Global vegetation and terrestrial carbon cycle changes after the last ice age


  • I. C. Prentice,

    1. Department of Biological Sciences, Macquarie University, North Ryde, NSW 2109, Australia
    2. Grantham Institute for Climate Change, and Division of Biology, Imperial College, Ascot SL5 7PY, UK
    Search for more papers by this author
  • S. P. Harrison,

    1. Department of Biological Sciences, Macquarie University, North Ryde, NSW 2109, Australia
    2. School of Geographical Sciences, University of Bristol, University Road, Bristol BS8 1SS, UK
    Search for more papers by this author
  • P. J. Bartlein

    1. Department of Geography, University of Oregon, Eugene, OR 97403-1251, USA
    Search for more papers by this author

Author for correspondence:
I. C. Prentice
Tel: +61 2 9850 4227


  • In current models, the ecophysiological effects of CO2 create both woody thickening and terrestrial carbon uptake, as observed now, and forest cover and terrestrial carbon storage increases that took place after the last glacial maximum (LGM). Here, we aimed to assess the realism of modelled vegetation and carbon storage changes between LGM and the pre-industrial Holocene (PIH).
  • We applied Land Processes and eXchanges (LPX), a dynamic global vegetation model (DGVM), with lowered CO2 and LGM climate anomalies from the Palaeoclimate Modelling Intercomparison Project (PMIP II), and compared the model results with palaeodata.
  • Modelled global gross primary production was reduced by 27–36% and carbon storage by 550–694 Pg C compared with PIH. Comparable reductions have been estimated from stable isotopes. The modelled areal reduction of forests is broadly consistent with pollen records. Despite reduced productivity and biomass, tropical forests accounted for a greater proportion of modelled land carbon storage at LGM (28–32%) than at PIH (25%).
  • The agreement between palaeodata and model results for LGM is consistent with the hypothesis that the ecophysiological effects of CO2 influence tree–grass competition and vegetation productivity, and suggests that these effects are also at work today.


Glacial periods have consistently been associated with low CO2 concentrations (Siegenthaler et al., 2005; Lüthi et al., 2008). They are not the cause of glaciations: glaciations are paced by orbital changes (Hays et al., 1976; Edwards, 2010), which produce changes in the seasonal timing and latitudinal distribution of insolation. However, CO2 concentrations of 170–200 ppm were repeatedly reached during the glacial maxima, and helped to ‘lock’ the Earth system – including tropical and Southern Hemisphere regions, remote from the climatic influence of the northern ice sheets – into a glacial mode. Although the mechanisms that lowered CO2 concentration during glaciations are still not established, there is some support for the hypothesis that increased dust transport from the continents reduced the CO2 concentration from early ice age intermediate levels to late ice age minimum levels by stimulating export production by phytoplankton in high-nitrogen, low-chlorophyll regions of the global ocean, particularly the Southern Ocean (Bopp et al., 2003; Kohfeld et al., 2005).

Whatever their causes, CO2 concentration changes during well-documented periods of recent Earth history provide opportunities to investigate the effects of CO2 on the biosphere. The increase in CO2 concentration during deglaciations encompassed all types of terrestrial vegetation, including tropical ecosystems, where direct experimental evidence for CO2 effects is limited or nonexistent. The most comprehensive palaeodata for any cold period are for the last glacial maximum (LGM): 20–26.5 ka (Clark et al., 2009). The Palaeoclimate Modelling Intercomparison Project (PMIP) Phase II ( has driven state-of-the-art climate models (as used in future projections for the Intergovernmental Panel on Climate Change, IPCC) with changes in ice sheet extent and orography, relative sea level, orbital parameters and glasshouse gas concentrations in order to represent the climate of LGM, conventionally defined as 21 ka (Braconnot et al., 2007).

Synthesis activities have meanwhile yielded global observational datasets of LGM land and sea surface conditions. The BIOME 6000 project, with subsequent regional updates, has provided a snapshot of biome distributions at LGM based on pollen and plant macrofossil data (Prentice et al., 2000 and references therein; Bigelow et al., 2003; Pickett et al., 2004; Marchant et al., 2009). Spatial coverage is far from uniform, with undersampling in most of the tropics – yet there are sufficient data points to document biome shifts across the tropics, and for the quantitative analysis of changes in forest cover (Harrison & Prentice, 2003).

These data are complemented by stable isotope measurements. Interpretations of changes in the triple isotope composition of atmospheric O2 (18O, 17O, 16O: Luz et al., 1999; Luz & Barkan, 2005) and the Dole effect (the δ18O offset between O2 and seawater: Bender et al., 1994; Hoffmann et al., 2004) depend on the unique capability of biological carbon cycling to transfer oxygen isotope signatures between water and O2. These measures provide information on global total (marine and terrestrial) gross primary production (GPP) (Bender et al., 1994; Blunier et al., 2002; Landais et al., 2007, 2010). The δ13C value of seawater, as preserved in the calcium carbonate shells of benthic marine foraminifera, indicates changes in global terrestrial carbon storage as a result of the fractionation in photosynthesis and the propagation of this signal into terrestrial biomass and soils (Shackleton, 1977; Bird et al., 1994; Prentice & Harrison, 2009). δ13C measurements from other materials and environments provide data on vegetation changes (a combined signal of ci : ca ratios of C3 plants and C3 vs C4 plant carbon) when measured in terrestrial sediments (Talbot & Johanessen, 1992; Giresse et al., 1994) and even offshore sediments, where compound-specific δ13C analysis has been applied to biomarkers originating from land plants (Rommerskirchen et al., 2006).

Two key findings already have unequivocal support from LGM data–model comparisons. First, it is necessary to invoke ecophysiological CO2 effects in order to account for the reduction in the tropical forest area at LGM as shown by pollen data. No climate model dries the continents sufficiently to produce such a reduction caused by climate change alone (Harrison & Prentice, 2003; Prentice & Harrison, 2009). Second, models forced by LGM climate alone produce an insufficient reduction of total land biosphere carbon storage, compared with the evidence from the marine δ13C record (Prentice & Harrison, 2009).

These findings were obtained using equilibrium biogeography models that translate potential net primary productivity (NPP) and bioclimatic limits of plant functional types (PFTs) into biomes using semi-empirical rules. Analyses of the contemporary carbon cycle require dynamic global vegetation models (DGVMs), which couple time-dependent changes in PFT abundances with ecophysiological and biogeochemical processes. DGVMs came to prominence with an analysis of future land carbon uptake projections (Cramer et al., 2001) highlighted in the IPCC Third Assessment Report (Prentice et al., 2001). DGVMs are now the main tool for modelling the land carbon cycle, but current models still generate different projections (see Friedlingstein et al., 2006; Sitch et al., 2008; and the IPCC Fourth Assessment Report, Denman et al., 2007). They also make assumptions about processes that are still highly controversial in the ecological literature. The disputed phenomena include woody thickening (the observed increase in woody plant cover in savannas across the global tropics and subtropics) and the terrestrial uptake of anthropogenic CO2 (the net land uptake of CO2, which continues despite tropical deforestation releasing CO2). DGVMs predict both phenomena as direct consequences of rising CO2 concentration. Woody thickening is predicted because of enhanced competition by woody (C3) plants against grasses. Net land uptake of CO2 is predicted because of CO2 fertilization: when C3 photosynthesis increases, NPP and litter production outpace the compensating increase in heterotrophic respiration from a growing soil carbon pool. Yet, both mechanisms have been claimed to be ineffective at the space and time scales modelled (e.g. Archer et al., 1995; Körner, 2006). If such critiques are well founded, the models must be incorrect, and their correct predictions coincidental.

Woody thickening and net CO2 uptake, however, also took place – over large spatial and long temporal scales – after LGM. It is therefore of considerable interest to run DGVMs under LGM conditions, including CO2 changes, and to ask whether the models correctly represent the nature and magnitude of subsequent changes in the state of the biosphere. If they do, this finding would support the way in which CO2 effects are treated in the models, and the implications for contemporary as well as palaeo-conditions.

We present DGVM simulations of global LGM and pre-industrial Holocene (PIH) vegetation and carbon cycling. We use the Land Processes and eXchanges (LPX) model (Prentice et al., unpublished), the most recent development of the Lund–Potsdam–Jena (LPJ) family of DGVMs (Sitch et al., 2003; Gerten et al., 2004). LPJ has previously been run for LGM using outputs from a mixed-layer ocean climate model (Kaplan et al., 2002) and one coupled atmosphere–ocean general circulation model (AOGCM) (Thonicke et al., 2005). We use outputs from PMIP II simulations made with four AOGCMs. Our objectives were to assess the realism of modelled LGM to PIH changes in vegetation distribution and carbon cycling resulting from known CO2 changes, combined with climate changes as simulated by state-of-the-art climate models, and to consider the implications for the simulation of carbon cycle processes in models.


The vegetation model

LPX (Prentice et al., unpublished) is a development from the LPJ SPread and InTensity of FIRE (LPJ-SPITFIRE) model (Thonicke et al., 2010), which, in turn, is based on the LPJ DGVM (Sitch et al., 2003; Gerten et al., 2004). The simulation of plant carbon and water exchanges in these models is identical with that in BIOME3 (Haxeltine & Prentice, 1996a) and BIOME4 (Kaplan et al., 2003). PFTs and their bioclimatic limits are derived from BIOME3 with minor modifications. LPJ adds dynamic representations of establishment, mortality, growth, carbon allocation (to fine roots, transport tissues and leaves), plant allometry (including relationships among height, diameter and crown diameter at the individual plant level) and dynamic competition among PFTs, whose abundance is defined in terms of foliage projective cover (FPC). LPJ-SPITFIRE and LPX represent the influence of potential ignition rates, vegetation properties and weather conditions on biomass burning. LPX uses Gerten et al.’s (2004) improved soil hydrology and a modification of Thonicke et al.’s (2010) fire scheme.

C3 photosynthesis is simulated in these models using a simplified implementation of the Farquhar et al. (1980) model. Leaf-level carboxylation capacity (Vcmax) is assigned dynamically based on an assumption of optimality (maximizing net assimilation rate over a 24-h period), given photosynthetically active radiation (PAR), temperature and ambient CO2 concentration ca (Haxeltine & Prentice, 1996b). A big-leaf approximation is used to integrate assimilation rates vertically through the canopy. Leaf internal CO2 concentration (ci) is a PFT-specific fixed fraction of ca under well-watered conditions and is drawn down when a water supply function falls below atmospheric demand.

The net effect of an increase in ca, if other environmental variables are constant, is the down-regulation of Vcmax, accompanied by an increase in daily net assimilation. Down-regulation is not mechanistically modelled, but arises from the optimality assumption. Given a conservative ci : ca ratio, enhancement of ca means that the value of Vcmax that allows the leaf to make full use of the available PAR is lower than at current ca. (A similar principle is implicit in recent theoretical treatments of CO2 effects: e.g. Franklin, 2007.) Because of this acclimation of Vcmax, the simulated response of photosynthesis under well-watered conditions varies with ca, approximately tracking the electron transport-limited rate. An increase in ca also produces a down-regulation of stomatal conductance (this follows from the assumption of a conservative ci : ca ratio combined with the ‘diminishing return’ of increasing ci), and thus a reduction in modelled plant water use.

The climate model simulations

The LGM climate simulations used to run LPX were provided by the FGOALS (China), HadCM3 (UK), IPSL (France) and MIROC (Japan) AOGCMs. These were the four models that provided all the variables required to run LPX at the time at which the PMIP II archive was accessed for this study. All the AOGCMs were run under the PMIP II protocol. Outputs were extracted for mean monthly values of daily minimum and maximum temperature, precipitation and fractional sunshine hours (48 variables altogether), at each climate model grid cell, in the LGM and PIH model runs. All outputs were converted to anomalies (differences between PIH and LGM values) for each climate model, grid cell and variable.

The vegetation modelling protocol

LPX simulations were performed on a 0.5° grid using the Climatic Research Unit TS2.1 dataset as the baseline climate (New et al., 2000). For the PIH control run of LPX, CO2 concentration was fixed at 280 ppm, and baseline climate data were provided for each grid cell as a repeated, detrended time series for 1948–2000. This approach assumes that the PIH climate was similar to the climate during the second half of the 20th century (in terms of mean values and the nature of interannual variability, which this protocol preserves), whilst avoiding the undesirable ‘sawtooth’ effect that would occur if no detrending were used. The model was spun up from a bare ground state until the soil carbon pool with the longest residence time reached stability.

For LGM simulations, it was necessary first to extend the baseline climate data out on to the continental shelf areas that were exposed at LGM. The time-varying, gridded baseline climate data were linearly extrapolated to −130 m elevation, applying a standard global −6 K km−1 lapse rate for temperatures, and empirically estimated global ‘lapse rates’: +2 mm km−1 and −0.01 km−1 for precipitation and fractional sunshine hours, respectively. LGM ocean and land ice grid points were then masked out, consistently with the PMIP II protocol. LPX was spun up at all of the additional grid points to provide a starting point for the LGM simulations. Then, for each LGM climate model and grid cell, the end point of the control simulation was used as the starting configuration and the model run was continued after a step-change to low CO2 (180 ppm) and an ‘LGM climate’. To obtain this LGM climate, the anomalies from each climate model were interpolated spatially on to the extended climate data grid (to deal with the lower resolution of the AOGCM grid relative to the climate data grid), and then the anomalies for each grid cell and variable were added to the baseline climate.

LPX internally generates ‘daily’ precipitation using a simple weather generator driven by precipitation (monthly total) and rain days (proportion of days with rain) (Gerten et al., 2004). As rain day outputs were not available from the climate models, rain day inputs to LPX were created using the assumption that the same covariation of precipitation and rain days (for each grid cell and month) holds for a change in climate as for interannual variability in the observed modern climate. A two-parameter function was fitted to this relationship as observed in the baseline climate data. The function is:

image(Eqn 1)

where f(P) is the proportion of days with rain (for any grid cell and month), P is the precipitation, and k and m are empirical constants. Eqn 1 was applied, with unchanged values of the constants, to infer LGM rain days from LGM precipitation for each grid cell and month of the simulations.

Assigning biomes to the LPX output

We used a method similar to that of Joos et al. (2004) to convert modelled vegetation properties to very broadly defined vegetation types (biomes) (Fig. 1). This coarse classification of world vegetation into biomes is appropriate for a global data–model comparison, although finer distinctions could be made on the basis of PFT abundances. The algorithm depends on modelled total FPC to distinguish between deserts, dry grass- and shrublands, and forests/savannas. Forests and savannas are separated by the height of the average individuals of woody PFTs (note that this modelled average height is much less than the maximum height that can be attained by large individuals in the real world). The structurally defined formations were subdivided to reflect differences between ‘cold’ biomes (with annual mean growing degree days above 5°C, GDD5 < 350 K days) and the rest, thus separating shrub tundra and tundra from their warmer climate counterparts. It should be noted that the distinction between shrub tundra and tundra implies that the latter term is used in a narrow sense, excluding shrub tundra (which has higher productivity and carbon storage). Within the forests and savannas, additional criteria based on the presence or dominance of particular woody PFTs were used to distinguish tropical, temperate and boreal biomes, as shown in Fig. 1. The resulting simulation of the PIH biome distribution is shown in Fig. 2.

Figure 1.

 The scheme used to assign biomes to Land Processes and eXchanges (LPX) output. The bottom panel represents a series of criteria that are evaluated in sequence from the top when the category assigned in the top panels is forest or savanna. FPC, foliage projective cover; GDD5, annual mean growing degree days above 5°C; PFTs, plant functional types.

Figure 2.

 The Land Processes and eXchanges (LPX) simulation of pre-industrial Holocene (PIH) biome distribution.

Representation of the outputs

As our focus here is not on the differences between AOGCMs, we present numerical results based on the averaging of the four resulting LPX outputs. To illustrate the uncertainty in these analyses, we also cite ranges in the text for key vegetation and carbon cycle quantities, representing the variation among results obtained with the four climate models. A simulated LGM biome map is shown based on the consensus among these four model runs (i.e. a biome is assigned only for grid cells in which all four runs agree).

This approach to the representation of the uncertainty of the different AOGCMs (summarizing results from several vegetation model runs) is preferred to the alternative of taking ‘average climate model’ results as a single driver for the DGVM, because of the potential for highly nonlinear (threshold) behaviour in large-scale vegetation patterns. For example, in the DGVM, several PFTs have winter temperature limits that represent each PFT’s characteristic mechanism of cold tolerance (Harrison et al., 2010). Tropical trees are assumed to be completely intolerant of frost, and so the expected effects of a large cooling include the elimination of tropical trees from regions in which they are found today because of the presence of frost in currently frost-free environments. This type of response could create discontinuities in aggregate values of variables, such as carbon storage, which the averaging of inputs to DGVM would conceal.


We compared simulated biome distributions at LGM, based on a consensus of LPX outputs among the four climate models, with the LGM palaeovegetation data from BIOME 6000 (Fig. 3). The data have been aggregated to the same set of biomes as the model output, using the definitions provided in Table 1. Simulated total biome areas, GPP and NPP, and carbon storages in biomass, detritus and soil organic matter were calculated for the PIH simulation (Table 2) and as mean values across the four LGM simulations (Table 3). (Note that the simulated values for soil carbon storage do not include additional carbon stored in peatlands or permafrost soils.)

Figure 3.

 Comparison of Land Processes and eXchanges (LPX)-simulated biome distribution at the last glacial maximum (LGM: top panel), based on LGM simulations from four climate models, with LGM biomes inferred from pollen and plant macrofossil records compiled by the BIOME 6000 project (bottom panel). The simulated biomes are shown as ‘uncertain’ at grid cells at which the four simulations do not yield the same biome.

Table 1.   Assignment of biomes classified by BIOME 6000 to the biomes mapped in Fig. 2
Land Processes and eXchanges (LPX) biome nameBIOME 6000 names
TundraTundra, erect dwarf shrub tundra, prostrate dwarf shrub tundra, cushion forb tundra, graminoid forb tundra, alpine grassland, moor
Shrub-tundraLow and high shrub-tundra
Dry grassland/shrublandTemperate xerophytic shrubland, xerophytic woods/scrub, temperate grassland, steppe, temperate grassland and xerophytic shrubland, cool grassland/shrubland, xerophytic shrubland, heathland
Boreal parklandCold deciduous forest
Temperate parklandTemperate evergreen needleleaf open woodland, open conifer woodland
Sclerophyll woodlandTemperate sclerophyll woodland and shrubland, dry sclerophyll forest/woodland, semi-arid woodland scrub
SavannaSavanna, tropical savanna, tropical deciduous broadleaf forest and woodland, tropical dry forest
Boreal forestCold evergreen needleleaf forest, taiga
Temperate forestCool evergreen needleleaf forest, cool conifer forest, cool mixed forest, cool-temperate rainforest, cool-temperate evergreen needleleaf and mixed forest, cold mixed forest, temperate evergreen needleleaf forest, temperate conifer forest, temperate deciduous broadleaf forest, temperate deciduous forest
Warm-temperate forestWarm-temperate broadleaf and mixed forest, warm-temperate broadleaf forest, warm-temperate rainforest, wet sclerophyll forest, warm mixed forest, broadleaved evergreen/warm mixed forest, warm evergreen forest
Tropical forestTropical evergreen broadleaf forest, tropical semi-evergreen broadleaf forest, tropical rainforest, tropical seasonal forest
Table 2.   Modelled land biosphere properties for the pre-industrial Holocene (PIH)a
BiomeLand area (Mm2)NPP, Pg C a−1 (g C m−2 a−1)GPP, Pg C a−1 (g C m−2 a−1)Biomass, Pg C (kg C m−2)Detritus, Pg C (kg C m−2)Soil, Pg C (kg C m−2)Total, Pg C (kg C m−2)
Above groundBelow groundTotalAbove groundBelow groundTotal
  1. GPP, gross primary production; NPP, net primary productivity.

  2. aCarbon cycle quantities are given as global values, and per unit area (in parentheses).

Total137.059.2 (432)124 (903)659 (4.8)112 (0.8)771 (5.6)142 (1.0)95 (0.7)236 (1.7)682 (5.0)1690 (12.3)
Tundra4.80.1 (11)0.1 (15)0 (0)0 (0)0 (0)1 (0.1)0 (0.1)1 (0.2)3 (0.6)4 (0.8)
Shrub-tundra5.81.6 (279)2.4 (408)3 (0.5)3 (0.5)6 (1.0)15 (2.6)11 (1.9)26 (4.5)77 (13.2)109 (18.7)
Desert17.40.5 (29)1.5 (87)0 (0)0 (0)0 (0)0 (0)1 (0)1 (0.1)3 (0.2)5 (0.3)
Dry grassland18.33.7 (202)8.8 (480)1 (0)1 (0.1)2 (0.1)3 (0.2)6 (0.3)9 (0.5)25 (1.4)37 (2.0)
Boreal parkland8.43.2 (377)5.7 (682)26 (3.0)17 (2.0)43 (5.0)23 (2.8)17 (2.0)40 (4.8)116 (13.8)199 (23.6)
Temperate parkland7.42.7 (364)5.9 (791)6 (0.8)4 (0.5)10 (1.3)5 (0.7)8 (1.1)13 (1.8)39 (5.2)62 (8.3)
Sclerophyll7.03.2 (462)7.5 (1060)3 (0.5)3 (0.4)6 (0.9)5 (0.7)7 (1.0)12 (1.7)33 (4.8)52 (7.3)
Savanna14.57.4 (515)17.9 (1240)7 (0.5)5 (0.3)12 (0.8)5 (0.3)7 (0.5)12 (0.8)34 (2.3)57 (4.0)
Boreal forest14.16.6 (469)14.6 (1030)159 (11.3)21 (1.5)180 (12.8)49 (3.5)19 (1.4)68 (4.8)194 (13.8)441 (31.3)
Temperate forest11.86.4 (543)15.7 (1340)103 (8.7)16 (1.3)118 (10.1)19 (1.6)10 (0.9)29 (2.5)85 (7.2)233 (19.8)
Warm-temperate forest4.63.3 (711)8.3 (1800)36 (0.8)7 (1.4)43 (9.3)4 (0.9)3 (0.6)7 (1.5)21 (4.5)71 (15.4)
Tropical forest23.020.5 (894)35.4 (1540)316 (13.7)36 (1.6)352 (15.3)12 (0.5)5 (0.2)18 (0.8)52 (2.3)422 (18.4)
Table 3.   Modelled land biosphere properties for the Last Glacial Maximum (LGM)a
BiomeLand area (Mm2)NPP, Pg C a−1 (g C m−2 a−1)GPP, Pg C a−1 (g C m−2 a−1)Biomass, Pg C (kg C m−2)Detritus, Pg C (kg C m−2)Soil, Pg C (kg C m−2)Total, Pg C (kg C m−2)
Above groundBelow groundTotalAbove groundBelow groundTotal
  1. NPP, net primary productivity; GPP, gross primary production.

  2. aCarbon cycle quantities are given as global values, and per unit area (in parentheses). All values that are greater for LGM than for the pre-industrial Holocene (PIH) are shown in bold.

Total132.041.0 (311)86.3 (655)340 (2.6)65 (0.5)405 (3.1)79 (0.6)68 (0.5)147 (1.1)520 (3.9)1070 (8.1)
Tundra17.80.1 (6)0.3 (19)0 (0)0 (0)0 (0)1 (0.1)1 (0.1)2 (0.1)90 (8.3)92 (8.5)
Shrub-tundra7.11.7 (241)2.9 (411)1 (0.1)1 (0.1)2 (0.3)16 (1.7)11 (1.6)27 (3.9)82 (12.9)111 (17.4)
Desert23.31.0 (43)2.4 (105)0 (0)1 (0)1 (0)1 (0)2 (0.1)2 (0.1)8 (0.4)11 (0.5)
Dry grassland21.44.1 (193)9.7 (453)1 (0.1)2 (0.1)3 (0.1)4 (0.2)8 (0.4)13 (0.6)39 (1.9)55 (2.6)
Boreal parkland7.62.2 (296)4.6 (611)8 (1.0)5 (0.7)13 (1.7)14 (1.9)13 (1.7)27 (3.6)80 (10.3)120 (15.6)
Temperate parkland6.02.2 (363)5.8 (959)6 (1.0)4 (0.7)10 (1.7)4 (0.7)5 (0.9)10 (1.6)28 (4.5)48 (7.6)
Sclerophyll9.64.1 (431)9.5 (989)6 (0.6)5 (0.5)10 (1.1)7 (0.7)9 (0.9)16 (1.6)45 (5.2)71 (8.1)
Savanna11.66.8 (591)14.8 (1280)10 (0.9)5 (0.3)15 (1.3)5 (0.4)7 (0.6)12 (1.0)34 (2.8)61 (5.0)
Boreal forest3.01.3 (423)3.0 (1020)31 (10.5)4 (1.4)36 (11.9)9 (3.1)4 (1.3)13 (4.4)38 (10.0)87 (22.8)
Temperate forest3.71.8 (467)5.7 (1500)31 (8.1)6 (1.6)36 (9.6)5 (1.2)2 (0.6)7 (1.8)20 (4.6)107 (14.2)
Warm-temperate forest2.11.4 (672)3.4 (1600)18 (8.7)3 (1.6)22 (10.2)2 (1.0)1 (0.6)3 (1.6)10 (0.4)36 (15.0)
Tropical forest18.914.3 (755)24.3 (1290)228 (12.1)29 (1.6)257 (13.6)10 (0.6)5 (0.3)15 (0.8)46 (2.3)318 (16.0)

Visual comparison of the LGM biome data and the simulated LGM vegetation maps (Fig. 3) shows agreement in major features. Both the model results and the data show the largest changes to have been in northern mid- to high latitudes, with dramatic reductions in the area occupied by boreal, temperate and warm-temperate forests at LGM. The lowland tundra region was greatly expanded at LGM, including vast areas in Eurasia at much lower latitudes than those at which lowland tundra is found today. There were also substantial, albeit less far-reaching, vegetation changes across the tropical regions. Savannas and grasslands encroached on the dry margins of today’s tropical forests in Amazonia, consistent with pollen evidence. Tropical forests were also partially replaced by savannas, sclerophyll woodlands or grasslands in Africa, southern China and, to a lesser extent, in South-East Asia, again consistent with pollen evidence (Fig. 3; see also Wurster et al., 2010). These retreats of tropical forest were partially compensated for in terms of the area occupied by the modelled occurrence of tropical forests on the considerable areas of exposed continental shelf. Although not documented in this dataset, there is additional pollen evidence from marine cores supporting the modelled expansion of tropical forests on to the continental shelf in South-East Asia (see, for example, Sun et al., 2000; Kershaw et al., 2001).

Total ice-free land area at LGM was reduced by 3.5% (Tables 2, 3). The reduction was relatively small because losses of available land because of ice-sheet extension were nearly compensated for by gains (mainly in the tropics) caused by lowered sea level. Overall, modelled global GPP and NPP were reduced by 30% (27–36% for GPP; 28–37% for NPP), and carbon storage in soils and vegetation by 37% (33–41%) or 620 Pg C (550–694 Pg C). The greater part of this modelled reduction comes from reduced biomass and detritus. Only 162 Pg C (112–196 Pg C) is from reduced soil carbon storage. The modelled total carbon storage reduction is similar to that obtained in an earlier study with LPJ (610 Pg C: Thonicke et al., 2005).

These changes can be compared with estimates based on global isotopic signals. The triple oxygen isotope method relies on the different modes of isotopic fractionation of the three isotopes in biological processes and in photochemical reactions involving exchanges with ozone in the stratosphere. Fractionation associated with primary production and respiration is mass dependent (so that fractionation against 18O is about twice as strong as fractionation against 17O), whereas fractionation in the stratosphere is mass independent. The resulting depletion of atmospheric 17O depends on the balance of biospheric vs stratospheric reactions and, with some assumptions, can be used to estimate total global GPP (Luz et al., 1999; Blunier et al., 2002; Luz & Barkan, 2005). The data for LGM, based on ice-core palaeoatmospheric measurements, indicate a reduction in global total GPP by 25–40% (Landais et al., 2007). This includes marine GPP, which is of a similar magnitude to terrestrial GPP and was probably not reduced at LGM (e.g. Bopp et al., 2003). The reduction in terrestrial GPP was therefore probably > 25–40%, and so the modelling procedure may even underestimate the magnitude of the reduction in GPP. Estimates of changes in terrestrial carbon storage were reviewed by Prentice & Harrison (2009), who noted that credible published estimates obtained with a variety of methods have continued to fall within the range 300–700 Pg C, as given by Bird et al. (1996). Thus, the isotopic analyses show changes of the same sign and, within their admittedly large uncertainties, generally similar magnitude to the LPX simulation of changes in total terrestrial carbon storage and GPP.

According to the model results, the tundra, desert, dry grassland/shrubland and sclerophyll biomes occupied larger areas at LGM than at PIH. By contrast, the area occupied by tropical forests was reduced by 18% (13–21%). Greater reductions were suffered by warm-temperate (54%, 42–61%), temperate (68%, 62–77%) and boreal (78%, 73–82%) forests. NPP per unit area, averaged over the (changing) distribution areas of each biome, was reduced, on average, by 16% (13–18%) for tropical forests, 5% (3–7%) for warm-temperate forests, 14% (11–19%) for temperate forests and 10% (7–13%) for boreal forests. By contrast, the modelled NPP of deserts and savannas (with a substantial C4 component) was increased at LGM because of reduced competition from C3 woody plants. The modelled increase in NPP of deserts was 46% (37–57%) and, of savannas, 15% (7–19%).

Vegetation carbon densities in forest biomes were generally reduced, but savanna, sclerophyll woodland, dry grassland and desert biomes showed increases. Soil carbon densities followed a qualitatively similar pattern to vegetation carbon densities with one notable exception: vegetation carbon density in the tundra biome was reduced by 46% (37–74%), whereas soil carbon density increased by a large and variable factor because of the strong inhibitory effect of low LGM temperatures on soil organic matter decomposition in mid- to high latitudes.

The modelled vegetation carbon density in tropical forests was reduced by 11% (10–13%), whereas soil carbon density was unchanged (−6 to +9%). This might be surprising at first sight (because NPP and litter production were reduced, which should lead to reduced soil carbon storage), but probably reflects a compensating effect of lower average temperatures, slowing soil decomposition. Although the tropical forest area was reduced, the modelled relative importance of tropical forests as a land carbon store was greater at LGM (30%; 28–32%) than at PIH (25%), primarily because of the major reductions in area of the other forest biomes.


Harrison & Prentice (2003) first used a process-based global biogeography model to show that the ecophysiological effects of CO2, acting through modification of the competitive equilibrium between woody plants and grasses – especially in the tropics, where the modelled negative effects of low CO2 on woody plant growth are greatest, whereas the effects on (C4) grasses are least – were required to account for the documented reduction of forest area at LGM. Effects of climate change alone were found to be inadequate, regardless of which model was used to provide the LGM climate. We have performed a comparable set of simulations here with more recent climate modelling results and using a DGVM, including both the CO2 and climate effects. We have compared the results with a more comprehensive set of pollen and plant macrofossil-based biome reconstructions (Fig. 3). Our results confirm that simulations including the effects of CO2 changes on tree–grass competition, as represented much more explicitly through dynamic competition and vegetation–fire interactions in DGVM, capture the broad features of the observed biome shifts (within the limitations imposed by the relatively sparse palaeoecological record from the tropics), and that they do so robustly despite differences in the LGM climate as represented by four AOGCMs (Fig. 3).

Thus, our results, when taken together with those of Harrison & Prentice (2003) using a model which simulates CO2 and climate effects on plant carbon and water exchanges identically to LPX, support the hypothesis that the contraction of tropical forest cover under LGM conditions was, to a large extent, a predictable outcome of the low CO2 concentration. This analysis is consistent with a CO2-based explanation for woody thickening today. There is also recent experimental evidence in favour of such an explanation (Kgope et al., 2010; Wigley et al., 2010), which can thus be considered both parsimonious and consistent with contemporary and palaeoecological evidence.

There remains a caveat that there are additional pressures that may favour woody thickening today, and so it may not be possible to attribute contemporary woody thickening uniquely to CO2 in any one case. However, this does not invalidate the general inference that rising CO2 is likely to be contributing to woody thickening in contemporary ecosystems, and that this process is likely to intensify as the CO2 concentration continues to rise.

By using a DGVM in this study, we have also been able to provide a complete global accounting of changes in GPP and carbon storage that is mutually consistent with the simulation of biome shifts. The modelled increases in GPP and carbon storage from LGM to PIH derive straightforwardly from the effect of low CO2 on C3 plant photosynthesis (Gerber et al., 2004). The magnitudes of the effects broadly agree with independent isotopic evidence. Extensive sensitivity experiments with the ‘parent’ LPJ DGVM (again, with an identical set-up to LPX for the simulation of CO2 and climate effects on productivity and water balance) show that GPP and carbon storage in the low (LGM to PIH) range of CO2 concentration respond much more strongly to CO2 than to commensurate changes in temperature (Gerber et al., 2004; see also simulations by Kaplan et al., 2002 and Joos et al., 2004). A parsimonious explanation thus attributes the LGM reductions in forest cover, production and carbon storage, especially in the tropics, primarily to downstream effects of photosynthetic physiology. The success of these simulations supports the implication that these same mechanisms are likely to be at work today.

There are important caveats regarding the time scale of the CO2 effects. The deglacial CO2 rise, and the associated increase in terrestrial biosphere carbon storage, took place over several thousand years. LPX does not model the effects of nitrogen limitation on plant productivity. The hypothesis of progressive nitrogen limitation (Luo et al., 2004) suggests that the near-term stimulatory effect of rising CO2 on NPP in temperate forest ecosystems [as observed over periods of several years in Free Air Carbon dioxide Enrichment (FACE) experiments, e.g. Norby et al., 2005] may decline on a multi-decadal time scale, as nitrogen availability falls behind the increasing plant demand for nitrogen set in train by the initial ecophysiological response to CO2. A decline in the CO2 fertilization effect and leaf nitrogen content during forest stand maturation has also been reported from a forest FACE study (Norby et al., 2010). It is plausible that nitrogen availability constraints to the CO2 response of plant and vegetation growth could more easily be overcome given thousands of years for a rebalancing of the total ecosystem nitrogen budget (e.g. through increased nitrogen fixation, natural nitrogen deposition and reduced nitrogen losses in the gaseous and aqueous phases). Our results do not rule out the possibility that nitrogen limitations constrain the realized CO2 response over time scales intermediate between the time scale of current FACE experiments and the much longer time scale of glacial–interglacial climate changes. Nor do they rule out a possible role of phosphorus limitation in constraining CO2 responses in tropical savannas and forests (e.g. Vitousek, 1984; Herbert & Fownes, 1995; Domingues et al., 2010). However, the evidence from LGM, as presented here, is inconsistent with the view espoused, for example, by Körner (2006) that the ecosystem-level response of carbon uptake and storage to CO2 concentration on long time scales must be negligible because of the constraint provided by the stoichiometry of plant biomass. Rather, the evidence suggests that a full biogeochemical analysis of ecosystem-level CO2 responses will depend on an improved understanding of the mechanisms that allow ecosystems to acquire and retain sufficient nutrients so that the photosynthetic response to CO2 can be translated into biomass growth.

At first glance, other models might be expected to behave in a generally similar way to LPX because of their common core of photosynthetic physiology. However, these models differ in many respects, such as whether (and, if so, how) they implement explicit nitrogen cycling constraints on carbon allocation and growth, whether carbon allocation to leaves, transport tissues and fine roots is influenced by CO2 concentration and/or nitrogen availability, their representation (or not) of fire–vegetation interactions and the particular assumptions they make about photosynthetic, respiratory and stomatal acclimation in the face of changes in CO2 concentration, temperature and water availability (Sitch et al., 2008). All of these differences could influence the modelled responses of ecosystem composition and carbon storage to CO2 changes. There is growing interest in the application of standard benchmarks for land models, exploiting a range of observations of the land surface and the carbon cycle in an attempt to narrow down some of the large uncertainties in carbon cycle modelling (Randerson et al., 2009). So far, these efforts have focused on contemporary observations, such as the seasonal cycles and interannual variability of atmospheric CO2, ‘greenness’ measures and runoff (see, for example, Abramowitz, 2005; Blyth et al., 2010). The use of multiple climate model results as a source of palaeoclimatic information for key times in the past, such as LGM, opens up new possibilities for the evaluation of terrestrial models under a more broadly inclusive range of environmental conditions (Abe-Ouchi & Harrison, 2009). Mid-Holocene and LGM climate model experiments, as developed by PMIP, are now designated as a priority for the IPCC Fifth Assessment Report, and will be included in the public archive of climate model results generated for IPCC. We suggest that the LGM modelling experiment described here could provide an additional valuable benchmark for terrestrial models, building on both PMIP and the solid body of evidence that now exists for the state of the biosphere at LGM.


This article is the outcome of an invitation to I.C.P. to give a keynote presentation at the 23rd New Phytologist Symposium in Guangzhou, China, in November 2009, which is gratefully acknowledged. It has benefited from the additional opportunity provided by Rich Norby and Ram Oren for I.C.P. to present these results at a symposium on FACE data–model comparisons at the Ecological Society of America Meeting, Pittsburgh, in August 2010. We thank Doug Kelley for performing the LPX model runs, Wang Han for assistance with GIS manipulations and mapping the model outputs and biome data, and the PMIP modelling groups for their work in developing and implementing simulation protocols and making their outputs freely available for research. The work has benefited from discussions with many colleagues, especially Philippe Ciais and Pierre Friedlingstein, and from the Palaeo Carbon Modelling Intercomparison Project (PCMIP: Abe-Ouchi & Harrison, 2009) Workshop sponsored by the Quantifying and Under-standing the Earth System (QUEST) programme of the UK Natural Environment Research Council (NERC). PCMIP is an activity of the International Geosphere–Biosphere Programme’s Analysis, Integration and Modelling of the Earth System (AIMES) core project, and is co-ordinated by Ayako Abe-Ouchi, Pierre Friedlingstein, S.P.H. and I.C.P.