global net primary productivity (NPP) of 64.5 PgC yr−1 is realistic and in the range of previous studies [Cramer et al., 1999], as are the simulated pre-industrial carbon stocks of 1073 PgC in vegetation and 1466 PgC in soils.
2.2. Experimental Set-up
 Coupled model experiments (Table 1) were designed to quantify the effects of vegetation dynamics, with and without the plant-physiological effects of high (quadrupled) CO2 concentration, on climate (after sufficient time for equilibration between vegetation and climate in the coupled model), and carbon storage (after a longer period of offline simulation with the LPJ model, to allow equilibration of carbon stores in the changed climate). CO2 quadrupling is comparable with concentrations approached by the end of the 21st century in “business as usual” scenarios with coupled climate-carbon cycle models [Cox et al., 2000; Friedlingstein et al., 2006]. We performed a pre-industrial control simulation with atmospheric CO2 concentration at 285 ppm (A1V1), a 4 × CO2 simulation with CO2 at 1141 ppm (A4V4), and a “climate change only” simulation in which the vegetation was permitted to adjust to climate change but with no physiological effects of elevated CO2, i.e., the vegetation model “sees” pre-industrial CO2 (A4V1). The coupled model was run for long enough to allow the simulated global vegetation patterns to stabilize. A1V1 and A4V4 were started from an unvegetated state, and required 150 years to achieve a stable vegetation pattern. A4V1 started from the final state of A4V4, and required a further 90 years to achieve a new stable vegetation pattern. In an additional 4 × CO2 climate simulation, the vegetation was fixed at the pre-industrial equilibrium state, as simulated in the control run.
[CO2]clim, CO2 concentration for the atmospheric model; [CO2]veg, CO2 concentration for the biosphere model; T, global mean temperature; ΔT, change in global mean temperature relative to A1V1; Cveg, global carbon storage in vegetation; Csoil, global carbon storage in soil; Ctot, global carbon storage in vegetation and soil (total carbon storage); ΔCtot, change in total carbon storage relative to A1V1. All carbon storages are values obtained at the end of a long (850–1000 years) offline simulation with LPJ, with climate prescribed from the end of the coupled model experiment.
 Because carbon stocks equilibrate more slowly than vegetation patterns [Sitch et al., 2008], steady-state vegetation and soil carbon storages were calculated for each experiment by continuing to run the LPJ model, offline, for 850–1000 years in the changed climate. In the fixed-vegetation case, two such offline runs were performed: one with 4 × CO2 maintained to allow physiological effects on carbon storage (A4V4F), and one with pre-industrial CO2 to eliminate these effects (A4V1F).
3.1. Global Climate and Vegetation
 The high-CO2 coupled model experiment (A4V4) produced a global mean temperature increase of 9.8 K relative to the control run, and a substantial change in vegetation distribution (Table 1 and Figure 1). Without vegetation dynamics (A4V4F), the warming was 8.7 K. In other words, vegetation dynamics amplified the global warming due to CO2 by 13%. The main cause of this amplification was reduced surface albedo (see Figure S1 of the auxiliary material). The vegetation simulated in A4V4 absorbs a greater proportion of shortwave radiation, because of increased forest area in high latitudes (boreal forest expanding into tundra [e.g., Levis et al., 1999]) and the encroachment of vegetation on sparsely vegetated areas in the subtropics (Figure 1).
 The simulated additional warming due to vegetation dynamics is concentrated in the northern mid- to high latitudes, reaching 5 K in the Arctic, and in semi-arid regions of the subtropics with a warming typically of 2–3 K (Figure 2a). This spatial pattern of additional warming closely follows the response of surface albedo changes (Figure S1) and consequent changes in the surface energy balance (Figure S2). The simulated surface albedo changes are greatest in the northern high latitudes because of the mutually reinforcing effects of increased snow masking by vegetation, reduced snow cover, and reduced Arctic sea ice cover.
3.2. Plant Physiological Response
 In semi-arid subtropical regions, the simulated changes in surface albedo are primarily forced by increases in vegetation cover promoted by high CO2 concentration, through the combination of CO2 fertilization (increased photosynthesis due to higher substrate concentration and reduced competition with O2 at the Rubisco reaction sites) and partial stomatal closure, which tends to conserve the ratio of leaf-internal to ambient CO2 concentration in the face of a diminishing biochemical return on internal CO2 concentration [Sellers et al., 1996]. Under high CO2, in conditions where water supply is limiting, a strong CO2 effect on plant growth is expected [Long et al., 2006] as increased photosynthesis and reduced water loss act in combination to increase the water use efficiency of vegetation. In seasonally dry climates the effect is to increase the growing season even if annual precipitation is unchanged, and to increase annual average NPP, leaf area index, and the fractional absorption of solar radiation, relative to dry and unvegetated soil surfaces.
 The importance of these plant-physiological responses to CO2 in producing the additional warming due to vegetation dynamics is illustrated by comparison of Figures 2a and 2b. The “climate change only” effect of vegetation dynamics (A4V1, Figure 2b) on the climate in the subtropics is slight. Many regions show a cooling, associated with reduced vegetation cover. The vegetation pattern simulated in A4V1 (Figure S3) differs notably from that simulated in A4V4, with extensive replacement of tropical forests by savannas; the physiological effects of high CO2 counteract tropical forest dieback in A4V4. In global average, the regional warming and cooling effects almost cancel, so that the global mean warming in A4V1 is barely (<0.1 K) greater than that in A4V4F (Table 1). In the northern high latitudes, however, the “climate change only” effect is substantial (1.2 K). Therefore, the simulated additional warming due to vegetation dynamics is robust for high northern latitudes even if the CO2 effect on vegetation cover is neglected, or if it is less than simulated by LPJ-MIROC.
 On the other hand, the full simulated effect of vegetation dynamics in high northern latitudes in A4V4 is about twice as large (2.3 K) as the “climate change only” effect, due to the remote climatic effect of CO2-induced vegetation changes on surface albedo in the subtropics. This is because change in surface albedo increase shortwave absorption and thereby increases the sensible heat flux, which leads to warming (Figure S2). Changes in longwave radiation and latent heat flux show smaller or opposite signals.
 The greatest climatic effect of plant-physiological responses to CO2 in our simulations is the effect mediated by vegetation structure and dynamics [Bala et al., 2006]. CO2-induced stomatal closure can also affect surface air temperature more directly, by reducing the latent heat flux [Sellers et al., 1996; Boucher et al., 2009]. We quantified this direct influence of stomatal conductance in an additional sensitivity experiment, in which the vegetation pattern of A4V4 was used but the physiological effect of CO2 was turned off (Figure S4). This experiment showed that the stomatal conductance effect contributes to the surface warming in high latitudes, but not or only slightly (<1 K) in the tropics and mid-latitudes.
3.3. Change in Carbon Storage
 We consider now the effects of the simulated climate changes on terrestrial carbon storage, after the carbon stores have equilibrated with the changed climate. The effects of climate change and CO2 change, with fixed vegetation, are opposite: climate change causes a loss of carbon from the land (−757 PgC in A4V1F) while climate and CO2 change together cause a net gain of carbon to the land (+218 in A4V4F), i.e., the CO2 fertilization effect dominates. This result is qualitatively in agreement with results obtained by 2100 in transient simulations [Cox et al., 2000; Friedlingstein et al., 2006; Sitch et al., 2008]. Some transient simulations of the 21st century have allowed for dynamic vegetation and resulting biogeophysical feedback, but the simulated vegetation patterns are not in equilibrium in 2100 [Cramer et al., 2001; Sitch et al., 2008]. Allowing the vegetation to fully adjust to new atmospheric conditions has a substantial impact on the simulated long-term carbon storage. Land carbon storage in A4V1 is reduced by a further 5% relative to A4V1F. More remarkably, land carbon storage in A4V4 is reduced by 12% relative to A4V4F. This change is large enough to overwhelm the CO2 fertilization effect, and thus to alter the sign of the net CO2 response from an increase (+218 PgC in A4V4F) to a decrease (−111 PgC in A4V4) in land carbon storage.
 Most of the reduction in land carbon storage in A4V4 relative to A4V4F occurs simply because the additional warming due to vegetation responses further stimulates the decomposition of soil organic matter in high northern latitudes, where low temperatures currently inhibit decomposition and allow the accumulation of large soil carbon stores, which are potentially vulnerable to warming. About three-quarters of the reduction comes from soil carbon, and the geographic patterns of carbon reduction for total carbon and soil carbon are similar (Figure S5). The additional carbon thus released to the atmosphere, due to vegetation dynamics and biogeophysical feedback, is equivalent to 154 ppm CO2, of which about half would be expected to remain in the atmosphere on a time scale of centuries [House et al., 2002]. If CO2 concentration had been allowed to “float” [e.g., Cox et al., 2000; Friedlingstein et al., 2006], the additional CO2 thus released would have further amplified the simulated warming.
 Ecosystem-level responses to a quadrupling of CO2 as modeled here cannot be directly verified, as ecosystem-level manipulations have not considered such high CO2 concentrations. However, Free Air Carbon Dioxide Enrichment (FACE) experiments in forests, where CO2 concentrations have been elevated by around 200 ppm, have yielded NPP enhancements in agreement with predictions of the LPJ model [Hickler et al., 2008]. In addition, analysis of last glacial maximum vegetation patterns has shown that the extent of reduction of tropical forests at that time cannot be accounted for without the effect of low (<200 ppm) ice-age CO2 concentration on plant water use efficiency and competition between C4 grasses and trees [Harrison and Prentice, 2003]. These findings are consistent with the expectation that high CO2 concentrations would shift low-latitude vegetation boundaries through physiological mechanisms, even in the absence of climate change.