Abstract

Future changes in atmospheric greenhouse gas concentrations and associated influences on climate could affect the future sustainability of tropical forests. The authors report on tropical forest projections from the new Hadley Centre Global Environmental Model version 2 Earth System configuration (HadGEM2-ES) and compare them to results from the previous generation model [third climate configuration of the Met Office Unified Model in lower resolution with carbon cycle (HadCM3LC)], which had projected near-complete dieback of the Amazon rain forest for a business as usual scenario. In contrast, HadGEM2-ES projects minimal change in Amazon forest extent. The main aim of this study is a preliminary investigation of this difference between the two models. It is found that around 40% of the difference in forest dieback projections is associated with differences in the projected change in dry-season length. Differences in control climatologies of temperature and dry-season length, projected regional warming, and the forest response to climate and CO2 also all contribute to the increased survival of forest in HadGEM2-ES. However, HadGEM2-ES does not invalidate HadCM3LC: Amazon dieback remains a possible scenario of dangerous change that requires further understanding. The authors discuss the relevance to assessments of dieback risk and future work toward narrowing uncertainty about the fate of the Amazon forest.

1. Introduction

One of the major environmental concerns associated with anthropogenic climate change is the potential for the significant loss of tropical forests, most notably the Amazon rain forest. Dieback of the Amazon forest was first reported in simulations of a Dynamic Global Vegetation Model (DGVM) driven by future climate change projected by the third climate configuration of the Met Office Unified Model (HadCM3; White et al. 1999) and has subsequently been simulated by a coupled climate–carbon cycle general circulation model (GCM), HadCM3 in lower resolution with carbon cycle (HadCM3LC; Cox et al. 2000, 2004). The phenomenon has since been found in simulations by a range of vegetation models driven by HadCM3LC climate patterns (Scholze et al. 2006; Sitch et al. 2008). Such dieback has been listed as a potential climate tipping point that may occur in the twenty-first century (Lenton et al. 2008) should emissions continue unabated. Positive land–atmosphere feedbacks could amplify the overall loss of tropical rain forests (Betts et al. 2004). For many reasons, such a major change of terrestrial biome is of obvious national and international concern.

Currently, tropical forests play a vital role by sustaining unique and diverse ecology and maintaining regional climate. Forest loss is also of concern because of the potential global feedback on climate through the release of stored carbon. There is increasing emphasis on assessing tropical forest resilience within the scientific community. This includes model studies for Amazonia in particular (Salazar et al. 2007; Scholze et al. 2006) and also large-scale, long-term observational campaigns in tropical forests across Amazonia (Phillips et al. 2009) and central Africa (Lewis et al. 2009). The 2005 and 2010 Amazon droughts showed that large-scale drying patterns were not necessarily limited to the models (Cox et al. 2008; Lewis et al. 2011; Marengo et al. 2008).

Projections of future forest stability make use of sophisticated coupled numerical modeling of the earth system. GCMs represent the most sophisticated tools for projections of future climate response to changes in atmospheric greenhouse gas and aerosol composition and land surface properties. DGVMs represent the processes that determine the growth and carbon storage of discrete plant functional types (PFTs), the competition between them, and the time scales of their response (Cramer et al. 2001). Coupled climate–vegetation simulations allow the representation of both future changes in vegetation in response to climate change (e.g., Jones et al. 2009) and also the impacts of changes in land surface on climate (Betts et al. 2004). In these simulations, CO2, rainfall, and temperature all affect the modeled vegetation distribution (e.g., Good et al. 2011). However, there is significant uncertainty in both the future climate change over the Amazon (e.g., Malhi et al. 2009) and the effects of CO2, temperature, and rainfall changes on vegetation (e.g., Galbraith et al. 2010; Lapola et al. 2009).

A later version of the Hadley Centre Global Environmental Model (version 1; HadGEM1) showed a less extreme drying of the Amazonian climate than its predecessors HadCM3 and HadCM3LC (Malhi et al. 2009). However, HadGEM1 did not include a dynamic vegetation model, so the impacts on forest cover were not simulated.

Here we present tropical forest simulations from the new Hadley Centre Global Environmental Model version 2 Earth System configuration (HadGEM2-ES). We compare the results with simulations from the previous GCM, HadCM3LC (Cox et al. 2000, 2004). The two models differ significantly in their future projections for the Amazon forest (Fig. 1). A preliminary investigation of this difference is the main aim of this study. Our principal tool for this purpose is the dry-season resilience (DSR) approach developed using HadCM3LC results (Good et al. 2011).

Fig. 1.

Broadleaf forest fraction at the end of the 1%to4x run for (top) HadCM3LC, (middle) HadGEM2-ES, and (bottom) the difference (HadGEM2-ES − HadCM3LC). The box over South America represents the area (12°S–3°N, 72°–48°W), as used by Malhi et al. (2009).

Fig. 1.

Broadleaf forest fraction at the end of the 1%to4x run for (top) HadCM3LC, (middle) HadGEM2-ES, and (bottom) the difference (HadGEM2-ES − HadCM3LC). The box over South America represents the area (12°S–3°N, 72°–48°W), as used by Malhi et al. (2009).

Processes of forest dieback are hard to diagnose in transient climate projections because of the substantial lag in forest response to climate changes (Jones et al. 2009). To simplify this, Good et al. (2011) developed a simple statistical model of the equilibrium tropical forest response to climate and CO2 in HadCM3LC (DSR method). They found that for HadCM3LC, the DSR method provided a good description of the large-scale patterns of a tropical forest in equilibrium across a large range of climates (from present day through 2100). Hence, the DSR method can be used to predict the equilibrium tropical forest state that HadCM3LC would simulate under alternative conditions of climate or CO2, or to understand existing HadCM3LC forest simulations. A secondary aim of the current study is to explore the application of the DSR method further, in a multi-GCM context.

2. Data and methods

a. HadCM3LC and HadGEM2-ES

1) HadCM3LC

HadCM3LC was one of the coupled climate–carbon cycle GCMs used in the Coupled Climate–Carbon Cycle Model Intercomparison Project (C4MIP; Friedlingstein et al. 2006) and includes the Met Office Surface Exchange Scheme, version 2 (MOSES2), the land surface model (Essery et al. 2003), coupled to a dynamic vegetation component, Top-down Representation of Interactive Foliage and Flora Including Dynamics (TRIFFID; Cox 2001). HadCM3LC has a spatial resolution of 2.5° latitude by 3.75° longitude in both the atmosphere and ocean components, with 19 and 20 vertical levels, respectively. The low-resolution ocean component necessitates the use of heat and freshwater flux corrections to prevent unacceptable climate drift.

MOSES2 has a four-layer soil model with a total soil depth of 3 m. TRIFFID represents five plant functional types: broadleaf trees, needleleaf trees, shrubs, and C3 and C4 grasses. This represents a substantial simplification compared to the large range of species in the real world, including both evergreen and dry deciduous broadleaf trees in the tropics. The regional extent of these vegetation types is determined by the net carbon balance of each PFT (accounting for the productivity and turnover of carbon) and competition between them. The Lotka–Volterra equations in TRIFFID are known to overestimate (underestimate) the fractional coverage of the dominant (subdominant) plant functional type (Arora and Boer 2006). Also, the model lacks fire disturbances, leading to too much forest cover in disturbance-dominated savanna regions. The vegetation responds to a wide array of environmental drivers, such as moisture availability, light, temperature, and carbon dioxide fertilization. The dependence of plant growth on nutrients (Thornton et al. 2009) or direct and diffuse light (Mercado et al. 2009) and the effects of fire (Malhi et al. 2009) or other disturbances are likely to be important but are not yet represented in this model.

2) HadGEM2-ES

The second model is the Hadley Centre Global Environmental Model version 2 Earth System configuration (Collins et al. 2011; Martin et al. 2011). Table 1 lists the main changes between HadCM3LC and HadGEM2-ES that are most likely to affect the regional Amazonian climate and/or the response of vegetation to climate. The largest changes are to the atmospheric model, which features a number of modifications that could strongly influence tropical climate (Martin et al. 2006). In particular, HadGEM2 features a new dynamical core, increased spatial resolution, a new boundary layer scheme, and major changes to the convection, cloud, and aerosol schemes (Johns et al. 2006; Martin et al. 2011; Collins et al. 2011). Taken together, these developments make the atmospheric component of HadGEM2 very different from its predecessor (Martin et al. 2006). HadGEM2 has a spatial resolution of 1.25° × 1.875° in the atmospheric component and 38 vertical levels, and 1° resolution in the ocean (increasing to ⅓° near the equator) and 40 vertical levels. Collins et al. (2011) show that the terrestrial carbon cycle in HadGEM2-ES performs better than in HadCM3LC despite HadGEM2-ES no longer requiring flux adjustments to correct for climate biases.

Table 1.

Major relevant changes from HadCM3LC to HadGEM2-ES.

Major relevant changes from HadCM3LC to HadGEM2-ES.
Major relevant changes from HadCM3LC to HadGEM2-ES.

HadGEM2-ES uses the same core land surface and vegetation schemes as HadCM3LC, but with some modifications. HadGEM2-ES features an improved representation of plant autotrophic respiration. In HadCM3LC, leaf respiration was represented as increasing exponentially with temperature. This meant that respiration could increase indefinitely with temperature leading to perhaps too strong a temperature sensitivity. Such an approach has been evaluated in laboratory work for short-term response but may not be appropriate for long-term vegetation response, which is thought to acclimate to higher temperatures (Atkin et al. 2005). In HadGEM2-ES, plant respiration has a temperature optimum above which respiration declines in line with photosynthesis (Clark et al. 2011).

The hydrology in HadGEM2-ES has been significantly improved. In the land surface model in HadCM3LC, free drainage occurs out of the bottom soil layer. As such there tends to be minimal vertical gradients in soil moisture. For HadGEM2-ES, a new large-scale hydrology (LSH) scheme (Clark and Gedney 2008; Gedney and Cox 2003) has been implemented. This scheme was developed to incorporate a more realistic representation of topographically driven surface runoff. It includes an additional deep soil layer within which the saturated conductivity of the soil reduces exponentially with depth. This means that drainage is impeded and instead the runoff is lost as lateral flow as an effective water table builds up. This results in an increase in soil moisture with depth. The new hydrology module showed substantial improvements when compared with observed streamflow in three catchments in southern France (Clark and Gedney 2008) and helped reduce surface temperature biases (Martin et al. 2011).

3) Transient and equilibrium vegetation configurations

To explore relationships in these two GCMs between forest cover and climate and CO2 fertilization, we make use of two configurations of the dynamic vegetation scheme (transient and equilibrium). In their standard configuration, both models simulate the transient (time evolving) forest extent in response to changes in climate variables and simulated vegetation productivity. A second configuration involves using the land surface scheme in “equilibrium” mode (for details see Cox 2001; supplementary information in Jones et al. 2009). This mode allows the forest to fully adjust to the prevailing climate by integrating the vegetation scheme for 10 000 yr for every 5 yr of the climate model run. Plots of regional vegetation fraction against time (supplementary information in Jones et al. 2009) demonstrate that 30-yr integrations in this mode are sufficient to achieve equilibrium. This simulates the “sustainable forest fraction.”

b. Experimental design

We use results from both transient and equilibrium simulations (unless otherwise stated, 30-yr means are used). Transient projections are first used to demonstrate the principal difference between HadCM3LC and HadGEM2-ES. Then equilibrium vegetation simulations are used to quantify the forest response to climate and CO2 in each model (the dry-season resilience method; section 2a). We also use a control integration with forcings fixed at preindustrial levels. The 230-yr control integration was initialized following a thorough spinup procedure (see Collins et al. 2011), which ensures that the long-time-scale components (including the ocean and vegetation) reach a stable state prior to fully coupled simulations.

The transient projections are initialized from the control simulation and are forced by CO2 increasing at 1% per year for 140 yr (reaching 4 times preindustrial levels, denoted 1%to4x). This simulation demonstrates a major difference in response between the two models, which needs to be explored before including extra complexity from non-CO2 forcing agents. For HadCM3LC we also repeated the 1%to4x integration, but with vegetation fixed at preindustrial levels (denoted 1%to4x-fixveg). This is used to eliminate the feedback of vegetation changes on regional climate.

Pairs of 30-yr equilibrium vegetation simulations (Table 2) were performed with HadGEM2-ES. Each pair was initialized from a point in the 1%to4x projection. In the first 30-yr simulation, the CO2 concentration was kept constant while the vegetation equilibrated with the prevailing climate and CO2. In the second simulation, the model was run for a further 30 yr, but with the CO2 concentration seen by the vegetation scheme set to preindustrial levels (i.e., different from that seen by the climate model). These pairs of simulations are important for distinguishing the effects of temperature and CO2 concentration on vegetation (otherwise, changes in temperature and CO2 are highly correlated). Results where both climate and vegetation schemes see 2 and 3 times preindustrial CO2 are not presented, but are used in the dry-season resilience derivation (section 2d).

Table 2.

List of equilibrium vegetation simulations used in the DSR method.

List of equilibrium vegetation simulations used in the DSR method.
List of equilibrium vegetation simulations used in the DSR method.

The equilibrium vegetation simulations with HadCM3LC (Table 2) are those described by Good et al. (2011). The philosophy is the same as for HadGEM2-ES, but with differences in the details. The HadCM3LC simulations were initialized from a slightly different CO2 scenario projection [based on the Special Report on Emissions Scenarios (SRES) A2 scenario] and are evaluated at different CO2 concentrations than the equivalent HadGEM2-ES. These differences do not affect our conclusions, since the dry-season resilience method, once tested (section 2e), allows us to compare equilibrium vegetation results at different climate and CO2 conditions (in fact, the method works by combining results from a very broad range of climate and CO2 conditions).

c. Observed data and CMIP3 multimodel results

Monthly gridded observations of surface temperature and precipitation are taken from the Climate Research Unit (CRU) version 2.1 dataset (Mitchell and Jones 2005) and averaged over 1950–2002.

To give context to the observational validation, we use preindustrial control simulations from multimodel results made available as part of the Climate Model Intercomparison Project (CMIP3). We show results from the following 17 models: BCCR-BCM2.0, CCSM3, CGCM3.1(T47), CGCM3.1(T63), CNRM-CM3, ECHAM5/MPI-OM, ECHO-G, FGOALS-g1.0, GFDL CM2.0, GFDL CM2.1, GISS-EH, GISS-ER, INGV-SXG, INM-CM3.0, MIROC3.2(hires), MIROC3.2(medres), and MRI CGCM2.3.2. Results are averaged over the entire control run for each model to remove noise from internal variability.

d. DSR method

The dry-season resilience method (Good et al. 2011) quantifies approximately the equilibrium tropical forest response to local surface climate and CO2 concentration in a given climate model. The DSR method is a simple linear function of local temperature, dry-season length, and CO2 concentration. For any given conditions of climate and CO2, this function predicts whether or not tropical forest is sustainable in the corresponding climate model.

The DSR method follows four main steps: equilibrium vegetation simulations, quantifying the climate and CO2 conditions, “bioclimatic zone plots,” and derivation of coefficients. These four steps are detailed below.

Equilibrium vegetation simulations (described above) give the sustainable forest fraction in the model under different climate–CO2 conditions. This is a key step, allowing direct links to be made between tropical forest and climate–CO2 drivers (by eliminating the complication of lagged forest responses to transient climate change).

We analyze the broadleaf forest fraction at each location between latitudes 20°S and 20°N.

To quantify the climate and CO2 drivers at each location, we use CO2 concentration, annual mean temperature, and annual dry-season length (the number of months with less than 100 mm of precipitation) (as defined and discussed by Good et al. 2011).

Bioclimatic zone plots are used to test the applicability of the DSR method (e.g., see Fig. 2a): for HadCM3LC, in plots of dry-season length against temperature, the equilibrium tropical forest fraction falls into two bioclimatic zones of low (high) forest fraction, characterized by high (low) temperature and a long (short) dry season (Good et al. 2011). The boundary between the two zones is approximately linear in these plots. This motivated the formulation of the linear DSR statistical model,

 
formula

where αT, αCO2, and c are constants with units of months (°C)−1, months ppmv−1, and months, respectively. Dry-season length (months), temperature T (°C), and CO2 concentration (ppmv) are the environmental drivers. DSR has units of months [because of the form of Eq. (1)].

Fig. 2.

Bioclimatic zone plots for (a),(c) HadCM3LC and (b),(d) HadGEM2-ES. Colors represent mean broadleaf fraction (brown: BLF < 0.2; orange: 0.2 > BLF > 0.4; light green: 0.4 < BLF < 0.6; green: BLF > 0.6; white: <2 data points in a pixel). Dashed line represents contour of DSR_HG2 = 0; dotted line represents contour of DSR_CM3 = 0. Results are included from all equilibrium vegetation runs where the CO2 concentration seen by the vegetation scheme is within 10 ppm of that in the title of the panel [(a) has two runs, (b) has four runs, and (c),(d) have one run]. Points are more sparse in (a),(c) because HadCM3LC has fewer grid points (lower spatial resolution).

Fig. 2.

Bioclimatic zone plots for (a),(c) HadCM3LC and (b),(d) HadGEM2-ES. Colors represent mean broadleaf fraction (brown: BLF < 0.2; orange: 0.2 > BLF > 0.4; light green: 0.4 < BLF < 0.6; green: BLF > 0.6; white: <2 data points in a pixel). Dashed line represents contour of DSR_HG2 = 0; dotted line represents contour of DSR_CM3 = 0. Results are included from all equilibrium vegetation runs where the CO2 concentration seen by the vegetation scheme is within 10 ppm of that in the title of the panel [(a) has two runs, (b) has four runs, and (c),(d) have one run]. Points are more sparse in (a),(c) because HadCM3LC has fewer grid points (lower spatial resolution).

For any combination of environmental conditions (of dry-season length, temperature, and CO2 concentration), DSR predicts which of the two bioclimatic zones these conditions reside in (i.e., whether high or low forest fraction is sustainable). Positive (negative) DSR suggests that tropical forest is (is not) sustainable in the model. Further, DSR predicts how far these environmental conditions are from the boundary between the two zones. That is, how much would the conditions have to change for the equilibrium forest state to change? DSR brings together the effects of changes in dry-season length, temperature, and CO2. The contour DSR = 0 is an estimate of the boundary between the two zones (see the dotted line in Fig. 2a).

The coefficients in Eq. (1) are derived using the same optimization procedure as Good et al. (2011), using data from all equilibrium vegetation simulations from the same climate model. This involved numerical maximization of the following quantity involving the broadleaf fraction (BLF): (BLF averaged over locations with DSR > 0) − (BLF averaged over locations with DSR < 0). This method tends to choose parameters such that the contour DSR = 0 lies along the boundary between the bioclimatic zones of high and low forest cover.

While DSR is derived from equilibrium vegetation simulations, it may be used to analyze projections of rainfall and temperature from transient simulations and thus to estimate the effect on long-term forest sustainability.

e. Can we use DSR to compare the two models?

Before using the DSR method (Good et al. 2011) to understand the different Amazon forest response in HadCM3LC and HadGEM2-ES, we address two questions. First, does the DSR approach (developed for HadCM3LC) also work in HadGEM2-ES? And second, does it make sense to compare values of DSR from the two models? We make use of equilibrium vegetation results from the entire tropics (not just the Amazon).

Figures 2b,d show that the basic DSR concept is applicable to HadGEM2-ES. This figure shows that in equilibrium vegetation simulations with HadGEM2-ES, bioclimatic zones of high (low) tropical forest fraction exist, which are characterized by low (high) dry-season temperature and length. The boundary between the two zones is approximately linear in these plots of dry-season length versus temperature. The location of the boundary depends also on the CO2 concentration seen by the vegetation (cf. Figs. 2a,c and 2b,d). Hence it makes sense to define DSR, which quantifies how far any given set of climate and CO2 conditions is from this boundary. Figure 2b combines results from four simulations (those where the vegetation sees 286-ppm CO2). The individual simulations each show similar patterns (not shown).

We derive a DSR function for HadGEM2-ES (denoted DSR_HG2) using the optimization procedure of Good et al. (2011). All of the HadGEM2-ES equilibrium vegetation simulation results were used in the optimization (Table 2). The resulting formula is

 
formula

For comparison, the DSR function for HadCM3LC (Good et al. 2011) is

 
formula

There are significant differences in the coefficients of DSR_HG2 and DSR_CM3, which we explore in the next section.

We first examine whether it makes sense to compare values of DSR from the two models. Figure 3 shows that the relationship between DSR_HG2 and equilibrium tropical broadleaf fraction is approximately the same as that found for HadCM3LC. The transition from high to low forest fraction (the point at which the red and green lines cross) is centered at almost the same value of DSR in both models (to within 0.25 months). In both models, the proportion of cells with high broadleaf fraction (the green lines) is small for DSR < −1.5 months. Similarly, in both models, the proportion with low broadleaf fraction (the red lines) is small for DSR > 1.5 months. That is, changes in DSR have approximately the same meaning in both models (although DSR is a different function of the climate and CO2 drivers). This simplifies the task of comparing HadGEM2 with HadCM3LC: from now on we just compare values of DSR from the two models.

Fig. 3.

The relationship between DSR (x axis) and tropical BLF (y axis). All results from all equilibrium vegetation simulations are included for (left) HadGEM2-ES and (right) HadCM3LC. The x axis for HadGEM2-ES (HadCM3LC) is calculated using the DSR_HG2 (DSR_CM3) function. Each symbol represents values for one tropical location in one simulation. Colored lines represent the fraction of grid cells with BLF < 0.05 (red), 0.05 < BLF < 0.4 (blue), and BLF > 0.4 (green).

Fig. 3.

The relationship between DSR (x axis) and tropical BLF (y axis). All results from all equilibrium vegetation simulations are included for (left) HadGEM2-ES and (right) HadCM3LC. The x axis for HadGEM2-ES (HadCM3LC) is calculated using the DSR_HG2 (DSR_CM3) function. Each symbol represents values for one tropical location in one simulation. Colored lines represent the fraction of grid cells with BLF < 0.05 (red), 0.05 < BLF < 0.4 (blue), and BLF > 0.4 (green).

Figure 3 also gives an indication of the uncertainty in DSR due to the effects of factors other than dry-season length and annual mean temperature (such as other changes to the rainfall or temperature seasonal cycles, or incident radiation). This may be inferred from the width (on the x axis) of the transition from high to low forest fraction (if DSR was a perfect predictor of forest fraction, this transition would have zero width). This uncertainty is approximately ±1.5 months.

3. Results

a. Tropical forest simulations from HadGEM2-ES and HadCM3LC

In the preindustrial control, the two GCMs have rather similar patterns of tropical forest (both in the Amazon and elsewhere; not shown). Their projections under the 1%to4x experiment (Fig. 1), outside of the Amazon, are also similar, but the two projections for the Amazon are very different. HadCM3LC shows the well-known Amazon dieback (Cox et al. 2004), while HadGEM2-ES shows minimal change. Here we report on a preliminary investigation of this difference.

This difference in Amazon forest projections could be due to one or both of two top-level factors: 1) differences in vegetation response to imposed surface climate and CO2, and 2) differences in projected Amazon surface climate (CO2 concentrations are prescribed in the 1%to4x experiment). Differences in projected surface climate may be further separated into two forms: differences in the control climate state and differences in the projected climate change.

The simplest explanation might be found if the projected global mean temperature was lower (implying weaker regional impacts) in HadGEM2-ES than in HadCM3LC. However, at the end of the 1%to4x experiment, HadGEM2-ES is slightly warmer (17.8°C for HadCM3LC and 18.4°C for HadGEM2-ES). These 30-yr mean differences are significant: the standard deviation of internal variability in 30-yr global mean temperatures is about 0.02°C. In the preindustrial control the two GCMs have very similar global temperatures (14.0°C for HadCM3LC and 13.7°C for HadGEM2-ES). This corresponds to 24% greater warming in HadGEM2-ES. The transient climate response (global warming at the time of CO2 doubling) is 1.9°C for HadCM3LC and 2.4°C for HadGEM2-ES.

Below we investigate the differences in more detail using the dry-season resilience approach (section 2d) as a comparison tool. Having first established that DSR is appropriate for this use, we examine GCM differences of the first type—how tropical vegetation responds to climate and CO2. We then compare the relative importance of the second form of difference in the surface climate over the Amazon.

b. How does the tropical vegetation response to climate and CO2 drivers differ between the two models?

The differences between DSR_HG2 and DSR_CM3 imply differences in how the land surface and vegetation schemes in the two GCMs respond to given inputs of rainfall, temperature, and CO2. We explore this, again using equilibrium vegetation results from the entire tropics. Figure 2 demonstrates the differing equilibrium forest response to climate and CO2 drivers, via contours of DSR_HG2 = 0 (dashed line) and DSR_CM3 = 0 (dotted line). The difference in the position of these contours between the top and bottom panels for each model shows the effect of changing the CO2 concentration seen by the vegetation scheme. The first key point of this figure is that for HadGEM2-ES (right column), tropical forest fraction is mostly high below the dashed line (i.e., for DSR_HG2 > 0) and low above it (DSR_HG2 < 0). This is not true for the dotted line (the contour of DSR_CM3 = 0). Hence the second key point of this figure is that there is a clear area in these panels where HadGEM2-ES sustains tropical forest for a set of environmental conditions where we would not expect HadCM3LC to do so (below the dashed line, where DSR_HG2 > 0, but above the dotted line, where DSR_CM3 < 0).

Figures 2a,c show comparable results for HadCM3LC. The boundary between high and low forest fraction is marked by the dotted line (DSR_CM3 = 0), and in the area between the dashed and dotted lines, forest is typically not sustained in HadCM3LC.

This difference in tropical vegetation response to the same environmental conditions provides a partial answer to our top-level question of why Amazon dieback is found in HadCM3LC but not HadGEM2-ES. Dieback is expected for less severe surface climate conditions in HadCM3LC than in HadGEM2-ES. The difference in forest response is minimal at the coolest locations but increases with temperature.

The levels of CO2 seen by the vegetation scheme are slightly different for HadCM3LC and HadGEM2-ES. For example, in Figs. 2a,c, the HadCM3LC results are for vegetation seeing year-2000 levels of CO2, while the HadGEM2-ES results are for vegetation seeing preindustrial levels of CO2. This does not affect our conclusions (although the dashed and dotted lines are in slightly different locations in Figs. 2a,c versus Figs. 2b,d because of the different CO2 levels).

We discuss the different vegetation responses further in section 4. In the next subsection we focus on results from the Amazon region, bringing in the other factor—model differences in surface climate.

c. What is the relative importance of differences in tropical forest response and in surface climate over the Amazon?

We now return to our main focus: the different Amazon forest response in the 1%to4x transient climate projections. To quantify the influence of the projected climate state on the vegetation distribution, we use the HadCM3LC 1%to4x-fixveg projection. This eliminates the feedback on climate from the Amazon forest dieback in HadCM3LC. We did not run a 1%to4x-fixveg projection for HadGEM2-ES because the Amazon forest change in the 1%to4x experiment is minimal in this model. Even the substantial dieback in HadCM3LC only increases the dry-season length by around 0.9 months, so corresponding effects in HadGEM2-ES will be small.

We make use of two facts established above (Fig. 2): that contours of DSR = 0 mark the boundary between conditions where tropical forest is sustainable and where it is not, and that DSR_HG2 and DSR_CM3 are different functions. As an analysis region (rectangle in Fig. 1) we combine the two regions of Malhi et al. (2009).

Figure 4 brings together both main forms of GCM differences: differences in surface climate and differences in tropical forest response to the given climate and CO2 conditions in the Amazon. The difference between the dashed and dotted lines represents how the GCMs have different tropical forest responses (at equilibrium) to the given climate and CO2 conditions (as in Fig. 2). The differences between Figs. 4a,c and 4b,d represent differences in climate between the two models.

Fig. 4.

Visualizing the various differences in DSR over the Amazon. Position of Amazon cells on biozone map for (a),(b) control climate and (c),(d) mean climate over last 30 yr of 1%to4x from (a),(c) HadCM3LC (1%to4x-fixveg simulation) and (b),(d) HadGEM2-ES. Lines represent contours of DSR = 0 for the LS–DGVM of HadCM3LC (dotted) and HadGEM2-ES (dashed).

Fig. 4.

Visualizing the various differences in DSR over the Amazon. Position of Amazon cells on biozone map for (a),(b) control climate and (c),(d) mean climate over last 30 yr of 1%to4x from (a),(c) HadCM3LC (1%to4x-fixveg simulation) and (b),(d) HadGEM2-ES. Lines represent contours of DSR = 0 for the LS–DGVM of HadCM3LC (dotted) and HadGEM2-ES (dashed).

In the preindustrial control climate (Figs. 4a,b), the two GCMs are rather similar in that forest is sustainable (below both the dashed and dotted lines) for most of the Amazon region (88% and 99% for HadCM3LC and HadGEM2-ES, respectively). This is consistent with the fact that the two GCMs have similar forest distributions in the preindustrial simulations.

At the end of the 1%to4x run (Figs. 4c,d), however, the differences are substantial. In HadCM3LC (Fig. 4c), forest is unsustainable across most of the region (82% of the crosses are above the dotted line). For HadGEM2-ES (Fig. 4d), forest is sustainable across much of the region (93% of the crosses are below the dashed line). We can quantify the overall GCM difference by the regional mean difference in DSR. This accounts for differences in both surface climate and vegetation response to climate and CO2. It is calculated as the regional mean of DSR_HG2 for the HadGEM2-ES climate minus the regional mean of DSR_CM3 for the HadCM3LC climate. Values for the control and future states of both models are plotted in Fig. 5. This shows that the difference in DSR between the two models at the end of the 1%to4x run (the two yellow lines in Fig. 5) is large: about 7 months. That is, the combined differences between the two projections are approximately equivalent to a 7-month difference in dry-season length.

Fig. 5.

As in Fig. 3(left), but with vertical lines marking the Amazon regional mean values of DSR for the control (purple) and end of the 1%to4x runs (yellow) for HadCM3LC (1%to4x-fixveg simulation; dotted) and HadGEM2-ES (dashed).

Fig. 5.

As in Fig. 3(left), but with vertical lines marking the Amazon regional mean values of DSR for the control (purple) and end of the 1%to4x runs (yellow) for HadCM3LC (1%to4x-fixveg simulation; dotted) and HadGEM2-ES (dashed).

For the HadCM3LC climate (Fig. 4c), 58% of the crosses are also above the dashed line. This means that substantial Amazon dieback would still be expected even if the slightly more resilient forest in HadGEM2-ES experienced the HadCM3LC climate. For the HadGEM2-ES climate (Fig. 4d), the reverse is true: forest is likely to be sustainable across much of the Amazon region even if this climate was experienced by the more sensitive HadCM3LC forest (86% of the crosses are below the dotted line). In summary, while differences in forest sensitivity (Fig. 4) contribute to the differing Amazon forest projections in the two GCMs, the major difference lies in the surface climate projections.

Differences in control state and climate change of temperature and dry-season length

The GCM differences in surface climate could be separated into four components: the control climate temperature, the control dry-season length, the change in temperature (from the control to the end of the 1%to4x run), and the change in dry-season length.

We illustrate (Fig. 6) how these four GCM differences (and the change in land surface/vegetation scheme) contributed to increased forest resilience in HadGEM2-ES. In Fig. 6, the symbols (representing the 1%to4x climate at each Amazon location) are progressively shifted from the HadCM3LC climate state (Fig. 6a) to the HadGEM2-ES climate state (Fig. 6e). (The HadCM3LC climate is interpolated linearly to the HadGEM2-ES spatial grid.) These do not represent new integrations: each panel shows the same number of symbols, plotted at different dry-season length–temperature values. Moving from Fig. 6a to 6b illustrates the GCM differences in control temperature. The symbols in Fig. 6b are plotted at temperatures given by the HadCM3LC 1%to4x temperature plus the (HadGEM2-ES − HadCM3LC) difference in control temperature. That is, the differences in temperature between Figs. 6b,a are the same as the differences in control climate temperature between HadGEM2-ES and HadCM3LC. These differences are rather small: most of the crosses remain above the contour DSR_CM3 = 0 (82% in Fig. 6a and 73% in Fig. 6b). Hence, differences in control climate temperature do not seem to be very important in explaining the difference in Amazon forest projections. Similarly, moving from Fig. 6b to 6c shows the GCM differences in control dry-season length. This causes a slightly larger change, moving the set of crosses closer to or past the dotted line (indicating greater forest resilience, with 62% of crosses above the line). Once we include the increment from the GCM differences in temperature change (Fig. 6d), the climate state has changed quite significantly compared to Fig. 6a (42% of crosses above the line). However, the largest individual GCM difference lies in the projected change in dry-season length (cf. Figs. 6e,d; only 14% of crosses above the line in Fig. 6e). Finally, the difference in forest response to climate and CO2 (cf. Figs. 6f,e) represents a further increment in forest resilience in HadGEM2-ES compared with HadCM3LC (leaving 7% of crosses above the line).

Fig. 6.

Visualizing the effect of gradually changing from HadCM3LC climate (1%to4x-fixveg simulation) to HadGEM2-ES 1%to4x climate (and then changing the vegetation response to climate). (a) Climate (symbols) and DSR = 0 contour (dotted line) as in HadCM3LC; (b) as in (a), but symbols are shifted on the x axis by the (HadGEM2-ES − HadCM3LC) difference in control temperature; (c) as in (b), but shifted vertically by the (HadGEM2-ES − HadCM3LC) difference in control dry-season length; (d) as in (c), but symbols plotted at HadGEM2-ES 1%to4x temperature; (e) as in (d), but plotted at HadGEM2-ES 1%to4x dry-season length; and (f) both climate and DSR = 0 (dashed line) as in HadGEM2-ES.

Fig. 6.

Visualizing the effect of gradually changing from HadCM3LC climate (1%to4x-fixveg simulation) to HadGEM2-ES 1%to4x climate (and then changing the vegetation response to climate). (a) Climate (symbols) and DSR = 0 contour (dotted line) as in HadCM3LC; (b) as in (a), but symbols are shifted on the x axis by the (HadGEM2-ES − HadCM3LC) difference in control temperature; (c) as in (b), but shifted vertically by the (HadGEM2-ES − HadCM3LC) difference in control dry-season length; (d) as in (c), but symbols plotted at HadGEM2-ES 1%to4x temperature; (e) as in (d), but plotted at HadGEM2-ES 1%to4x dry-season length; and (f) both climate and DSR = 0 (dashed line) as in HadGEM2-ES.

A regional mean perspective is given in Fig. 7. This bar chart represents the regional mean differences in DSR between HadCM3LC and HadGEM2-ES, including both differences in climate and differences between DSR_CM3 and DSR_HG2. The total difference is about 7 months. Each segment of the bar shows how DSR is affected by one of the five forms of model difference explored in Fig. 6 (control temperature and dry-season length, changes in temperature and dry-season length, and differences between DSR_CM3 and DSR_HG2). The relative contributions of these differences will be sensitive to various factors, including the choice of geographical region and the projection emissions scenario. A slightly different balance is also obtained if the components are switched to a different order (because the vegetation has differing temperature sensitivities in the different GCMs). However, in either case we can conclude that differences in the projected change in dry-season length account for around 40% of the total difference (the length of the ΔDSL segment in Fig. 7 as a fraction of the total length of the bar), while the other components all have the same sign and so all contribute to the increased resilience in HadGEM2-ES. There is some uncertainty in this 40% figure, as factors other than dry-season length and annual mean temperature could affect the total difference in forest resilience between the two models. These other factors cause an overall uncertainty in DSR at any given location of around 1.5 months (Fig. 3), giving an uncertainty of around 2 months for a difference between two independent DSR values. Accounting for this implies a range of around 30%–55% in the relative contribution of dry-season length (this range is conservative, since the uncertainty in regional mean DSR will be less than that at individual grid locations). Choosing different “Amazon” regions could change the GCM difference in dry-season length change by up to ∼1 month, which implies a similar range: 26%–53%.

Fig. 7.

Regional mean differences in DSR between HadCM3LC (1%to4x-fixveg simulation) and HadGEM2-ES. Each segment of the bar shows how DSR is affected by one of the five forms of model difference explored in Fig. 6. The “DSR” segment represents the effect of changing the land surface and vegetation schemes.

Fig. 7.

Regional mean differences in DSR between HadCM3LC (1%to4x-fixveg simulation) and HadGEM2-ES. Each segment of the bar shows how DSR is affected by one of the five forms of model difference explored in Fig. 6. The “DSR” segment represents the effect of changing the land surface and vegetation schemes.

d. Observational comparison

Having used plots of dry-season length against temperature to explore the top-level differences between HadGEM2-ES and HadCM3LC, we now use similar plots for a brief observational assessment of the two models (Fig. 8). The aim is to examine whether HadGEM2-ES or HadCM3LC can be demonstrated to be significantly more or less reliable. We provide context using results from 17 CMIP3 GCMs.

Fig. 8.

Amazon regional mean values of dry-season length and temperature. Monthly gridded observations of surface temperature and precipitation taken from the CRU version 2.1 dataset (Mitchell and Jones 2005) and averaged over 1950–2002 (blue asterisk); HadCM3LC (red square); HadGEM2-ES (orange square); and CMIP3 models (excluding HadCM3, HadGEM1; black diamonds). Uncertainty from internal variability is less than the symbol size. Model data are averaged over the whole of each preindustrial control experiment. Observed temperature has a small adjustment to make it comparable with the GCM preindustrial control results (given by the CMIP3 ensemble mean temperature difference between the control and historical simulations for the observation period). Simulated historical changes in dry-season length are negligible compared to the spread across different models.

Fig. 8.

Amazon regional mean values of dry-season length and temperature. Monthly gridded observations of surface temperature and precipitation taken from the CRU version 2.1 dataset (Mitchell and Jones 2005) and averaged over 1950–2002 (blue asterisk); HadCM3LC (red square); HadGEM2-ES (orange square); and CMIP3 models (excluding HadCM3, HadGEM1; black diamonds). Uncertainty from internal variability is less than the symbol size. Model data are averaged over the whole of each preindustrial control experiment. Observed temperature has a small adjustment to make it comparable with the GCM preindustrial control results (given by the CMIP3 ensemble mean temperature difference between the control and historical simulations for the observation period). Simulated historical changes in dry-season length are negligible compared to the spread across different models.

The first key point from this plot is that HadGEM2-ES provides very good historical simulations of these two climate variables. The bias with respect to the observations is small compared to the distance from the estimated thresholds of forest resilience (contours DSR_CM3 = 0 and DSR_HG2 = 0). It is also small compared to the range of results from CMIP3. HadGEM2-ES represents an improvement over HadCM3LC in these terms, although this difference between the two models is small compared to the CMIP3 range, and HadCM3LC also has a relatively realistic control climate state.

The range of control dry-season length simulations across the CMIP3 ensemble is very large—more than 5 months—which is similar to the change projected by HadCM3LC. There is also a clear ensemble bias toward long dry-season lengths in this region. Substantial uncertainty and bias in this ensemble have previously been found (Malhi et al. 2009) using a different metric—dry-season water deficit. These results imply a need to be cautious in using this generation of GCMs for assessing future rainfall changes, although it has been suggested that there is some consistency in projected increases in the dry-season water deficit for eastern Amazonia (Malhi et al. 2009).

4. Discussion

We now discuss the model differences in a bit more detail. Figure 9 shows spatial maps of dry-season length and temperature over tropical South America, under the preindustrial control and the end of the 1%to4x experiment. These maps show both similarities and differences between the two GCMs. In the preindustrial control, the two GCMs have similar patterns of both temperature and dry-season length. They also show some similarities in the future projection. Both models show an expansion of the northeastern region of long dry-season length, consistent with the relatively robust drying signal in eastern Amazonia found in CMIP3 models (Malhi et al. 2009), and both show high temperatures toward the Amazon interior, with temperatures exceeding 30°C across most of our analysis region. In the control, however, HadGEM2-ES maintains the area of short dry-season length (<4 months) farther toward the east near the equator and is 1°–2°C warmer across much of this region. Large differences in dry-season length (>4 months) at the end of the 1%to4x experiment are focused in an approximately latitudinal band centered just south of the equator.

Fig. 9.

Surface climate over tropical South America in (left) HadCM3LC, (middle) HadGEM2-ES, and (right) the difference. (top rows) Dry-season length and temperature for the control experiment, respectively. (bottom rows) Dry-season length and temperature averaged over the last 30 yr of the 1%to4x experiment, respectively.

Fig. 9.

Surface climate over tropical South America in (left) HadCM3LC, (middle) HadGEM2-ES, and (right) the difference. (top rows) Dry-season length and temperature for the control experiment, respectively. (bottom rows) Dry-season length and temperature averaged over the last 30 yr of the 1%to4x experiment, respectively.

Model differences in dry-season precipitation may largely be attributed to differences in sea surface temperature (SST) changes in the tropical Atlantic. Both models exhibit a north–south dipole pattern of SST changes (Figs. 10a,b), which could affect dry-season precipitation over the south Amazon (Good et al. 2008). However, the sign of change in the north–south SST gradient is opposite in the two models, with a much larger change in HadCM3LC. We use an SST gradient index and its relationship to the June–August precipi tation change over the southern Amazon basin (Fig. 10c) previously established by Good et al. (2008). We find (Fig. 10d) that the SST changes in the tropical Atlantic can explain quantitatively the differences between HadCM3LC (fixveg) and HadGEM2-ES in dry-season precipitation change over the south Amazon River basin [a slightly different geographical region and measure of precipitation than elsewhere in the current study, for consistency with Good et al. (2008)]. The dry-season precipitation change over the region in Fig. 1 is slightly negative for HadGEM2-ES, as found for dry-season length. Further, the main differences in future dry-season length just south of the equator (Fig. 9) fall in a region identified as sensitive to changes in the tropical Atlantic SST gradient (Good et al. 2008).

Fig. 10.

Atlantic SST changes and their relationship with southern Amazon precipitation. (a),(b) Spatial pattern of SST changes (K; 1%to4x; 30-yr mean anomaly relative to control minus mean SST change over 30°S–30°N). (c) Reproduced from Good et al. (2008), the relationship between June–August changes in southern Amazon precipitation (y axis) and their index of the tropical Atlantic cross-equatorial SST gradient (x axis). Solid line represents relationship derived from interannual variability in an atmosphere-only model; symbols represent multimodel coupled climate projections. (d) As in (c), but the symbols show changes (end of 1%to4x minus control) for HadCM3LC fixveg (*) and HadGEM2-ES (+).

Fig. 10.

Atlantic SST changes and their relationship with southern Amazon precipitation. (a),(b) Spatial pattern of SST changes (K; 1%to4x; 30-yr mean anomaly relative to control minus mean SST change over 30°S–30°N). (c) Reproduced from Good et al. (2008), the relationship between June–August changes in southern Amazon precipitation (y axis) and their index of the tropical Atlantic cross-equatorial SST gradient (x axis). Solid line represents relationship derived from interannual variability in an atmosphere-only model; symbols represent multimodel coupled climate projections. (d) As in (c), but the symbols show changes (end of 1%to4x minus control) for HadCM3LC fixveg (*) and HadGEM2-ES (+).

The different sensitivity of tropical forest to climate and CO2 in the two models may be due to differences in both the vegetation and hydrology treatments. Plant respiration in HadGEM2-ES is less sensitive to high temperatures than in HadCM3LC. Also, offline studies with the land surface model forced with present-day climate show that there is more water available (i.e., above the vegetation wilting point) when the LSH module is used (as in HadGEM2-ES). Both of these changes may contribute to the increased forest resilience in the HadGEM2-ES runs. The increased water availability with the LSH module may also help maintain a shorter dry season in HadGEM2-ES.

The difference between the 1%to4x and 1%to4x-fixveg projections of Amazonian dry-season length for HadCM3LC allows us to quantify vegetation–climate feedbacks from the Amazon dieback in this model. The regional mean difference is 0.9 months, which is of comparable magnitude to that found by previous studies (Costa and Pires 2010; Good et al. 2011).

Caution is implied in using simple observational comparisons, such as Fig. 8 for weighting different model projections in probabilistic projections. The difference in control climate between HadCM3LC and HadGEM2-ES is small compared to the CMIP3 spread. However, these two models show very different future climates. That is, assessing the accuracy of the simulated control climate may not constrain future change effectively. Also, HadCM3LC shows a rather extreme future projection, so the outcome of a probabilistic assessment would be rather sensitive to the weighting given to this model.

Regarding future work, the dry-season resilience approach is a useful starting point for understanding differences in forest projections. As in some other studies (e.g., Malhi et al. 2009), we have emphasized the contribution of precipitation to uncertainty in Amazon forest projections. However, major challenges exist in understanding in more detail the range of projections from different GCMs. First there is a need to better understand the different indices of rainfall, such as dry-season length, maximum climatological water deficit, or annual mean rainfall (e.g., Good et al. 2011; Malhi et al. 2009; Salazar et al. 2007), in terms of their relevance to forest impacts, their sensitivity to model uncertainty, and how well they characterize projected changes in tropical rainfall. As drivers of uncertainty in Amazon rainfall change, patterns of sea surface temperature in the tropical Pacific and Atlantic have been found to be important (Good et al. 2008; Harris et al. 2008). However, GCM differences in the two-dimensional spatial patterns of SST variability change make it hard to quantify such effects (Good et al. 2008). Progress here is required to understand the causes of change in SST patterns (and in associated model uncertainty), which could have multiple origins (Good et al. 2009). Further such physical understanding is required to formulate more relevant observational assessments of model reliability. There is also a clear need for more GCMs with good historical simulations of the Amazon climate, and the addition of key missing processes such as fire.

A more detailed understanding of model differences in the vegetation response is also required. Significant differences between different vegetation models have been found in terms of their response to climate and CO2 (e.g., Galbraith et al. 2010; Lapola et al. 2009). The dry-season resilience approach could be applied to other land surface and vegetation schemes. Other issues like the influence of fire (Golding and Betts 2008) and multiple stable forest states (Sternberg 2001) also need further examination.

5. Conclusions

In summary, we find that in both HadGEM2-ES and HadCM3LC, two tropical zones can be identified: one where forest is sustainable and another where it is unsustainable. The boundary between the two zones is predicted by a simple function of climate and CO2 conditions. In a climate change simulation, forest dieback is possible when climate and/or CO2 change such that a previously forested region becomes part of the zone of unsustainable forest. For the Amazon this happens in HadCM3LC but not in HadGEM2-ES (for CO2 up to 4 times preindustrial levels). There are two aspects to this difference between the two models: 1) the location of the boundary (i.e., under what climate–CO2 conditions forest is sustainable), and 2) the projected climate over the Amazon. The differences in projected climate come from differences in both the control climate and the projected change in climate. We find that all of these factors contribute toward an intact future Amazon in HadGEM2-ES: tropical forest is sustainable at warmer and drier climates in HadGEM2-ES, the Amazon is farther from the zone of unsustainable forest in the control simulation, and climate change over the Amazon is much less severe (i.e., much less movement toward the zone of unsustainable forest). The latter is the dominant difference between the two models.

The dominant single factor we identified was the difference in projected change in dry-season length, accounting for about 40% of the total difference in future forest resilience. This may be largely attributed to differences in tropical Atlantic SST change. Differences in control climatologies of temperature and dry-season length, in the projected regional warming, and in the forest response to climate and CO2 also all contribute to increased future forest resilience in HadGEM2-ES.

A primary question might be, does this new result from HadGEM2-ES change our assessed risk of Amazon dieback? At present, however, dieback risk is very hard to quantify, given the shortcomings in precipitation simulations from the CMIP3 generation of climate models (the major difference between HadGEM2-ES and HadCM3LC lies in the projected change in dry-season length). Therefore it is hard to assert that this risk has changed. Second, a lack of understanding of the detailed physical mechanisms behind the different rainfall projections is a major barrier to weighing the reliability of different model projections for specific cases like the Amazon. This makes it hard to identify the core model elements that need to be well formulated to give good Amazon rainfall projections. HadGEM2-ES includes many changes in detail that would be expected to give an overall improvement in global climate projections. However, it is not currently possible to say how significant these improvements might be for the specific case of Amazon forest projections. Thus, the new HadGEM2-ES result does not invalidate the HadCM3LC dieback projection. Indeed, the latter remains a possible scenario of dangerous change, which requires further understanding. An interesting result from HadCM3LC (Good et al. 2011) that still holds in HadGEM2-ES is that the tropical mean negative effect of warming on the forest is approximately balanced by the positive effect of CO2 increases.

Acknowledgments

This work was supported by the Joint U.K. DECC/Defra Met Office Hadley Centre Climate Programme (GA01101). We acknowledge the two modeling groups—the Program for Climate Model Diagnosis and Intercomparison (PCMDI) and the WCRP’s Working Group on Coupled Modelling (WGCM)—for their roles in making available the WCRP CMIP3 multimodel dataset. Support of this dataset is provided by the Office of Science, U.S. Department of Energy. We acknowledge suggestions from two anonymous reviewers that significantly improved the manuscript.

REFERENCES

REFERENCES
Arora
,
V. K.
, and
G. J.
Boer
2006
:
Simulating competition and coexistence between plant functional types in a dynamic vegetation model. Earth Interact., 10. [Available online at http://EarthInteractions.org.]
Atkin
,
O.
,
D.
Bruhn
, and
M.
Tjoelker
,
2005
:
Response of plant respiration to changes in temperature: Mechanisms and consequences of variations in Q10 values and acclimation
.
Adv. Photosynth. Respir.
,
18
,
95
135
,
doi:10.1007/1-4020-3589-6_7
.
Betts
,
R. A.
,
P. M.
Cox
,
M.
Collins
,
P. P.
Harris
,
C.
Huntingford
, and
C. D.
Jones
,
2004
:
The role of ecosystem-atmosphere interactions in simulated Amazonian precipitation decrease and forest dieback under global climate warming
.
Theor. Appl. Climatol.
,
78
,
157
175
.
Clark
,
D. B.
, and
N.
Gedney
,
2008
:
Representing the effects of subgrid variability of soil moisture on runoff generation in a land surface model
.
J. Geophys. Res.
,
113
,
D10111
,
doi:10.1029/2007JD008940
.
Clark
,
D. B.
, and
Coauthors
,
2011
:
The Joint UK Land Environment Simulator (JULES), model description – Part 2: Carbon fluxes and vegetation dynamics
.
Geosci. Model Dev.
,
4
,
701
722
.
Collins
,
W. J.
, and
Coauthors
,
2011
:
Development and evaluation of an Earth-system model – HadGEM2
.
Geosci. Model Dev. Discuss.
,
4
,
997
1062
.
Costa
,
M. H.
, and
G. F.
Pires
,
2010
:
Effects of Amazon and Central Brazil deforestation scenarios on the duration of the dry season in the arc of deforestation
.
Int. J. Climatol.
,
30
,
1970
1979
.
Cox
,
P. M.
,
2001
:
Description of the TRIFFID Dynamic Global Vegetation Model. Met Office Hadley Centre Tech. Note 24, 17 pp
.
Cox
,
P. M.
,
R. A.
Betts
,
C. D.
Jones
,
S. A.
Spall
, and
I. J.
Totterdell
,
2000
:
Acceleration of global warming due to carbon-cycle feedbacks in a coupled climate model
.
Nature
,
408
,
184
187
.
Cox
,
P. M.
,
R. A.
Betts
,
M.
Collins
,
P. P.
Harris
,
C.
Huntingford
, and
C. D.
Jones
,
2004
:
Amazonian forest dieback under climate-carbon cycle projections for the 21st century
.
Theor. Appl. Climatol.
,
78
,
137
156
.
Cox
,
P. M.
, and
Coauthors
,
2008
:
Increasing risk of Amazonian drought due to decreasing aerosol pollution
.
Nature
,
453
,
212
215
.
Cramer
,
W.
, and
Coauthors
,
2001
:
Global response of terrestrial ecosystem structure and function to CO2 and climate change: Results from six dynamic global vegetation models
.
Global Change Biol.
,
7
,
357
373
.
Essery
,
R. L. H.
,
M. J.
Best
,
R. A.
Betts
,
P. M.
Cox
, and
C. M.
Taylor
,
2003
:
Explicit representation of subgrid heterogeneity in a GCM land surface scheme
.
J. Hydrometeor.
,
4
,
530
543
.
Friedlingstein
,
P.
, and
Coauthors
,
2006
:
Climate–carbon cycle feedback analysis: Results from the C4MIP model intercomparison
.
J. Climate
,
19
,
3337
3353
.
Galbraith
,
D.
,
P. E.
Levy
,
S.
Sitch
,
C.
Huntingford
,
P.
Cox
,
M.
Williams
, and
P.
Meir
,
2010
:
Multiple mechanisms of Amazonian forest biomass losses in three dynamic global vegetation models under climate change
.
New Phytol.
,
187
,
647
665
.
Gedney
,
N.
, and
P. M.
Cox
,
2003
:
The sensitivity of global climate model simulations to the representation of soil moisture heterogeneity
.
J. Hydrometeor.
,
4
,
1265
1275
.
Golding
,
N.
, and
R.
Betts
,
2008
:
Fire risk in Amazonia due to climate change in the HadCM3 climate model: Potential interactions with deforestation
.
Global Biogeochem. Cycles
,
22
,
GB4007
,
doi:10.1029/2007GB003166
.
Good
,
P.
,
J. A.
Lowe
,
M.
Collins
, and
W.
Moufouma-Okia
,
2008
:
An objective tropical Atlantic sea surface temperature gradient index for studies of south Amazon dry-season climate variability and change
.
Philos. Trans. Roy. Soc.
,
B363
,
1761
1766
.
Good
,
P.
,
J. A.
Lowe
, and
D. P.
Rowell
,
2009
:
Understanding uncertainty in future projections for the tropical Atlantic: Relationships with the unforced climate
.
Climate Dyn.
,
32
,
205
218
.
Good
,
P.
,
C.
Jones
,
J.
Lowe
,
R.
Betts
,
B.
Booth
, and
C.
Huntingford
,
2011
:
Quantifying environmental drivers of future tropical forest extent
.
J. Climate
,
24
,
1337
1349
.
Harris
,
P. P.
,
C.
Huntingford
, and
P. M.
Cox
,
2008
:
Amazon Basin climate under global warming: The role of the sea surface temperature
.
Philos. Trans. Roy. Soc.
,
B363
,
1753
1759
.
Johns
,
T. C.
, and
Coauthors
,
2006
:
The new Hadley Centre Climate Model (HadGEM1): Evaluation of coupled simulations
.
J. Climate
,
19
,
1327
1353
.
Jones
,
C.
,
J.
Lowe
,
S.
Liddicoat
, and
R.
Betts
,
2009
:
Committed terrestrial ecosystem changes due to climate change
.
Nat. Geosci.
,
2
,
484
487
.
Lapola
,
D. M.
,
M. D.
Oyama
, and
C. A.
Nobre
,
2009
:
Exploring the range of climate biome projections for tropical South America: The role of CO2 fertilization and seasonality
.
Global Biogeochem. Cycles
,
23
,
GB3003
,
doi:10.1029/2008GB003357
.
Lenton
,
T. M.
,
H.
Held
,
E.
Kriegler
,
J. W.
Hall
,
W.
Lucht
,
S.
Rahmstorf
, and
H. J.
Schellnhuber
,
2008
:
Tipping elements in the Earth’s climate system
.
Proc. Natl. Acad. Sci. USA
,
105
,
1786
1793
.
Lewis
,
S. L.
, and
Coauthors
,
2009
:
Increasing carbon storage in intact African tropical forests
.
Nature
,
457
,
1003
1006
.
Lewis
,
S. L.
,
P. M.
Brando
,
O. L.
Phillips
,
G. M. F.
van der Heijden
, and
D.
Nepstad
,
2011
:
The 2010 Amazon drought
.
Science
,
331
,
554
.
Malhi
,
Y.
, and
Coauthors
,
2009
:
Exploring the likelihood and mechanism of a climate-change-induced dieback of the Amazon rainforest
.
Proc. Natl. Acad. Sci. USA
,
106
,
20 610
20 615
.
Marengo
,
J. A.
, and
Coauthors
,
2008
:
The drought of Amazonia in 2005
.
J. Climate
,
21
,
495
516
.
Martin
,
G. M.
,
M. A.
Ringer
,
V. D.
Pope
,
A.
Jones
,
C.
Dearden
, and
T. J.
Hinton
,
2006
:
The physical properties of the atmosphere in the new Hadley Centre Global Environmental Model (HadGEM1). Part I: Model description and global climatology
.
J. Climate
,
19
,
1274
1301
.
Martin
,
G. M.
, and
Coauthors
,
2011
:
The HadGEM2 family of Met Office Unified Model climate configurations
.
Geosci. Model Dev. Discuss.
,
4
,
723
757
.
Mercado
,
L. M.
,
N.
Bellouin
,
S.
Sitch
,
O.
Boucher
,
C.
Huntingford
,
M.
Wild
, and
P. M.
Cox
,
2009
:
Impact of changes in diffuse radiation on the global land carbon sink
.
Nature
,
458
,
1014
1017
.
Mitchell
,
T. D.
, and
P. D.
Jones
,
2005
:
An improved method of constructing a database of monthly climate observations and associated high-resolution grids
.
Int. J. Climatol.
,
25
,
693
712
.
Phillips
,
O. L.
, and
Coauthors
,
2009
:
Drought sensitivity of the Amazon rainforest
.
Science
,
323
,
1344
1347
.
Salazar
,
L. F.
,
C. A.
Nobre
, and
M. D.
Oyama
,
2007
:
Climate change consequences on the biome distribution in tropical South America
.
Geophys. Res. Lett.
,
34
,
L09708
,
doi:10.1029/2007GL029695
.
Scholze
,
M.
,
W.
Knorr
,
N. W.
Arnell
, and
I. C.
Prentice
,
2006
:
A climate-change risk analysis for world ecosystems
.
Proc. Natl. Acad. Sci. USA
,
103
,
13 116
13 120
.
Sitch
,
S.
, and
Coauthors
,
2008
:
Evaluation of the terrestrial carbon cycle, future plant geography and climate-carbon cycle feedbacks using five Dynamic Global Vegetation Models (DGVMs)
.
Global Change Biol.
,
14
,
2015
2039
.
Sternberg
,
L. D. L.
,
2001
:
Savanna–forest hysteresis in the tropics
.
Global Ecol. Biogeogr.
,
10
,
369
378
.
Thornton
,
P. E.
, and
Coauthors
,
2009
:
Carbon-nitrogen interactions regulate climate-carbon cycle feedbacks: Results from an atmosphere-ocean general circulation model
.
Biogeosciences
,
6
,
2099
2120
.
White
,
A.
,
M. G. R.
Cannell
, and
A. D.
Friend
,
1999
:
Climate change impacts on ecosystems and the terrestrial carbon sink: A new assessment
.
Global Environ. Change
,
9
(
Suppl.
),
21
30
.