Climate Sensitivity Estimated from Temperature Reconstructions of the Last Glacial Maximum

See allHide authors and affiliations

Science  09 Dec 2011:
Vol. 334, Issue 6061, pp. 1385-1388
DOI: 10.1126/science.1203513


Assessing the impact of future anthropogenic carbon emissions is currently impeded by uncertainties in our knowledge of equilibrium climate sensitivity to atmospheric carbon dioxide doubling. Previous studies suggest 3 kelvin (K) as the best estimate, 2 to 4.5 K as the 66% probability range, and nonzero probabilities for much higher values, the latter implying a small chance of high-impact climate changes that would be difficult to avoid. Here, combining extensive sea and land surface temperature reconstructions from the Last Glacial Maximum with climate model simulations, we estimate a lower median (2.3 K) and reduced uncertainty (1.7 to 2.6 K as the 66% probability range, which can be widened using alternate assumptions or data subsets). Assuming that paleoclimatic constraints apply to the future, as predicted by our model, these results imply a lower probability of imminent extreme climatic change than previously thought.

Climate sensitivity is the change in global mean near-surface air temperature ΔSAT caused by an arbitrary perturbation ΔF (radiative forcing) of Earth’s radiative balance at the top of the atmosphere with respect to a given reference state. The equilibrium climate sensitivity for a doubling of atmospheric carbon dioxide (CO2) concentrations (ECS2xC) from preindustrial times has been established as a well-defined standard measure (1). Moreover, because transient (disequilibrium) climate change and impacts on ecological and social systems typically scale with ECS2xC, it is a useful and important diagnostic in climate modeling (1). Initial estimates of ECS2xC = 3 ± 1.5 K suggested a large uncertainty (2), which has not been reduced in the past 32 years despite considerable efforts (110). On the contrary, many recent studies suggest a small possibility of very high (up to 10 K and higher) values for ECS2xC (310), implying extreme climate changes in the near future, which would be difficult to avoid. Efforts to use observations from the past 150 years to constrain the upper end of ECS2xC have met with limited success, largely because of uncertainties associated with aerosol forcing and ocean heat uptake (8, 9). Data from the Last Glacial Maximum (LGM), 19,000 to 23,000 years ago, are particularly useful to estimate ECS2xC because large differences from preindustrial climate and much lower atmospheric CO2 concentrations [185 parts per million (ppm) versus 280 ppm preindustrial] provide a favorable signal-to-noise ratio, both radiative forcings and surface temperatures are relatively well constrained from extensive paleoclimate reconstructions and modeling (1113), and climate during the LGM was close to equilibrium, avoiding uncertainties associated with transient ocean heat uptake.

Here, we combine a climate model of intermediate complexity with syntheses of temperature reconstructions from the LGM to estimate ECS2xC using a Bayesian statistical approach. LGM, CO2 doubling, and preindustrial control simulations are integrated for 2000 years using an ensemble of 47 versions of the University of Victoria (UVic) climate model (14) with different climate sensitivities ranging from ECS2xC = 0.3 to 8.3 K considering uncertainties in water vapor, lapse rate and cloud feedback on the outgoing infrared radiation (fig. S1), as well as uncertainties in dust forcing and wind-stress response. The LGM simulations include larger continental ice sheets, lower greenhouse gas concentrations, higher atmospheric dust levels (fig. S2), and changes in the seasonal distribution of solar radiation [see supporting online material (SOM)]. We combine recent syntheses of global sea surface temperatures (SSTs) from the Multiproxy Approach for the Reconstruction of the Glacial Ocean (MARGO) project (12) and SATs over land, based on pollen evidence (13) with additional data from ice sheets and land and ocean temperatures (figs. S3 and S4; all reconstructions include error estimates). The combined data set covers over 26% of Earth’s surface (Fig. 1A).

Fig. 1

Annual mean surface temperature (SST over oceans and SAT over land) change between the LGM and modern. (A) Reconstructions of ΔSSTs from multiple proxies (12), ΔSATs over land from pollen (13), and additional data (SOM). (B) Best-fitting model simulation (ECS2xC = 2.4 K).

Figure 2 compares reconstructed zonally averaged surface temperatures with model results. Models with ECS2xC < 1.3 K underestimate the cooling at the LGM almost everywhere, particularly at mid-latitudes and over Antarctica, whereas models with ECS2xC > 4.5 K overestimate the cooling almost everywhere, particularly at low latitudes. High-sensitivity models (ECS2xC > 6.3 K) show a runaway effect resulting in a completely ice-covered planet. Once snow and ice cover reach a critical latitude, the positive ice-albedo feedback is larger than the negative feedback because of reduced longwave radiation (Planck feedback), triggering an irreversible transition (fig. S5) (15). During the LGM, Earth was covered by more ice and snow than it is today, but continental ice sheets did not extend equatorward of ~40°N/S, and the tropics and subtropics were ice free except at high altitudes. Our model thus suggests that large climate sensitivities (ECS2xC > 6 K) cannot be reconciled with paleoclimatic and geologic evidence and hence should be assigned near-zero probability.

Fig. 2

Zonally averaged surface temperature change between the LGM and modern. The thick black line denotes the climate reconstructions, and the gray shading the ±1, 2, and 3 K intervals around the observations. Modeled temperatures, averaged using only cells with reconstructions (see Fig. 1), are shown as colored lines labeled with the corresponding ECS2xC values.

Posterior probability density functions (PDFs) of the climate sensitivity are calculated by Bayesian inference, using the likelihood of the observations ΔTobs given the model ΔTmod(ECS2xC) at the locations of the observations. The two are assumed to be related by an error term ε; ΔTobs = ΔTmod(ECS2xC) + ε, which represents errors in both the model (endogenously estimated separately for land and ocean; fig. S13) and the observations (fig. S3), including spatial autocorrelation. A Gaussian likelihood function and an autocorrelation length scale of λ = 2000 km are assumed. The choice of the autocorrelation length scale is motivated by the model resolution and by residual analysis. (See sections 5 and 6 in the SOM for a full description of the statistical method, assumptions, and sensitivity tests.)

The resulting PDF (Fig. 3), considering both land and ocean reconstructions, is multimodal and displays a broad maximum with a double peak between 2 and 2.6 K, smaller local maxima around 2.8 and 1.3 K, and vanishing probabilities below 1 K and above 3.2 K. The distribution has its mean and median at 2.2 and 2.3 K, respectively, and its 66 and 90% cumulative probability intervals are 1.7 to 2.6 K and 1.4 to 2.8 K, respectively. Using only ocean data, the PDF changes little, shifting toward slightly lower values (mean 2.1 K, median 2.2 K, 66% 1.5 to 2.5 K, and 90% 1.3 to 2.7 K), whereas using only land data leads to a much larger shift toward higher values (mean and median 3.4 K, 66% 2.8 to 4.1 K, and 90% 2.2 to 4.6 K).

Fig. 3

Marginal posterior PDFs for ECS2xC. (A) Estimated from land and ocean, land-only, and ocean-only temperature reconstructions using the standard assumptions (1 × dust, 0 × wind stress, 1 × sea-level correction of ΔSSTSL = 0.32 K) (see SOM). (B) Estimated under alternate assumptions about dust forcing, wind stress, and ΔSSTSL using land and ocean data.

The best-fitting model (ECS2xC = 2.4 K) reproduces well the reconstructed global mean cooling of 2.2 K (within two significant digits), as well as much of the meridional pattern of the zonally averaged temperature anomalies (correlation coefficient r = 0.8) (Fig. 2). Substantial discrepancies occur over Antarctica, where the model underestimates the observed cooling by almost 4 K, and between 45° to 50° in both hemispheres, where the model is also too warm. Simulated temperature changes over Antarctica show considerable spatial variations (Fig. 1B), with larger cooling of more than 7 K over the West Antarctic Ice Sheet. The observations are located along a strong meridional gradient (fig. S7). Zonally averaged cooling of air temperatures is about 7 K at 80°S, more consistent with the reconstructions than the simulated temperature change at the locations of the observations. Underestimated ice sheet height at these locations could be a reason for the bias (16), as could deficiencies of the simple energy balance atmospheric model component. Underestimated cooling at mid-latitudes for the best fitting model also points to systematic model problems, such as the neglect of wind or cloud changes.

Two-dimensional features in the reconstructions are less well reproduced by the model (r = 0.5) (Fig. 1). Large-scale patterns that are qualitatively captured (Fig. 1) are stronger cooling over land than over the oceans, and more cooling at mid- to high latitudes (except for sea ice covered oceans), which is contrasted by less cooling in the central Pacific and over the Southern Hemisphere subtropical oceans. Continental cooling north of 40°N of 7.7 K predicted by the best-fitting model is consistent with the independent estimate of 8.3 ± 1 K from inverse ice-sheet modeling (17).

Generally, the model solution is much smoother than the reconstructions, partly because of the simple diffusive energy balance atmospheric model component. The model does not simulate warmer surface temperatures anywhere, whereas the reconstructions show warming in the centers of the subtropical gyres in parts of the northwest Pacific, Atlantic, and Alaska. It systematically underestimates cooling over land and overestimates cooling of the ocean (fig. S8). The land-sea contrast, which is governed by less availability of water for evaporative cooling over land compared with the ocean (18), is therefore underestimated, consistent with the tension between the ECS2xC inferred from ocean-only versus land-only data (Fig. 3). A possible reason for this bias could be overestimated sea-to-land water vapor transport in the LGM model simulations, perhaps due to too high moisture diffusivities. Other model simplifications, such as neglecting changes in wind velocities and clouds or errors in surface albedo changes in the dynamic vegetation model component, could also contribute to the discrepancies. The ratio between low latitude (40° S to 40° N) land and sea temperature change in the best-fitting model is 1.2, which is lower than the modern ratio of 1.5 found in observations and modeling studies (19).

Despite these shortcomings, the best-fitting model is within the 1σ-error interval of the reconstructed temperature change in three quarters (75%, mostly over the oceans) of the global surface area covered by reconstructions (fig. S8). The model provides data-constrained estimates of global mean (including grid points not covered by data) cooling of near-surface air temperatures ΔSATLGM = –3.0 K [66% probability range (–2.1, –3.3), 90% (–1.7, –3.7)] and sea surface temperatures ΔSSTLGM = –1.7 K [66% (–1.1, –1.8), 90% (–0.9, –2.1)] during the LGM (including an increase of marine sea and air temperatures of 0.3 and 0.47 K, respectively, due to 120-m sea-level lowering; otherwise, ΔSATLGM = –3.3 K and ΔSSTLGM = –2.0 K).

The ranges of 66 and 90% cumulative probability intervals, as well as the mean and median ECS2xC values, from our study are considerably lower than previous estimates. The most recent assessment report from the Intergovernmental Panel on Climate Change (6), for example, used a most likely value of 3.0 K and a likely range (66% probability) of 2 to 4.5 K, which was supported by other recent studies (1, 2023).

We propose three possible reasons that our study yields lower estimates of ECS2xC than previous work that also used LGM data. First, the new reconstructions of LGM surface temperatures show less cooling than previous studies. Our best estimates for global mean (including grid points not covered by data) SAT and SST changes reported above are 30 to 40% smaller than previous estimates (21, 23). This is consistent with less cooling of tropical SSTs (–1.5 K, 30°S to 30°N) in the new reconstruction (12) compared with previous data sets (–2.7 K) (24). Tropical Atlantic SSTs between 20°S and 20°N are estimated to be only 2.4 K colder during the LGM in the new reconstruction compared with 3 K used in (23), explaining part of the difference between their higher estimates of ECS2xC and ΔSATLGM (–5.8 K).

The second reason is limited spatial data coverage. A sensitivity test excluding data from the North Atlantic leads to more than 0.5 K lower ECS2xC estimates (SOM section 7 and fig. S14 and S15). This shows that systematic biases can result from ignoring data outside selected regions, as done in previous studies (22, 23), and implies that global data coverage is important for estimating ECS2xC. Averaging over all grid points in our model leads to a higher global mean temperature (SST over ocean, SAT over land) change (–2.6 K) than using only grid points where paleoclimate data are available (–2.2 K), suggesting that the existing data set is still spatially biased toward low latitudes and/or oceans. Increased spatial coverage of climate reconstructions is therefore necessary to improve ECS2xC estimates.

A third reason may be the neglect of dust radiative forcing in some previous LGM studies (21) despite ample evidence from the paleoenvironmental record that dust levels were much higher (25, 26). Sensitivity tests (Fig. 3) (SOM section 7) show that dust forcing decreases the median ECS2xC by about 0.3 K.

Our estimated ECS2xC uncertainty interval is rather narrow, <1.5 K for the 90% probability range, with most (~75%) of the probability mass between 2 and 3 K, which arises mostly from the SST constraint. This sharpness may imply that LGM SSTs are a strong physical constraint on ECS2xC. However, it could also be attributable to overconfidence arising from physical uncertainties not considered here, or from misspecification of the statistical model.

To explore this, we conduct sensitivity experiments that perturb various physical and statistical assumptions (Fig. 3 and figs. S14 and S15). The experiments collectively favor sensitivities between 1 and 3 K. However, we cannot exclude the possibility that the analysis is sensitive to uncertainties or statistical assumptions not considered here, and the underestimated land/sea contrast in the model, which leads to the difference between land- and ocean-based estimates of ECS2xC, remains an important caveat.

Our uncertainty analysis is not complete and does not explicitly consider uncertainties in radiative forcing due to ice-sheet extent or different vegetation distributions. Our limited model ensemble does not scan the full parameter range, neglecting, for example, possible variations in shortwave radiation due to clouds. Nonlinear cloud feedback in different complex models make the relation between LGM and CO2 doubling–derived climate sensitivity more ambiguous than apparent in our simplified model ensemble (27). More work, in which these and other uncertainties are considered, will be required for a more complete assessment.

In summary, using a spatially extensive network of paleoclimate observations in combination with a climate model, we find that climate sensitivities larger than 6 K are implausible, and that both the most likely value and the uncertainty range are smaller than previously thought. This demonstrates that paleoclimate data provide efficient constraints to reduce the uncertainty of future climate projections.

Supporting Online Material

Materials and Methods

SOM Text

Figs. S1 to S16

Table S1

References (28103)

References and Notes

  1. Acknowledgments: This work was supported by the Paleoclimate Program of the National Science Foundation through project PALEOVAR (06023950-ATM). Thanks to S. Harrison and two anonymous reviewers for thoughtful and constructive comments that led to substantial improvements in the paper.
View Abstract

Navigate This Article