Multiple mechanisms of Amazonian forest biomass losses in three dynamic global vegetation models under climate change


Author for correspondence:
David Galbraith
Tel: +44 (0) 131 445 8594


  • The large-scale loss of Amazonian rainforest under some future climate scenarios has generally been considered to be driven by increased drying over Amazonia predicted by some general circulation models (GCMs). However, the importance of rainfall relative to other drivers has never been formally examined.
  • Here, we conducted factorial simulations to ascertain the contributions of four environmental drivers (precipitation, temperature, humidity and CO2) to simulated changes in Amazonian vegetation carbon (Cveg), in three dynamic global vegetation models (DGVMs) forced with climate data based on HadCM3 for four SRES scenarios.
  • Increased temperature was found to be more important than precipitation reduction in causing losses of Amazonian Cveg in two DGVMs (Hyland and TRIFFID), and as important as precipitation reduction in a third DGVM (LPJ). Increases in plant respiration, direct declines in photosynthesis and increases in vapour pressure deficit (VPD) all contributed to reduce Cveg under high temperature, but the contribution of each mechanism varied greatly across models. Rising CO2 mitigated much of the climate-driven biomass losses in the models.
  • Additional work is required to constrain model behaviour with experimental data under conditions of high temperature and drought. Current models may be overly sensitive to long-term elevated temperatures as they do not account for physiological acclimation.

Centro de Previsão de Tempo e Estudos Climáticos


Climate Research Unit


Amazonian vegetation carbon


The change (2002–2100) in Amazonian vegetation carbon


Dynamic Global Vegetation Model


Free-air CO2 Enrichment


General Circulation Model


Intergovernmental Panel for Climate Change


Hadley Centre Coupled Model, version 3




Met Office Surface Exchange Scheme


factor by which respiration increases following a 10-degree increase in temperature


Special Report on Emissions Scenarios


throughfall exclusion


Top-down Representation of Interactive Foliage and Flora Including Dynamics


The possibility of substantial loss of Amazonian rainforest cover and carbon as a consequence of climate change (‘Amazon die-back’) was first reported by White et al. (1999) following offline simulations performed with the Hybrid dynamic global vegetation model (DGVM). The loss of the Amazonian rainforest was also a key part of the amplifying climate–carbon cycle feedback simulated by Cox et al. (2000). Given the importance of the Amazon Basin in carbon storage (Melillo et al., 1993) and in moisture and heat exchange with the atmosphere (Gash et al., 2004), extensive loss of Amazonian rainforest could have pronounced effects on global climate, and may constitute a ‘tipping point’ in the Earth System (Lenton et al., 2008).

Changes to the Amazonian climate which might reduce rainforest cover in favour of savanna are predicted by a number of global climate models (Salazar et al., 2007; Malhi et al., 2009a). The most severe losses of Amazonian rainforest have been simulated using the HadCM3 (Hadley Centre Coupled Model, version 3) model (Cox et al., 2000, 2004; Betts et al., 2004), which predicts particularly strong reductions in rainfall and increases in temperature over Amazonia. Although it predicts extreme changes in Amazonian climate in the 21st century, the HadCM3 model simulates key aspects of tropical climate variability more realistically than most other global climate models (Cox et al., 2004, 2008; Li et al., 2006). The reduction in rainfall predicted by HadCM3 has generally been considered to be the major cause of simulated rainforest loss (Betts et al., 2004; Huntingford et al., 2008). However, the role of rainfall relative to other climatic factors has not been analysed explicitly. Several other environmental factors will influence modelled Amazonian carbon stocks, including temperature, humidity and atmospheric CO2 concentration. The relative importance of each of these (and their interactions) is unclear.

The means by which these environmental drivers affect plant functioning in models may also be complex. For example, rising temperature affects net primary productivity (NPP) directly through effects on photosynthesis and respiration, but also through increases in leaf-to-air vapour pressure deficit (VPD) (Lloyd & Farquhar, 2008) (Fig. 1). Increased VPD can reduce photosynthesis in the short-term through reduced stomatal conductance or in the longer term through changes in plant water balance as a result of increased evapotranspiration. The relative importance of each of these temperature mechanisms in causing simulated changes in Amazonian vegetation carbon stocks has not been investigated.

Figure 1.

 Schematic illustration of the mechanistic pathways, both direct and indirect, through which increasing temperatures can affect net primary productivity (NPP). The exact mechanisms differ from model to model, but the major processes are shown here.

This study sought to analyse the mechanisms underlying the predicted change in rainforest vegetation structure and function across the Amazon in three DGVMs: Hyland (Levy et al., 2004), Lund–Potsdam–Jena (LPJ) (Sitch et al., 2003) and MOSES-TRIFFID, a combination of the Met Office Surface Exchange Scheme (MOSES) and the TRIFFID (Top-down Representation of Interactive Foliage and Flora Including Dynamics) vegetation model. We designed a factorial experiment to quantify the relative contributions of CO2, humidity, precipitation and temperature (and their interactions) to changes in Amazonian vegetation carbon. We also conducted further simulations to investigate the importance of direct and indirect effects of temperature. We conclude by discussing whether the representation of the response of Amazonian rainforest to climatic change in these DGVMs is likely to be accurate.

Materials and Methods


We define the Amazon region as being a rectangle, extending from 3.0°N to 12.5°S and from 70°W to 48°W; hereafter, all values refer to this domain unless stated otherwise. We subdivide this domain into eastern and western regions along the 60°W longitude. We denote the carbon present in the biomass of Amazonian vegetation as Cveg in Pg C, and the change in this quantity between 2003 and 2100 as ΔCveg.

Vegetation models

Three dynamic vegetation models, developed primarily for global and regional applications, were used in this study. The models were selected as they showed varying decreases in Amazonian Cveg in a recent model intercomparison study by Sitch et al. (2008). Here we provide only a brief description of key processes in each model for understanding changes in vegetation carbon, the key variable of interest in this study. For more comprehensive mathematical descriptions of the models, the reader is referred to the original papers (Friend et al., 1997; Cox, 2001; Sitch et al., 2003).

Hyland is a DGVM largely based on the Hybrid model (Friend et al., 1997; Friend & White, 2000) and modified according to Levy et al. (2004). The model simulates three plant functional types (PFTs; evergreen tree, deciduous tree and C3 grass), which compete with each other for light. Soil moisture availability is simulated using a one-layer bucket model, where the soil water holding capacity is a dynamic function of soil organic matter (as a surrogate for rooting depth). Soil moisture status, temperature, CO2 and VPD have a direct effect on stomatal conductance, which is based on the Jarvis–Stewart multiplicative model (Jarvis, 1976; Stewart, 1988). The canopy energy balance is solved (as a function of air temperature, isothermal net radiation, humidity, and resistances to heat and water transfer assuming a constant wind speed) to give surface temperature and evapotranspiration rate (Friend, 1995). Photosynthesis is based on Farquhar et al. (1980), with modifications made by Friend (1995). Plant respiration is taken to be a constant fraction (0.5) of gross primary production (GPP) (Levy et al., 2004). The input variables to Hyland are air temperature, precipitation, humidity, downward shortwave radiation and the atmospheric CO2 concentration.

The LPJ DGVM (Sitch et al., 2003) simulates competition for water and light between nine plant functional types. Soil moisture availability is simulated using a two-layer soil model with a total depth of 2 m. Photosynthesis is based on the scheme of Collatz et al. (1991, 1992), itself a modification of Farquhar et al. (1980), as implemented by Haxeltine & Prentice (1996). The impact of drought on photosynthesis is via a canopy conductance feedback which takes into account both the atmospheric demand and the supply of water from the roots. Stomatal conductance is reduced through increases in atmospheric CO2, but does not affect the surface temperature which is assumed to be equal to the air temperature. The temperature response of plant respiration is independent of that of photosynthesis and follows an Arrhenius function, based on Lloyd & Taylor (1994). In this version, the inputs to LPJ are rainfall, temperature, the monthly averaged daily percentage of sunshine hours and the atmospheric CO2 concentration.

The TRIFFID model (Cox, 2001) was run coupled to the MOSES land surface scheme (Cox et al., 1998) and the MOSES-TRIFFID combination mimics that used by Cox et al. (2000), where widespread loss of the Amazon rainforest was first reported with the model. TRIFFID simulates five plant functional types (broadleaf tree, evergreen tree, shrub, C3 grass and C4 grass) which compete with each other following Lotka–Volterra dynamics (Cox, 2001). A four-layer soil model is simulated with a total depth of 3.0 m, although individual plant functional types differ in their rooting depths. The soil moisture status directly scales net leaf photosynthesis, which is calculated based on Collatz et al. (1991, 1992). The surface energy balance is calculated separately for each plant functional type (PFT) using a variant of the Penman–Monteith equation (Essery et al., 2003), which diagnoses the evapotranspiration and surface temperature given the aerodynamic resistance and stomatal conductance for each PFT. Stomatal conductance is sensitive to temperature, vapour pressure deficit, soil moisture status and photosynthetically active radiation, and reduces with increasing CO2 concentration (Cox et al., 1998). As a result this model is known to produce significant reductions in evapotranspiration and therefore increases in runoff (Gedney et al., 2006) and surface temperature (Betts et al., 2004) under increasing CO2. Plant respiration increases exponentially with temperature in this version of the model, according to a simple "Q10" dependency where Q10 = 2.0 (see Cox, 2001). In this offline study, the inputs to MOSES-TRIFFID are: downward shortwave and longwave radiation at the surface; rainfall; humidity and temperature at screen level (1.5 m); windspeed at 10 m; and the atmospheric CO2 concentration.

Climate data

For the historical period (1901–2002), we used the observed climatology from the Climate Research Unit (CRU) data set (New et al., 2002), aggregated to HadCM3 resolution (3.75° longitude × 2.5° latitude as described by Sitch et al., (2008). A near-present day control climatology was defined as the 1983–2002 period within these data. This was used for control simulations with no mean change in future climate, but with interannual variability as observed during this period. Future climate (2003–2100) was based on a pattern-scaling approach to reproduce HadCM3 simulations (Huntingford & Cox, 2000) to enable comparison with previous studies (Cox et al., 2004; Huntingford et al., 2004; Huntingford et al., 2008; Sitch et al., 2008). Simulations were conducted for four IPCC SRES scenarios (A1FI, A2, B1 and B2). Anomalies from the HadCM3 simulation were added to the control climatology, thereby maintaining the same interannual variability as the control but with a change in the mean. Changes in the mean value of each of the climate variables across the Amazon region, and for each scenario, are shown in Fig. 2. The four scenarios are associated with prescribed CO2 concentrations based on Nakicenovic et al. (2000). CO2 concentrations in 2100 ranged from 531 ppm (B1 scenario) to 925 ppm (A1FI scenario). Relative to the final year of the historical simulation (2002), mean annual temperature increased by 2.76°C in the most conservative scenario, and by 7.16°C in the most severe scenario. Precipitation decreased from an annual mean value of 2138 mm in 2002 to 1457–1867 mm yr−1 in 2100, while mean annual relative humidity decreased from 77.8% in 2002 to 53.8–68.1% in 2100.

Figure 2.

 Predictions of change in key environmental factors over the Amazon region, derived by adding anomalies from the HadCM3 climate model to the 1983–2002 mean climatology.

Factorial simulations

Each model was run to its preindustrial equilibrium state using data from the first decade of the CRU data set (1901–1910), as described in Sitch et al. (2008). Models were then run from their preindustrial equilibrium in 1901 through to 2002. To investigate the contribution of environmental input variables to variation in ΔCveg, a four-factor two-level full factorial experimental design was used for each SRES scenario over the future climate period (2003–2100). The four factors were precipitation, temperature, relative humidity and CO2. The two levels corresponded to either the control climatology or the HadCM3 simulation of climate change. All combinations of the environmental variables of interest were included, giving a total of 16 simulations for each SRES scenario for Hyland and TRIFFID (Table 1) (LPJ does not take humidity as an input and thus only eight simulations were conducted for the LPJ model for each SRES scenario).

Table 1.   Summary of all simulations conducted as part of this study
Factorial simulations (all models)
SimulationDescriptionSRES scenario
  1. NA, not applicable; CRU, Climate Research Unit; GPP, gross primary productivity; VPD, vapour pressure deficit.

ControlNo future climate change, but 1983–2002 variability maintainedNA
PPrecipitation varies according to SRES; other variables according to control climatologyA1FI, A2, B1, B2
CCO2 varies according to SRES; other variables according to control climatologyA1FI, A2, B1, B2
HHumidity varies according to SRES; other variables according to control climatologyA1FI, A2, B1, B2
TTemperature varies according to SRES; other variables according to control climatologyA1FI, A2, B1, B2
PCPrecipitation and CO2 vary according to SRES; other variables according to control climatologyA1FI, A2, B1, B2
PHPrecipitation and humidity vary according to SRES; other variables according to control climatologyA1FI, A2, B1, B2
PTPrecipitation and temperature vary according to SRES; other variables according to control climatologyA1FI, A2, B1, B2
CHCO2 and humidity vary according to SRES; other variables according to control climatologyA1FI, A2, B1, B2
CTCO2 and temperature vary according to SRES; other variables according to control climatologyA1FI, A2, B1, B2
HTHumidity and temperature vary according to SRES; other variables according to control climatologyA1FI, A2, B1, B2
PCHPrecipitation, CO2 and humidity vary according to SRES; other variables according to control climatologyA1FI, A2, B1, B2
PCTPrecipitation, CO2 and temperature vary according to SRES; other variables according to control climatologyA1FI, A2, B1, B2
PHTPrecipitation, humidity and temperature vary according to SRES; other variables according to control climatologyA1FI, A2, B1, B2
CHTCO2, humidity and temperature vary according to SRES; other variables according to control climatologyA1FI, A2, B1, B2
PCHTPrecipitation, CO2, humidity and temperature vary according to SRES; other variables according to control climatologyA1FI, A2, B1, B2
Simulations with additional input variables (TRIFFID only) – complement to factorial analysis
TRI-ALLFull simulation with all variables according to SRES: includes additional input variables not considered in the factorial analysis – wind, longwave radiation and surface pressureA1FI, A2, B1, B2
Simulations to evaluate the importance of individual mechanisms to temperature-induced Amazonian carbon losses (all models)
T-INDTemperature associated with VPD varies according to SRES; temperature associated with direct photosynthesis and respiration from control runA2 (only temperature effect considered)
T-PHOTTemperature associated with photosynthesis varies; temperature associated with VPD and respiration from control runA2 (only temperature effect considered)
Simulations with alternate prescriptions of plant respiration (all models)
RESP-FIXPlant respiration taken to be a fixed fraction (0.5) of GPP; this is already the default prescription in HylandA2 (only temperature effect considered)
RESP-EXPPlant respiration taken to be an exponential function of temperature in TRIFFID (Cox et al., 2000) and Hyland (based on Ryan et al., 1991) and Arrhenius function in LPJ (Lloyd & Taylor, 1994). These correspond to the default prescriptions in TRIFFID and LPJ and to the prescription used in an earlier version of HylandA2 (only temperature effect considered)
Simulation of Tapajos and Caxiuanã drought experiment conditions
TFEStepped (50%) reduction in incident rainfall applied to CRU climatology in 2002 and continued for 10 yr. Only the grid cells containing the coordinates of the drought experiment sites in Tapajós (3°04′S, 54°95′W) and Caxiuanã (1°43′ S, 51°27′ W) were considered. Models were run with their default, global parameter settingsNA

Evaluation of simulated current vegetation carbon

Simulated vegetation carbon at the end of the historical period (1983–2002) was compared against a data set derived by remote sensing produced by Saatchi et al. (2007), which estimates aboveground living biomass for Amazonia at 1-km resolution. The Saatchi et al. (2007) data were aggregated to the resolution used in the study (3.75° longitude × 2.5° latitude) for comparison with the model output. We also compare the regional biomass values simulated by each model with a range of literature estimates. The vegetation models used in this study do not distinguish between aboveground and belowground woody biomass, grouping both pools into a single metric of vegetation carbon (Cveg). To enable comparison with model output, we assumed that belowground biomass in Amazonia is equal to 21% of aboveground living biomass (Malhi et al., 2006; Saatchi et al., 2007), and thus multiplied the Saatchi et al. (2007) estimates by 1.21.

Analysis of variance (ANOVA)

To quantify the effects of individual environmental drivers and their interactions on ΔCveg, we applied factorial analysis of variance (ANOVA) to the model output. Comprehensive descriptions of the application of ANOVA to sensitivity analysis of model experiments have been provided by previous authors (e.g. Campolongo & Saltelli, 2000; Chan et al., 2000) and here we provide only a brief overview of its application in this study. ANOVA has traditionally been applied to experimental field studies, where natural variability in measurements may be high and replication is necessary to provide confidence in the statistical significance of a particular result (e.g. the difference between two treatment groups). However, ANOVA is increasingly being used to analyse output from deterministic models (e.g. Stevens et al., 1996; Cameron et al., 2005; Hodson & Sutton, 2008; Shuman & Shugart, 2009) where there is no stochastic element and thus replication is not relevant. In such models, such as the DGVMs used in this study, a common use of ANOVA is to assess the average contribution of each input factor (main effect) and their interactions to the overall outcome variance, defined as the sum of the squared differences between each individual simulation and the overall mean (Campolongo & Saltelli, 2000). In the context of this experiment, for factor x, the main effect is equivalent to ΔCveg in the simulation where only factor x is varied, minus ΔCveg from the control simulation. Interactions are present where the effect of one factor depends on the level of another factor, and ‘interaction effects’ quantify the additional combined effects of factors on the response variable. The main effects and interactions of all factors were estimated for each DGVM and scenario, and these are presented both in terms of ΔCveg and in terms of the fraction of overall variance accounted for.

Effect of other variables not included in factorial analysis

In addition to the factorial simulations, one further simulation (TRI-ALL in Table 1) for each scenario was run for the TRIFFID model to incorporate the additional effect of other variables predicted to change by HadCM3, but not used as inputs in the other two models. These variables included the windspeed and downward longwave radiation which are used to calculate the surface energy balances and surface temperatures in MOSES-TRIFFID. To simplify the analysis, these additional variables were treated together. Their combined effect was calculated as the difference relative to the simulation with only the four main factors included.

Partitioning of direct and indirect effects of temperature

Additional simulations were conducted to separate the effect of temperature into direct and indirect effects on plant physiology (Fig. 1). In Hyland and TRIFFID, the indirect effect of temperature via increased VPD was explicitly represented, while in LPJ, the indirect effect was via the effect of temperature on equilibrium evapotranspiration rates. To separate these effects, we performed an additional simulation (T-IND in Table 1) for each model, whereby we supplied the plant physiology components of the model with temperatures as calculated in the control simulation, while calculating VPD (Hyland and TRIFFID)/equilibrium evapotranspiration rates (LPJ) with temperature from HadCM3 simulations for the A2 scenario. All other environmental variables were held at their control values.

We also conducted another simulation (T-PHOT in Table 1) for each model which sought to separate the contributions of plant respiration and of photosynthesis. In this case, we calculated respiration components using temperatures from the control simulation, while photosynthesis was calculated using temperatures from HadCM3 simulations for the A2 scenario.

Sensitivity of ΔCveg to temperature dependence of plant respiration

To examine the sensitivity of vegetation carbon change to the prescription of plant respiration dependence on temperature, we conducted new simulations (RESP-FIX and RESP-EXP in Table 1), including spin-up and historical runs (1901–2002) with two alternate representations of plant respiration, either an explicit temperature dependence independent of that of photosynthesis or a constant fraction (0.5) of GPP. Both these simulations used temperatures from the HadCM3 simulation for the A2 scenario, with all other environmental variables held at control climatology values. Table 1 provides further details of these simulations.

Simulation of Amazonian throughfall exclusion (TFE) experiments

Runs were conducted with each DGVM where we attempted to simulate the throughfall exclusion (TFE) experiments at Caxiuanã National Forest and at Tapajós National Forest. In these runs, we applied the same spin-up procedure and ran the models as described above but with CRU climate data (1901–2002) for the two grid cells containing these sites. We then applied a 50% stepped reduction in incident rainfall for a period of 10 yr. The default configurations of the models were used for these runs (i.e. no site-specific parameterizations were applied). We compared the simulated changes in biomass with observed values at each site (Brando et al., 2008; da Costa et al., 2010).


Evaluation of simulated current biomass

Hyland and LPJ both simulated a total present-day vegetation carbon of c. 115 Pg C over the Amazonian region used in this study (c. 6 million km2), while TRIFFID simulated a total current vegetation carbon of c. 77 Pg C. These values are within the published range of Amazonian biomass estimates (Table 2), from c. 60 to 140 Pg C for a similar area. Fig. 3 shows a comparison between the Saatchi et al. (2007) estimates and our modelled values. The models generally overestimated the Saatchi et al. (2007) values, particularly in areas with low biomass. This may partly be a result of the models’ relative insensitivity to low rainfall, as discussed in the section on the simulation of the Amazonian drought experiments. However, some of the discrepancy will be attributable to the fact that the models represent potential vegetation, while Saatchi et al. (2007) depicts actual vegetation. The medians of the observed and modelled data sets are rather close. Median simulated vegetation carbon ranged from 14.9 kg C m−2 in TRIFFID to 19.9 kg C m−2 in Hyland. Median observed vegetation carbon ranged from 15.1 kg C m−2 (Saatchi et al., 2007) to 18.0 kg C m−2 across 227 Amazonian plots compiled by Malhi et al. (2006).

Table 2.   Comparison of modelled estimated of total Amazonian biomass with a number of literature estimates
SourceVegetation carbon (Pg C)Area (km2)Methodology
  1. Vegetation carbon estimates include belowground vegetation carbon stocks but exclude dead biomass components.

  2. DGVM , dynamic global vegetation model.

Hyland DGVM (this study)1156 000 000Model – potential vegetation only
LPJ DGVM (this study)1156 000 000Model – potential vegetation only
TRIFFID DGVM (this study)776 000 000Model – potential vegetation only
Saatchi et al. (2007)60–908 235 450Remote sensing and field observations – includes forest and nonforest biomes and managed land
Malhi et al. (2006)85–1405 760 000Interpolation of field estimates
Houghton et al. (2001)39–934 000 000Seven approaches including field estimates, environmental gradients and remote sensing
Figure 3.

 Comparison of simulated biomass with remotely sensed estimates. The data on the x-axis are the mode of the 1-km biomass values from Saatchi et al. (2007) data for each of the 52 grid cells comprising the Amazonian region. The Saatchi et al. (2007) data were increased by 21% to account for belowground carbon stocks, to enable direct comparison with dynamic global vegetation model (DGVM) biomass estimates.

Factorial simulations with HadCM3 climate data

With all factors included, all models simulated net losses in Cveg between 2003 and 2100, except for Hyland in the A2, B1 and B2 scenarios (the net effect in Fig. 4). CO2 had the largest effect of any single variable, and consistently increased Cveg (Table 3, Fig. 4). Gains in Cveg under increased CO2 were highest in the Hyland model (24.8–47.8 Pg C; 21.4–41.3%) and lowest in the LPJ model (16.5 s–33.9 Pg C; 14.2–29.1%). Cveg decreased in all DGVM/SRES scenario combinations where CO2 was held constant. Temperature change acted to decrease Cveg across all models, though in a rather more variable way, and had the greatest effect in TRIFFID, where increased temperature led to a decrease in Cveg of 20–52 Pg C (26.0–67.5% relative to 2002). In Hyland, the contribution of temperature to ΔCveg varied widely between emission scenarios, ranging from −2.9 to −34.1 Pg C (2.5–29.4%). In LPJ, rising temperatures accounted for a loss of 13.0–27.3 Pg C (11.3–23.8%). In LPJ, the magnitude of the effects of precipitation and temperature increase were similar, causing decreases in Cveg in all four scenarios (loss of 10.1–25.0 Pg C; 8.8–21.7%).

Figure 4.

 Contribution of environmental factors to simulated changes in Amazonian vegetation carbon (Cveg) for four SRES scenarios in three dynamic global vegetation models (DGVMs). Main effects and the sum of all interaction terms, as quantified in the factorial ANOVA, are shown. The overall net effect of including all factors is shown as the inner grey bar.

Table 3.   Percentage of variance in the change in carbon present in the biomass of Amazonian vegetation (ΔCveg) explained by each factor and their interactions
  1. Results are the average from ANOVA analyses of all four IPCC scenarios. The factors are: P, precipitation; C, carbon dioxide; H, humidity; T, temperature.

P × C0.080.150.00
P × H0.070.000.02
P × T0.060.130.00
C × H0.040.000.05
C × T0.210.021.41
H × T0.980.000.34
P × C × H0.000.000.00
P × C × T0.000.010.00
P × H × T0.020.000.00
C × H × T0.140.000.01
P × C × H × T0.000.000.00
% variance explained by main effects98.3699.7098.16
% variance explained by interaction effects1.640.301.84

Of particular note is the finding that precipitation change contributed little to overall ΔCveg in Hyland and TRIFFID, being directly responsible for mean losses of 5.4 (4.6%) and 2.9 Pg C (3.8%), respectively, across all scenarios. In the case of TRIFFID, this is a magnitude less than the effect of temperature. In Hyland, decreases in humidity further reduced Cveg in all four scenarios (losses of 4.4–10.1 Pg C; 3.9–8.7%). In TRIFFID, other variables not included in the factorial analysis and which were not inputs to other models (see the section on ‘Effect of other variables not included in factorial analysis’) also contributed significantly to losses of ΔCveg (loss of 10.6–20.5 Pg C; 13.8–21.6%) and thus were important in determining the net ΔCveg (Fig. 4).

In terms of variance accounted for, interaction terms among variables were relatively unimportant, explaining < 2% of the total variance in ΔCveg (Table 3). Interaction terms had the largest combined effect in the TRIFFID model (Fig. 4), where a considerable positive interaction term between CO2 and temperature was observed (i.e. the net effect when both varying temperature and varying CO2 were included in the simulations was more positive than the added individual effects of CO2 and temperature for the simulations where only one of the variables was changed). In Hyland, the combined effect of the interaction terms was negative in the more severe (A1FI and A2), but positive in the less extreme (B1 and B2) scenarios. This is because the positive interaction between CO2 and temperature is more important in the more conservative scenarios, whereas the negative interaction term between humidity and temperature becomes more important in the more severe scenarios. Interaction terms had virtually no effect in the LPJ model.

The geographical distribution of ΔCveg is shown in Fig. 5 for a subset of the A1FI (most extreme) factorial simulations. Increased CO2 on its own led to increased Cveg across all Amazonian grid cells in all models (Fig. 5a). In Hyland, decreased relative humidity reduced Cveg in all Amazonian grid cells, although this effect was slightly more pronounced in southeastern Amazonia (Fig. 5b). In TRIFFID, decreasing relative humidity had a small but variable effect.

Figure 5.

 Geographical distribution of changes in Amazonian vegetation carbon as simulated in a subset of factorial runs (HadCM3, A1FI scenario). The top four rows show results from simulations where all inputs were held constant at baseline values except CO2 (a), relative humidity (b), precipitation (c) and temperature (d). The final row shows the results when all of these were varied (PCHT simulation; Table 1). Red indicates loss of carbon (% of original vegetation carbon), green indicates gain, and white indicates little or no change. The rectangle shows the Amazonian domain as defined in this study.

With decreased precipitation (Fig. 5c), Hyland and TRIFFID showed only small losses (generally < 10%) in Cveg across most Amazonian grid cells. In fact, substantial losses of vegetation carbon as a result of rainfall reduction alone were largely restricted to grid cells in southeast Amazonia. These are located close to the semi-arid caatinga region of northeast Brazil, where all models simulate large losses of Cveg in response to decreasing precipitation. In both Hyland and TRIFFID, West Amazonia was found to be very insensitive to reduced rainfall, even under the most extreme SRES scenario. Although some grid cells in West Amazonia received up to 50% less rainfall in 2100 than in the historical period (Fig. 6a,c), Cveg in West Amazonia was only reduced by an average of 6% in Hyland and 3% in TRIFFID, compared to average reductions of 10 and 13%, respectively, in East Amazonia. LPJ simulated ΔCveg of −21 and −26% in West and East Amazonia, respectively (Figs 5c, 6b, A1FI scenario).

Figure 6.

 Relationship between simulated change in vegetation carbon (2003–2100) (kg m−2) and change in precipitation (left-hand panels) and temperature (right-hand panels) for all Amazonian grid cells. Data from all four SRES scenarios are shown. All other variables, including atmospheric CO2 concentrations, are held at baseline values (simulations P and T in Table 1).

The geographical pattern of the response to temperature increase was generally consistent in all three models, with slightly greater loss of vegetation carbon in West Amazonia than in the East, reflecting HadCM3 predictions of higher temperatures in West Amazonia compared with East Amazonia. TRIFFID in particular simulated very high temperature-driven losses of Cveg across all Amazonian grid cells, with over 85% of grid cells losing > 50% of original Cveg in the A1FI scenario (Fig. 5d). Across all Amazonian grid cells, a much closer correspondence was observed between ΔCveg and change in temperature than change in precipitation (Fig. 6). This demonstrates that DGVM forests respond to precipitation reductions once a threshold (or tipping point) has been exceeded. In LPJ and TRIFFID, any rise in temperature, without changes in other environmental factors, was sufficient to induce losses in Cveg, even in the most conservative (B1) scenario where some grid cells experienced an increase in temperature of < 2°C. In Hyland, the average ΔCveg caused by increased temperature was very small (−3.7% compared to −10.1% for LPJ and −26.6% for TRIFFID) in the B1 scenario and a few grid cells did not show reductions in Cveg. With all factors included, losses in Cveg were larger in West Amazonia than in East Amazonia in all three models (Fig. 5e).

Indirect vs direct temperature effects

The indirect effect of temperature (via increased VPD or transpiration) was found to be a minor contributor to temperature-induced ΔCveg in the TRIFFID and LPJ models (Fig. 7), where it was responsible for 2% (1.3 out of 54 Pg C) and 6% (1.4 out of 22 Pg C) of the total temperature-induced ΔCveg in the A2 scenario. In the Hyland model, the indirect effect of temperature was found to be more important than the direct effect of temperature on plant physiology, being responsible for approximately two-thirds (12 out of 19 Pg C) of temperature-induced ΔCveg in the A2 scenario. Analysis of stomatal conductance, evapotranspiration and soil moisture output for this model showed that the indirect effect of temperature on reduced GPP is a result of the effect of VPD on stomatal conductance, rather than changes in evapotranspiration and soil moisture status (data not shown). Separation of the direct effect of temperature on ΔCveg into the direct physiological effects on photosynthesis and respiration revealed that the two processes contributed approximately equally to temperature-driven ΔCveg in the TRIFFID model. In the LPJ model, reduced photosynthesis was the primary cause of temperature-driven ΔCveg, being responsible for approximately two-thirds of the total reduction, while increasing respiration was responsible for one-third.

Figure 7.

 Contributions of a direct effect of temperature on plant physiology (black bars) and an indirect effect via increased vapour pressure deficit (VPD; grey bars) and/or transpiration to simulated changes in Amazonian vegetation carbon (Cveg). Other environmental factors (e.g. rainfall and CO2) are held constant (simulation T-IND in Table 1).

Effect of assumptions regarding temperature dependence of plant respiration

The chosen parameterization of plant respiration dependence on temperature had significant consequences for predicted ΔCveg in all three models (Fig. 8). In the simulation where temperature was changed according to the A2 scenario and all other variables held at control values, Cveg decreased by 96% by 2100 in the model configuration where plant respiration increases exponentially with temperature. This compares to only a 14% loss in the default configuration where plant respiration is a fixed fraction of GPP. With the same forcing data, LPJ simulated a 19% loss of Cveg in the model configuration based on Lloyd & Taylor (1994), compared to a 10% loss in the ‘fixed fraction’ simulation. TRIFFID simulated a loss of 51% in Cveg in the configuration where respiration varies exponentially with temperature while only resulting in a loss of 9.5% when plant respiration was considered to be a fixed fraction of GPP.

Figure 8.

 Impact of representing plant respiration as an exponential function of temperature (broken lines) or fixed fraction of productivity (solid lines) on simulated changes in Amazonian vegetation carbon (Cveg). In these simulations, all factors are held at present-day values (simulations RESP-FIX and RESP-EXP in Table 1).

Simulation of biomass changes at Amazonian TFE experiments

All three models were found to be very insensitive to the simulated reductions in rainfall analogous to the conditions of the Amazonian TFEs (i.e. a stepped 50% reduction in rainfall, Fig. 9). The simulated TFE treatment had little effect in the Hyland simulations, while causing small reductions in biomass in LPJ (c. 5% loss of biomass after 10 yr of drought) and in TRIFFID (c. 2–3% loss of biomass). These simulated reductions are much lower than the reported losses of biomass of c. 20% following 7 yr of TFE treatment at Caxiuanã (da Costa et al., 2010) and c. 25% reduction following 4 yr of TFE treatment at Tapajós (Brando et al., 2008).

Figure 9.

 Simulated effect of a precipitation reduction similar to that of experimental drought studies in Amazonia (i.e. a 50% reduction in incident rainfall from year 0 onwards) (simulation TFE in Table 1). The graphs show the response of aboveground biomass to imposed drought at (a) Caxiuanã National Forest and (b) Tapajós National Forest.


Modelled mechanisms of Amazon rainforest ΔCveg

This study found that, in the DGVMs considered, rainfall was not the most important driver of modelled Amazonian biomass reduction. Indeed, two of the DGVMs (Hyland and TRIFFID) were very insensitive to future reductions in rainfall. In these DGVMs, even in some grid cells receiving c. 50% less rainfall in 2100, precipitation-driven losses of Cveg were minimal (< 10%), with substantial losses only occurring in grid cells receiving an annual rainfall in 2100 below a threshold value of c. 700 mm. This threshold was frequently exceeded in previous work using the fully coupled Hadley Centre GCM (Betts et al., 2004), because of much lower simulated present-day precipitation over the Amazon and the effects of land–atmosphere feedbacks. Therefore, the coupled model may be more sensitive to future precipitation reduction than our DGVM simulations.

LPJ showed the greatest sensitivity to reduced precipitation. Analysis of modelled soil moisture output of the three models revealed that soil moisture was depleted more for a given reduction in precipitation than in Hyland and TRIFFID. Thus, differences in the parameterization of soil hydraulic properties across models may partially explain the differences in sensitivity to reduced precipitation. Harris et al. (2004) found that calibrating soil parameters in the TRIFFID model with local data greatly reduced the amount of plant available water in the model and increased modelled soil moisture stress at two Amazonian sites. It is therefore feasible that use of more appropriate soil parameters (Harris et al., 2004; Fisher et al., 2008) could lead to greater drought sensitivity and increased losses of Cveg in TRIFFID. Soil hydraulic parameters such as soil water holding capacity and soil moisture content at wilting point (the point where stomata close completely under water stress) are generally derived from studies in temperate ecosystems and may be unsuitable for tropical ecosystems. For instance, field studies have shown that tropical clay soils generally have lower bulk density, higher permeability and lower available water capacity than temperate soils (Tomasella & Hodnett, 2004).

Temperature was found to be a major factor determining ΔCveg in all three models. The effect of temperature was especially pronounced in the TRIFFID model, where it was responsible for up to 12 times more loss of Cveg than precipitation, but was important across all DGVMs. Interestingly, the mechanisms determining temperature-induced losses of Cveg varied across models (Fig. 7). In Hyland, the indirect effect of temperature via increased VPD was the dominant mechanism, while in LPJ and TRIFFID the direct effects of temperature on plant physiology predominated. Moreover, in LPJ, temperature-induced decline in photosynthetic rate was more important in causing loss of Cveg than temperature-driven increases in respiration, while in TRIFFID the two processes contributed approximately equally to biomass losses.

Differences between models can be understood in terms of the underlying representations of the temperature dependences of photosynthesis and respiration in the models. In the current version, Hyland does not explicitly model plant respiration but assumes that it is a fixed fraction (0.5) of GPP. Application of this approach to the other two DGVMs greatly reduced the amount of modelled biomass loss. In turn, representing plant respiration as an exponential function of temperature in Hyland led to much increased losses of biomass in that model (Fig. 8). A common feature of the DGVMs in this study is that the response of photosynthesis to temperature is modelled with a fixed temperature optimum, and losses in Cveg would be expected to occur once the optimal temperature is surpassed. The exact location of this fixed optimum varies between models (Fig. 10), but is particularly low in LPJ, where present-day temperatures over Amazonia already exceed the photosynthetic optimum. In LPJ, the optimum leaf temperature for photosynthesis in the dominant tree PFTs over Amazonia is 21°C, compared to 28°C in TRIFFID and 32°C in Hyland.

Figure 10.

 Response of gross photosynthetic rate to leaf temperature for the dominant Amazonian tree plant functional type in three dynamic global vegetation models (DGVMs). CO2 concentrations are set at 380 ppm, irradiance is fixed at 1000 W m−2, and no water stress is assumed. All rates are expressed as a fraction of the maximum rate. Dashed lines indicate mean Amazonian air temperature from 1983 to 2002, and mean Amazonian air temperature in 2100 under the most conservative (B1) and most severe (A1FI) scenarios considered in this study.

Humidity was only found to be an important contributor to ΔCveg in the Hyland model, where it is a more important driver of ΔCveg than precipitation. In the other models used in this study, the effect of humidity on stomatal conductance was either nonexistent (LPJ) or very weak (TRIFFID). Stomatal conductance in Hyland is modelled according to the Jarvis–Stewart multiplicative model (Jarvis, 1976; Stewart, 1988), where VPD has a strong effect on stomatal conductance, independent of the effects of soil moisture and temperature per se. Rising temperatures also contribute to VPD-induced stomatal closure through the Jarvis–Stewart scheme, possibly explaining why the indirect effect of temperature is most important in Hyland.

All models showed high CO2 fertilization responses, resulting in gains in forest biomass of up to 40% in 2100 relative to present-day biomass. This CO2 effect mitigated most of the biomass loss attributable to climate change in our simulations. This is similar to the result of Lapola et al. (2009), who forced the CPTEC potential vegetation model (version 2) with climate data from an ensemble of IPCC AR4 models and reported that simulations without an included CO2 fertilization response invariably led to shifts to drier and less productive biomes.

Are modelled mechanisms of ΔCveg realistic?

Our results raise the question of whether the modelled mechanisms driving losses of forest biomass are realistic. Here we review the evidence from observational studies in the region to gain insights into the plausibility of model predictions and reach four broad conclusions.

Conclusion 1: the DGVMs used in this study may be underestimating the effect of soil moisture stress  The two throughfall exclusion experiments in Amazonia, where throughfall reaching the soil was reduced by c. 50%, indicate that the Amazon is more sensitive to decreases in precipitation than the DGVMs in this study suggest. In fact, the models used in this study were unable to capture the considerable (20–30%) reductions in standing biomass observed at both Amazonian TFE experiments following 4–7 yr of drought treatment (Fig. 9). Further evidence of the vulnerability of Amazonian rainforest to reduced precipitation comes from studies of natural drought events. Phillips et al. (2009) reported that the 2005 drought that affected large areas of Amazonia was in fact sufficient to reverse the regional long-term carbon sink. Interestingly, the HadCM3LC coupled climate–carbon cycle model appears to be able to reproduce 2005-like droughts (Cox et al., 2008), although our results suggest that its land model (MOSES-TRIFFID) may in fact be undersensitive to rainfall reductions. Other studies examining the effect of El Niño events in Amazonia also show almost immediate effects on tree growth and mortality, although these are often short-lived (Williamson et al., 2000). We note that this apparent insensitivity to drought occurs despite the absence of a number of phenomena which may confer additional resilience to seasonal and multi-year drought, such as hydraulic redistribution (Lee et al., 2005; Oliveira et al., 2005) and deep rooting depths (Williams et al., 1998; Baker et al., 2008; Poulter et al., 2009).

Conclusion 2: the DGVMs used in this study may be overestimating the effect of temperature on Amazonian carbon balance  Little experimental work has been carried out in Amazonia to assess the mid- to long-term impact of elevated temperatures on the Amazonian rainforest carbon balance. No large-scale warming experiments, analogous to the TFE experiments, exist in the Amazon. However, increasing temperature has been associated with reductions in tree growth in Central American and Malaysian rainforests (Clark et al., 2003; Feeley et al., 2007) and has also been implicated as an important causal factor of increasing tree mortality in western North America (van Mantgem et al., 2009).

The photosynthetic temperature responses in the DGVMs closely approximate the limited field measurements existing for the Amazon and other tropical regions. For example, Tribuzy (2005) and Doughty & Goulden (2008) report a decline in leaf photosynthetic rates at c. 30°C, which is in agreement with the temperature optima in Hyland and TRIFFID and with other tropical studies (Koch et al., 1994; Keller & Lerdau, 1999; Graham et al., 2003; Leakey et al., 2003). Most of these studies, however, have generally focused on short-term fluctuations in photosynthetic rate at the individual leaf level. None of the models used in this study accounts for the possibility of acclimation of both photosynthesis and respiration to rising temperatures over longer time frames and thus could be overestimating the effect of temperature on Amazonian carbon balance. There is substantial empirical evidence for both photosynthetic acclimation (Hikosaka et al., 2006; Way & Sage, 2008; Kositsup et al., 2009) and respiratory acclimation (Atkin & Tjoelker, 2003; Atkin et al., 2005) to high temperatures. The respiration dependences on temperature examined in this paper are the two most commonly used approaches whereby plant respiration is an exponential function of temperature or is directly coupled to plant productivity (i.e. the temperature dependence of respiration and photosynthesis is shared). Although respiration does increase exponentially with temperature over very short time frames (minutes to hours), this approach is almost certainty inadequate over longer time frames (months to years) (King et al., 2006). Over such time frames, there is some empirical support for constancy in the ratio of respiration to photosynthesis (Gifford, 1995; Dewar et al., 1999; Waring et al., 1998). There is also evidence that this ratio is not constant for all forest types, varying, for example, with forest age (de Lucia et al., 2007). The often-used ratio of 0.5 may not be applicable to Amazonian rainforests, which generally have higher respiration costs per unit of assimilated carbon than temperate forests, although carbon use efficiency does seem to vary across forest types (Malhi et al., 2009b). However, although homeostasis of respiration and photosynthesis is apparently maintained across a moderate range of growth temperatures, this homeostasis may break down at higher growth temperatures (Atkin et al., 2007). The temperature limit to which this homeostasis is maintained remains an open question.

Conclusion 3: modelled CO2 responses appear to concur with the results of observational studies of biomass gain but considerable uncertainties persist in the future extent of CO2-driven biomass gains  Several studies using a range of approaches (e.g. Bartak et al., 1999; Long et al., 2004; Kimball et al., 2007) have shown that high atmospheric CO2 concentrations stimulate plant growth. Rising CO2 concentrations have been invoked as an explanatory mechanism of the observed pan-Amazonian increases in biomass over recent decades (Baker et al., 2004; Phillips et al., 2008). DGVMs, including those used in this study, reproduce observed pan-tropical increases in forest biomass at the end of the 20th century fairly well (Lewis et al., 2009). However, several uncertainties remain about the magnitude of the CO2 fertilization effect in the future. Free-air CO2 enrichment (FACE) experiments, carried out under CO2 concentrations of 550 ppm, reveal an average increase in NPP of 23% across a range of sites (all temperate). Sitch et al. (2008) found close agreement between these values and those simulated by the DGVMs in this study for the same conditions. Hickler et al. (2008), however, suggest that the effects of CO2 fertilization observed in experimental studies in temperate zones may not be representative of tropical regions, as higher temperatures might lead to higher CO2 fertilization rates. However, the CO2 fertilization response may be constrained in future by other mechanisms, such as nutrient limitation, which are not included in the DGVMs used in this study. Furthermore, increased CO2 may lead to changes in species composition that favour the growth of lianas and fast-growing tree species (Korner, 2004, Phillips et al., 2004). Bunker et al. (2005) suggest that such changes could result in substantially less biomass storage by the Amazon rainforest.

Conclusion 4: a number of processes currently not included in DGVMs could affect losses of biomass from Amazonian forests under climatic change  In reality, any future changes in large-scale biomass storage in the Amazon rainforest are likely to be driven by complex interactions among processes that the current generation of DGVMs is only beginning to address or has not yet included (Table 4). At the ecological scale, these include interactions between fire (Bond et al., 2005), drought-induced tree mortality (Sperry et al., 2002; McDowell et al., 2008), land use (Levy et al., 2004), nutrient cycling (Xu-Ri & Prentice, 2008) and vegetation succession (Sato et al., 2007; Huntingford et al., 2008). Many of these drivers are currently being incorporated within DGVMs (Table 4), although important challenges such as simulation of nutrient-land cover feedbacks (e.g. Ostle et al., 2009; Senna et al., 2009) still remain. Several physiological mechanisms, besides thermal acclimation responses, presently not addressed in DGVMs could affect future projections of Amazonian ΔCveg under climate change. For example, the response of mitochondrial respiration to drought may be more complex than current models suggest, as water stress may affect plant tissues differently (Atkin & Macherel, 2009). For example, root respiration is generally found to be more inhibited that leaf respiration (Atkin & Macherel, 2009), which in some cases has even been shown to increase under drought (Metcalfe et al., 2010). However, developments in this field are still hampered by a lack of clear process understanding (Meir et al., 2008). Further uncertainty also exists in the response of respiration to increasing CO2 (e.g. Gonzalez-Meler et al., 2004). Better mechanistic understanding of these processes, coupled with their incorporation in vegetation models, will help to reduce the uncertainty surrounding predictions of the response of Amazon rainforest to climate change.

Table 4.   Nonexhaustive list of potentially important processes for modelling climate change impacts on Amazonian biomass that are not generally included in the current generation of dynamic global vegetation models (DGVMs)
ProcessScientific consensusModel development statusIncreased/decreased sensitivity of Cveg to changes in climate and atmospheric composition (relative to non-inclusion in DGVMs)Key references
  1. Cveg, carbon present in the biomass of Amazonian vegetation; GPP, gross primary productivity; NPP, net primary productivity; N, nitrogen; P, phosphorus.

Physiological/physical processes
Thermal acclimation of photosynthesisSupported by a number of studies, although greater process-based understanding is required Limited data available for tropical ecosystemsNot incorporatedDecreased: GPP maintained at higher temperaturesMedlyn et al. (2002); Hikosaka et al. (2006); Kattge & Knorr (2007)
Acclimation of photosynthetic capacity to high CO2FACE studies support acclimation of photosynthetic capacity to CO2, but this has a very limited impact on carbon gains under high CO2. No FACE experiments exist for tropical ecosystemsNot incorporatedIncreased: any acclimation should reduce carbon gain but the magnitude of the effect is uncertainLeakey et al. (2009)
Thermal acclimation of respirationSupported by a number of studies, although greater process-based understanding is required (Atkin et al., 2005)Included in a small number of ecosystem models (Wythers et al., 2005 and King et al., 2006) and in one DGVM (Atkin et al., 2008)Decreased: NPP would be maintained at higher temperaturesAtkin & Tjoelker (2003); Atkin et al. (2005); Atkin et al. (2008), King et al. (2006), Wythers et al. (2005)
Response of plant respiration components to droughtRespiration generally inhibited less than photosynthesis under drought; responses of different tissues may vary; root respiration decreases more than leaf respiration, which may even increase under droughtOften no explicit effect of drought on plant respiration components. However, there are indirect effects. First, via reduced respiratory substrate as a result of reduced photosynthesis and second via changes in leaf temperature due to changes in stomatal conductance affecting surface energy balance.Possibly increased: respiration may increase more under moisture stress than currently simulated, resulting in lower NPPFlexas et al. (2006); Meir et al. (2008); Atkin & Macherel (2009); Metcalfe et al. (2010)
Nutrient limitation on CO2 fertilizationGrowth of Amazon rainforests is thought to be limited by phosphorus (Quesada et al., 2009). However, there is little experimental evidence linking P availability and CO2 fertilization. Nitrogen believed to be nonlimiting in Amazonia apart from regrowing forestsFully interactive N cycle included in small number of models (e.g. Xu-Ri & Prentice, 2008) but P not incorporatedPossibly increased: nutrient limitation (especially P) may decrease uptake rate of CO2 in Amazonia. However, some mechanisms exist that might enhance P uptake under high CO2 (Lloyd et al., 2001)McKane et al. (1995); Lloyd et al. (2001); Xu-ri & Prentice (2008); Quesada et al. (2009); Wang & Houlton (2009)
Hydraulic redistributionObserved across range of ecosystems including AmazoniaIncluded in a small number of ecosystem models (Lee et al., 2006; Baker et al., 2008Decreased (short-term): might lead to short-term increase in water/nutrient availabilityOliveira et al. (2005); Li et al. (2005), Baker et al. (2009)
Ecological processes
Drought-induced tree mortality (especially of large canopy trees) through xylem cavitation or carbon starvationStrong evidence that mortality of large trees is the major driver of biomass reductions during drought in Amazonia but the physiological mechanism for this still requires investigationDrought mortality processes generally not mechanistically incorporatedPossibly increased: may lead to faster declines in biomass than current model approachesNepstad et al. (2007); McDowell et al. (2008); da Costa et al. (2010)
Fire dynamicsClear linkage between increased drought and fireIncorporated in some DGVMs (e.g. Thonicke et al. 2001, 2010; Venevsky et al. 2002; Bond et al., 2005)Increased: would enhance biomass lossAragão et al. (2008); Thonicke et al. (2001, 2010); Bond & Woodward (2005)
Deforestation/loggingDeforestation/logging will reduce carbon stocks directly and indirectly through synergisms with drought and fireIncorporated in small number of models (Levy et al., 2004)Increased: would enhance biomass lossLevy et al. (2004); Nepstad et al. (2008)
Plant community compositionVariable response of plant groups to climate changeCrudely incorporated via plant functional type (PFT) schemesVariable: the PFT assemblage can have a strong influence on simulated biomass (see Fisher et al., 2010).Bunker et al. (2005)Fisher et al. (2010)


This study was supported by the NERC-funded QUERCC Project, part of the QUEST Programme, the University of Edinburgh School of Geosciences and CEH Edinburgh. We would like to thank Ben Booth, Rosie Fisher, Toby Marthews, Marcel van Oijen, Ben Poulter and many others for insightful discussions during the development of this work.