Search tips
Search criteria 


Logo of transbThe Royal Society PublishingPhilosophical Transactions BAboutBrowse By SubjectAlertsFree Trial
Philos Trans R Soc Lond B Biol Sci. 2016 September 19; 371(1703): 20150311.
PMCID: PMC4978869

The future distribution of the savannah biome: model-based and biogeographic contingency


The extent of the savannah biome is expected to be profoundly altered by climatic change and increasing atmospheric CO2 concentrations. Contrasting projections are given when using different modelling approaches to estimate future distributions. Furthermore, biogeographic variation within savannahs in plant function and structure is expected to lead to divergent responses to global change. Hence the use of a single model with a single savannah tree type will likely lead to biased projections. Here we compare and contrast projections of South American, African and Australian savannah distributions from the physiologically based Thornley transport resistance statistical distribution model (TTR-SDM)—and three versions of a dynamic vegetation model (DVM) designed and parametrized separately for specific continents. We show that attempting to extrapolate any continent-specific model globally biases projections. By 2070, all DVMs generally project a decrease in the extent of savannahs at their boundary with forests, whereas the TTR-SDM projects a decrease in savannahs at their boundary with aridlands and grasslands. This difference is driven by forest and woodland expansion in response to rising atmospheric CO2 concentrations in DVMs, unaccounted for by the TTR-SDM. We suggest that the most suitable models of the savannah biome for future development are individual-based dynamic vegetation models designed for specific biogeographic regions.

This article is part of the themed issue ‘Tropical grassy biomes: linking ecology, human use and conservation’.

Keywords: savannah, dynamic global vegetation model, statistical distribution model, global change, CO2 fertilization, tropics

1. Introduction

Within savannahs there is enormous scope for variation in ecosystem structure and function. Tree cover can vary from near-closed canopies to almost treeless, productivity from close to zero up to 12 tonnes C ha−1 yr−1, while fires can occur annually or not at all [13]. The complexity of factors interacting to determine the relative abundance of trees and grasses within savannahs, or biome shifts from savannah to grassland or forest, makes modelling their distribution and projecting global change impacts challenging. Factors such as temporal separation in phenological niches, spatial separation of hydrological niches, demographic bottlenecks caused by frequent fire and herbivores have variously been invoked to explain the existence of savannahs and their extensive distribution [4].

Perhaps unsurprisingly, many early models produced poor simulations of the distribution of the savannah biome, or conveniently ignored its existence [5]. Savannahs cover 20% of the land surface, store over 15% of terrestrial above-ground carbon, and are relied upon to support the livelihoods of millions of people through agriculture, resource extraction and tourism [3]. It is, therefore, imperative to improve their representation in global models of plant biogeography and biogeochemical cycles, and produce improved projections of their response to global environmental change [6].

Two classes of models are typically used to project the future distribution of vegetation units in a changing environment, dynamic vegetation models (DVMs) and statistical distribution models (SDMs). SDMs (variously called species distribution models, niche models or habitat suitability models) correlate the current distribution of species, vegetation types or theoretically any biotic entity with present conditions based on the presence/absence data and geographical maps of environmental variation. Lehmann et al. [7] used SDMs to study the environmental limits of savannah in Africa, Australia and South America, showing that the contemporary climate range of savannahs differs on each continent. The limitations of SDMs are well known, and include the assumption that observed distributions are in equilibrium with the environment, unreliable projections in novel environmental conditions and exclusion of competitive interactions [8]. These drawbacks have prompted the development of models which explicitly represent physiological processes, the outcomes of which ultimately determine species distributions. Once such model developed for plants by Higgins et al. [9], uses the Thornley transport resistance statistical distribution model (TTR-SDM) [10] to describe plant physiological performance and links parameters with abiotic environmental factors. While other physiologically motivated distribution models require detailed physiological data for any species or biological entity, the TTR-SDM inversely estimates these parameters, requiring only distribution data and environmental conditions as input. Therefore, it can theoretically be fitted for a wide array of plant species or vegetation units.

Despite progress in the methods used to project climate change impacts on vegetation using statistical methods, no model of this class can account for the impacts of rising atmospheric CO2 concentrations on plant functioning and distributions. The lack of spatial variation in atmospheric CO2 concentrations does not allow for the space-for-time substitutions typically employed to project forward in time, and CO2 manipulations on whole plants or ecosystems cannot be adequately replicated to cover the full range of other confounding factors. Photosynthetic response to CO2 concentration is, however, well understood at the leaf level and carbon gain in response to variation in temperature, water availability and CO2 well described with process-based equations [11,12]. DVMs link these models of plant physiology with models of vegetation dynamics and biogeochemistry with the climate system, scaling from the leaf level to whole plants and ecosystems [13]. DVMs have been used to investigate the impacts of global change on biome distributions, the carbon cycle and feedbacks between vegetation and the climate system [5,14,15].

DVMs typically represent plant diversity using a limited suite of plant functional types (PFTs), differing in key parameters often relating to phenology, fire resistance and photosynthetic type [16]. This simplification has attracted much criticism [17], and motivated the development of a new generation of DVMs which allow plant traits to vary in response to environmental conditions [1719]. This results in a greater possible diversity of plant forms and vegetation types. However, it presents the significant challenge of correctly parametrizing how the environment selects for particular combinations of traits. When attempting to model global vegetation patterns this challenge is further compounded by biogeographic and phylogenetic variation in the relationships among plant traits and the environment [2022].

Here we use both classes of models described above to investigate the potential impacts of climatic changes and increasing atmospheric CO2 on the future distribution of the savannah biome. Regional differences in savannah structure and function are well described [20,23], and hence we use three variants of one DVM each designed and parametrized separately for tropical systems in Africa, Australia and South America. Similarly, we fit the TTR-SDM independently to savannah distributions in Africa, Australia and South America. We, additionally, extrapolate a DVM and the TTR-SDM for Africa to Australia and South America to investigate the biases introduced when inappropriately calibrated models are used globally. The DVM we use does not define fixed PFTs with predetermined traits and the TTR-SDM is not purely correlative, attempting to project the distribution of the savannah biome based on physiological principles. Using these state-of-the-art modelling techniques we try to separate the impact of climatic changes and increasing CO2 and provide our best attempt at projecting the future distribution of the savannah biome.

2. Material and methods

(a) Savannah extent

To validate and calibrate our models of savannah extent, we first need a common definition of what constitutes the savannah biome; thereafter we need a map of its current extent. To define the savannah biome, we follow Ratnam et al. [24] in limiting our circumscription of savannahs to systems with an understory of C4 grasses and a discontinuous layer of savannah trees with a suite of traits distinguishing their ecology from that of forest trees (e.g. fire resistance, open canopies).

The savannah biome map we use was described by Lehmann et al. [7] and has been used to calibrate SDMs for tropical savannahs. It is constructed through the fusion of multiple regional vegetation maps. Vegetation types described and mapped for each region were reclassified according to the criteria of Ratnam et al. [24] as either savannah or non-savannah. The fused vegetation maps extend from 30° S to 30° N and include Australia, Africa and the main savannah regions of South America by covering Venezuela and Brazil. We limit our analyses to the more natural domain of the tropics (23.44° S–23.44° N), and exclude regions above 2000 m elevation, as the DVMs we use are designed specifically for tropical biomes and do not include the plant types or processes necessary to simulate tropical montane regions.

(b) Thornley transport resistance model

The TTR-SDM is based on Thornley's theory [10] as implemented by Higgins et al. [9]. The TTR-SDM as developed by Thornley [10] is an ordinary differential equation model that considers how plant biomass growth is influenced by carbon uptake, nitrogen uptake and by the allocation of carbon and nitrogen between roots and shoots. It explicitly separates the physiological processes of resource uptake from biomass growth. Higgins et al. [9] used the model to relate physiological rates to spatial variation in environmental conditions. Specifically, the model considers how carbon uptake might be limited by temperature, soil moisture, solar radiation and shoot nitrogen; how nitrogen uptake might be limited by temperature, soil moisture and soil nitrogen; and how growth and respiration might be influenced by temperature. The model runs on a monthly time step which allows it to explicitly consider how seasonal fluctuations in the environmental forcing variables interactively influence plant resource uptake and growth. In this study, we use the same model version as used in Higgins et al. [9]. While the TTR-SDM is a plant growth model, parameters are not measured but rather calibrated against distribution data and used to predict distribution patterns and not growth.

The TTR-SDM uses data on soil water, solar radiation (monthly estimates from Trabucco & Zomer [25]), mean, maximum and minimum temperature (monthly estimates from Hijmans et al. [26]) and soil nitrogen (from Shangguan et al. [27]). All variables are available at an approximate grid resolution of 1 km (30 arc-seconds).

The TTR-SDM predicts the potential biomass of a hypothetical plant as forced by the environmental conditions at a site. To relate this to the distributions used for calibration and validation, we follow Higgins et al. [9]. We assume that pi, the probability of the savannah biome being present at site i, is described by the complementary log–log of the modelled plant biomass at site i and that the likelihood of observing the presence/absence data (yi) at site i is described by the Bernoulli distribution. To estimate the parameters, we used a differential evolution optimization algorithm [28] to find the set of parameters that maximize this likelihood over all sites. The predicted probability of savannah being present was converted to a presence/absence map of savannah distribution by using a probability threshold chosen to maximize the agreement between observed and modelled savannah distribution.

(c) Adaptive dynamic global vegetation model

We used three different versions of a dynamic vegetation model originally described by Scheiter & Higgins [29], the adaptive dynamic global vegetation model (aDGVM). Each version is tailored for a single continent. For Africa we used the original aDGVM, with updates described by Scheiter et al. [30], here referred to as aDGVM1-Africa. aDGVM1-Africa models representative of 1 ha stands model plant physiology according to a set of principles employed by many other DVMs [31]. Leaf-level carbon gain is modelled using the method of Collatz et al. [12], and photosynthesis coupled to stomatal conductance is estimated using the method of Ball et al. [32]. aDGVM1-Africa is individual-based, tracking the size and carbon status of all trees. Grasses are represented as superindividuals with either C3 or C4 photosynthesis. Most plant traits are fixed; however, leaf phenology and carbon allocation are adapted in response to environmental conditions and the carbon status of individual plants. Thus, theoretically every plant can have a different phenological cycle and carbon allocation strategy. Fire intensity is estimated using equations from Higgins et al. [33] parametrized with data from 200 experimental fires in Africa. Each day, the potential fire intensity is calculated based on fuel biomass, fuel moisture and wind speed. If an ignition source is present and the potential fire intensity exceeds a threshold, then the fire spreads, burning a proportion of the modelled pixel depending on its intensity. Two tree types are represented—a fire-tolerant, shade-intolerant savannah tree and a fire-intolerant, shade-tolerant forest tree. If a tree is exposed to fire, topkill (loss of all above-ground biomass) occurs depending on tree size, tree type and fire intensity. aDGVM1-Africa has been benchmarked using data on above-ground biomass in South African savannahs and biome distributions in Africa, and performs better than other dynamic global vegetation models (DGVMs) in reproducing observed African biome distributions [29,30].

A modified version of the original aDGVM model is described by Scheiter et al. [34], here referred to as aDGVM1-Oz. New parameter values were chosen to describe a ‘typical’ Australian savannah tree and grass type, and include increased tree sensitivity to shading, altered allometries and decreased shading of grasses by trees. Electronic supplementary material, table S1, provides a full list of parameters changed between aDGVM1-Africa and aDGVM1-Oz. These include refitted tree allometries, shade tolerances of trees and grasses, fire resistance and mortality. Other than the changed parameters, the structure of aDGVM1-Oz is identical to that of aDGVM1-Africa. aDGVM1-Oz was compared with data collected at sites in the OzFlux transect across northern Australia and shown to adequately simulate observed gross primary productivity, ecosystem respiration, evapotranspiration and tree demography [34].

Despite allowing variation in phenological strategies and carbon allocation, all other plant traits in aDGVM1 are fixed. Scheiter et al. [17] describe conceptually a new version of the aDGVM (referred to here as aDGVM2) that allows multiple plant traits to vary. In aDGVM2, plant traits influence growth, mortality and reproduction through direct effects on processes such as leaf-level carbon gain, water uptake and competition. Individuals possessing trait combinations that confer success in a particular environment produce more offspring and pass their trait combinations to subsequent generations while those that perform poorly are ultimately out-competed. Trade-offs among traits constrain the available trait space, precluding the emergence of unrealistic trait strategies. aDGVM2 is based on similar physiological principles to aDGVM1. Individual woody plants or grasses can theoretically have any combination of the variable traits, though in reality, unrealistic combinations and those not suited to the environment are quickly removed from the population. Because it is not feasible to correctly parametrize all the trade-off surfaces, link traits to all relevant process and include all traits relevant to every biome in every region across the world, aDGVM2 has been initially developed and parametrized to simulate tree and grass biomass in the South American tropics.

Our intention in comparing models designed for specific continents is not to identify precisely the mechanisms responsible for divergence in model projections among continents. Rather we quantify the level of agreement among models compared with observed savannah distributions, and in the simulated impact of climate change and increasing atmospheric CO2 concentrations. All the models compared are designed for tropical savannah systems. Our goal is to examine the biases that are possible when a model appropriate for the biome under investigation is applied to an inappropriate biogeographic region.

(d) Simulations

We simulate savannah vegetation at 0.5 degree resolution for the tropics of continental Africa, Australia and South America. Future climate conditions and trajectories of atmospheric CO2 concentration change are based on downscaled output from a single CMIP5 model—MPI-ESM-LR—forced using a single emissions scenario—RCP4.5—[35]. RCP4.5 describes a pathway of greenhouse gas emissions stabilization by mid-twenty-first century and fall sharply thereafter. We increase CO2 concentrations required for DVM simulations in line with RCP-4.5 projections. The average standardized error of mean climate simulated by MPI-ESM-LR is 50% of CMIP3 models, and simulated climate from earlier model versions was shown to be among the most reliable for the tropics [35,36]. MPI-ESM is one of the few general circulation models (GCMs) with a dynamic representation of natural vegetation changes, thus simulating feedbacks of vegetation changes to climate [37]. We note that there are significant differences in projected future climate changes among CMIP5 models, but choose to use only a single model forced by a single emission scenario to focus on differences in savannah extent and global change impacts simulated among vegetation models.

For all aDGVM1 simulations, we use a 100-year spin-up with preindustrial (1860) climate and CO2 and then simulate forward to 2070, expressing climatic change as monthly temperature and precipitation anomalies from prevailing twentieth century conditions given by New et al. [38]. aDGVM2 simulations were identical to those of aDGVM1-Africa and aDGVM1-OZ apart from an extended spin-up of 631 years due to the longer time needed for convergence in this more complex model. Two sets of simulations were conducted to separate the role of changing CO2 and climate in driving future vegetation change: the first with changing climate and increasing CO2 according to RCP4.5 and the second with changing climate and increasing CO2 following historical trends to 1990, after which CO2 levels remain fixed. We output and analyse savannah distributions simulated in 1990 and 2070. Simulated vegetation is defined as savannah when tree canopy cover is between 20 and 80% and above-ground grass biomass exceeds 0.5 tonnes ha−1 with C4 grasses being dominant.

For projecting the distribution of savannahs under future climates with the TTR-SDM temperature and precipitation, projections for 2070 are downscaled using Hijmans et al. [26] and future soil water is recalculated using the methods described in Trabucco & Zomer [39]. It was assumed that solar radiation and soil nitrogen levels remain at ambient levels. The model does not consider how atmospheric CO2 might influence plant growth. TTR-SDM analyses were run at 30 arc-seconds (approx. 1 km) resolution; maps were, however, plotted at 0.5-degree resolution.

3. Results

Savannah distributions simulated by the TTR-SDM for contemporary conditions (figure 1) agree far better with observed savannah distribution than DVM simulations for 1990 (figures 2 and and3).3). This is unsurprising given that the TTR-SDM is fitted using knowledge of savannah distribution (the same knowledge used to evaluate the model's performance), whereas the DVMs are parametrized using physiological and ecological knowledge.

Figure 1.
Savannah distribution under contemporary conditions simulated using the Thornley transport resistance statistical distribution model (TTR-SDM) calibrated and projected separately for each continent (a) and calibrated using savannah distribution in Africa ...
Figure 2.
Savannah distribution in 1990 simulated using individual-based dynamic vegetation models. (a) A separate continent-specific model version is applied to Australia (aDGVM1-Oz) and South America (aDGVM2). (b) A single model version, parametrized and calibrated ...
Figure 3.
Observed and predicted savannah occurrence modelled using the Thornley transport resistance statistical distribution model (TTR-SDM) and dynamic vegetation models (DVM). Models were calibrated using savannah distribution in Africa and projected to Africa, ...

Similar to the result found by Lehmann et al. [7] when extrapolating a model calibrated using a single continent across the tropics (figure 1b), a poorer fit is obtained relative to a model fitted separately for each continent—though only slightly (single model κ = 0.61; multiple model κ = 0.62, figure 3). Savannahs in Africa have been shown to extend into arid environments beyond the lower precipitation limits of savannahs elsewhere [7]. Hence the model fitted to savannah distributions in Africa overestimates the distribution of savannahs in the semi-arid east of Brazil and overestimates the southern limit of savannahs in western and central Australia. The best fit model—obtained by fitting the TTR-SDM separately to each continent and then combining the results—generally overpredicts savannahs in Australia and Africa, and underpredicts their range in South America (figure 1a).

Extrapolating the DVM for Africa (aDGVM1-Africa) to South America and Australia results in similar bias as seen when extrapolating the TTR-SDM (figure 2b). Savannahs are overpredicted in the semi-arid areas of eastern Brazil and at their southern range limit in Australia. Underprediction occurs in the wettest region of Australia, where aDGVM1-Africa incorrectly projects forest occurrence, and at the southern limit of Cerrado in Brazil. Applying separate models to each continent improves overall model fit (single model κ = 0.35; multiple model κ = 0.42, figure 2a). Systematic overprediction no longer occurs in low rainfall regions of Australia (less than 400 mm MAP—mean annual precipitation) and Brazil (less than 800 mm MAP), where the continent-specific models predict grasslands and dry forests, respectively (electronic supplementary material, figure S1). This, however, does not resolve the problem of underprediction in the southern Cerrado. At both the upper (approx. 1500 mm MAP) and lower (approx. 200 mm MAP) rainfall limits of savannahs in Africa, occurrence is underestimated. Savannahs are incorrectly projected in the African Rift Valley and the Horn of Africa. A similar problem of overprediction in this region occurs with the TTR-SDM. Here it is likely that complex topography and climate variation—with bimodal annual precipitation in some regions—complicate simulation of vegetation patterns.

Across all models, the general direction of change from contemporary conditions to 2070 using the RCP4.5 emission pathway is towards an overall reduction in the extent of savannahs (figures 4 and and5).5). The TTR-SDM predicts contraction of savannahs in Africa, Australia and South America at the arid end of their current range limit as a result of global temperature increase and resulting water stress. Projections of change are given by comparing simulated savannah distribution under contemporary conditions to the distribution simulated under 2070 conditions. Thus, projections of change in missclassified areas should be interpreted with caution across all models (e.g. loss of savannahs at the southern range limit in Australia using the TTR-SDM, figure 4a).

Figure 4.
Savannah distribution in 2070 simulated using the Thornley transport resistance statistical distribution model (TTR-SDM—a) and individual-based dynamic vegetation models (DVM—b,c), parametrized and calibrated separately for each continent. ...
Figure 5.
The frequency of shifts in savannah extent simulated using the Thornley transport resistance statistical distribution model (TTR-SDM) and individual-based dynamic vegetation models (DVM) calibrated separately for Africa, Australia and South America. For ...

In contrast with the TTR-SDM, when forcing the DVMs with the same scenario of climate and precipitation change and keeping CO2 constant at 1990 levels, savannah distributions are projected to remain stable in Africa and Australia (figure 4b). The difference between models in these simulations may reflect estimation errors introduced by extrapolating TTR-SDM projections beyond the climate range used for model fitting. These projections also reflect instantaneous vegetation change in response to changing climate, whereas the DVM simulations have yet to fully respond, with vegetation change lagging behind climatic change. Disequilibrium between savannah vegetation and climate in many regions suggests that change may be nonlinear, with low initial change followed by abrupt shifts to forest or grassland states [40,41].

Finally, when simulating global change impacts on savannah distribution in 2070 using separate DVMs for each continent forced by changing temperature, precipitation and rising atmospheric CO2, the trend simulated by the TTR-SDM is reversed (figure 4c). Now, in Africa and South America, rather than contract at their arid range limit, savannahs expand in some regions. Increased plant water-use efficiency (WUE) as a result of rising atmospheric CO2 decreases drought stress and allows trees to expand into arid grasslands, and grasslands to expand into deserts and semi-deserts (electronic supplementary material, figure S3). Savannahs are projected to spread in the Horn of Africa in response to precipitation increase. Distribution change in response to rainfall changes should, however, be interpreted with caution. We use only a single GCM forced with a single emissions scenario. Rainfall increase in the Horn of Africa is simulated by multiple earth system models (ESMs); however, this result is not robust and is reversed when downscaled using some regional models [42].

Most of the simulated change between 1990 and 2070 results from CO2 fertilization of tree growth in Africa and South America, facilitating the spread of forests into areas previously occupied by savannahs (electronic supplementary material, figure S3). A total of 45% and 57% of the areas projected as savannahs in 1990 in Africa and South America, respectively, might convert to other biomes by 2070 (figure 5). Fire plays a key role in limiting the distribution of forests in areas climatically suitable to both forests and savannahs. When fire is switched off in all variants of aDGVM, forest distributions expand and savannahs contract [17,29,34]. With increasing CO2, faster tree growth rates increase the probability of forest seedlings and saplings attaining sizes invulnerable to fire, subsequently suppressing grass growth and decreasing fire activity (electronic supplementary material, figure S4). This contradicts the TTR-SDM projection where savannahs are most vulnerable at their boundary with grasslands and deserts. By contrast, Australian savannahs remain relatively stable, with only 5% projected to convert by 2070 (figure 5). aDGVM1-Oz does simulate increasing tree density and overground carbon in response to rising CO2 and climatic change [34]; however, this does not lead to conversion to forest as is projected in Africa using aDGVM1-Africa, but rather denser savannahs (electronic supplementary material, figure S5).

4. Discussion

We find striking differences in projections of the future distribution of the savannah biome among DVMs and the TTR-SDM. This is not surprising given that only the DVMs include the impact of rising atmospheric CO2 and competitive interactions between grasses and trees. Even in DVM simulations with fixed CO2, projections for savannah extent in 2070 differ from the TTR-SDM. This difference is partially due to the TTR-SDM projections reflecting instantaneous vegetation change following climatic change, whereas the DVM explicitly stimulates the rate of change and resulting lag. The importance of demographic processes for determining tree/grass ratios and biome extent in savannahs suggests that even complex SDMs—such as the TTR-SDM—cannot represent savannah sensitivity to climatic change over realistic timescales.

Scheiter & Higgins [43] show how competitive interactions between grasses and trees define the conditions under which savannahs can occur, and February et al. [44] show experimentally that grasses compete more effectively than trees at increased rainfall in dry climates (approx. 550 mm MAP), even though at much higher rainfall trees would outcompete grasses. The individual-based structure of all aDGVM versions allows for above- and below-ground competitive interactions among trees and grasses, and demographic impacts of fire to be represented. This is not to say that the simulations provided by DVMs can be assumed to be inherently superior to SDM projections and hence reliable. Incorrect parametrization, or the absence of other mechanisms modulating responses to rising CO2 and changing water availability will bias projections. For example, the aDGVM does not include a nitrogen cycle interaction with plant growth. Nitrogen availability alters tree/grass ratios in savannahs and reduces the CO2 fertilization effect on tree growth [45,46]. Indeed, Higgins et al. [47] have shown in a field experiment that tree-dominated experimental plots have lower rates of N-mineralization than grass-dominated plots, suggesting that progressive nitrogen limitation may act as a negative feedback on CO2 fertilization-fuelled expansion of trees in savannahs.

The biased projections of contemporary savannah distribution in Australia and South America obtained when applying aDGVM1-Africa to these continents show the impact of parameter misspecification. It is common practice when projecting vegetation change in response to environmental change globally to assume that biomes and the PFTs from which they are assembled are globally coherent units [13]. Biogeographic variation within biomes is ignored, or assumed to be small relative to among-biome differences. This simplification has convincingly been shown to be invalid for tropical savannahs. Lehmann et al. [23] showed that regional variation in the functional relationships among woody vegetation, fire and climate would underpin regional differences in the response of woody plant structure and above-ground carbon to global change. These intercontinental functional differences are partially attributable to differences in the architecture, competitive interactions and fire resistance of regional floras [20,48].

Here for the first time, we provide projections of future savannah biome distribution using different, regionally appropriate model versions for the same biome using a DGVM (figure 4c). African and South American savannahs are projected to decrease in extent as increasing atmospheric CO2 concentration drives forest expansion into savannahs. In Africa and Australia, some savannah spread is projected into areas currently too arid to support savannah trees—driven by increased WUE and water availability [49]. This occurs in concert with grassland expansion into deserts and semi-deserts. Grass and tree expansion in aridlands is not simulated in South America as these climates do not occur within the regions to which this analysis is restricted. The directionality of these changes is supported by numerous experimental and observational studies that have documented forest expansion in savannahs in South America and Africa, and increasing tree and grass cover in certain African aridlands [5053].

Directly transposing our understanding of savannah sensitivity to increasing atmospheric CO2 concentration from Africa—synthesized in aDGVM1-Africa—to Australia would suggest that forest encroachment into mesic savannahs is a likely outcome. However, this scenario is not projected by aDGVM1-Oz. Rather, savannahs are projected to remain stable in 2070 under the given climate change and emission scenario. Limited evidence exists for a coherent trend of forest expansion into Australian savannahs [54]. Here parametrizing Australian trees as more sensitive to shading than African savannah trees and reducing the shading effect of trees on grasses in Australia is partially responsible for the future resilience of Australian savannahs. Grasses are able to persist even under dense stands and self-shading enforces a negative feedback on tree cover increase. High simulated fire activity in Australia may further mitigate forest expansion (electronic supplementary material, figure S4). This example shows the value of applying regionally appropriate models. Most databases of plant properties used to parametrize global vegetation in models of the carbon cycle and earth system are extremely biased in their geographical sampling [55], and hence this type of misspecification is likely to be the rule rather than the exception.

Over large areas of the tropics, climatic conditions do not predetermine the dominant biome [41,56]. In areas where multiple stable biomes are possible, alternative states are maintained by feedbacks among tree cover, grass growth and fire activity. Which of the potential biomes is indeed found is determined by historical events which initialize and maintain these feedback loops [57]. This creates a challenge when attempting to model realized biome distributions, as model fit will be contingent upon historical events and initial conditions [58]. Regional differences in palaeoclimate and palaeovegetation as well as disturbance history will compound differences in the contemporary relationship between environmental conditions and vegetation patterns, in addition to regional differences due to floristics [59]. Using realistic, regionally appropriate initializations in DVMs can remedy this problem. However, this undermines any attempt to model savannah distributions using correlative SDMs which do not include a representation of physiological and ecological processes, as differences between fundamental and realized climate niches might be exacerbated when alternative stable states are possible.

Given that the projected changes in vegetation are driven by a single scenario for future climate and atmospheric CO2 change, it is probable that the magnitude and geographical distribution of changes will differ from the projections presented here. Furthermore, we do not model human land-use and land-use change, which will alter the patterns and direction of change projected here through desertification and deforestation. Nonetheless, the projection of approximately half the current extent of savannahs in Africa and South America converting to other biomes by 2070 is concerning. The disruption to entrenched agricultural practices and ecotourism would incur enormous economic cost, and negative impacts on savannah biodiversity would be substantial [60,61]. However, the potential to respond to and mitigate these changes exists. African savannahs are unique in maintaining a relatively intact mammalian megafauna. This megafauna can alter woody plant abundance at regional scales [62]. Human extraction of fuelwood occurs extensively in Africa, supporting human livelihoods and bringing significant economic benefits [63]. The management of this human resource use and animal densities can be altered to encourage or discourage woody plant spread depending on desired outcomes.

Savannah distributions are notoriously hard to predict using climate alone. Over much of the tropics of Africa, Australia and South America, the environmental template does not predetermine whether savannahs, forests or grasslands will occur [41,64]. Intercontinental differences in savannah function limit the applicability of DVMs with globally defined homogeneous biomes. The challenge of designing and parametrizing regionally appropriate DVMs incorporating the complexity necessary to capture savannah dynamics is great. But given the accumulation of evidence suggesting that savannahs are at risk from human-induced climate change and increasing atmospheric CO2 concentrations, this challenge must be met.

Supplementary Material

Supplementary figures and tables:


We thank Caroline Lehmann for providing us with savannah distribution maps and encouraging this study.

Data accessibility

All simulated current and future Savannah distributions are freely available to download in raster format from the Dryad Digital Repository:

Authors' contributions

G.R.M. led the study and writing. L.L., S.I.H., S.S. and A.T. contributed to writing. G.R.M., L.L., S.I.H., S.S. and A.T. ran models and provided data.

Competing interests

We declare we have no competing interests.


G.R.M. is funded by the National Research Foundation (NRF) of South Africa's PDP programme. S.S. and L.L. were funded by the Deutsche Forschungsgemeinschaft (DGF) grants SCHE 1719/1-1, SCHE 1719/2-1.


1. Higgins SI, Bond WJ, Trollope WS 2000. Fire, resprouting and variability: a recipe for grass–tree coexistence in savanna. J. Ecol. 88, 213–229. (doi:10.1046/j.1365-2745.2000.00435.x)
2. Sankaran M, et al. 2005. Determinants of woody cover in African savannas. Nature 438, 846–849. (doi:10.1038/nature04070) [PubMed]
3. Grace J, Josè JS, Meir P, Miranda HS, Montes RA 2006. Productivity and carbon fluxes of tropical savannas. J. Biogeogr. 33, 387–400. (doi:10.1111/j.1365-2699.2005.01448.x)
4. Sankaran M, Ratnam J, Hanan NP 2004. Tree–grass coexistence in savannas revisited–insights from an examination of assumptions and mechanisms invoked in existing models. Ecol. Lett. 7, 480–490. (doi:10.1111/j.1461-0248.2004.00596.x)
5. Cramer W, et al. 2001. Global response of terrestrial ecosystem structure and function to CO2 and climate change: results from six dynamic global vegetation models. Glob. Change Biol. 7, 357–373. (doi:10.1046/j.1365-2486.2001.00383.x)
6. Baudena M, et al. 2015. Forests, savannas, and grasslands: bridging the knowledge gap between ecology and Dynamic Global Vegetation Models. Biogeosciences 12, 1833–1848. (doi:10.5194/bg-12-1833-2015)
7. Lehmann CER, Archibald S, Hoffmann WA, Bond WJ 2011. Deciphering the distribution of the savanna biome. New Phytol. 191, 197–209. (doi:10.1111/j.1469-8137.2011.03689.x) [PubMed]
8. Guisan A, Thuiller W 2005. Predicting species distribution: offering more than simple habitat models. Ecol. Lett. 8, 993–1009. (doi:10.1111/j.1461-0248.2005.00792.x)
9. Higgins SI, et al. 2012. A physiological analogy of the niche for projecting the potential distribution of plants. J. Biogeogr. 39, 2132–2145. (doi:10.1111/j.1365-2699.2012.02752.x)
10. Thornley JHM. 1998. Modelling shoot:root relations: the only way forward? Ann. Bot. 81, 165–171. (doi:10.1006/anbo.1997.0529)
11. Farquhar GD, von Caemmerer SV, Berry JA 1980. A biochemical model of photosynthetic CO2 assimilation in leaves of C3 species. Planta 149, 78–90. (doi:10.1007/BF00386231) [PubMed]
12. Collatz G, Ribas-Carbo M, Berry J 1992. Coupled photosynthesis-stomatal conductance model for leaves of C4 plants. Funct. Plant Biol. 19, 519–538.
13. Prentice IC, Bondeau A, Cramer W, Harrison SP, Hickler T, Lucht W, Sitch S, Smith B, Sykes MT 2007. Dynamic global vegetation modeling: quantifying terrestrial ecosystem responses to large-scale environmental change. In Terrestrial ecosystems in a changing world (eds Canadell JG, Pataki DE, Pitelka LF). Global Change—The IGBP Series, pp. 175–192. Berlin, Germany: Springer.
14. Sitch S, et al. 2008. Evaluation of the terrestrial carbon cycle, future plant geography and climate carbon cycle feedbacks using five Dynamic Global Vegetation Models (DGVMs). Glob. Change Biol. 14, 2015–2039. (doi:10.1111/j.1365-2486.2008.01626.x)
15. Huntingford C, et al. 2013. Simulated resilience of tropical rainforests to CO2-induced climate change. Nat. Geosci. 6, 268–273. (doi:10.1038/ngeo1741)
16. Sitch S, et al. 2003. Evaluation of ecosystem dynamics, plant geography and terrestrial carbon cycling in the LPJ dynamic global vegetation model. Glob. Change Biol. 9, 161–185. (doi:10.1046/j.1365-2486.2003.00569.x)
17. Scheiter S, Langan L, Higgins SI 2013. Next-generation dynamic global vegetation models: learning from community ecology. New Phytol. 198, 957–969. (doi:10.1111/nph.12210) [PubMed]
18. Sakschewski B, von Bloh W, Boit A, Rammig A, Kattge J, Poorter L, Peñuelas J, Thonicke K 2015. Leaf and stem economics spectra drive diversity of functional plant traits in a dynamic global vegetation model. Glob. Change Biol. 21, 2711–2725. (doi:10.1111/gcb.12870) [PubMed]
19. Verheijen LM, Brovkin V, Aerts R, Bönisch G, Cornelissen JHC, Kattge J, Reich PB, Wright IJ, van Bodegom PM 2013. Impacts of trait variation through observed trait–climate relationships on performance of an Earth system model: a conceptual analysis. Biogeosciences 10, 5497–5515. (doi:10.5194/bg-10-5497-2013)
20. Moncrieff GR, Lehmann CE, Schnitzler J, Gambiza J, Hiernaux P, Ryan CM, Shackleton CM, Williams RJ, Higgins SI 2014. Contrasting architecture of key African and Australian savanna tree taxa drives intercontinental structural divergence. Glob. Ecol. Biogeogr. 23, 1235–1244. (doi:10.1111/geb.12205)
21. Heberling JM, Fridley JD 2012. Biogeographic constraints on the world-wide leaf economics spectrum. Glob. Ecol. Biogeogr. 21, 1137–1146. (doi:10.1111/j.1466-8238.2012.00761.x)
22. Mitchell N, et al. 2015. Functional traits in parallel evolutionary radiations and trait-environment associations in the Cape Floristic region of South Africa. Am. Nat. 185, 525–537. (doi:10.1086/680051) [PubMed]
23. Lehmann CER, et al. 2014. Savanna vegetation-fire-climate relationships differ among continents. Science 343, 548–552. (doi:10.1126/science.1247355) [PubMed]
24. Ratnam J, Bond WJ, Fensham RJ, Hoffmann WA, Archibald S, Lehmann CER, Anderson MT, Higgins SI, Sankaran M 2011. When is a ‘forest’ a savanna, and why does it matter? Glob. Ecol. Biogeogr. 20, 653–660. (doi:10.1111/j.1466-8238.2010.00634.x)
25. Trabucco A, Zomer RJ 2010. Global soil water balance geospatial database. CGIAR Consortium for Spatial Information. See
26. Hijmans RJ, Cameron SE, Parra JL, Jones PG, Jarvis A 2005. Very high resolution interpolated climate surfaces for global land areas. Int. J. Climatol. 25, 1965–1978. (doi:10.1002/joc.1276)
27. Shangguan W, Dai Y, Duan Q, Liu B, Yuan H 2014. A global soil data set for earth system modeling. J. Adv. Model. Earth Syst. 6, 249–263. (doi:10.1002/2013MS000293)
28. Mullen K, Ardia D, Gil DL, Windover D, Cline J 2011. DEoptim: an R package for global optimization by differential evolution. J. Stat. Softw. 40, 1–26. (doi:10.18637/jss.v040.i06)
29. Scheiter S, Higgins SI 2009. Impacts of climate change on the vegetation of Africa: an adaptive dynamic vegetation modelling approach. Glob. Change Biol. 15, 2224–2246. (doi:10.1111/j.1365-2486.2008.01838.x)
30. Scheiter S, Higgins SI, Osborne CP, Bradshaw C, Lunt D, Ripley BS, Taylor LL, Beerling DJ 2012. Fire and fire-adapted vegetation promoted C4 expansion in the late Miocene. New Phytol. 195, 653–666. (doi:10.1111/j.1469-8137.2012.04202.x) [PubMed]
31. Prentice IC, Harrison SP, Bartlein PJ 2011. Global vegetation and terrestrial carbon cycle changes after the last ice age. New Phytol. 189, 988–998. (doi:10.1111/j.1469-8137.2010.03620.x) [PubMed]
32. Ball JT, Woodrow IE, Berry JA 1987. A model predicting stomatal conductance and its contribution to the control of photosynthesis under different environmental conditions. In Progress in photosynthesis research, pp. 221–224. Berlin, Germany: Springer.
33. Higgins SI, Bond WJ, Trollope WS, Williams RJ 2008. Physically motivated empirical models for the spread and intensity of grass fires. Int. J. Wildland Fire 17, 595–601. (doi:10.1071/WF06037)
34. Scheiter S, Higgins SI, Beringer J, Hutley LB 2015. Climate change and long-term fire management impacts on Australian savannas. New Phytol. 205, 1211–1226. (doi:10.1111/nph.13130) [PubMed]
35. Giorgetta MA, et al. 2013. Climate and carbon cycle changes from 1850 to 2100 in MPI-ESM simulations for the Coupled Model Intercomparison Project phase 5. J. Adv. Model. Earth Syst. 5, 572–597. (doi:10.1002/jame.20038)
36. Gleckler PJ, Taylor KE, Doutriaux C 2008. Performance metrics for climate models. J. Geophys. Res. Atmos. 113, 104 (doi:10.1029/2007JD008972)
37. Taylor KE, Stouffer RJ, Meehl GA 2012. An overview of CMIP5 and the experiment design. Bull. Am. Meteorol. Soc. 93, 485–498. (doi:10.1175/BAMS-D-11-00094.1)
38. New M, Lister D, Hulme M, Makin I 2002. A high-resolution data set of surface climate over global land areas. Clim. Res. 21, 1–25. (doi:10.3354/cr021001)
39. Trabucco A, Zomer RJ, Bossio DA, van Straaten O, Verchot LV 2008. Climate change mitigation through afforestation/reforestation: a global analysis of hydrologic impacts with four case studies. Agric. Ecosyst. Environ. 126, 81–97. (doi:10.1016/j.agee.2008.01.015)
40. Higgins SI, Scheiter S 2012. Atmospheric CO2 forces abrupt vegetation shifts locally, but not globally. Nature 488, 209–212. (doi:10.1038/nature11238) [PubMed]
41. Staver AC, Archibald S, Levin SA 2011. The global extent and determinants of savanna and forest as alternative biome states. Science 334, 230–232. (doi:10.1126/science.1210465) [PubMed]
42. Niang I, Ruppel OC, Abdrabo MA, Essel A, Lennard C, Padgham J, Urquhart P 2014. Africa. In Climate Change 2014: impacts, adaptation, and vulnerability. Part B: Regional Aspects. Contribution of Working Group II to the Fifth Assessment Report of the Intergovernmental Panel on Climate Change (eds Barros VR, editor. et al.), pp. 1199–1265. Cambridge, UK: Cambridge University Press.
43. Scheiter S, Higgins SI 2007. Partitioning of root and shoot competition and the stability of savannas. Am. Nat. 170, 587–601. (doi:10.1086/521317) [PubMed]
44. February EC, Higgins SI, Bond WJ, Swemmer L 2013. Influence of competition and rainfall manipulation on the growth responses of savanna trees and grasses. Ecology 94, 1155–1164. (doi:10.1890/12-0540.1) [PubMed]
45. Sankaran M, Ratnam J, Hanan N 2008. Woody cover in African savannas: the role of resources, fire and herbivory. Glob. Ecol. Biogeogr. 17, 236–245. (doi:10.1111/j.1466-8238.2007.00360.x)
46. Norby RJ, Warren JM, Iversen CM, Medlyn BE, McMurtrie RE 2010. CO2 enhancement of forest productivity constrained by limited nitrogen availability. Proc. Natl Acad. Sci. USA 107, 19 368–19 373. (doi:10.1073/pnas.1006463107) [PubMed]
47. Higgins SI, Keretetse M, February EC 2015. Feedback of trees on nitrogen mineralization to restrict the advance of trees in C4 savannahs. Biol. Lett. 11, 20150572 (doi:10.1098/rsbl.2015.0572) [PMC free article] [PubMed]
48. Dantas VdL, Pausas JG 2013. The lanky and the corky: fire-escape strategies in savanna woody species. J. Ecol. 101, 1265–1272. (doi:10.1111/1365-2745.12118)
49. Morgan JA, et al. 2004. Water relations in grassland and desert ecosystems exposed to elevated atmospheric CO2. Oecologia 140, 11–25. (doi:10.1007/s00442-004-1550-2) [PubMed]
50. Mitchard E, Saatchi S, Lewis SL, Feldpausch T, Woodhouse I, Sonkè B, Rowland C, Meir P 2011. Measuring biomass changes due to woody encroachment and deforestation/degradation in a forest–savanna boundary region of central Africa using multi-temporal L-band radar backscatter. Remote Sens. Environ. 115, 2861–2873. (doi:10.1016/j.rse.2010.02.022)
51. Silva LCR, Sternberg L, Haridasan M, Hoffmann WA, Miralles-Wilhelm F, Franco AC 2008. Expansion of gallery forests into central Brazilian savannas. Glob. Change Biol. 14, 2108–2118. (doi:10.1111/j.1365-2486.2008.01637.x)
52. Masubelele ML, Hoffman MT, Bond WJ, Gambiza J 2014. A 50 year study shows grass cover has increased in shrublands of semi-arid South Africa. J. Arid Environ. 104, 43–51. (doi:10.1016/j.jaridenv.2014.01.011)
53. Stevens N, Erasmus BFN, Archibald S, Bond WJ 2016. Woody encroachment over 70 years in South African savannahs: overgrazing, global change or extinction aftershock? Phil. Trans. R. Soc. B 371, 20150437 (doi:10.1098/rstb.2015.0437) [PMC free article] [PubMed]
54. Murphy BP, Lehmann CE, Russell-Smith J, Lawes MJ 2014. Fire regimes and woody biomass dynamics in Australian savannas. J. Biogeogr. 41, 133–144. (doi:10.1111/jbi.12204)
55. Kattge J, et al. 2011. TRY–a global database of plant traits. Glob. Change Biol. 17, 2905–2935. (doi:10.1111/j.1365-2486.2011.02451.x)
56. Hirota M, Holmgren M, Van Nes EH, Scheffer M 2011. Global resilience of tropical forest and savanna to critical transitions. Science 334, 232–235. (doi:10.1126/science.1210657) [PubMed]
57. Moncrieff GR, Scheiter S, Bond WJ, Higgins SI 2014. Increasing atmospheric CO2 overrides the historical legacy of multiple stable biome states in Africa. New Phytol. 201, 908–915. (doi:10.1111/nph.12551) [PubMed]
58. Perring MP, De Frenne P, Baeten L, Maes SL, Depauw L, Blondeel H, Carón MM, Verheyen K 2015. Global environmental change effects on ecosystems: the importance of land-use legacies. Glob. Change Biol. 22, 1361–1371. (doi:10.1111/gcb.13146) [PubMed]
59. Moncrieff GR, Bond WJ, Higgins SI 2016. Revising the biome concept for understanding and predicting global change impacts. J. Biogeogr. 43, 863–873. (doi:10.1111/jbi.12701)
60. Gray EF, Bond WJ 2013. Will woody plant encroachment impact the visitor experience and economy of conservation areas? Koedoe 55, 9 (doi:10.4102/koedoe.v55i1.1106)
61. Parr CL, Gray EF, Bond WJ 2012. Cascading biodiversity and functional consequences of a global change-induced biome switch. Divers. Distrib. 18, 493–503. (doi:10.1111/j.1472-4642.2012.00882.x)
62. Asner GP, Levick SR, Kennedy-Bowdoin T, Knapp DE, Emerson R, Jacobson J, Colgan MS, Martin RE 2009. Large-scale impacts of herbivores on the structural diversity of African savannas. Proc. Natl Acad. Sci. USA 106, 4947–4952. (doi:10.1073/pnas.0810637106) [PubMed]
63. Cline-Cole RA, Main HAC, Nichol JE 1990. On fuelwood consumption, population dynamics and deforestation in Africa. World Dev. 18, 513–527. (doi:10.1016/0305-750X(90)90068-9)
64. Bond WJ, Woodward FI, Midgley GF 2005. The global distribution of ecosystems in a world without fire. New Phytol. 165, 525–538. (doi:10.1111/j.1469-8137.2004.01252.x) [PubMed]

Articles from Philosophical Transactions of the Royal Society B: Biological Sciences are provided here courtesy of The Royal Society