We study the response of the land biosphere to climate change by coupling a climate general circulation model to a global carbon cycle model. This coupled model was forced by observed CO2 emissions for the 1860–1990 period and by the IPCC SRES-A2 emission scenario for the 1991–2100 period. During the historical period, our simulated Net Primary Production (NPP) and net land uptake (NEP) are comparable to the observations in term of trend and variability. By the end of the 21st century, we show that the global NEP is reduced by 56% due to the climate change. In the tropics, increasing temperature, through an increase of evapotranspiration, acts to reduce the soil water content, which leads to a 80% reduction of net land CO2 uptake. As a consequence, tropical carbon storage saturates by the end of the simulation, some regions becoming sources of CO2. On the contrary, in northern high latitudes, increasing temperature stimulates the land biosphere by lengthening the growing season by about 18 days by 2100 which in turn leads to a NEP increase of 11%. Overall, the negative climate impact in the tropics is much larger than the positive impact simulated in the extratropics, therefore, climate change reduce the global land carbon uptake. This constitutes a positive feedback in the climate-carbon cycle system.
If you can't find a tool you're looking for, please click the link at the top of the page to "Go to old article view". Alternatively, view our Knowledge Base articles for additional help. Your feedback is important to us, so please let us know if you have comments or ideas for improvement.
 As a result of human activities (fossil fuels combustion and land use change), the atmospheric concentration of carbon dioxide has increased by about 30% since 1860 (Figure 1). The induced positive radiative forcing tends to warm the surface. Indeed, the global temperature has risen up by 0.6°C over the same period. This increase of temperature is modulated by other greenhouse gases and aerosols changes, but it is most likely that if the atmospheric CO2 continue to increase, a larger climate change would occur in the future [Intergovernmental Panel on Climate Change (IPCC), 2001].
 Over the last decade, the annual increase of atmospheric carbon reached 2.9 ± 0.1 Gt C yr−1 [IPCC, 2001]. This rate of increase is less than half of the estimated anthropogenic CO2 emissions (fossil fuels = 6.3 ± 0.4 Gt C yr−1 and deforestation = 1.6 ± 0.8 Gt C yr−1 during the 1990–1998 period, [IPCC, 2001]), the other part being absorbed by the oceans and by the terrestrial biosphere. Land and ocean carbon cycles are, among other things, controlled by climate and atmospheric CO2. There is therefore an intimate coupling between climate change and the global carbon cycle and it is thus very important to characterize and quantify this coupling in the context of future climate prediction.
 The way both carbon cycle and climate system will interact in the future are however sources of a lot of uncertainties, namely, the climate sensitivity to greenhouse gases (GHG) and aerosols [Cubasch and Fisher-Bruns, 2000; Meehl et al., 2000], the global carbon cycle and its response to climate change [Cramer et al., 2001] and the economical uncertainty due to the GHG emissions scenarios [Nakicenovic et al., 2000]. Several modeling studies estimate the impact of increasing atmospheric CO2 and future climate change on the terrestrial carbon cycle. To do so, they usually force their terrestrial model with future climates computed by a Global Circulation Model (GCM) forced either by a prescribed transient increase of atmospheric CO2 [Cao and Woodward, 1998; Cramer et al., 2001; Friedlingstein et al., 2001] or by a doubling of the CO2 concentration [Betts et al., 1997; King et al., 1997; Melillo et al., 1993; Neilson and Drapek, 1998]. The main outcome of these studies is that the terrestrial biosphere is strongly affected by the climate change. Therefore, future atmospheric CO2 will be driven not only by the emission scenario, but also by the efficiency of land and ocean to absorb CO2 under the future climate. Hence, a consistent estimate of future atmospheric CO2 and future climate requires a fully coupled climate-carbon model. In a recent study using a GCM coupled to carbon cycle models, Cox et al.  found that the global warming will be amplified because of the positive feedback between terrestrial carbon cycle and the climate change. An important point highlighted by Cox et al.  is that the land biosphere will become a large source of carbon to the atmosphere after 2050 induced by the dieback of Amazonian rain forests and by a global increase of soil carbon respiration.
 Here, we use the IPSL fully coupled climate-carbon cycle simulation [Dufresne et al., 2002] to analyze in details the impact of future climate change on the terrestrial carbon cycle. We first describe the models used and the simulations performed. Then, we evaluate the model performance against the historical period. Finally, we quantify the response of the terrestrial biosphere to rising CO2 and changing climate over the 21st century. To do so, we perform additional simulations in order to identify which climate variables control the simulated response.
2. Models and Coupling Description
2.1. Description of the Climate and Ocean Carbon Models
 The climate model used is a 3-dimensional coupled Ocean-Atmosphere General Circulation Model (OAGCM), composed of an ocean circulation model (OPA-ICE [Delecluse et al., 1993]) and a atmospheric model (LMD-5.3 [LeTreut and Li, 1991]). A coupler (OASIS [Terray et al., 1995]) is used to ensure spatial interpolation and time synchronization when exchanging variables between the atmosphere and the ocean.
 The oceanic carbon cycle model, HAMOCC3 [Maier-Reimer, 1993] computes ocean biological carbon fluxes based on dissolved phosphates, and links marine export production to phosphate utilization. It has been adapted to OPA-ICE OGCM [Aumont, 1998] and is forced by monthly mean fields of ocean circulation, temperature, salinity, wind, sea ice and fresh water fluxes at surface.
2.2. Description of the Terrestrial Biosphere Model
 The terrestrial carbon cycle model, SLAVE [Friedlingstein et al., 1995], runs at the same horizontal spatial resolution than the Atmospheric GCM. It accounts for nine natural ecosystems and croplands [Friedlingstein et al., 1995]. In SLAVE, terrestrial carbon cycling is driven by the GCM monthly fields of surface temperature, precipitation, solar radiation and by the annual atmospheric CO2 concentration which directly influences Net Primary Productivity (NPP). The model computes the water budget, net primary productivity, allocation, phenology, biomass, litter and soil carbon budgets. Carbon assimilated through NPP is allocated to three phytomass pools: leaves, stems and roots. Litter and soil carbon pools are both divided into metabolic and structural components. NPP is computed by SLAVE as a function of the three climatic variables following an ε formulation of NPP as developed in the CASA model [Potter et al., 1993]. NPP is proportional to absorbed photosynthetically active radiation (APAR) and is scaled down by temperature and precipitation stress factors.
APAR is deduced from incoming solar radiation and from Leaf Area Index (LAI) [Sellers et al., 1996], LAI being diagnosed from the calculated leaf biomass. ε is the maximum light use efficiency. The two temperature stress factors, Tε1 and Tε2, depress NPP at very high and low temperature and when the temperature is above or below the optimum temperature [Potter et al., 1993]. Note that we assume here that the optimum temperature for photosynthesis does not change in the future as we use a fixed vegetation map. The precipitation stress term, Wε, is function of evapotranspiration and varies from 0 in very dry ecosystems to 1 in very wet ecosystems.
 The soil water content Qm is function of monthly precipitation (PPTm), of potential evapotranspiration (PETm), which is calculated following Thornthwaite formulation [Thornthwaite, 1948; Thornthwaite and Mather, 1957], and of runoff (RUNm). If monthly precipitation is larger than monthly evapotranspiration, the surface water content is increased by the difference between precipitation and evapotranspiration and the excess water runs off. On the contrary, if precipitation are insufficient to satisfy plant request, the plant uses the totality of precipitation and a fraction of soil water content.
 For each of the four litter and soil pools, heterotrophic respiration (RH) is calculated as the product of the pool carbon content by a pool specific decomposition rate (equation (2)). This rate depends on soil moisture and temperature (equation (3)) reflecting the fact that decomposition is favored by an increase of temperature or soil humidity.
Ci represents the amount of carbon in soil and in litter pools. Ki,max, the optimal decomposition rate depends on the pool : litter Kmax is equal to 10.4 yr−1 for metabolic fraction and 0.58 yr−1 for structural fraction and soil Kmax is equal to 0.006 yr−1. fT is a Q10 function : a temperature increase of 10°C induces an increase of a factor of Q10 of the decomposition rate, with Q10 being fixed to 2 for each pool. The water function, fH2O, first increases when soil water content increases. Above a threshold, it decreases to account for reduced decomposition under anaerobic conditions [Parton et al., 1993]. Finally, the net CO2 flux between the land biosphere and the atmosphere, called the Net Ecosystem Production (NEP), is simply computed as the difference between NPP and RH.
 SLAVE suffers from three major limitations in this study. First, it does not account for land use changes such as deforestation or forest regrowth. Croplands are fixed at their present-day distribution based on the map of Matthews . Second, it does not take into account of natural vegetation dynamic. As for croplands, we use the present-day vegetation distribution [Matthews, 1983] across the entire simulation. Third, anthropogenic nitrogen deposition on land, a process which is believed to play a role in the present carbon balance, is not included in the model.
 We know these limitations may be of importance, however, the role of land cover changes and nitrogen deposition in the current carbon budget are still highly uncertain [Caspersen et al., 2000; Holland et al., 1997; Joos et al., 2002; Nadelhoffer et al., 1999; Pacala et al., 2001; Schimel et al., 2001]. Moreover, future changes in land cover may very likely to be mainly driven by the direct human activity rather than by climate change. In this pioneer study, where our main focus is the future response of the biosphere, we decided not to include these processes and to concentrate on the two major impacts of rising CO2 and climate change.
2.3. Description of the Carbon-Climate Simulations
 Coupling between the carbon cycle models and the climate model is asynchronous, with a coupling time step of 1 year. The oceanic and terrestrial carbon models are driven by monthly climatic fields simulated by the OAGCM which itself is forced by annual atmospheric CO2 concentration calculated the year before by the carbon models following equation (4).
where EMIi, BIOi and OCEi represent the anthropogenic emissions, land biospheric and oceanic uptakes of CO2 (in Gt C yr−1) and CO2i is the CO2 atmospheric concentration (in ppmv) of year i. The next year concentration CO2i +1 is given to the OAGCM to simulate the next climate year. CO2 is initialized at 286 ppmv for year 1860.
 Two 241-year long (from 1860 to 2100) coupled simulations were carried out (Table 1). The first simulation is a control simulation where anthropogenic CO2 emissions are set to zero. Equation (4) becomes then CO2i +1 = CO2i − (BIOi + OCEi)/2.12, that is atmospheric CO2 concentration is only a function of uptake or release of carbon by land biosphere and ocean. The Control simulation is performed to assess the stability of the coupling, the absence of drift and to quantify the impact of internal climate variability on carbon fluxes. The second simulation, called Coupled Scenario run, uses prescribed anthropogenic CO2 emissions following historical data before 1990 [Andres et al., 1996] and the IPCC/SRES-A2 emission scenario from 1990 up to 2100 [Nakicenovic et al., 2000]. Note that SRES-A2 emissions account for both fossil fuel emissions and deforestation fluxes. In the Coupled Scenario simulation, land and ocean carbon cycles respond to the atmospheric CO2 increase and to the consequent climate change.
Table 1. Simulations Description
computed (no emissions)
Coupled Scenario run
computed (A2-SRES scenario)
computed (A2-SRES scenario)
taken from the Control run
Climate Impact run
taken from the Fertilization run
taken from the Coupled Scenario run
Temperature Impact run
taken from the Fertilization run
taken from the Control run, except temperature taken from Coupled Scenario run
Precipitation Impact run
taken from the Fertilization run
taken from the Control run, except precipitation taken from Coupled Scenario run
Radiation Impact run
taken from the Fertilization run
taken from the Control run, except radiation taken from Coupled Scenario run
Soil Water Content Impact run
taken from the Fertilization run
taken from the Control run, except soil water content taken from Coupled Scenario run
 In order to analyze the feedback between climate and carbon cycle, we perform two others simulations, the Fertilization and the Climate Impact simulations (Table 1). In the fertilization run, we use the same IPCC emission scenario, but we suppress the influence of the climate change on the carbon cycle. That is to say, OCEi and BIOi are calculated using the climate of the Control simulation. The difference between this simulation and the Control one isolates the impact of CO2 increase only on land.
 The difference in simulated uptakes between the Coupled Scenario run and the Fertilization run are primarily due to climate change (which tends to reduce the uptakes) but also, as a direct consequence, to higher atmospheric CO2 (which tends to enhance the uptakes). The Climate Impact run uses the climate of the Coupled Scenario simulation and the atmospheric CO2 of the Fertilization simulation. This last simulation allows us to separate the climate impact from the impact of additional atmospheric CO2 on the carbon fluxes (Table 1). This simulation should be seen as the worst case scenario as the carbon fluxes are negatively impacted by the climate change but are not positively impacted by this additional atmospheric CO2.
 In our modeling framework, the ocean and land carbon cycle responses are intimately coupled. For instance, if the land biosphere releases CO2 to the atmosphere, then the ocean uptake will be stronger. In the following, we analyze only the response of the land biosphere, but it should be kept in mind that this response implicitly and indirectly is function of the role of the ocean. A symmetrical analysis of the ocean carbon cycle in the coupled carbon cycle-climate model is presented by Bopp et al. .
3. Control Simulation and Evaluation of the Coupled Scenario Simulation Over the Historical Period
 The results from the Control simulation does not show any significant drift over the duration of the simulation (1860–2100). The variability of the calculated atmospheric concentration of CO2 is about 3 ppmv around the initial value of 286 ppmv and does not show any trend larger than 1 ppmv per century (Figure 2). The simulated net terrestrial biospheric flux varies between −2.1 Gt C yr−1 and +2.3 Gt C yr−1, with a mean value close to 0 Gt C yr−1 (Figure 3). Simulated NPP amounts to 57.5 ± 2.5 Gt C yr−1. Carbon contents of phytomass, litter and soil are 547 ± 3 Gt C, 103 ± 1 Gt C and 1165 ± 4 Gt C, respectively.
 To evaluate the results of the Coupled Scenario simulation, we used data set available along the historical period (1860–1999). We compare the simulated atmospheric CO2 concentration over the historical period to Antarctic ice core measurements (DE08 and DSS [Etheridge et al., 1996], H15 [Kawamura et al., 1997], Siple [Friedli et al., 1986]) combined with Mauna Loa measurements [Keeling et al., 1995] as shown in Figure 1. The simulation compares well with observations, however, we underestimate atmospheric CO2 during the period 1900–1940 [IPCC, 2001]. After 1960, the simulated CO2 concentration matches well the atmospheric observations from Mauna Loa with a maximum difference of less than 1 ppmv. The simulated atmospheric CO2 depends on the CO2 emissions but also on the oceanic and land carbon uptakes. These latter are also simulated in a realistic amount. Indeed, the net land biospheric uptake reaches 1.9 Gt C yr−1 for the 1980s with a interannual variability of about 2.0 Gt C yr−1 over the 1980s. We simulate a NPP of 64 Gt C yr−1 in 2000 (Figure 3). These values are in the range of the recent estimations [IPCC, 2001; Keeling et al., 1996; Saugier and Roy, 2001].
 In the Northern Hemisphere, the amplitude of the seasonal cycle of atmospheric CO2, which is mainly driven by the seasonality of NEP, has increased by 22% at Mauna Loa (20°N) since 1960 [Keeling et al., 1996]. We compared the amplitude change of atmospheric CO2 at Mauna Loa with the amplitude change of the NEP of the Coupled Scenario simulation between 20°N and 90°N (Figure 4). The simulated NEP amplitude change reaches 12%, which is lower than the observations.
4. Land Responses Under Atmospheric Increasing CO2 Only
 In this section, we analyze results from the Fertilization simulation where atmospheric CO2 depends on IPCC emissions and oceanic and land CO2 uptakes, these latter being calculated with the Control climate (Table 1). By 2100, cumulated fossil fuel and land use emissions reach 2190 Gt C, only 980 Gt C remains in atmosphere (Figure 2), the rest (1310 Gt C) being absorbed by the ocean and by the land biosphere. Compared to 1860, terrestrial carbon increases by 640 Gt C by the end of the simulation (Table 2). More than half of this excess carbon (57%) is stored in the phytomass. One third accumulates in soils, the rest being stored in litter (Table 2). Sizewise, phytomass and litter pools increase by 65% by 2100 while soil carbon increases only by 18%. Within the living phytomass, as can be expected, most of carbon accumulates in wood. However, the rate of increase of excess carbon starts to decrease in wood around 2010, whereas it still increases in soils by 2100. It means that phytomass carbon begins to saturate, as opposed to soil carbon which has a long turnover time and accumulates an excess carbon over longer timescales.
Table 2. Global Carbon Fluxes and Pools for the Control, the Coupled Scenario, the Fertilization, and the Climate Impact Runsa
Coupled Scenario Run
Climate Impact Run
Values in parentheses are for the 90°S–30°N, 30°N–50°N and 50°N–90°N latitudinal bands, respectively. Fluxes are in Gt C yr−1 and pools are in Gt C.
 Regionally, two thirds of the carbon storage (416 Gt C in 2100) occurs in the tropics and southern extratropics (90°S–30°N) lands (Table 2). The last third is essentially stored in Northern Hemisphere deciduous and coniferous forests.
5. Terrestrial Responses Under Increasing CO2 and Climate Change
 In the Coupled Scenario simulation, the overall effect of climate change (climate change alone + additional atmospheric CO2) is to reduce terrestrial carbon uptake, the oceanic uptake being almost unchanged [Dufresne et al., 2002].
 The cumulated land uptake reaches 465 Gt C by 2100, compared to 640 Gt C in the Fertilization run (Table 2). As a result, atmospheric CO2 reaches 778 ppmv in 2100 in the Coupled Scenario run, compared to 700 ppmv in the Fertilization simulation (Figure 2). This positive climate feedback on atmospheric CO2 is analyzed in detail by Dufresne et al. . Here we concentrate on analyzing the different mechanisms responsible for the global reduction of land uptake.
5.1. Global and Regional Responses
 The NPP simulated by the Coupled Scenario run reaches 82 Gt C yr−1 by 2100, compared to 94 Gt C yr−1 in the Fertilization run. This reduction of 12 Gt C yr−1 is the result of two opposite effects: (1) a reduction of NPP due to the climate change (Figure 5a), and (2) an increase of NPP due to the additional atmospheric CO2. As mentioned in section 2.3., we use the Climate Impact simulation to separate these two effects. In this last run, NPP does not benefit from a higher atmospheric CO2 and only reaches 78 Gt C yr−1 (Table 2).
 In the following we compare results from the Climate Impact and the Fertilization runs. As these two simulations are forced by the same atmospheric CO2, they allow us to isolate the climate impact alone on the terrestrial biosphere.
 In the Climate Impact run, the increase of terrestrial carbon pool amounts to 412 Gt C by 2100 (to be compared with 640 Gt C in the Fertilization run, Table 2). Around 2100, NEP only amounts to 2.9 Gt C yr−1 in the Climate Impact run (to be compared with 7.7 Gt C yr−1 in the Fertilization run) (Figure 3). Actually, by 2060, the NEP saturates around 4 Gt C yr−1 and begins to decrease after 2090, whereas NEP simulated by the Fertilization run continues to increase in 2100. Most of this climate impact occurs in the tropics, where NEP only reaches 0.6 Gt C yr−1 in the Climate Impact run, whereas it reaches 4.9 Gt C yr−1 in the Fertilization run by 2100 (Table 2). In the Northern Hemisphere middle and high latitudes (30°N–90°N), the impact of climate change is not so strong. Tundra and conifer forests show a small increase in NEP in response to climate change, while deciduous forests, grasslands and croplands shows the opposite.
 Changes in NEP are controlled by changes in NPP and in RH. In the model, RH is a function of the soil and litter carbon content and of the decomposition rate (equations (2) and (3)). Therefore, climate can affect RH directly through the decomposition rate, but also indirectly through changes in carbon pools size (subsequent to changes in NPP and litterfall).
 In the tropics, litter and soil carbon are drastically reduced (Table 2) mainly because of the strong reduction of carbon input. This directly leads to a quasi proportional reduction of RH. The direct climate impact on RH is relatively small: The decomposition rate increase following the warming is partially reduced by a decrease due to the tropical soil drying (Figure 5b).
 In the Northern Hemisphere, the small climate induced increase of tundra and conifers NPP does not translate into any increase in litter and soil carbon. The climate induced increase in RH is therefore due to an acceleration of the decomposition rate which follows the large increase of surface temperature. Deciduous forests do not show any increase of NPP, but they show a large increase in decomposition rate which leads to an increase of RH and a decrease in litter and soil carbon (Table 2). Temperate grasslands and croplands have a behavior similar to the one observed in the tropics : a decrease of NPP which drives a decrease in pool size and therefore in RH.
5.2. Analysis of the Mechanism Driving the Climate Induced Changes In Land Carbon Cycle
 In this section, we highlight the mechanisms that drive the ecosystems response to climate change. Our terrestrial biospheric model is driven by three external climate variables: surface temperature, precipitation, and incoming solar radiation (Figure 6). To separate the impact of each variable on the terrestrial biosphere, we performed three offline simulations (Table 1, Temperature Impact, Precipitation Impact and Radiation Impact simulations) where two of the forcing variables are set to their control value and the third one is taken from the Climate Impact run. In those simulations, atmospheric CO2 is also set equal to the one of the Climate Impact run.
 As shown in Figure 6, temperature has a direct control on growth and decomposition rate but also an indirect control on these fluxes through soil water content (equations (1) and (3)). Indeed, soil water content is calculated in SLAVE from a balance between precipitation, evapotranspiration, which directly depends on temperature, and runoff. A decrease in soil water content will tend to reduce NPP and soil decomposition, whereas a large increase in soil water content may also decrease decomposition if soils become water saturated. To separate the overall impact of temperature change on land biospheric fluxes from the one specifically due to change in soil humidity, we ran a fourth simulation where all climate variables are set to their Control value except for the calculation of soil water budget where we used temperature and precipitation from the Climate Impact run (Soil Water Content Impact simulation, Table 1).
5.2.1. Tropics and Southern Hemisphere
 In the tropics, we find that future evolution of both temperature and precipitation have a negative impact on tropical NPP, temperature being the main driver for the tropical NPP response to climate change (Table 3). The Temperature Impact and the Soil Water Content Impact simulations show basically the same response for NPP, meaning that the impact of higher temperature is primarily indirect, via the reduction of soil water content. Elevated temperatures lead to an increase in water stress on tropical vegetation (Figure 5b), explaining the reduction of NPP. This effect is common to seasonal and evergreen tropical forests and savannas (Figure 7), where loss of soil water is of 20% by 2100. Over evergreen tropical forests, we simulate a slight increase in precipitation, which is too low to counterbalance the large temperature effect (Table 3). We also find that change in radiation alone has a negligible positive impact on tropical productivity (Table 3).
Table 3. Change in NPP Attributable to Increased Atmospheric CO2, Climate Change and Change in Specific Climate Variablea
Tropics (90°S 30°N)
North (30°N 50°N)
North (50°N 90°N)
Deciduous Forests (Trop/Temp)
Seasonnal Tropical Forests
Evergreen Tropical Forests
See Table 1 for more details. Values in parentheses are for tropical regions (30°S–30°N) and for temperate regions (30°N–50°N).
−0.10 (−0.33/+ 0.23)
−4.00 (−4.2/+ 0.2)
Soil water content effect
 As mentioned before, the response of RH follows the one of NPP. Overall, the strong NEP reduction we simulate in the tropics is mainly due to the indirect impact of temperature on NPP through a reduction of soil water content.
5.2.2. Middle and High Northern Latitudes
 In the Northern Hemisphere middle and high latitudes, for all ecosystems, the direct impact of elevated temperature on NPP is quite important. Tundras and coniferous forests experience a longer growing season with favorable temperature. For these ecosystems, this mechanism explains the overall increase of NPP under climate change (Figure 7). For deciduous forests, temperate grasslands and croplands, the direct effect of temperature on NPP is also present, however, it is counterbalanced by the impact of reduced soil moisture on NPP (Table 3 and Figure 7). This latter is due to the combination of higher surface temperature and reduction of summer precipitation. We also find that the increase of incoming radiation during the summer tends to increase NPP.
 Concerning tundras and conifers, the NEP changes follow changes in NPP, that is a slight increase because of higher temperature. For deciduous forests, where NPP does not increase, the decrease in NEP is explained by an increase of RH as the decomposition rate is favored by warmer temperature. The temperate grasslands and croplands NEP reduction follows the NPP one, as RH is reduced since the litter and soil carbon pool size are decreased.
5.3. Response of the Seasonal Amplitude of Land Biospheric Fluxes to Climate Change
 Careful analysis of the seasonal behavior of NEP in our simulation makes it possible to derive additional information on the respective role of NPP and RH. The evolution of the amplitude of NEP seasonal cycle in mid and high latitude bands of the Northern Hemisphere is represented in Figure 8.
5.3.1. High Northern Latitudes
 At high northern latitudes (>50°N) the NEP seasonal peak to peak amplitude in the Climate Impact run increases by 55% in 2100 (Figure 8). This increase is mainly due to rising atmospheric CO2 (by 52%) stimulating NPP (Figure 9). Temperature change increases NPP and RH amplitudes by direct effect but decreases them by indirect effect via soil water content. Precipitation change which increases soil water content leads to a small amplitude increase. Overall, climate change has a negligible impact on the change of amplitude of NEP as both the NPP and RH amplitude are affected the same way by climate.
5.3.2. Middle Northern Latitudes
 The model results in the mid northern latitudes are drastically different from the high latitudes (Figure 8). Between 30°N and 50°N, the increase in NEP amplitude of the Climate Impact run is greater than north of 50°N and reaches 95% by 2100. The fertilization effect and the effect of temperature change on NEP amplitude are of the same order, 50% of this increase is due to the increase of atmospheric CO2 and 45% is due to temperature change. This climate induced increase in NEP amplitude mainly occurs in the late summer as the terrestrial biosphere becomes a strong source of carbon (Figure 9).
 The simulated seasonal amplitude of NPP decreases under climate change (Figure 9). NPP is reduced in summer as a result of two factors: (1) warmer temperature and reduced precipitation which increases water stress on productivity and (2) warmer temperature which directly increases the heat stress on productivity. Indeed, during this period, simulated temperature exceed optimum temperature (18°C) (Figure 10). As a result, the maximum of NPP occurs in May in the Climate Impact run, whereas it occurs in June in the Fertilization run.
 RH amplitude is not affected by the climate change. Indeed, it is stimulated by the increasing temperature (especially in spring and summer) but reduced by the reduced soil water content (during the whole year) (Figure 9).
 Overall, the large NPP reduction in summer induces a strong NEP source in July and August (Figure 9). This explains the climate induced increase of NEP seasonal cycle amplitude.
5.4. Response of the Growing Season Length to Climate Change
 We define the length of the growing season net flux as the period where NPP exceeds RH.
5.4.1. High Northern Latitudes
Figure 11 represents the evolution of the onset and the offset of the growing season for the high latitudes (>50°N). Though the increasing atmospheric CO2 has shown an impact on NEP seasonal amplitude, it shows no impact on the duration of the growing season (comparison between Fertilization and Control runs). The simulated growing season is lengthened by about 18 days by 2100, because of an earlier onset (12 days) and of a later termination (6 days).
 The earlier onset is essentially due to changes in temperature. As mentioned in the previous section, the maximum of NPP and hence NEP occurs earlier in the year as spring temperature are closer to the optimum temperature in the Climate Impact run (Figure 10). The end of the growing season is controlled by two opposite phenomena: It is delayed by about 7.5 days because of warmer temperatures but it is advanced by about 1.5 days because of lower soil water content (Figure 10).
5.4.2. Middle Northern Latitudes
 In 30°N–50°N, the growing season net flux is more difficult to evaluate because of the shape of the NEP signal. We simulate a net carbon uptake from March to June and a release from July to September, then a net carbon uptake between October and November and finally a release from December to February (Figure 11). The beginning of the growing season is advanced by 15 days. It can be explained by the same reason than for the high latitudes: direct impact of higher temperatures on spring NPP. Because of the particular shape of NEP in late summer and fall, we cannot estimate any change in the offset of the growing season.
6. Discussion and Conclusions
 To understand the role of the terrestrial biosphere in controlling future atmospheric CO2 and indirectly future climate change, we analyzed several simulations from a coupled climate-carbon model forced by an SRES-A2 IPCC scenario of fossil fuel and deforestation CO2 emissions [Dufresne et al., 2002].
 For the coming century, the simulated increase of terrestrial uptake due to higher atmospheric CO2 is partially reduced because of climate change (difference of 228 Gt C between the Fertilization and the Climate Impact simulations). The global land biospheric response is largely driven by the tropical ecosystems (decrease in storage by 215 Gt C), which are strongly affected by drought stress. After 2050, some regions in the tropics even become sources of carbon. Only at high latitudes of the Northern Hemisphere, our model predicts that photosynthesis will, on the contrary, be stimulated by the increase of surface temperature which extends the growing season of the terrestrial biosphere.
 The global reduction of carbon uptake induces an increase of atmospheric CO2 (78 ppmv by 2100) which in turn enhances the terrestrial uptake. Overall, this “additional” fertilization effect slightly reduces the negative impact of climate on the terrestrial storage. The climate change alone induces a global reduction of 228 Gt C while the additional CO2 leads to an extra storage of 53 Gt C.
 However, the result described above are very likely to be climate and carbon models dependent. Indeed, the OAGCMs vary considerably in their ability to simulate climate change. Previous climate change OAGCMs intercomparison studies [IPCC, 2001; Meehl et al., 2000; Räisäinen, 2001] have shown that the OAGCMs agreement to simulate temperature change is very good in most of the world whereas there is very little agreement in the simulation of precipitation change (especially in tropics). As tropics are the regions where we simulate the largest carbon uptake reduction, we can expect our results to be somehow dependent on the OAGCM used. However, the Hadley OAGCM that has been used in the same context shows results consistent with ours: a reduction of tropical NEP because of climate change [Cox et al., 2000].
 As we use different climate and carbon cycle models in this study, the uncertainty in the results can come from both sides. To evaluate roughly the consequences of the way we model the climate-carbon cycle system, we compare the results of our study to the Cox et al.  coupled simulation.
Cox et al.  used the Hadley model to investigate the climate-carbon cycle feedback. By 2100, atmospheric CO2 reaches 980 ppmv in the Hadley coupled simulation, whereas it only amounts to 770 ppmv in our coupled simulation. Reasons for such a large difference are mainly due to the land biosphere response to climate change. The IPSL cumulated terrestrial net storage amounts to 465 Gt C in 2100 whereas the Hadley model simulates a net source to the atmosphere of 100 Gt C.
 In the IPSL simulation, the terrestrial biosphere does not become a global source of carbon. We only locally simulate net carbon release to the atmosphere (in South America) due to the soil drying which induces a saturation of NPP. Several differences between the IPSL and Hadley simulations require further attention. First, the IPSL partitioning of excess carbon due to the increase of atmospheric CO2, without climate change, is greater in the vegetation (400 Gt C in 2100) than in soils (200 Gt C) whereas Hadley terrestrial biosphere has the opposite behavior. The partitioning of carbon in above versus below ground pools in the IPSL coupled simulation is similar to the one of previous simulations conducted offline [Cramer et al., 2001]. Whether the excess carbon will, in the future, go preferentially in the vegetation or in the soil is still an open question. Second, the impact of climate change on soils is much larger in the Hadley simulation (reduction of 550 Gt C in 2100) than in the IPSL one (reduction of 100 Gt C) or in the various land biosphere models analyzed by Cramer et al.  (reduction of 150 Gt C averaged across models). This could partly be due to the large storage in the Hadley soil model under increased CO2 only. We also found in our simulations that soil respiration generally decreases because NPP and hence soil carbon input decreases under climate change, a mechanism not seen in the Hadley simulations. This may be due to different assumptions in soil turnover times in the two models. Third, another difference between the Hadley framework and ours is the inclusion of a dynamic vegetation model in Hadley model: drought stress initiates loss of Amazonian forest, which also contribute to release carbon to the atmosphere. In the IPSL simulation, the land biosphere model uses a fixed distribution of ecosystems, thus neglecting the impact of climate on the vegetation distribution and also neglecting the importance that such processes may have in the future, for example by reducing plants population or NPP [Cox et al., 2000; Cramer et al., 2001]. As explained in section 2.2, we decided not to account for future scenario of land use and therefore we preferred not to account for vegetation dynamic at all. It is worth noting that both simulations suffer from severe simplifications.
 Regarding the modeling of climate change, we do not account for changes in aerosols emissions that are very likely to come along with CO2 emissions. Atmospheric chemistry as a whole is not accounted for in such simulations. Future works should consider the evolution of aerosols and other greenhouse gases in the atmosphere and their control on the Earth radiative budget [Haywood and Boucher, 2000; Jacobson, 2001].
 Regarding the carbon cycle, we manage to simulate annual atmospheric CO2 from 1860 to 2000, however, we do not include some terrestrial processes that are very likely to play an important role in the net land uptake. Indeed, we assume that the uptake is only driven by climate and atmospheric CO2, whereas anthropogenic nitrogen deposition and land cover changes are also believed to play a large role. Several studies estimated the importance of anthropogenic nitrogen deposition on terrestrial carbon cycle [Holland et al., 1997; Nadelhoffer et al., 1999]. The role played by anthropogenic nitrogen remains largely uncertain but could be significant and should be included in future climate-carbon cycle studies. Another potentially large component of the present-day terrestrial uptake is due to forest regrowth in the temperate regions Caspersen et al., 2000; Joos et al., 2002; Pacala et al., 2001; Schimel et al., 2001. Such processes are also not included in the present climate-carbon simulations. Future simulations, should clearly use historical land cover changes data and future land cover changes scenarios.
 In summary, even without considering all the processes that can affect land carbon cycle, our results highlight the crucial role that the land biosphere may play in future climate change. Despite large sources of uncertainties, it is generally found that land biospheric carbon pools may be affected by the climate change leading to a positive feedback between carbon cycle and climate system. However, its amplitude is still uncertain. Intercomparison project of climate-carbon cycle models should help to better understand the mechanisms involved in the feedback [Cox et al., 2002].
 It is also clear that such experiments are still very idealized, rather than credible simulations of what will happen in the future. Even if the results of this study are roughly consistent with observations during the historical period, it is likely that not considering all involved processes may lead to uncertainties and that improving terrestrial carbon cycle models is of the uppermost importance, just like improving GCM predictions.
 This work was part of the IPSL modeling pole and is supported by PNEDC Program. Computer service was provided by IDRIS/CNRS and by DKRZ.