A global coupled atmosphere/vegetation model and a dynamic ice sheet model were employed to study the impact of climate-vegetation interactions on the onset of the Antarctic ice sheet during the Eocene-Oligocene transition. We found that the CO2 threshold for Antarctic glaciation is highly sensitive to the prevailing vegetation. In our experiments, the CO2 threshold is less than 280 ppm if the Antarctic vegetation is dominated by forests and between 560 and 1120 ppm for tundra and bare ground conditions. The large impact of vegetation on inception is attributed to the ability of canopies to shade the snow-covered ground, which leads to a weaker snow albedo feedback and higher summer temperatures. However, the overall effect of canopy shading on the Antarctic climate also depends on features like local cloudiness and atmospheric meridional heat transport. Our results suggest that vegetation feedbacks on climate are crucial for the timing of the Antarctic glaciation.
The widespread expansion of the Antarctic ice sheet during the Eocene-Oligocene Transition (EOT; ∼ 34 Ma) constitutes one of the most striking reorganizations of the global climate that is seen in our geological records [e.g., Zachos et al., 1996, 2001; Lear et al., 2000]. The traditional hypothesis is that Antarctic glaciation was caused by changes in the regional ocean circulation resulting from opening of Southern Ocean seaways [Kennett, 1977; Barker et al., 2007]. More recently, however, it has been argued that the onset of the Antarctic ice sheet resulted from a cooling due to declining atmospheric CO2 levels [DeConto and Pollard, 2003; Pagani et al., 2011], possibly in combination with orbital conditions favoring cool summers [Coxall et al., 2005]. The CO2 threshold at which point Antarctica becomes glaciated is poorly constrained among climate models, but it generally varies between 560 and 924 ppm depending on the climate model as well as on other features like the atmospheric lapse rate and the bedrock topography [Gasson et al., 2013].
In addition to changes in internal or external climate forcing agents, the possibility of ice sheet inception is influenced also by the local land surface conditions. For example, it is well known that changes in vegetation characteristics may influence the regional climate, and consequently also the environmental conditions for ice sheet inception [Crucifix and Loutre, 2002; Meissner et al., 2003; Kageyama et al., 2004; Colleoni et al., 2009]. Vegetation influences the surface energy balance primarily through the surface albedo and the sensible and latent heat fluxes [Bonan, 2002]. The effect of vegetation on the surface albedo appears to be particularly strong in snow-covered high-latitude regions, where forests have the potential to shade the snow-covered ground, thus effectively increasing the surface albedo [Robinson and Kukla, 1985; Bonan et al., 1992, 1995]. Bonan et al.  found that the effect of vegetation masking of snow albedo serves to increase the surface spring temperature by as much as 14°C in present boreal regions.
Given the large impact of vegetation on the climate of high latitudes, the vegetation evolution across the EOT may have played an important role for the onset of the Antarctic ice sheet. The evolution of Antarctic vegetation during the Cenozoic, as derived from fossil plant records, is summarized in Thorn and DeConto . They suggested that during the Paleocene and most of the Eocene (65–37 Ma), Antarctic vegetation consisted mainly of mixed broadleaf and needleleaf temperate forests. From the late Eocene to the earliest Miocene (37–24 Ma), the temperate forests were gradually replaced by lower diversity vegetation, ultimately ending up with tundra. There is not much information on the vegetation prevailing at the time of glaciation. The very few plant fossils from this time indicate that forests were still present, although mixed with an increasing amount of lower vegetation types, like ferns [Thorn and DeConto, 2006; Francis et al., 2009].
The climate response to different Antarctic vegetation types was examined by Thorn and DeConto  using a global atmospheric general circulation model (AGCM). They found that a completely forested Antarctica serves to increase the surface temperature by up to 8°C during summer compared to if the continent was covered by tundra. Thorn and DeConto focused primarily on the sensitivity on the Antarctic climate to vegetation extremes. However, the impact of climate-vegetation interactions on the inception of the Antarctic ice sheet remains an open question. In this study, we use a coupled global atmosphere/dynamic vegetation model along with a stand-alone dynamic ice sheet model to examine the impact of climate-vegetation interactions on the inception of the Antarctic ice sheet. More specifically, we are interested to find out to which extent the CO2 threshold for Antarctic glaciation is sensitive to vegetation changes.
2 Models and Experimental Design
2.1 Atmosphere/Vegetation Model
The AGCM in this study is the National Center for Atmospheric Research (NCAR) Community Atmospheric Model 3.1 (CAM) [Collins et al., 2006]. We perform the simulations using a T42 horizontal resolution (∼2.8°×2.8°) and 26 hybrid levels in the vertical. In all simulations, the model is coupled to slab ocean model with prescribed ocean heat transport and mixed-layer depth. Land surface processes in CAM are computed by the Community Land Model 3 (CLM) [Oleson et al., 2004], which can handle up to four different land cover types within a grid cell (glacier, lake, wetland, and vegetation). The vegetated fraction of each grid point is represented by the fractional coverage of, in total, 15 plant functional types (PFTs). To allow for changing vegetation, we use CAM coupled to a Dynamic Global Vegetation Model (DGVM) [Levis et al., 2004], which computes the spatial and temporal evolution of the PFTs annually.
First we carry out a spin-up simulation with the coupled atmosphere/vegetation model (CAM-DGVM) starting from bare ground conditions and using pCO2=4480 ppm. Subsequent to the spin-up simulation, two sets of four steady state simulations with successively decreasing CO2 (2240, 1120, 560, and 280 ppm) are conducted. In the first set of simulations (1), we allow for dynamic vegetation using CAM-DGVM, whereas in the second set (2), we prescribe the vegetation cover from the spin-up (4480 ppm) simulation. Hence, the difference between the simulations in (1) and (2) for a certain CO2 yields the isolated changes caused by vegetation dynamics. Apart from the CO2, all other trace gases are set to preindustrial values.
The experiments are carried out with Eocene [Sewall et al., 2000] as well as with modern land-sea distribution and topography. In the simulations with modern boundary conditions (BCs), we set the glacial fraction to zero everywhere and replace the current Antarctic topography with the isostatically relaxed present-day bedrock topography. The location of Antarctica is similar in both BCs, but the Antarctic land area is about 30% smaller in the Eocene simulations than in the modern ones [see, e.g., Goldner et al., 2013]. Furthermore, the Antarctic topography in the Eocene case is generally lower than in the modern (310 m versus 710 m mean elevation). The prescribed ocean heat fluxes and mixed-layer depths in the Eocene simulations have been derived from a set of simulations with the corresponding fully coupled model (CCSM3) using the same CO2 levels as in our experiments [Liu et al., 2009]. The ocean heat fluxes and mixed-layer depths from the Liu et al.  simulations were also used in the Eocene slab ocean experiments in Caballero and Huber  and Goldner et al. . For further details, see Liu et al.  and Goldner et al. . In the modern simulations, the ocean heat fluxes and mixed-layer depths have been derived from present-day conditions. In all simulations, we use a globally uniform soil texture (43% sand and 18% clay). The spin-up are 200 and 270 model years for the Eocene and modern cases, respectively (Figure S1 in the supporting information). All other simulations are run for between 100 and 200 years, depending on how quickly the vegetation reaches a new equilibrium state. The last 20 model years are used in the analysis in all simulations.
2.2 Ice Sheet Model
To simulate the onset of the Antarctic ice sheet, we conduct a set of equilibrium simulations using the land-based dynamic/thermodynamic ice sheet model (ISM) Simulation Code for Polythermal Ice Sheets (SICOPOLIS) [Greve, 1997]. The model is run with 40 km horizontal resolution and 81 vertical levels. We use the model's default parameter values for Antarctica, except for the lapse rate which we set to the global mean (6.5°C km−1) for all seasons. For simplicity, the sea level is kept constant in all simulations, and instant calving is assumed to occur immediately once the ice sheet reaches the coastline. As bedrock topography, we use the maximum extent of the reconstructed Antarctic topography at the EOT from Wilson et al. , which has a mean elevation of 810 m. The surface temperature and precipitation fields in SICOPOLIS are linearly interpolated from the CAM/CAM-DGVM simulations using the global-mean lapse rate to correct for elevation differences between the CAM and Wilson et al.  topographies.
3.1 Vegetation and Temperature Changes
The simulated Antarctic vegetation with Eocene topography is shown in Figure 1. In the spin-up (4480 ppm) simulation, Antarctic vegetation is primarily dominated by deciduous broadleaf forests. The vegetation response to declining CO2 levels is roughly consistent with the Cenozoic vegetation evolution reported in Thorn and DeConto ; The modeled forests under high CO2 levels (≥2240 ppm) is consistent with the early-to-middle Eocene vegetation, whereas the mixed forests/grasslands state at about 1120 ppm characterizes the long-term vegetation transition from the Eocene forests to the tundra conditions during the late Oligocene/early Miocene. At low CO2 levels (≤560 ppm), the vegetation vanishes almost completely. The vegetation evolution using modern BCs is qualitatively similar to the one with Eocene BCs (Figure S2). The most prominent difference is that the major vegetation transitions occur at higher CO2 in the modern case (Figure S2).
To locate possible regions of ice sheet inception, Figures 1 and S2 (red contours) depict areas covered by perennial snow in the CAM-DGVM simulations. The fractional coverage of perennial snow on Antarctica in all simulations is summarized in Table 1. In the simulations with prescribed 4480 ppm vegetation, there is no perennial snow, not even in the 280 ppm simulations. Further, there is generally more perennial snow in the modern simulations than in the Eocene ones.
Table 1. Simulated Fraction Coverage of Perennial Snow on Antarctica fs, and East Antarctic Averages of Simulated Annual Mean Temperature , Annual Temperature Range (Temperature of the Warmest Month ), and Precipitation PEAa
PEA (m yr−1)
The values in the left column (CAM) below each quantity are derived from the simulations with prescribed vegetation, whereas the right column (CAM-DGVM) contains values derived from simulations with dynamic vegetation. To account for the elevation differences between the modern and Eocene topography, , , and PEA have been interpolated to the ISM grid prior to the spatial averaging. Furthermore, this enables a direct comparison with other GCMs [see Gasson et al., 2013, Table 2].
Eocene 1120 ppm
Eocene 560 ppm
Eocene 280 ppm
Modern 1120 ppm
Modern 560 ppm
Modern 280 ppm
Figure 2 depicts the simulated zonal mean surface temperature Ts for all the Eocene simulations with prescribed vegetation cover (Figures 2a–2c) as well as the anomalies induced by vegetation changes ΔTsv (Figures 2d–2f). Compared to the present-day control simulation (Figures 2a–2c), there is a substantial temperature increase over Antarctica owing to the reduced albedo and elevation associated with the removal of the ice sheet. The magnitude of this temperature increase (30–40°C) is consistent with other model simulations of the Eocene [Lunt et al., 2012]. Vegetation changes are associated with a substantial temperature decrease at high latitudes (Figures 2d–2f). For Antarctica, this is especially evident during the austral summer (Figure 2f). Using modern BCs, the austral summer ΔTsv is even further reduced over Antarctica (Figure S3), which coincides with a larger perennial snow cover in those simulations (Table 1).
The fact that the modern simulations have more perennial snow and need higher CO2 levels to sustain the Antarctic forests suggests that the Antarctic background climate is generally colder in the modern simulations than in the Eocene ones. This discrepancy is partly attributed to the elevation differences of the Antarctic topography between the two BCs. However, even after being interpolated to the ISM grid (and thus “brought” to the same elevation), large differences of the climate characteristics between the modern and Eocene simulations remain (Table 1). Since the Antarctic land area is larger in the modern BCs than in the Eocene ones, the Antarctic climate in the modern simulations has a greater continental influence, which is characterized by a larger annual temperature range (in Table 1). As a result, despite colder annual mean conditions in the modern simulations ( in Table 1), the summer temperatures are slightly higher than in the Eocene ones.
3.2 One-Dimensional Energy Balance Model
To understand the reasons behind the ΔTsv responses, we use a latitude (φ) dependent one-dimensional energy balance model (EBM) to estimate the relative contributions of various components on the total ΔTsv response. In steady state, the EBM is given by
where is the zonal mean net vertical energy flux at the top of the atmosphere, which is equal to the sum of the total heating rates and the total meridional heat flux. Because the heating rate anomalies caused by vegetation changes are negligible over Antarctica, we only consider the heat transport changes in the subsequent analysis. Note that the slab ocean model requires prescribed ocean heat fluxes, which means that all heat transport changes reported here occur in the atmosphere. All other notations in equation (1) are standard [see, e.g., Heinemann et al., 2009]. Since the temperature influences glaciation mainly through summer melting, we use the vertical energy fluxes from the CAM/CAM-DGVM simulations averaged over the austral summer to force the EBM. Further, we apply the EBM only to the 280 ppm and 560 ppm CAM/CAM-DGVM simulations.
With the EBM we are able to almost exactly reproduce the austral summer ΔTsv from the CAM/CAM-DGVM simulations (Figure 3, thick green lines). Most of the vegetation-induced cooling on Antarctica stems from changes in the planetary albedo α along with a smaller contribution from decreasing emissivity ε (Figure 3). Further, in all simulations, there is a strong negative feedback associated with increased atmospheric heat transport, which serves to increase ΔTsv (Figure 3). In the Eocene 560 ppm simulation, most of the albedo-induced cooling stems from cloud changes (compare the full- and clear-sky planetary albedo responses in Figure 3a). In the other three simulations, cloud changes have only a small effect on the planetary albedo (Figures 3b–3d). However, also in these simulations cloudiness increases over Antarctica, but because the surface is predominately covered by snow (Table 1), which has a similar albedo to clouds, changing cloudiness has only a minor effect on the regional energy balance.
3.3 Sensitivity Experiments
Owing to the uncertainty associated with the prevailing Antarctic vegetation during the EOT, we conduct five sensitivity experiments (using modern BCs and pCO2=280 ppm) with different prescribed vegetation densities on Antarctica. In these simulations, we scale the Antarctic leaf and stem area indices (L and S, respectively) from the spin-up simulation by 0, 20, 40, 60, and 80%, yielding mean summer Antarctic L+S of 0, 0.81, 1.63, 2.44, and 3.26, respectively. To isolate the impact of different feedbacks on the snow cover, we calculate the ablation using the positive degree-day (PDD) method [Calov and Greve, 2005] with the same degree-day coefficient for snow as in SICOPOLIS (3 mm d−1 °C−1). The results are shown in Figure 4 as spatial averages over Antarctica. Extensive coverage of perennial snow (Figure 4a) exists only when the annual ablation drops below the annual accumulation (Figure 4b), which occurs only for low vegetation densities in our simulations. Similar ablation rates are obtained using Ts derived from the EBM (Figure 4b).
Next, we use the EBM to estimate the effect of vegetation on the ablation through changing the albedo. The fraction shortwave radiation absorbed by vegetation during summer fSW increases with L+S essentially according to the Beer-Lambert law fSW∼1− exp(−kv(L+S)) (Figure 4a), where kv is a constant (kv≈0.55 in our case). The effect of vegetation on the surface albedo αs is given by αs=αsvfSW+αsg(1−fSW), where αsv and αsg are the albedos of vegetation and bare ground, respectively. Assuming that all snow accumulates on the ground, and that the bottom of the canopies are higher than the snow depth, we can use the EBM to crudely estimate the effect of vegetation masking of snow albedo on ablation by setting αsg to the albedo of snow (∼0.7), and αsv to the albedo of vegetation (∼0.2). Hence, for simplicity we assume that there is no canopy interception of snow in our EBM. In CAM, however, canopy interception of precipitation and the associated effects on albedo are accounted for. The contribution of the surface albedo to the planetary albedo is calculated using the method in Donohoe and Battisti [2011, equation 5], taking the atmospheric reflection and absorption as well as the heat transports, heating rates and emissivities from the sensitivity experiment with L+S=0.
The resulting ablation due to vegetation masking of snow albedo is shown in Figure 4c (green line). Due to the nature of the Beer-Lambert law, vegetation masking serves to substantially increase the ablation rates already for small L+S values (>1 m yr−1 at L+S=0.81). Taking into account also the effect of cloudiness, which decreases with L+S in our case (Figure 4a), the ablation rates are substantially enhanced (Figure 4c, blue line). Hence, the strength of the vegetation masking of snow albedo is sensitive to the prevailing cloudiness. This is because changing cloudiness (i.e., the atmospheric albedo) effectively alters the surface contribution to the planetary albedo [Donohoe and Battisti, 2011]. Note that when keeping the vegetation density at zero, the surface albedo remains high, which implies that isolated changes in cloudiness have only a small effect on the ablation (Figure 4c, red line). Finally, adding also the effects of changing heat transport and emissivity to the vegetation and cloud feedbacks, the ablation decreases (Figure 4c, black line). This ablation decrease is caused by the strong negative feedback associated with the increased atmospheric heat transport (see Figure 3).
3.4 Antarctic Glaciation
Using the climate from the simulations with prescribed vegetation (“CAM” in Table 1), Antarctica remains practically ice free even for the lowest CO2 levels (Figure 5, dashed lines). On the contrary, the colder conditions caused by vegetation changes (Table 1, “CAM-DGVM”) yield extensive glaciation from 560 ppm (Figure 5, solid lines). Thus, using the climate from the 560 ppm CAM-DGVM simulation with Eocene topography, Antarctica becomes glaciated in the ISM although there is almost no perennial snow in the AGCM (Table 1). This discrepancy results from the lapse-rate correction used to account for the different topographies in SICOPOLIS and CAM [Gasson et al., 2013]. The spatial distributions of the ice sheets at equilibrium resulting from the Eocene and modern climate simulations are shown in Figures S4 and S5, respectively.
4 Discussion and Conclusions
In this study, we used global AGCM coupled to a DGVM along with a dynamic ISM to examine the impact of climate-vegetation interactions on the onset of the Antarctic ice sheet during the EOT. A major limitation of our study is that we use a mixed-layer ocean model and thereby omit potential changes in the ocean dynamics. Various studies have shown that changes in the ocean dynamics across the EOT may influence the Antarctic climate conditions as well as the total meridional heat transport [e.g. Huber and Nof, 2006; Cristini et al., 2012]. In our model, the atmospheric meridional heat transport to Antarctica during summer increases with the local meridional temperature gradient, which is maintained predominately by the Antarctic surface conditions. Although this response is consistent with the first-order theory of the atmospheric meridional heat transport [e.g., Budyko, 1969], using a fully coupled atmosphere-ocean model rather than an AGCM would possibly yield a different response.
Our main result is that the CO2 threshold for Antarctic glaciation is highly sensitive to the prevailing vegetation cover. In our ISM experiments, the CO2 threshold varies between <280 ppm if the Antarctic vegetation is dominated by forests, and 560–1120 ppm for tundra and bare ground conditions. The latter values lie within the proxy-derived estimates of the CO2 levels during the EOT [Beerling and Royer, 2011] suggesting that close to bare ground conditions are necessary for glaciation. Hence, our results suggest that a large-scale transition from the Eocene forests to tundra or bare ground conditions should have preceded the EOT. However, the modeled CO2 threshold in the various studies depends on what criterion is used to define glacial inception. In Gasson et al. , the use of Glimmer ISM allows for inception already at 1120 ppm in many models, even though the average climate conditions derived from those simulations are similar to the 1120 ppm Eocene simulations shown here (compare our Table 1 to their Table 2). A large part of the literature about glacial inception relies on the fact that a perennial snow cover develops in the GCMs. Therefore, a good match between perennial snow in the GCM and ice cover in the ISM strengthens the probability that inception is not model dependent. In our simulations, perennial snow in the AGCM is consistent with ice cover in the ISM in five of the six experiments. Gasson et al.  do not show the simulated perennial snow cover. Furthermore, their ice sheet experiments were conducted without basal sliding, which greatly helps the thickening of the ice cover and therefore might have triggered inception earlier with respect to our experiments.
The large impact of vegetation on Antarctic glaciation is attributed to the ability of canopies to shade the snow-covered ground, resulting in increased summer temperatures mediated by a weaker snow albedo feedback. The ability of vegetation to mask the snow albedo increases with the vegetation density according to the Beer-Lambert law, which is a robust feature in high latitudes [Bonan et al., 1992, 1995; Thorn and DeConto, 2006]. However, in addition we found that the magnitude of the vegetation-induced temperature anomaly is also sensitive to the local atmospheric albedo, which is primarily controlled the prevailing cloud conditions. This is because clouds effectively regulate the amount of shortwave radiation reaching the surface [Donohoe and Battisti, 2011]. In our experiments, the local albedo due to clouds varies between 0.22 and 0.25 over a vegetated Antarctica. Various studies have found large cloud changes as a consequence of Antarctic glaciation [Heinemann et al., 2009; Lunt et al., 2012; Goldner et al., 2013], but only Heinemann et al.  reported on the cloud conditions prior to glaciation. In their Eocene simulation, the albedo due to clouds over Antarctica is about 0.4, which is higher than in our study and should therefore result in a weaker impact of vegetation on the planetary albedo. Our sensitivity experiments suggest that even sparse forests, which probably existed on Antarctica during the late Eocene, played a key role in the timing of the glaciation. However, to constrain the role of Antarctic vegetation during the EOT, it is necessary that the Eocene cloud conditions, along with their contributions to the planetary albedo, are evaluated also in other AGCMs and coupled atmosphere-ocean models.
We acknowledge support from the research funding programme LOEWE of Hesse's Ministry of Higher Education. LOEWE also provided financial support for the simulations carried out at the LOEWE Frankfurt Centre for Scientific Computing. The AGCM used in this study (CAM3) was developed by NCAR, which is funded by National Science Foundation.
The Editor thanks Luisa Cristini and an anonymous reviewer for their assistance in evaluating this paper.