Fire and fire-adapted vegetation promoted C4 expansion in the late Miocene

Authors


Author for correspondence:
Simon Scheiter
Tel: +49 (0)69 798 40167
Email: scheiter@em.uni-frankfurt.de

Summary

  • Large proportions of the Earth’s land surface are covered by biomes dominated by C4 grasses. These C4-dominated biomes originated during the late Miocene, 3–8 million years ago (Ma), but there is evidence that C4 grasses evolved some 20 Ma earlier during the early Miocene/Oligocene. Explanations for this lag between evolution and expansion invoke changes in atmospheric CO2, seasonality of climate and fire. However, there is still no consensus about which of these factors triggered C4 grassland expansion.
  • We use a vegetation model, the adaptive dynamic global vegetation model (aDGVM), to test how CO2, temperature, precipitation, fire and the tolerance of vegetation to fire influence C4 grassland expansion. Simulations are forced with late Miocene climates generated with the Hadley Centre coupled ocean–atmosphere–vegetation general circulation model.
  • We show that physiological differences between the C3 and C4 photosynthetic pathways cannot explain C4 grass invasion into forests, but that fire is a crucial driver. Fire-promoting plant traits serve to expand the climate space in which C4-dominated biomes can persist.
  • We propose that three mechanisms were involved in C4 expansion: the physiological advantage of C4 grasses under low atmospheric CO2 allowed them to invade C3 grasslands; fire allowed grasses to invade forests; and the evolution of fire-resistant savanna trees expanded the climate space that savannas can invade.

Introduction

Almost 30% of the Earth’s land surface is dominated by grasses that use the C4 photosynthetic pathway (Grace et al., 2006). Current knowledge suggests that C4 grasses rapidly invaded and replaced large areas of C3 grasslands and forests from the late Miocene into the Pliocene, 3–8 million years ago (Ma) (Cerling et al., 1993, 1997; Ehleringer et al., 1997). However, both molecular evidence (Christin et al., 2008; Vicentini et al., 2008) and fossil evidence (Jacobs et al., 1999; Strömberg, 2011) suggest that C4 plants evolved and diversified in multiple grass lineages some 20 Ma earlier during the early Miocene/Oligocene. Explanations for this time lag have invoked the physiological advantage of C4 plants over C3 plants at high temperature and low atmospheric CO2 concentrations (Ehleringer et al., 1997). Specifically, it has been hypothesized that atmospheric CO2 fell below a critical threshold in the late Miocene, and thereby triggered C4 dominance and expansion (Cerling et al., 1997; Ehleringer et al., 1997). However, evidence is now accumulating that atmospheric CO2 concentrations did not decline substantially in the period between the origin and expansion of C4 grasses (Beerling & Royer, 2011). Indeed, it appears that the critical CO2 threshold had already been crossed during the Oligocene (25–30 Ma), which implies that CO2 may have been a selective force for the evolution of C4 photosynthesis, but did not drive the assembly of the C4 grassland biome (Beerling & Osborne, 2006).

Alternative explanations for the C4 expansion have invoked changes in the seasonality of climate and in fire regimes (Keeley & Rundel, 2003, 2005; Beerling & Osborne, 2006; Osborne, 2008). It has been hypothesized that both the uplift of the Tibetan Plateau c. 8 Ma ago and the reduction of the Paratethys Sea intensified the seasonal Indian monsoon climate, primarily in South Asia (Ramstein et al., 1997; Zhisheng et al., 2001; Zhang et al., 2007), and that the more seasonal climate favored grasses over trees (Beerling & Osborne, 2006). That is, the warm wet season promotes abundant grass production, whereas the dry season increases tree sapling mortality and lowers the moisture content of dead biomass, which allows more frequent and intense fires. The involvement of fire is well supported by charcoal depositions in marine sediments, indicating increases in fire activity in the late Miocene (Herring, 1985; Morley & Richards, 1993; Jia et al., 2003; Keeley & Rundel, 2003, 2005). Because it is known that tropical forest trees have high susceptibility, even to low-intensity fires (Cochrane et al., 1999; Cochrane, 2003), the incidence of fires may have led to the retreat of forests. In addition, low atmospheric CO2 concentrations would have reduced tree growth rates, making trees more susceptible to fire (Bond & Midgley, 2001). Such seasonal environments may have selected for fire- and drought-tolerant tree types that invest more carbon to storage compartments and fire protection (Hoffmann et al., 2003, 2005; Ratnam et al., 2011). Hence, it is likely that fire-resistant and drought-tolerant trees played a role in the Miocene expansion of C4 biomes.

It has been argued that C4 grasses benefit more strongly than C3 grasses from seasonal and open environments because C4 grasses have higher water use efficiencies, higher temperature optima for photosynthesis and are shade intolerant, with photosynthesis saturating at higher irradiance (Ehleringer et al., 1997; Knapp & Medina, 1999). However, Edwards et al. (2010) argued that the success of C4 grasses is only indirectly linked to leaf physiology, and that other traits must be invoked to explain C4 expansion. Candidate traits are related to fire. For example, Ripley et al. (2010) observed that foliage of the C4 subspecies of Alloteropsis semialata is more flammable than that of the C3 subspecies, and that the C4 type invests more carbon in belowground biomass compartments that facilitate rapid regrowth after disturbances such as fire. However, Martin (2010) observed that flammability is more closely linked to the phylo-genetic lineage than to the photosynthetic pathway. Hence, the extent to which the evolution of fire-promoting traits in grasses influenced the potential of grasses to invade forests remains uncertain.

In this study, we test the conditions that promote the invasion of C4 ecosystems into C3 ecosystems using a dynamic vegetation model (the adaptive dynamic global vegetation model (aDGVM); Scheiter & Higgins, 2009). The aDGVM imitates ecophysiological processes, competition and demography to simulate the dynamics of C3 and C4 grasses and fire-tolerant and fire-sensitive trees. In addition to differences in leaf physiology, the aDGVM considers differences in the fire and shade tolerance of these functional types. The forcing variables explored are CO2, temperature and fire. We use late Miocene climate simulations from the Hadley Centre coupled ocean–atmosphere–vegetation general circulation model HadCM3L for 180, 280 and 400 ppm CO2 (C. Bradshaw et al., unpublished; Beerling et al., 2012). These CO2 concentrations span current estimates of Miocene atmospheric CO2 concentration (Pagani et al., 2009; Beerling & Royer, 2011) and allow us to explore how CO2 concentrations influence C4 expansion. We test four specific hypotheses: (1) the outcome of asymmetric competition driven by physiological differences between C3 and C4 photosynthesis can explain C4 expansion in the late Miocene; (2) fire is a crucial driver of C4 expansion; (3) fire-promoting traits in grasses influence the potential of C4 grasses to invade forests; and (4) fire-resistant savanna trees influence the range of environmental conditions that support C4-dominated biomes. In this study, we only conduct simulations for Africa in which the model has been parameterized and previously evaluated (Scheiter & Higgins, 2009).

Description

Model description

We use the aDGVM (see Scheiter & Higgins, 2009 for a full model description), a dynamic vegetation model for tropical grass–tree systems. The aDGVM integrates plant physiological processes generally used in dynamic global vegetation models (DGVMs, Prentice et al., 2007) with novel processes that allow plants to dynamically adjust leaf phenology and carbon allocation to environmental conditions. The aDGVM projects, for instance, longer growing seasons in response to anticipated climate change and higher carbon allocation to root biomass in more water-limited environments (Scheiter & Higgins, 2009). The aDGVM is individual based, which means that it keeps track of state variables, such as the biomass, height and photosynthetic rates of individual plants. This approach allows the impacts of herbivores (Scheiter & Higgins, 2012) and fire (Scheiter & Higgins, 2009) on vegetation structure and demography to be modeled as a function of plant height. Each grass type (C3 or C4) is represented by two super-individuals, representing grasses beneath or between tree canopies. The aDGVM only requires generally available environmental input data and typically simulates vegetation in 1-ha stands.

In the aDGVM, fire intensity is modeled as a function of fuel loads, fuel moisture and wind speed (Higgins et al., 2008). Fire spreads when the fire intensity exceeds a threshold value of 300 kJ m−1 s−1, the daily fire ignition probability pfire (1%) is exceeded and an ignition takes place. Ignition sequences, which indicate when ignitions take place, are randomly generated. This fire model ensures that fire regimes are influenced by fuel biomass and climate; however, fire ignitions and the ignition probability are not linked to environmental conditions. Fire removes aboveground grass biomass, whereas the response of trees to fire is a function of tree height and fire intensity (the ‘topkill’ effect; Higgins et al., 2000). Seedlings and juveniles (< 2 m) in the flame zone are damaged by each fire, whereas adult trees (> 2 m) are almost fire resistant and are only damaged by intense fires. Grasses and topkilled trees can regrow from root reserves after fire (Bond & Midgley, 2001). Fire influences tree mortality indirectly by shifting the carbon balance. In the aDGVM, a negative carbon balance increases the probability of mortality.

The performance of the aDGVM was evaluated by Scheiter & Higgins (2009), where it was shown that the aDGVM can simulate the current distribution of vegetation in Africa better than alternative dynamic vegetation models, and that the aDGVM can simulate biomass observed in a long-term fire manipulation experiment in the Kruger National Park (Experimental Burn Plots; Higgins et al., 2007).

The original version of the aDGVM, as described by Scheiter & Higgins (2009), only simulates fire-resistant savanna trees and C4 grasses. For this study, we introduce two additional plant functional types: fire-sensitive forest trees and C3 grasses. As we aim to test how fire influences C4 grassland expansion, the parameterization of plant functional types is mainly based on differences in fire tolerance. Compared with the original version of the model, these changes improve slightly the agreement between simulated biomes and biome distribution patterns obtained from remote sensing (Supporting Information Table S1). Figure S1 shows the biome distribution patterns simulated for current climate conditions (CRU data; New et al., 2002).

Tree types

Forest and savanna trees differ in their responses to shading and fire (Bond, 2008; Ratnam et al., 2011). The model forest tree type is assumed to be shade tolerant, which allows recruitment and growth in dense forest stands in which light availability is low. In addition, high carbon allocation to aboveground biomass compartments and a higher specific leaf area (SLA) allow forest trees to capture light efficiently. By contrast, the model savanna tree type is assumed to be shade intolerant, which means that it cannot recruit, regenerate and grow rapidly in shade, instead prefering the high light availability of nonforested, open landscapes.

The savanna tree type is assumed to be fire tolerant, whereas the forest tree type is assumed to be fire sensitive. Savanna trees typically allocate more carbon to bark and nonstructural carbon reserves. A thick bark is known to reduce fire injury (Hoffmann et al., 2003), and belowground carbon reserves allow savanna trees to resprout vigorously after the removal of aboveground biomass by fire (Bond & Midgley, 2001). By contrast, forest trees have a thinner bark and often have lower resprouting vigor, traits that make them susceptible, even to low-intensity fires (Cochrane et al., 1999; Cochrane, 2003). In the aDGVM, we simply assume that forest trees have higher topkill probabilities and lower resprouting probabilities than savanna trees, as we do not explicitly simulate bark and storage compartments.

The assumed differences in shade tolerance and fire response between forest and savanna trees create a trade-off. Forest trees can survive in undisturbed and dense forest stands, but cannot survive in more fire-driven systems because of their low resprouting vigor. By contrast, savanna trees are outcompeted in dense tree stands because of their shade intolerance, but they can survive in more open, fire-driven systems. Details of how forest and savanna trees are parameterized in the aDGVM are provided in Table 1.

Table 1.   Parameterization of forest and savanna trees in the model (differences between these tree types are based on Ratnam et al., 2011).
ParameterForest treeSavanna treeHypothesis
Allocation to rootsLow, 0.15High, 0.35Forest trees allocate more carbon to aboveground growth, whereas savanna trees allocate more carbon below ground. High root : shoot ratios and belowground storage have been observed in savanna tree saplings (Hoffmann et al., 2005). This difference in allocation has, in the model, the consequence that forest trees can grow taller than savanna trees
SLAHigher, 12 m2 kg−1Lower, 10 m2 kg−1References in: Prior et al. (2003); Hoffmann et al. (2005); Rossatto et al. (2009);  Ratnam et al. (2011)
Shade toleranceHighLowForest trees can establish, regenerate and survive in dense forest stands; savanna trees cannot. The parameterization of the light competition model is changed to reflect this
Topkill probabilityHighLowSavanna trees invest more in fire protection, whereas forest trees do not and can be topkilled by low-intensity fires
Resprouting probabilityLow, 75%High, 99%Savanna trees have a high resprouting probability after fire (Hoffmann et al., 2005)
Canopy radius to height ratioHigher, 0.37Lower, 0.32Savanna trees have narrower canopy diameters for a given basal area than forest trees (Hoffmann et al., 2005)

Grass types

The primary difference between model C3 and C4 grasses is the difference in leaf physiology. In addition, grasses can differ in traits related to fire. There are, however, surprisingly few data on the differences in flammability and fire response of different grasses. Ripley et al. (2010) observed more dead biomass, a lower moisture content and higher flammability in C4 than C3 subspecies of Alloteropsis semialata, suggesting that these C4 grasses promote fire. For certain lineages, particularly the Panicoideae subfamily, similar findings have been reported for a larger sample of closely related C3 and C4 grasses drawn from a southern African regional species pool (Martin, 2010). Flammability is, however, not restricted to C4 grasses and has also been observed in various C3 grasses (e.g. Arundo, Coffman et al., 2010; and Ampelodesmos, Grigulis et al., 2005). Martin (2010): concluded that the flammability of grasses is primarily defined by the phylogenetic lineage and not by the photosynthetic pathway.

In the model, differences in fire-related traits are aggregated by simply assuming different curing rates of grasses at the end of the growing season. More specifically, we simulate that rapid-curing grasses cure by 80% within the first 30 d after the end of the growing season (Cheney & Sullivan, 1997) and that slow-curing grasses cure by 20% under the same conditions. This assumption ensures that rapid-curing grasses produce drier fuel and therefore higher fire intensities than slow-curing grasses. To test the sensitivity of vegetation patterns to fire-related traits, we conduct simulations with different curing rates of C3 and C4 grasses (see section on ‘Simulation experiments’). We assume as a baseline scenario that C4 grasses of open, tropical and subtropical environments cure rapidly, whereas C3 grasses cure slowly. It should be noted that the differences between the C3 and C4 photosynthetic pathways only influence fire behavior indirectly by producing different amounts of fuel biomass.

C3 and C4 grasses also differ in their shade tolerance. C4 grasses are typically more shade intolerant than C3 grasses and are excluded from dense forest stands (Ehleringer, 1978; Pearcy & Ehleringer, 1984; Osborne & Freckleton, 2009). By contrast, the tropical C3 ancestors of C4 grasses probably survived under dense tree stands (Edwards & Smith, 2010), and a number of modern tropical C3 lineages remain shade specialists (Osborne & Freckleton, 2009). In the aDGVM, light competition parameters are modified for C3 and C4 grasses to mimic these differences in shade tolerance.

Biome classification

We classify vegetation into biome types because this aggregated variable simplifies the interpretation of simulation results. The classification scheme is summarized in Fig. S2. Low grass biomass together with low tree cover defines the desert biome. When the tree cover is low and the grass biomass exceeds a threshold, the vegetation, depending on the ratio of C3 to C4 grasses, is classified as C3 or C4 grassland. At intermediate tree cover, the ratio of C3 to C4 grasses and the cover of savanna trees are used for the classification. Vegetation is classified as a woodland when the forest tree cover exceeds the savanna tree cover, whereas vegetation is classified as forest when the tree cover exceeds 80%. Savannas occur when the savanna tree cover exceeds the forest tree cover and C4 grasses dominate over C3 grasses. In addition, we define a ‘C3 savanna’ biome in which the savanna tree cover exceeds the forest tree cover and C3 grasses dominate.

Late Miocene climatology

We used climatologies simulated for the late Miocene by the Hadley Centre coupled ocean–atmosphere–vegetation general circulation model HadCM3L (C. Bradshaw et al., unpublished; Beerling et al., 2012). Late Miocene paleogeography is based on the Markwick (2007) reconstruction and the model was integrated over 2100 yr. Climate parameters were estimated by taking the mean of the last 50 yr of the simulations. Climate was simulated at three different CO2 concentrations, 180, 280 and 400 ppm, to span the range of uncertainty in Miocene atmospheric CO2 concentrations estimated from proxy data (Freeman & Hayes, 1992; Kürschner et al., 1996, 2008; Pagani et al., 1999a,b, 2009; Pearson & Palmer, 2000; Demicco et al., 2003). For each CO2 concentration, we used monthly means for temperature, precipitation and relative humidity. These data were given on a 3.75 × 2.5 grid. Temperature and precipitation means for Africa are provided in Table 2. Climate reconstructions for increasing CO2 concentrations from 180 to 400 ppm generate a decrease in mean annual precipitation (MAP) from 653 to 589 mm, whereas the mean temperature increases from 22.2 to 28.6°C (Table 2).

Table 2.   Mean precipitation and temperature for Africa simulated by the Hadley Centre coupled ocean–atmosphere–vegetation general circulation model (GCM) for the late Miocene at three different CO2 concentrations (180, 280, 400 ppm)
ScenarioIndexCO2 concentration (ppm)Precipitation (mm yr−1)Temperature (°C)
  1. For comparison, conditions for the reference period between 1961 and 1990 (New et al., 2002) are provided (current conditions).

Late Miocene 1118065322.2
Late Miocene 2228062725.6
Late Miocene 3340058928.6
Current conditions438762927.4

Simulation experiments

The following paragraphs describe the simulation experiments used to test the hypotheses presented in the Introduction. For each hypothesis, we conduct simulations at the continental scale (summary provided in Table 3). We initialize the model without fire and C4 grasses to imitate a C4-free state in the Oligocene/early Miocene. To suppress fire, we simply set the ignition probability pfire to zero because, in the aDGVM, pfire and fire ignitions are not linked to environmental conditions. After this model spin-up, C4 grasses and fire are introduced and the change in the biome distribution is recorded. Vegetation is classified using the classification scheme shown in Fig. S2. Simulations are repeated for the three late Miocene climatologies to test how C4 expansion is related to CO2 conditions. For hypotheses 2 and 4, we additionally conduct sensitivity analyses at three study sites. These sensitivity analyses do not use the late Miocene climatologies.

Table 3.   Hypotheses tested in this study and summary of simulation experiments.
  Simulation experiment
 HypothesisActorsSpin-up1st period2nd period
  1. The spin-up phase consists of years 1–150, the first period of years 151–600 and the second period of years 601–900. The vegetation state after 600 yr is also denoted as the ‘transient state’, and the vegetation state after 900 yr is denoted as the ‘equilibrial state’. The ‘+’ sign indicates that simulations were conducted in the presence of the actor.

Hypothesis 1Physiological differences between C3 and C4 photosynthesis can explain C4 expansion in the late MioceneC3 grass, slowly curing+++
C4 grass, rapidly curing++
Savanna tree type
Forest tree type+++
Fire
Hypothesis 2Fire triggers C4 expansion
 Case 1: Fire introduced before C4 grassesC3 grass, slowly curing+++
C4 grass, rapidly curing+
Savanna tree type++
Forest tree type+++
Fire++
 Case 2: Fire introduced after C4 grassesC3 grass, slowly curing+++
C4 grass, rapidly curing++
Savanna tree type++
Forest tree type+++
Fire+
 Case 3: Fire and C4 grasses introduced simultaneouslyC3 grass, slowly curing+++
C4 grass, rapidly curing++
Savanna tree type++
Forest tree type+++
Fire++
Hypothesis 3Fire-promoting traits in grasses influence C4 expansion
 Case 1: C3 and C4 grasses cure rapidlyC3 grass, rapidly curing+++
C4 grass, rapidly curing++
Savanna tree type++
Forest tree type+++
Fire++
 Case 2: Only C3 grasses cure rapidlyC3 grass, rapidly curing+++
C4 grass, slowly curing++
Savanna tree type++
Forest tree type+++
Fire++
 Case 3: C3 and C4 grasses cure slowlyC3 grass, slowly curing+++
C4 grass, slowly curing++
Savanna tree type++
Forest tree type+++
Fire++
Hypothesis 4The savanna tree type extends the range of environmental conditions in which C4 biomes occurC3 grass, slowly curing+++
C4 grass, rapidly curing++
Savanna tree type
Forest tree type+++
Fire++

Hypothesis 1: to test whether physiological differences between C3 and C4 photosynthesis can explain C4 expansion in the late Miocene, we conduct invasion experiments for Africa in the absence of fire (pfire = 0%). After a 150-yr model spin-up in the absence of C4 grasses, C4 grasses are introduced and simulations are conducted for a further 900 yr.

Hypothesis 2: to test whether fire is necessary for C4 expansion, we conduct invasion experiments for Africa in the presence of fire. To test whether the distribution of C4 biomes is influenced by the order of how fire and C4 grasses are introduced, we consider three different cases: (1) fire was in the system before C4 grasses evolved; (2) fire ignition probability increased after C4 evolution; and (3) fire and C4 grasses appeared synchronously. For case (1), we run a 150-yr model spin-up in the absence of fire (pfire = 0%) and C4 grasses. After the spin-up, fire is switched on (pfire = 1%) and, after a further 450 yr, C4 grasses are introduced. Simulations are conducted for a further 450 yr. For case (2), we run a 150-yr model spin-up in the absence of C4 grasses and fire. After the spin-up, we introduce C4 grasses and, after a further 450 yr, fire. For case (3), we run a 150-yr model spin-up, and then introduce both fire and C4 grasses, and run the model for a further 900 yr.

For three specific study sites, we investigate the C4 invasion process in more detail and examine whether fire-driven C4 biomes can occupy the environmental space in which the model projects C4 dominance. We conduct these simulations for a savanna study site in the Kruger National Park, South Africa (25°S, 31°35′E, 570 mm MAP, hereafter ‘low-rainfall site’), for a study site at the savanna–forest border in the Digya National Park, Ghana (7°30′N, 0°E, 1320 mm MAP, hereafter ‘intermediate-rainfall site’) and for a rainforest study site in the Salonga National Park, Democratic Republic of the Congo (2°S, 20°E, 1930 mm MAP, hereafter ‘high-rainfall site’). For each study site, simulations are conducted for a range of different CO2 concentrations and temperatures spanning conditions from the late Miocene into the future as projected by the IPCC (2007), holding MAP constant at the values given above. More specifically, CO2 concentrations are varied from 100 to 1000 ppm in 25-ppm steps and temperatures are varied from − 6°C to + 6°C in 0.5°C steps, relative to the site-specific temperature as provided by New et al. (2002). We run a 150-yr spin-up in the absence of C4 grasses and fire, and then introduce C4 grasses and fire. Vegetation is simulated for a further 1000 yr to ensure that vegetation is in equilibrium with the environmental conditions. We record the biome type, the fractional cover of forest and savanna trees, the C3 to C4 grass ratio and fire frequency.

Hypothesis 3: to test whether fire-promoting traits in grasses influence their potential to invade woodlands and forests irrespective of the photosynthetic pathway, we conduct invasion experiments for Africa: (1) where both C3 and C4 grasses cure rapidly (and thereby both promote fire); (2) where C3 grasses cure rapidly, whereas C4 grasses cure slowly; and (3) where both C3 and C4 grasses cure slowly. The case in which C4 grasses cure rapidly, whereas C3 grasses cure slowly, is used to test hypotheses 1 and 2. We run a model spin-up for 150 yr without C4 grasses and without fire, and then fire and C4 grasses are introduced. Simulations are conducted for a further 900 yr.

Hypothesis 4: to test whether fire-tolerant savanna trees extend the range of environmental conditions in which C4 biomes exist, we simulate the biome distribution in Africa in the absence of the savanna tree type. We run a model spin-up for 150 yr in the absence of fire and C4 grasses, and then fire and C4 grasses are introduced.

To illustrate the role of the savanna tree type in more detail, we simulate the probability of C4 dominance along a CO2 gradient for the three specific study sites (low, intermediate and high rainfall) in the presence and absence of the savanna tree type. Within a simulation run, CO2 is increased from 190 to 800 ppm within 350 yr, and then decreased from 800 to 190 ppm within 350 yr; that is, transient dynamics are considered. This range spans the CO2 concentrations from the Miocene to those projected by the IPCC (2007) to occur in 2100. We conduct 100 simulation runs for each study site to account for the stochastic effects of fire and rainfall. For each study site, we calculate the fraction of simulation runs in which vegetation is classified as a C4-dominated biome.

Model evaluation

Pound et al. (2011) provide paleobotanical data for the late Miocene. These data are classified into biome types generally used in dynamic vegetation models. We reclassified the biome types used by Pound et al. (2011) into the biome types used in this study (Table S2). We used 19 sites from this dataset for a data–model comparison.

Results

Assuming a late Miocene climate, no C4 grasses and no fire, the model projects that the vegetation cover in Africa is dominated by woodlands and forests, covering c. 50% of the land surface (Fig. 1a–c). C3 grasslands cover c. 12% and deserts cover c. 36%. In these cases, the biome boundaries are defined by precipitation. At high-rainfall sites, woody vegetation outcompetes grasses. Tree dominance decreases as precipitation decreases. Low-rainfall sites are dominated by grasses and arid sites are deserts. Thus, 90% of the woodlands and forests are found at sites that receive > 240 mm MAP, 80% of the grasslands are found between 65 and 380 mm MAP and 90% of the deserts are found at sites that receive < 100 mm MAP.

Figure 1.

Biome cover in simulations designed to test hypotheses 1 and 2. Simulations were initialized without C4 grasses and fire (a–c), and then C4 grasses (d–f, hypothesis 1) or C4 grasses and fire (g–i, hypothesis 2, case 1) were introduced. The maps show vegetation at the end of the 900 yr simulation period (here denoted as ‘equilibrial state’, Table 3). The numbers in the graphs show the percentage of grid cells covered by different biome types; the arrows indicate transitions between the scenario without C4 grasses and fire (a–c) and the two other scenarios. The width of the arrows represents the number of transitions. Triangles in (g–i) represent paleovegetation as provided by Pound et al. (2011).

This biome pattern is relatively insensitive to the CO2 concentration, because, in the late Miocene climatologies, precipitation decreases as CO2 increases (Table 2). Hence, higher vegetation growth rates at higher CO2 concentrations are counteracted by low precipitation. However, a sensitivity analysis in which CO2 concentrations were modified, with precipitation and temperature held constant, produced a strong vegetation response, because of increasing water use efficiency at higher CO2 concentrations (higher photosynthetic rate, lower stomatal conductance). In this sensitivity analysis, the areas covered by deserts in the Sahel/Sahara region (north of 10°N) are 72%, 63% and 55% for CO2 concentrations of 180, 280 and 400 ppm, respectively.

Introducing C4 grasses does not change this pattern dramatically, except that large proportions of the C3 grasslands shift to being either mixed C3–C4 grasslands or C4-dominated grasslands because of the physiological advantages of C4 grasses (Fig. 1d–f). However, C4 grasslands cannot invade woodlands or forests because the modeled C4 grasses are shade sensitive and cannot survive in the dense forest stands. Hypothesis 1 (Table 3) is therefore not supported.

Fire dramatically changes the vegetation distribution and allows C4 grasslands and savannas to invade woodlands and forests (Fig. 1g–i). This result supports hypothesis 2 (Table 3). The area covered by C4 grasslands and savannas, depending on the CO2 concentration, is between 40% and 54%. The area covered by woodlands and forests decreases to values between 11% and 24% in the presence of fire. Differences in the atmospheric CO2 concentration influence the biome distribution, with a trend towards more woodlands and forests at higher CO2 concentrations. Differences in CO2 concentrations do not, however, change the fundamental result that C4-dominated vegetation can only invade woodlands and forests in the presence of fire.

The simulated biome distribution broadly agrees with paleovegetation data provided by Pound et al. (2011). Most data points are correctly reproduced by at least one simulation scenario (with or without fire, 180, 280 or 400 ppm; Fig. 1g–i). However, there is no perfect agreement between paleodata and one specific simulation scenario. The best agreement is achieved in simulations with fire at 280 ppm CO2, where 26% of the data points are simulated correctly.

The equilibrium distribution of C4-dominated biomes is not influenced by whether fire or C4 grasses are introduced first. Pairwise comparison of the equilibrium distributions in scenarios in which C4 grasses are introduced before fire, fire is introduced before C4 grasses or fire and C4 grasses are introduced synchronously shows high agreement (kappa values between 0.73 and 0.8). The transient states do, however, differ strongly (Fig. 2). When C4 grasses and fire are introduced synchronously, C4 biomes quickly reach the equilibrium state (Fig. 2a–c). When C4 grasses appear earlier than fire, C4 grasses invade C3 grasses, but the invasion of woodlands and forests is delayed until fire is introduced (Fig. 2d–f). When C4 grasses are introduced after fire, the model projects that c. 16% of the land surface is covered by C3 grasslands and that 2–14% is covered by C3 savannas, that is, by a mixture of C3 grasses and savanna trees (Fig. 2g–i). However, because of the physiological advantage of C4 grasses, these C3 grasslands and C3 savannas are immediately invaded by C4 grasses as soon as they are introduced.

Figure 2.

Vegetation states after 600 yr (here denoted as ‘transient state’, Table 3) for different ways in which fire and C4 grasses were introduced into the tests of hypothesis 2 (Table 3): C4 grasses and fire synchronously (a–c, case 3); C4 grasses then fire (d–f, case 2); fire then C4 grasses (g–i, case 1). The graphs indicate transitions between the scenario without C4 grasses and fire and the three different scenarios.

The potential of C4 grasslands and savannas to invade woodlands and forests depends on site-specific climatic conditions (Fig. 3). At lower CO2 concentrations and higher temperatures, C4 grasses rapidly invade C3 grass communities. C4 grasses, in turn, increase fire activity and thereby facilitate the invasion of C4 grasses into woodlands and the replacement of the fire-sensitive forest tree type with the fire-resistant savanna tree type. At higher CO2 concentrations and lower temperatures, C3 plants (both grasses and trees) have a competitive advantage over the C4 grasses, and the C4 grasses and savanna trees cannot invade woodlands and forests.

Figure 3.

Results from invasion experiments for different CO2 and temperature levels: (a) at a low-precipitation study site in the Kruger National Park (South Africa, 570 mm MAP) and (b) at an intermediate-precipitation study site in Ghana (1320 mm MAP). Simulations were initialized without C4 grasses, and then C4 grasses were introduced and vegetation was simulated for 1000 yr. Numbers (1–4) in the panels ‘Biome type’ indicate the different climatic conditions as provided in Table 2; the black lines indicate the simulated boundary between C3 and C4 grass dominance.

The boundary between C3 and C4 biomes, and the ability of fire-driven C4 biomes to occupy the whole environmental space in which C4 photosynthesis has a physiological advantage over C3 photosynthesis, is influenced by precipitation. At the low-rainfall site, C4 biomes dominate under environmental conditions in which C4 photosynthesis has a physiological advantage (Fig. 3a). At the intermediate-rainfall site, C3 biomes may dominate under environmental conditions in which C4 photosynthesis has a physiological advantage over C3 photosynthesis (Fig. 3b). At the high-rainfall site, C4 grasses and savannas cannot establish and rainforests persist under all combinations of CO2 and temperature (not shown).

We now turn to the hypothesis that the fire-promoting traits of grasses influence the potential of C4 biomes to invade woodlands and forests (hypothesis 3, Table 3). When both C3 and C4 grasses cure rapidly, the areas covered by different biomes only change by 1–2% compared with the situation in which only C4 grasses cure rapidly (Figs 1g–i, 4a–c, kappa values between 0.76 and 0.82), simply because C4 grasses dominate over C3 grasses; that is, the differential curing rates of C3 and C4 species, assumed when testing hypothesis 2, do not affect the outcome. When it is assumed that C4 grasses cure slowly, whereas C3 grasses cure rapidly, C4 grasses can still expand into woodlands and forests. However, the area covered by savannas is reduced from between 22% and 28% to c. 18% for different CO2 concentrations (Fig. 4d–f). Therefore, the area covered by woodlands and forests increases to between 22% and 28%. When both C3 and C4 grasses cure slowly, the biome distribution is similar to the situation in which only C3 grasses cure rapidly (not shown). These findings support the hypothesis that fire-promoting traits of grasses influence C4 expansion. This pattern is also influenced by the physiological advantage of C4 grasses under the late Miocene climatic conditions. C4 grasses produce more biomass than C3 grasses, and thereby promote fire activity at the continental scale (area burnt, mean fire intensity and mean fire frequency, Fig. 5).

Figure 4.

Biome cover in simulation scenarios in which C3 and C4 grasses cure rapidly (a–c) and C4 grasses cure slowly whereas C3 grasses cure rapidly (d–f, hypothesis 3, Table 3). The graphs show transitions between the two scenarios and the scenario in which C4 grasses cure rapidly whereas C3 grasses cure slowly (Fig. 1g–i).

Figure 5.

Fire activity in the presence and absence of C4 grasses and with slow- or rapid-curing grasses at the continental scale. Data are simulated for the three late Miocene climatologies.

The simulation results support the hypothesis that the evolution of the fire-resistant savanna tree type has an impact on the distribution of fire-driven biomes (hypothesis 4, Table 3). In the presence of fire-resistant savanna trees, C4 grasslands and savannas dominate under environmental conditions in which, in the absence of savanna trees, forests and woodlands would dominate (Fig. 6). Simulations for the low-rainfall study site show that, in the presence of the savanna tree type, the vegetation state moves between a C4 grassland/savanna state and a forest state when the CO2 concentrations cross a threshold of between 550 and 650 ppm (Fig. 7a). In the absence of the savanna tree type, the dominance of C4 biomes is restricted to CO2 concentrations < 450 ppm. At the intermediate-rainfall site, the savanna tree type is essential for the existence of C4-dominated biomes (Fig. 7b). In the presence of the savanna tree type, C4-dominated biomes can occur over the entire gradient of CO2 concentrations, whereas, in the absence of the savanna tree type, vegetation is always dominated by C3 biomes. At the high-rainfall study site, C4 biomes cannot dominate regardless of the presence or absence of savanna trees (result not shown).

Figure 6.

Expansion of C3- and C4-dominated biomes in the presence and absence of the savanna tree type (STT; hypothesis 4, Table 3). The red area indicates regions in which the savanna tree type is required for the existence of C4-dominated biomes.

Figure 7.

Probability of C4 dominance for increasing and decreasing levels of CO2 (indicated by arrows) at the low-rainfall (570 mm MAP) study site (a) and the intermediate-rainfall (1320 mm MAP) study site (b) in presence (black lines) and absence (gray lines) of the savanna tree type.

The effect of the savanna tree type on the extent of the C4 savanna biome is a result of the shade and fire tolerances of the modeled functional types. In the absence of the savanna tree type, C4 grasses can exclude forest trees at low CO2 concentrations because of high fire activity and the low growth and resprouting rates of forest trees. At higher CO2 concentrations, trees grow sufficiently rapidly to suppress grass growth through light competition and to suppress fire. At intermediate CO2 concentrations, savanna trees can persist, because of their fire tolerance, in the presence of C4 grasses; their low leaf area index, however, ensures that they do not exclude the C4 grasses, which, in turn, serves to maintain fire in the system. This prevents the invasion of forest trees which would, due to their high leaf area, exclude the shade-intolerant C4 grasses.

Discussion

The broad vegetation patterns simulated by the aDGVM for the late Miocene agree with previous vegetation reconstructions and simulation results for that period (François et al., 2006; Micheels et al., 2007; Pound et al., 2011). However, compared with previous reconstructions, the aDGVM overestimates the area covered by deserts in the Sahara regions. In previous reconstructions, northern Africa is covered by grasslands and semi-desert (François et al., 2006), by xerophytic shrublands and deserts (Pound et al., 2011) or even fully covered by grasslands and savannas (Micheels et al., 2007). These differences can be explained by uncertainties in both the vegetation and the climate models. How ecological processes are modeled, the plant functional types included and the biome classification schemes differ between dynamic vegetation models (Prentice et al., 2007). An intercomparison of six DGVMs for ambient climate conditions showed that the simulated covers of deserts, grasslands and savannas in Africa varied between 30% and 49%, 21% and 35% and 4% and 19%, respectively (Scheiter & Higgins, 2009; based on data provided by Cramer et al., 2001), and one would expect similar variability under late Miocene conditions. Another source of uncertainty is the difference between late Miocene climatologies. While François et al. (2006) and Micheels et al. (2007) used the ECHAM4 model coupled to a mixed-layer ocean model, we used climatology from the Hadley Centre coupled ocean–atmosphere–vegetation general circulation model (HadCM3L; C. Bradshaw et al., unpublished). Simulations with the Sheffield Dynamic Global Vegetation Model (SDGVM) (Woodward et al., 1995; Beerling & Woodward, 2001) coupled to the Hadley Centre climatology produced Sahara desert and grassland distributions comparable with the aDGVM simulations presented here (Beerling et al., 2012; Fig. S3). In addition, the ECHAM4 climatologies used by François et al. (2006) and Micheels et al. (2007) were generated by assuming a vegetated Sahara, which implies that the reconstructions also project more vegetation in the Sahara region than simulated here.

Our results provide quantitative support for the idea that the late Miocene expansion of C4-dominated ecosystems (C4 grasslands and savannas) into woodlands and forests was the outcome of three interacting ecological processes. First, C4 grasses invaded C3 grasslands, transforming them into mixed C3–C4 grasslands and C4 grasslands. Differences in the leaf-level physiology of C3 and C4 types under hot, open environmental conditions (Ehleringer et al., 1997) are sufficient to explain this process. This mechanism operated only in arid regions that could not support forests and, in this case, precipitation was important in maintaining biome boundaries. However, model simulations suggest that the continental-scale significance of this mechanism in Africa would have been rather limited in the absence of fire. Based on this evidence, climatic drying therefore seems an unlikely direct driver of C4 grassland expansion (Sepulchre et al., 2006). Secondly, changes in fire regimes allowed grasses to invade woodlands and forests, and fire was required to maintain biome boundaries (Sankaran et al., 2005; Scheiter & Higgins, 2007). As the model simulates C4 grasses outcompeting C3 species in the hot, open conditions after fire, this process leads to the invasion of forest or woodland by C4 grasses (Keeley & Rundel, 2003, 2005). Thirdly, our model simulations support the hypothesis that the evolution of fire-tolerant savanna trees (Simon et al., 2009) expanded the climate space occupied by fire-driven C4 biomes into areas previously occupied by forests. The co-existence of savanna trees and grasses is permitted because the open canopy of fire-tolerant trees allows shade-intolerant grasses to persist, and the survival in fires allows trees to occur in areas with a continuous flammable grass cover (Ratnam et al., 2011). Based on this evidence from simulation modeling, we propose that the continental extent of C4 grasses was limited during the early and mid-Miocene by the limited occurrence of fires.

Paleoevidence suggests that forests and woodlands dominated in the early and mid-Miocene, whereas grasses existed only in forest–grassland mosaics (Strömberg, 2011). There is no evidence for fire-prone grasslands with low tree cover until the late Miocene (Strömberg, 2011). The model can only mimic this delayed expansion of grasslands by delaying the incidence of fire to the late Miocene. The model can simulate transitions from woodlands and forests to open, fire-prone grasslands and savannas dominated by C3 rather than C4 grasses when it is assumed that fire regimes changed before C4 grass evolution. Although transitions to open grasslands may have already occurred in the early and mid-Miocene, for instance in North America (Edwards et al., 2010), charcoal records do not support an involvement of fire at this early stage in Africa (Strömberg, 2011). These findings support the first two ecological mechanisms proposed. However, the model cannot explain why fire regimes changed in the late Miocene or whether these changes were triggered by changes in the seasonality of climate. Our third mechanism requires that fire-tolerant savanna trees evolved in fire-driven, open biomes. According to phylogenetic analyses, fire-adapted trees evolved repeatedly in multiple forest lineages in response to the expansion of C4 grasslands (Simon et al., 2009). However, further studies are required to assess whether savanna trees evolved in response to the expansion of fire-driven ecosystems, or whether they were involved in the construction of the fire-driven ecosystems (Olding-Smee et al., 2003). As the timing of the three proposed mechanisms differs with continent, future studies should aim to understand the causes of these differences.

The aDGVM projects that C4 grassland expansion was positively associated with fire-promoting traits. Nonetheless, C4 grass invasion of forests was simulated even if it was assumed that the flammability of C4 grasses is low. This suggests that fire, rather than fire-promoting traits, is the necessary prerequisite for C4 expansion. Grasses are, to a certain extent, pre-adapted to fire and a certain level of fire tolerance is essential for the dynamics analyzed in this study (all grasses in these simulations survived and resprouted after fires). Our simulations also suggest that fire-adapted trees extend the range of environmental conditions in which C4 biomes can dominate. Hence, whether C4 grasslands and savannas occur in all regions in which the Ehleringer model (Ehleringer et al., 1997) projects C4 dominance is not only a function of fire and environmental conditions (Sankaran et al., 2005), but is also related to the presence of fire-adapted vegetation types. Previous modeling studies have neglected the importance of fire adaptation for the expansion of C4-dominated systems, and a better knowledge of fire-related traits of different vegetation types would improve the robustness of our analysis.

The model projects that fire-driven ecosystems with a C3 grass cover and with fire-resistant savanna trees (C3 savannas) can persist in the absence of C4 grasses. Although these systems cannot persist in the presence of C4 grasses under late Miocene or ambient conditions, because of the physiological advantage of C4 grasses, they may persist under future climatic conditions. At elevated CO2, the Ehleringer model projects C3 dominance. This implies that C3 grasses have the potential to invade C4 grasses, provided that there is a sufficiently large pool of fire-tolerant C3 grass species.

In conclusion, our model analysis is consistent with the operation of three ecological mechanisms promoting the invasion of C3 vegetation by C4 grasses. First, C4 grasses invade C3 grasslands and dominate via direct competition in hot, open environments. Secondly, fires allow grasses to invade forests. Thirdly, fire adaptation in grasses and trees expands the climate space occupied by C4-dominated systems. The extent to which C4 grasses can occupy the entire climate space predicted by the Ehleringer model depends on the fire-promoting traits of vegetation. Physiological differences between C3 and C4 grasses are necessary, but not sufficient, to explain C4 expansion. That is, models that consider only leaf physiology cannot explain important aspects of savanna biome assembly. This study highlights that dynamic vegetation models need to consider how life history strategies respond to resource availability and to disturbances such as fire (Osborne & Beerling, 2006; Bond, 2008).

Acknowledgements

Financial support was provided by Hesse’s Landes-Offensive zur Entwicklung Wissenschaftlich-ökonomischer Exzellenz (LOEWE) and by the Deutsche Forschungsgemeinschaft (DFG). Catherine Bradshaw was funded by an Natural Environment Research Council in the UK (NERC) PhD Studentship.

Ancillary