High sensitivity of future global warming to land carbon cycle processes

Carbon cycle processes control the relative partitioning of emitted carbon dioxide between the atmosphere, land surface and the ocean [1, 2]. The recent explicit modelling of these processes in Atmosphere Ocean General Circulation Models (AOGCMs) coupled with carbon cycle models, predicts a range of future atmospheric CO₂ concentrations, and hence global warming, for a given emissions scenario [3, 4]. Land processes, that currently serve to provide a net sink from the atmosphere of CO₂ emissions, have the potential to become dramatically less efficient under future climate conditions [5], with major implications for international agreements to mitigate warming. Yet estimates of the magnitude of the reduction disagree [3, 5].

Within the climate research community there is increasing emphasis on understanding and quantifying uncertainty in future projections of climate change. So far this has mainly focused on physical climate feedbacks [6–9] through variables such as the climate sensitivity (CS) [10] or the transient climate response (TCR) [6, 11] or considered emergent constraints arising from limiting total emissions of CO₂ within simpler modelling frameworks [12–14].

Important steps forward have also been made looking at the relative contribution of uncertainty arising from carbon cycle processes. The C⁴MIP ensemble (Coupled Climate Carbon Cycle Model Intercomparison Project [5]) demonstrated a significant range of uncertainty in predicted temperature and atmospheric CO₂. Although C4MIP contains a range of land carbon processes, distinguishing carbon cycle and physical processes requires either a simplified metric [5] or simplified carbon cycle models [15]. Results from [15] suggest that within C4MIP simulations, carbon cycle (land and ocean) processes represent only 40% of the contribution from physical process uncertainty.

Here we make a targeted re-evaluation of contribution to the overall uncertainty in future climate (physical and carbon cycle), arising specifically from the land component of the carbon cycle. We use a parameter perturbation approach as previously used to quantify the impact of atmospheric process uncertainty on CS [10] and atmospheric and ocean process uncertainty on TCR [11, 16]. We apply this methodology to the land carbon cycle component of a fully coupled climate–carbon cycle AOGCM (supplementary information available at stacks.iop.org/ERL/7/024002/mmedia), whilst keeping all other components (ocean carbon cycle and physical processes) invariant across the ensemble. The land surface configuration utilized for this study, models the exchange of energy, water and carbon fluxes, representing how plants and soils respond to changes in temperature, moisture and CO₂. This model framework explicitly represents dynamical changes in vegetation cover in response changes in climate conditions. This enables vegetation changes to move beyond projections of equilibrium responses, that the previous generation of land carbon models relied on, to explore transient vegetation changes to rapidly evolving climate conditions [17, 18]. The other advance of dynamical vegetation models, compared to previous equilibrium classifications, is inclusion of CO₂ fertilization effects which act to mitigate plant temperature and water stress [18].

We vary parameters relating to the sensitivities of plant photosynthesis and soil respiration to temperature; stomatal conductance; soil water availability and surface evaporation; and plant competition The six parameters and their ranges are as follows: the optimal temperature values for photosynthesis [19], T_opt(K) was varied ±5° around the standard value of each plant functional type, similar to the range used in [20]; top leaf nitrogen concentration (as a fraction of carbon content) [19], N_LO (kg N/kg C), was varied within a range (for each PFT) determined by available plant trait information [21] and simulated historical forest distributions; the maximum ratio of the internal to external CO₂ concentration, which affects the maximum stomatal conductance [19], f₀, was varied based on ranges estimated from calibration of land surface models against simultaneous fluxes of energy, water and carbon from a number of sites; the minimum leaf area index (LAI) that a plant functional type (PFT) needs to achieve before it begins to compete for space [22] minLAI, not being directly observable its range was determined by its ability to reproduce observed forest distributions; the critical level of the volumetric soil moisture concentration, below which soil moisture limits both plant photosynthesis [19] and surface evaporation, θ_crit (m³ water/m³ soil), exploring a broad range from moisture limitations between the wilting and the saturation points and finally the 'Q₁₀' temperature dependence of soil respiration [19] Q₁₀ explored a range based on previous published estimates. 17 individual historical and future (using the A1B SRES scenario [23]) simulations were performed using a coupled carbon cycle climate model (HadCM3C), each exploring a different space in the parameter uncertainty ranges. The parameter values selected for each simulation and the over all ranges are detailed in table 1. The parameters, land surface equations and ranges are further documented in the supplementary information (available at stacks.iop.org/ERL/7/024002/mmedia) and [24].

Our predictions of CO₂ concentrations simulated across the ensemble from 1850 to 2100 are shown in figure 1. While there are some differences in the simulations of historical global atmospheric CO₂, the models show marked divergence as the 21st century progresses, as evidenced by a significant spread between models of 461 ppm (669–1130 ppm) by year 2100. Putting this range into context: the C⁴MIP ensemble produced a range of atmospheric CO₂ in 2099 of only 282 ppm (727–1009 ppm for an alternative future emission scenario SRES A2 [23]; see vertical bar on figure 1). Figure 1 also compares our uncertainty using a full AOGCM, with the conventional uncertainty estimates from two simpler Earth system Models of Intermediate Complexity (EMIC) [1], namely ISAM and Bern-CC. The greater future spread within the AOGCM cannot solely be attributed to greater freedom across this ensemble to diverge from observed trends during the historical period. The two simulations with the two lowest CO₂ concentrations in 2000, have the lowest and highest concentrations in the future respectively. This was despite the C⁴MIP ensemble and the EMIC estimates also including ocean carbon cycle and physical climate uncertainties, in addition to the land carbon cycle processes illustrated here. The spread of CO₂ concentrations arising from land carbon cycle uncertainties (461 ppm) is greater than full spread of future SRES concentration scenarios [23] (418 ppm) when carbon cycle uncertainties are neglected (which range from 540 ppm (B1) to 958 ppm (A1FI) in 2100, see figure 1). Moreover, the upper end of CO₂ concentrations in 2100 for our ensemble from a single mid-range emission scenario (A1B) is as much as 18% greater than the highest SRES concentration scenario (when carbon cycle uncertainty is neglected).

Zoom In Zoom Out Reset image size

The findings presented in figure 1 demonstrate the need to factor in uncertainty between prescribed CO₂ emissions and atmospheric concentrations when understanding the full global warming implications of different emissions socio-economic 'storylines'. Inverting, this work demonstrates that given this relationship is currently subject to quite considerable uncertainty, then any prescription of future CO₂ concentration thresholds might be achievable across a range of diverse emissions scenarios. Thus, as well as choice of socio-economic storyline, land surface parameterizations are critical factors in determining the fidelity of our predictions of the future evolution of atmospheric CO₂ concentrations. This has important implications for policy and decision relevant climate risk assessments as well as integrated assessment tools (such as [25] and [26]).

In figure 2 we compare each of the simulations against observed estimates of the amount of emitted carbon, between 1980 and 1999, taken up by the land and ocean. This enables us to incorporate information arising from uncertainties in the historical land use contribution to CO₂ emissions [2] (which would not be accounted for by comparing atmospheric concentrations (figure 1) with historical measurements [27]). This figure shows the comparison between simulated (black) and our observed estimate (green) (including C4MIP in blue). Looking at the future projections (red), all models show a reduction in the land surface's ability to draw down CO₂, with a corresponding increase in ocean and (more significantly from a global warming perspective) atmosphere fractions. It is the degree to which this happens which varies between simulations. A large proportion of the simulations suggest that the land surface will represent a net source of CO₂, in addition to anthropogenic emissions, by the end of the century. There is no relationship, however, between the future airborne and land fractions and the simulated magnitude during the historical period, suggesting that the observed carbon budgets do not reduce the range of future projections. Indeed, we see this manifested in figure 1, where the two ensemble members with the lowest atmospheric CO₂ in 2000 have the lowest and highest in 2100 (figure 1).

Zoom In Zoom Out Reset image size

There are other ways with which these simulations point back to our observations of the real world. In figure 3 we explore relationships between the parameter uncertainty, in land surface processes, and the magnitude of the modelled response. Specifically figure 3 illustrates the relationship between the year 1860-to-2100 atmospheric CO₂ change and each of the six land carbon cycle parameters varied across our 17 ensemble members. This shows that the impact of uncertainty in the high temperature optima for photosynthesis dominates the in future land surface responses, with weaker correlations arising from the other parameters, consistent with the earlier findings of [19] within a similar model framework. It is the broadleaf trees that represent large tropical carbon stocks—where temperatures are most likely to exceed T_opt values in a warming climate. For broad-leafed trees, reported optimum temperatures cover a broad range [28]. Optimal temperatures for electron transport capacity is likely to be the dominant, under high light conditions and future increases in CO₂, of the three temperature dependent pathways which determine the temperature dependence of photosynthesis. For broadleaf trees, these values range between 19° and 39°, with over 90% within the 28°–38° range sampled by this study [28]. In general we expect ecosystems to be optimized to the long-term climate state, in which case T_opt values which are close to the mean growing season temperature would be more likely. The implication of uncertainty in T_opt are likely to extend beyond this study, across current carbon cycle GCMs, as current state of the art terrestrial biosphere models share similar formulation for temperature controls on photosynthesis [29].

Zoom In Zoom Out Reset image size

The dominance of uncertainty in optimal temperatures for photosynthesis is interesting (figure 3), as it highlights a potential epistemic uncertainty in current formulation in land surface models. The region where in a warming world ambient temperatures have the potential to persistently exceed optimal photosynthetic temperatures is in the tropics. However there remains considerable controversy and debate on the likely vulnerability of many tropical rain forest species to a future warming [30–32], although [30] suggest damage to photosynthetic machinery could occur in the near future at current rates of warming. To understand future responses of ecosystems to climate change, consideration of the extent to which plants can acclimate both photosynthesis and respiration to increasing temperatures is critical [31]. In the circumstances of strong ability to acclimate, then parameter T_opt would become a function of warming extent and thus time. Nevertheless, there are very few studies of temperature acclimation responses for tropical trees [33, 34]. In the absence of sufficient data to constrain the high temperature plant responses, particularly in tropical ecosystems [30–34], the under-constrained response of terrestrial carbon cycling will remain a leading order uncertainty for future projections of atmospheric CO₂ concentrations.

In figure 4, we present the impact of our parameter uncertainty on the effect that the associated predicted trajectories of atmospheric CO₂ have on global warming (figure 4, green plume). This corresponds to a large temperature range of 3.3–5.7 °C (standard deviation 0.61 °C) by year 2099 with respect to the 1980–99 mean. As for the CO₂ responses presented in figure 1, this variation is due to uncertainty in modelled land carbon cycle processes. To place this range in context, we can compare this variation directly with that from another ensemble using a version of the Hadley Centre GCM (figure 4, blue plume), but in that instance designed to explore uncertainties in the atmospheric feedback mechanisms such as cloud and convection processes [11, 35]. By the end of the century, such modelled perturbations to the atmospheric physics give rise to a comparable uncertainty range of 2.4–5.0 °C (standard deviation 0.69 °C) for the A1B scenario. The range of responses can also be compared atmospheric responses within the multimodel archive (WCRP CMIP3) which suggests a range of 2.0–4.4 (standard deviation 0.59 °C) at the same point. This comparison confirms the expected previously unrecognized importance of land carbon cycle uncertainties for predictions of future temperatures which are directly comparable in magnitude to those explored through ensembles sampling uncertainties in atmospheric physics.

Zoom In Zoom Out Reset image size

Figure 4 also shows not only that explicit modelling of uncertainty in land carbon cycle processes produces both more spread but also that it tends to increase the average magnitude of the climate response. This is related to the spread of CO₂ (figure 1) explored across the models, which extends the upper bound of the range of 2100 CO₂ concentrations expected from emission driven C⁴MIP simulations for this scenario (supplementary information available at stacks.iop.org/ERL/7/024002/mmedia). Indeed, the reference atmospheric CO₂ time series (based on the stand configuration of Bern-CC), which is used in most current state of the art AOGCMs (including the atmospheric physics ensemble, blue plume figure 4), lies in the lower end of the land carbon ensemble range (figure 1). This suggests that the inclusion of carbon cycle uncertainties into conventional climate model projections is likely to extend the upper temperature bound of the warming we did expect for a given future scenario.

The analysis presented here demonstrates that the plausible range of land carbon cycle feedbacks on climate change is possibly greater than previously estimated and with this directly translating to particularly large uncertainty in future levels of global warming. At present, land carbon feedbacks are not well constrained by observations highlighting the need to identify robust measurements which are capable of constraining the land carbon cycle's response to global warming and physiological response to raised CO₂ concentrations. The exact nature of the temperature dependences of photosynthetic metabolism represents the most important uncertainty identified.

Acknowledgments