The dominant vegetation over much of the global land surface is not predetermined by contemporary climate, but also influenced by past environmental conditions. This confounds attempts to predict current and future biome distributions, because even a perfect model would project multiple possible biomes without knowledge of the historical vegetation state.
Here we compare the distribution of tree- and grass-dominated biomes across Africa simulated using a dynamic global vegetation model (DGVM). We explicitly evaluate where and under what conditions multiple stable biome states are possible for current and projected future climates.
Our simulation results show that multiple stable biomes states are possible for vast areas of tropical and subtropical Africa under current conditions. Widespread loss of the potential for multiple stable biomes states is projected in the 21st Century, driven by increasing atmospheric CO2. Many sites where currently both tree-dominated and grass-dominated biomes are possible become deterministically tree-dominated.
Regions with multiple stable biome states are widespread and require consideration when attempting to predict future vegetation changes. Testing for behaviour characteristic of systems with multiple stable equilibria, such as hysteresis and dependence on historical conditions, and the resulting uncertainty in simulated vegetation, will lead to improved projections of global change impacts.
Understanding the factors that influence the distribution of major vegetation formations is a fundamental challenge in ecology and biogeography. Major works on the subject have related vegetation distributions to climate and – to a lesser extent – soils (Holdridge, 1967; Walter, 1968; Whittaker, 1975; Archibold, 1995). The prevailing idea is that given a set of climatic and edaphic conditions, the dominant biome can be predicted. Within such systems, each biome occupies a unique region of climate space – typically ordered along axes of rainfall and temperature. Authors have, however, begun to highlight that multiple biomes are often possible for a given set of conditions, and that vegetation can modify local and regional climate conditions. Multiple stable biome states are possible when land–atmosphere feedbacks result in multiple stable vegetation–climate equilibria (Margulis & Lovelock, 1974; Wang & Eltahir, 2000; Oyama & Nobre, 2003). At smaller scales, sharp localised boundaries force local micro-climates, reinforcing the differences across boundaries (Uhl & Kauffman, 1990; Ray et al., 2005). However, even for a given prescribed climatology, disturbance regimes maintained by internal feedbacks and the characteristics of dominant vegetation can exclude other vegetation formations that although climatically suited, are unable to tolerate the disturbance regime (Staver et al., 2011a; Hoffmann et al., 2012a; Murphy & Bowman, 2012).
Where the prevailing climate allows multiple biome states, these biomes often exist as mosaics with starkly contrasting properties (Woodward et al., 2004; Bond, 2005; Staver et al., 2011b; Scheffer et al., 2012). Boundaries between them can be sharp, with high species turnover and large differences in ecosystem structure (Parr et al., 2012; Dantas et al., 2013). The current distribution of biomes and location of boundaries will be determined partly by historical conditions, as these have allowed the initialization of the feedbacks that maintain a particular biome. Shifts in the location of boundaries, the relative frequency of vegetation types or even switching to a single dominant biome at the landscape level can be forced by a variety of factors. Imposed changes in climate and nutrient concentrations can force major reorganization of vegetation patterns, with even small shifts in conditions having major impacts if critical thresholds are crossed (Scheffer et al., 2001). Moreover, even in a constant environment, disruption of internal feedbacks by land use change, grazing pressure or modification of fire regimes can provoke switching between biomes (Louppe et al., 1995; Zimov et al., 1995; Favier et al., 2012).
The savanna complex represents the most extensive ensemble of biomes with potential for alternate stable states (Hirota et al., 2011; Staver et al., 2011b). Vast regions of the tropics and sub-tropics in Africa, Australia and South America are covered by biomes dominated by either trees or C4 grasses, or a mixture of the two. Although maximum tree cover is limited by moisture availability in low rainfall areas, in areas of intermediate rainfall actual tree cover varies greatly and multiple biome states are possible (Sankaran et al., 2005; Higgins et al., 2010; Staver et al., 2011a). When biomes dominated by C4 grasses occur (grasslands and savannas), tree cover is kept low by periodic fires causing mortality of aboveground organs (Bond, 2008). When high tree cover biomes (forests and woodlands) occur, dense canopies prevent light-demanding C4 grasses from growing and thus exclude fires. The feedback between fire, tree cover and C4 grasses is, however, vulnerable to disruption by changes in climate, atmospheric CO2 concentrations and fire regimes (Bond & Midgley, 2012).
African savannas and grasslands support vast numbers of large herbivores and sustain the livelihoods of millions of people (Grace et al., 2006). The forests and woodlands of Africa are highly biodiverse and store billions of tonnes of aboveground carbon (Lewis et al., 2009). Biome switches between grassy and forested states disrupt entrenched agriculture, biogeochemical cycles and community structure (Parr et al., 2012). Large changes in tree cover have already been observed across Africa with several studies reporting increases, as well as switching from grassy to forested biome states (Mitchard et al., 2009; Wigley et al., 2009; Buitenwerf et al., 2012). Future projections for African biomes predict major expansion of forests at the expense of savannas and grasslands (Sitch et al., 2008; Scheiter & Higgins, 2009). Experimental, observational and model-based exploration of the drivers of observed and projected changes in biome distributions all implicate increasing atmospheric CO2 concentrations as the major factor preferentially favouring trees over C4 grasses (Kgope et al., 2010; Bond & Midgley, 2012; Higgins & Scheiter, 2012).
The possibility of alternate biome states confounds our ability to predict current and future vegetation distribution. In regions where the environment alone does not predetermine the biome and a variety of outcomes are possible, the observed biome will depend on whether feedbacks maintaining the system in a particular state have been initialized and maintained. Thus, even with a perfect model for projecting biome distributions, without prior knowledge of the initial system state, the best possible prediction would still allow for multiple possible outcomes. Indeed, the distribution of African savannas – which are not the only possible stable biome over much of their range – has proved to be amongst the most challenging to model using dynamic global vegetation models (Cramer et al., 2001). The existence of alternate biome states suggests that initializing the system in an open canopy state will favour grasslands and savannas, with the reverse applying to a closed canopy initialization and forests. Although intrinsic dependence on initial conditions presents a challenge for modelling biome distribution, it does however allow evaluation of the climatic and geographic space in which multiple stable biomes can be expected by comparing vegetation modelled using alternate initializations.
Here we compare the distribution of tree- and grass-dominated biomes across Africa simulated using a dynamic global vegetation model (DGVM) developed and benchmarked specifically for tropical tree–grass biomes. We compare the projected distribution of tree- and grass-dominated biomes using identical model structures and environmental inputs, but with alternate initial conditions designed to test for multiple stable biome states. Thus we begin with a treeless or grassless African continent and observe the outcome of simulation runs given these initial conditions. The drivers of alternate stable states are investigated by comparing fire impacts and the relationship between precipitation and tree cover. Based on previous studies in savannas (Sankaran et al., 2005; Higgins et al., 2010; Staver et al., 2011a), we expect that the greatest differences in biome distributions between initializations to occur in zones of intermediate rainfall where aridity or high rainfall does not predetermine the dominant biome. In accordance with previous studies, we expect that any changes between current and future patterns of stability will largely be the result of changes in atmospheric CO2 concentration rather than projected changes in temperature and precipitation patterns.
We simulate African vegetation using a DGVM called the ‘aDGVM’ (adaptive dynamic global vegetation model, Scheiter & Higgins, 2009), developed and tested for tropical grass–tree systems. A full model description is provided in Scheiter & Higgins (2009) with the updated version, used here, described in Scheiter et al. (2012). The aDGVM combines plant physiology normally used in DGVMs (Prentice et al., 2007) with adaptive dynamic carbon allocation and leaf phenology, and an improved representation of fire impacts (Scheiter & Higgins, 2009). The aDGVM simulates vegetation in representative 1-ha stands and is used in individual-based-tracking state variables such as biomass, height and photosynthetic rates of individual plants. Carbon allocation within an individual is adapted in response to environmental conditions, with carbon preferentially allocated to the biomass compartment relevant to the resource most limiting growth. Leaf bud-burst and abscission are determined by a plant's carbon status and not by predetermined environmental thresholds.
The individual-based approach also allows various disturbances, such as the impacts of herbivores (Scheiter & Higgins, 2012) or fire (Scheiter & Higgins, 2009), to be modelled as a function of plant height, which determines vulnerability. Grasses are represented as super-individuals beneath and between tree canopies. Fire in the aDGVM is modelled based on the equations of Higgins et al. (2008). Fuel loads, fuel moisture and wind speed determine fire intensity, with parameters estimated using data from southern African and Australian savannas. Fire spreads when the potential fire intensity exceeds a threshold value of 300 kJ m−1 s−1, and a randomly generated ignition event occurs. Fire removes aboveground grass biomass, and causes the death of aboveground tree organs (‘topkill’). Topkill is a function of tree height and fire intensity, with small individuals (< 2 m) topkilled by most fires, and large trees affected by only the most intense fires (Higgins et al., 2000; Higgins et al.,2012). An updated version of the aDGVM was presented in Scheiter et al. (2012). This version includes two trees types; a shade tolerant, fire intolerant forest tree and a shade intolerant, fire tolerant savanna tree (Bond, 2008; Ratnam et al., 2011). As a consequence of these different life history traits, savanna trees dominate in open, frequently burnt environments, whereas forest trees dominate when fire is rare and tree cover high. Two grass types are also included in the updated version. C3 and C4 grass types are differentiated by their leaf level physiology, rates of curing and shade tolerance (Osborne & Freckleton, 2009). The aDGVM was shown to simulate the current distribution of vegetation in Africa better than alternative dynamic vegetation models, and could also simulate biomass change and responses to long-term fire manipulation in the Kruger National Park, South Africa (Higgins et al., 2007; Scheiter & Higgins, 2009). The updated version slightly improves the agreement between simulated and observed biome distribution patterns (Scheiter et al., 2012).
We simulate the vegetation of continental Africa between 1850 and 2100 at a resolution of 1 degree. Each site is simulated 100 times to account for model stochasticity. Snapshots of vegetation from 2000 and 2100 are then used for further analysis. Simulations are forced using projected changes in climate given by the Max Planck Institute for Meteorology's (Hamburg) ECHAM5 IPCC (2007) projections with atmospheric CO2 from IPCC (2007) SRES A1B projections (Roeckner, 2005). In order to test the sensitivity of vegetation and biome stability to changes in CO2 we also ran simulations using unchanging temperature and precipitation from the Climate Research Unit's empirical climate data (New et al., 2002), with atmospheric CO2 from IPCC (2007) SRES A1B projections. Additional simulations were run for selected sites for longer time series (500 yr) in order to evaluate the long term stability of simulated vegetation. These simulations also used unchanging temperature and precipitation from the Climate Research Unit's empirical climate data and ambient CO2 concentrations, with 20 replicate simulations per site. An independent, randomly generated ignition sequence was used for each model run.
Previous studies using the aDGVM have initialized the model using a 100 yr spin-up under 1850 conditions with both trees and grasses present (Scheiter & Higgins, 2009). Because we were interested in the effects of initial vegetation on the proportions of forest vs grasslands, we analysed and compared vegetation simulated using two alternate initializations. In the ‘grassland’ initialization, we introduce C3 and C4 grasses and spin-up the model for 100 yr under 1850 conditions while excluding trees. This allows grass biomass to reach its maximum potential in the absence of tree competition before trees are introduced. For the ‘forest’ initialization, we introduce 100 small and medium sized trees, randomly assigned to forest or savanna tree types and again spin-up the model for 100 yr under 1850 conditions, this time while excluding grasses. Thus trees establish and grow in the absence of grass competition and vegetation is unaffected by fire. After the spin-up period in each initialization, the excluded plant type is introduced.
We focus on the changes in the distribution of biomes dominated by trees, grasses or both. The vegetation simulated by the aDGVM is translated into a biome type using the same classification scheme described in Scheiter et al. (2012) and Higgins & Scheiter (2012). When grass biomass exceeds a threshold (0.5 t ha−1) and tree cover is low (< 10%), vegetation is classified as either C3 or C4 grassland. When tree cover is at intermediate percentages (> 10% and < 80%), but canopies are not closed or near closed, vegetation is classified as either savanna or woodland. Woodlands occur when forest trees dominate and savannas when savanna trees dominate. Vegetation is classified as forest when tree cover exceeds 80%. The stability of biome distributions is evaluated by comparing the extent of grass-dominated (grasslands and savannas) and tree-dominated (woodland and forest) biomes between initializations. Because tree cover and fire influence biome type by controlling the feedbacks that maintain biomes in their current state when alternate states are possible, we also compare how these factors change in response to different initializations under current and future conditions. Data from Sankaran et al. (2005) are compared to simulated tree cover patterns.
Results differed little between simulations using changing climate and CO2 or constant climate and changing CO2 (Supporting Information Fig. S1). This suggests that differences between vegetation modelled in 2000 and 2100 are due principally to the direct and indirect effects of projected increases in atmospheric CO2 concentrations on vegetation rather than temperature and precipitation changes. We present and discuss results from simulations accounting for both changing climate and CO2.
Simulation results from 2000 show large differences in the projected distribution of grass- and tree-dominated biomes between initializations (Fig. 1). These differences are not transient, but rather represent different stable states that persist over long periods under current conditions (Fig. 2). When initialized in the ‘forest’ state, high probabilities of forest and woodland biomes occur over almost all of tropical and subtropical Africa (Fig. 1). Savannas and grasslands are restricted to arid and semi-arid regions on the fringe of their current actual distribution. When initialized in the ‘grassland’ state, biome distributions resemble actual patterns more closely (κ-value ‘forest’ initialization = 0.41, κ-value ‘grassland’ initialization = 0.54, Fig. S2). Forest and woodland covers the Congo basin and coastal West Africa, with savannas and grassland covering most of southern Africa and the Sahel with high probability. Dependence on initial conditions is highest in seasonally dry areas of intermediate rainfall (700–1000 mm MAP) in central, southern and west Africa. Observed differences in biome distribution relate well to fire frequency patterns. Initialization in the ‘forest’ state reduces modelled fire frequency across large areas of central and southern Africa, with the largest reductions in fire frequency occurring approximately where the biome state is most dependent on initialization (Fig. 1).
Biome distributions simulated for 2100 differ far less between the two initializations (Fig. 3). Tree-dominated biomes are projected to dominate throughout tropical and subtropical Africa, regardless of the initialization, indicating that this distribution is stable and resistant to invasion by grass-dominated biomes. Grass-dominated biomes are restricted to southern Africa and the Sahel. The spatial extent and frequency of fire is greatly reduced compared to 2000, particularly in southern Africa, with little systematic difference between initializations.
The simulated pattern of tree cover increase with increasing rainfall in 2000 agrees well with field data from Sankaran et al. (2005) (Fig. 4a). Tree cover is deterministically low at very low rainfall, where deserts and arid grasslands occur. At intermediate rainfall (300–600 mm) tree cover is restricted below a threshold set by rainfall and arid or semi-arid savannas occur. Above this threshold, closed canopies are possible. Actual tree cover is often much lower though, with a bimodal pattern of a high frequency of intermediate tree cover in savannas and closed canopies in forests (Fig. S3). It is in this range of 700–1300 mm that initializing the model in the ‘grassland’ or ‘forest’ state most affects tree cover patterns. The ‘forest’ initialization reduces the number of sites simulated with intermediate tree cover and increases the proportion of high tree cover sites. The ‘forest’ initialization does not, however, cause all or even the majority of intermediate tree cover sites to convert to closed canopies. When rainfall is higher than 1600 mm almost all sites have closed canopies, regardless of initial conditions.
Tree cover simulated for 2100 responds very differently to rainfall (Fig. 4b). Closed canopies are possible at 400 mm, with the range in which tree cover is restricted much reduced. The climate space for savannas and grasslands is thus much narrower. Most sites above 1000 mm now have deterministically closed canopies, with no bimodality in the distribution of tree cover (Fig. S3). Whether the model is initialized as ‘grassland’ or ‘forest’ causes little systematic difference across all precipitation zones.
Our analysis shows that the distribution of African biomes is highly sensitive to initial conditions. Large areas of central, southern and west Africa are not deterministically dominated by grassy or woody biomes, but rather can persist in either state depending on the initial conditions. Shifts from savannas and grassland to forests and woodland between initializations are driven by increases in tree cover and changes in the dominant tree type. These changes suppress grass growth and hence reduce fire, allowing the system to persist in alternate stable states. However, multiple stable biome states are only possible under a limited range of environmental conditions, with very arid, very wet and aseasonal environments unlikely to support different biome states under current conditions.
The potential for alternate stable biomes is projected to be almost completely lost in 2100, driven largely by projected increases in atmospheric CO2 concentration. C3 photosynthesis in trees is not carbon saturated under current and past 20th century CO2 concentrations, whereas C4 photosynthesis in tropical grasses is, thus the relative performance of trees increases into the 21st century in our model (Ehleringer et al., 1997; Jolly & Haxeltine, 1997). This shifts the competitive balance in favour of trees, resulting in increased tree cover and forest tree dominance at lower rainfall as well as a reduction in fire impacts on trees (Scheiter & Higgins, 2009; Higgins & Scheiter, 2012). Moreover, the maximum potential tree cover for a given amount of precipitation increases, such that areas previously too arid for forests can support closed canopies in 2100. In addition to experimental studies showing improved tree performance with CO2 enrichment (Kgope et al., 2010; Quirk et al., 2013) supporting these findings, observational studies have linked increasing tree cover and decreasing fire impacts in the 20th century with CO2 increase (Wigley et al., 2009; Buitenwerf et al., 2012; Tng et al., 2012).
The widespread potential for multiple stable biome states simulated under current conditions suggests that switching between stable states is currently possible if the system is forced beyond a tipping point through interventions such as the introduction of severe fires or the encouragement of tree establishment (Parrotta et al.,1997; Silvério et al., 2013). However, with the projected switching of many areas into the tree-dominated state in the 21st century, reversion back to the grass-dominated state will become increasingly unlikely, as the expansion of tree-dominated biomes occurs regardless of initial conditions. In our simulations, by 2100 the distribution of tree- and grass -dominated biomes becomes pre-defined by the prevailing climate. Alteration of the dominant vegetation would thus require continuous intervention to achieve stasis, confounding attempts to manage landscapes for biodiversity or carbon storage, as a tipping point, beyond which the system can be forced into an alternate stable biome state, no longer exists. Our findings suggest that analytical models of stability in the forest–savanna–grassland biome complex, such as those of Hirota et al. (2011) and Staver & Levin (2012), should include atmospheric CO2 concentration and temperature as additional resource axes. This would allow more thorough estimation of the environmental conditions under which these biomes can persist as alternate stable states and the thresholds that, when crossed, cause the system to bifurcate though the loss of stable equilibria.
Our projections of changes in forest distribution are made using the output of a single climate model (ECHAM5) forced with a single emissions scenario (SRES A1B). Moreover, in the real world, vegetation feeds back to alter the regional climate which introduces additional uncertainty in projections of future biome stability, potentially reducing the likelihood of multiple stable biomes states in some regions and enhancing it in others (Charney, 1975; Brovkin et al., 1998; Oyama & Nobre, 2003). Incorporating uncertainty in climate change projections for Africa by using output from an ensemble of climate models forced by a range of emissions scenarios would improve the reliability of our projections for future biome stability. Additionally, coupling of land surface models providing projections of biome stability to atmospheric simulation models could incorporate the potential modulating role of land–atmosphere feedbacks. Multiple vegetation models forced by a range of climate models have, however, projected increases in tree-dominated biomes in Africa in response to climate change and CO2 increase (Sitch et al., 2008; Doherty et al., 2010; Huntingford et al., 2013), and several lines of empirical evidence (Mitchard et al., 2009; Wigley et al., 2009) suggest that switches to more tree-dominated biome states are already underway. The consequences of these switches for biodiversity and carbon storage are likely to be significant (Bond & Parr, 2010; Parr et al., 2012).
It is interesting to note that the actual current distribution of biomes is better approximated by initializing the simulations in a treeless state. Given the considerable palaeo-evidence for grassland expansion and forest retreat to small refugia during glacial periods when atmospheric CO2 concentrations were half those of today, it is likely that the vegetation we observe today developed from a low tree cover landscape (Dupont et al., 2000; Prentice & Jolly, 2000; Harrison & Prentice, 2003; Prentice et al., 2011). Thus the current dominance of savanna and grassland over large tracts of Africa may partially be the legacy of Pleistocene conditions. The dependence of accurate simulations of current biome distributions on historical climate and landscape conditions in this study has implications for the development and testing of global vegetation models. Because the potential for alternate stable biome states is an intrinsic feature of many systems, testing for behaviour characteristic of systems with multiple stable equilibria, such as hysteresis and dependence on historical conditions (Beisner et al., 2003; Scheffer & Carpenter, 2003) should be an important part of model development and analysis. Many models are designed and their initializations implemented to avoid any residual effect of the initial system state. It is likely that simulations initialized with vegetation informed by historical and current conditions will produce more reliable projections. The potential for DGVMs to be used as tools for evaluating where, and under which environmental conditions, one is most likely to find alternative biome states, as well as the effect of global changes on the future patterns of biome stability, remains largely unexplored.
A growing body of evidence suggests that the vegetation of Africa is not predetermined by climate over much of the continent (Bond et al., 2003; Sankaran et al., 2005; Staver et al., 2011a). Here we have provided additional evidence from a DGVM. Vast areas of other continents have also been identified as having multiple possible biome states (Bond et al., 2005; Hirota et al., 2011; Staver et al., 2011b; Hoffmann et al., 2012b; Murphy & Bowman, 2012; Scheffer et al., 2012). Thus it appears that this is an ubiquitous and important feature to consider when attempting to predict global biome distribution patterns. The re-evaluation of the one climate–one biome philosophy behind attempts to predict future changes in large-scale vegetation patterns (Bergengren et al., 2011) and explicit exploration of uncertainty in modelled biome states will lead to an improved ability to predict and respond to vegetation change in the 21st century.
Financial support was provided by Hesse's Landes-Offensive zur Entwicklung Wissenschaftlich okonomischer Exzellenz (LOEWE). G.M. was funded by a Deutsche Akademische Austauschdienst (DAAD) PhD Scholarship. We thank C. Huntingford and three anonymous reviewers for helpful comments on this manuscript.