Large uncertainties in future biome changes in Africa call for flexible climate adaptation strategies

Anthropogenic climate change is expected to impact ecosystem structure, biodiversity and ecosystem services in Africa profoundly. We used the adaptive Dynamic Global Vegetation Model (aDGVM), which was originally developed and tested for Africa, to quantify sources of uncertainties in simulated African potential natural vegetation towards the end of the 21st century. We forced the aDGVM with regionally downscaled high‐resolution climate scenarios based on an ensemble of six general circulation models (GCMs) under two representative concentration pathways (RCPs 4.5 and 8.5). Our study assessed the direct effects of climate change and elevated CO2 on vegetation change and its plant‐physiological drivers. Total increase in carbon in aboveground biomass in Africa until the end of the century was between 18% to 43% (RCP4.5) and 37% to 61% (RCP8.5) and was associated with woody encroachment into grasslands and increased woody cover in savannas. When direct effects of CO2 on plants were omitted, woody encroachment was muted and carbon in aboveground vegetation changed between –8 to 11% (RCP 4.5) and –22 to –6% (RCP8.5). Simulated biome changes lacked consistent large‐scale geographical patterns of change across scenarios. In Ethiopia and the Sahara/Sahel transition zone, the biome changes forecast by the aDGVM were consistent across GCMs and RCPs. Direct effects from elevated CO2 were associated with substantial increases in water use efficiency, primarily driven by photosynthesis enhancement, which may relieve soil moisture limitations to plant productivity. At the ecosystem level, interactions between fire and woody plant demography further promoted woody encroachment. We conclude that substantial future biome changes due to climate and CO2 changes are likely across Africa. Because of the large uncertainties in future projections, adaptation strategies must be highly flexible. Focused research on CO2 effects, and improved model representations of these effects will be necessary to reduce these uncertainties.


| INTRODUC TI ON
Climate change is expected to drive changes in ecosystem structure and functioning as well as geographical shifts of ecosystems and biomes (Engelbrecht & Engelbrecht, 2016;Hoegh-Guldberg et al., 2018;Niang et al., 2014). Such ecosystem changes will impact the potential for future land uses and the livelihoods of people in Africa, where agriculture accounted for ~50% of employment in 2019 (ILO, 2019). The critical ecosystem services provided vary for different biomes (Chapin et al., 2011, p. 428), such as carbon sequestration (forests), and pasture for grazing (grasslands and savannas, Naidoo et al., 2008). Biome changes and shifts therefore impact ecosystem services (Gonzalez et al., 2010). Furthermore, many animal species are strongly associated with certain biome types (Jetz & Fine, 2012), which together with vegetation changes implies large potential impacts on prevailing biodiversity.
Biome shifts attributed to climate change have already been observed in Africa (Niang et al., 2014), but do not show a consistent pattern. For instance, declines in tree density and changes in species composition have led to a southward shift of the savanna vegetation zone in West Africa (Gonzalez et al., 2012) into previously more humid areas. At the same time, increased woody cover in savannas has been reported at many sites across Africa, including West Africa.
Increasing atmospheric CO 2 concentrations ([CO 2 ]), changes in land management and altered rainfall patterns were identified as likely drivers of this woody vegetation increase (Stevens et al., 2017).
With the widespread mixture of grasses with C 4 photosynthesis and woody plants with C 3 photosynthesis across African savannas, plant-physiological effects of increasing [CO 2 ] might change the competitive balance between grasses and trees and thus play a key role for future ecosystem changes (Bond & Midgley, 2000;Midgley & Bond, 2015;Osborne et al., 2018). Furthermore, changes in plant growth interact with changes in fire regimes, and direct enhancement of tree sapling growth rates under elevated [CO 2 ] (eCO 2 ) increases the likelihood that young tree individuals escape the 'firetrap'. This might lead to a positive feedback, where an initial increase in woody vegetation suppresses grasses, thereby reduces fire activity, which in turn benefits woody plants (Bond & Midgley, 2012;Midgley & Bond, 2015).
However, the magnitude of eCO 2 effects, both directly on plant growth through CO 2 fertilization of photosynthetic efficiency (Long et al., 2004) and indirectly on ecosystem hydrology (through a reduction of stomatal conductance, g s ; Ainsworth & Rogers, 2007;Medlyn et al., 2001), is still debated (Körner, 2015;Medlyn et al., 2015;Zaehle et al., 2014). Free-Air carbon dioxide enrichment (FACE) experiments suggest that an increase in carbon assimilation in C 3 plants does not necessarily scale to increased plant growth (e.g. Körner et al., 2005;Medlyn et al., 2015), particularly when limited by nutrient availability (e.g. Jiang, Caldararu, et al., 2020;Jiang, Medlyn, et al., 2020;Norby et al., 2010). For ecosystem hydrology and water use efficiency (WUE), eCO 2 effects translate into a balance of decreased transpiration and water demand due to reduced g s (De Kauwe et al., 2013) and increased water demand following increased net primary production (NPP; Warren et al., 2011).
Dynamic Global Vegetation Models (DGVMs) are a widely used tool to project future ecosystem changes and to attribute ongoing changes to different drivers and mechanisms (Prentice et al., 2007).
DGVMs simulate ecophysiological processes, such as photosynthesis and plant growth, vegetation dynamics and structure, geographical distribution of plant biomes and biogeochemical cycles (e.g. water and carbon), in particular in response to climate change (e.g. Prentice et al., 2007;Sitch et al., 2008). However, the effects of eCO 2 on plants and interactions with nutrient limitations in DGVMs are still uncertain Medlyn et al., 2015). If key plant demographic processes such as mortality, recruitment of tree saplings and fire impacts on plant individuals are implemented in DGVMs (Fisher et al., 2018), they can capture complex dynamics in savanna ecosystems.
In this study, we simulated the impacts of climate change and eCO 2 on carbon stocks, WUE and biome distribution of potential natural vegetation (PNV) in Africa using the adaptive Dynamic Global Vegetation Model (aDGVM; Scheiter & Higgins, 2009).
The aDGVM was originally developed for Africa and its savannas (Scheiter & Higgins, 2009) and has been applied and tested in several Africa-focused case studies (e.g. Scheiter & Savadogo, 2016). It simulates woody plant demography for individual trees and this allows fire impacts to be conditioned on individual tree size. Advancing on earlier studies with the aDGVM, we used an ensemble of regional climate model (RCM) data based on six downscaled general circulation models (GCMs; Archer et al., 2018;Davis-Reddy et al., 2017;Engelbrecht et al., 2015) and two Representative Concentration Pathway (RCP) scenarios (RCPs 4.5 and 8.5, Stocker et al., 2013). This is the first time that this ensemble of downscaled GCM data has been used as a climate driver for a DGVM covering Africa. We ran the aDGVM with and without eCO 2 effects enabled to assess uncertainty related to plant-physiological CO 2 effects and to identify important drivers of vegetation change. Our study quantifies how uncertainty in projections caused by CO 2 , in particular concerning CO 2 effects on WUE and biomass, interacts with uncertainty due to the choice of GCM and RCP.

| The aDGVM
For this study, we used the well-tested aDGVM, a regionally adapted DGVM (e.g. Scheiter & Higgins, 2009;Scheiter et al., 2015;Scheiter & Savadogo, 2016). The aDGVM was developed for tropical and subtropical grass-tree ecosystems (for details, see Scheiter & Higgins, 2009;Scheiter et al., 2012). It incorporates ecophysiological processes that are commonly implemented in DGVMs (Prentice et al., 2007). State variables such as photosynthetic rates, biomass or height are simulated for individual plants depending on environmental conditions. For each plant the aDGVM dynamically simulates leaf phenology and flexibly determines carbon allocation to plant biomass compartments (roots, stem or leaf biomass). aDGVM prioritizes carbon allocation to compartments that are most limiting for plant growth based on the constraining factors water, light or photosynthesis. Physiological processes, such as photosynthesis, g s and transpiration, are simulated for each individual plant based on environmental and plant individual state variables such as light availability. Stomatal conductance is represented by the model from Ball et al. (1987) and is directly proportional to relative humidity, and the ratio of photosynthesis and [CO 2 ] at the leaf surface (Medlyn et al., 2001).
The aDGVM simulates 1-ha plots that are assumed to be representative for the simulated grid cell. Grasses with C 3 and C 4 photosynthesis are each implemented as two types of super-individuals that represent grasses growing beneath tree canopies and in tree canopy gaps. The aDGVM distinguishes savanna and forest trees  as two distinct tree types that differ in their fire and shade tolerance (Bond & Midgley, 2001;Ratnam et al., 2011 Fire in the aDGVM is determined by fuel load and fuel moisture, both dependent on biomass growth and thus indirectly influenced by climate, and wind speed following Higgins et al. (2008).
Ignitions are simulated as random events decoupled from climatic or regional factors. Fire disturbance is therefore one of the factors causing some stochasticity in the model results. Whether fire spreads after a stochastic ignition event depends on fire intensity and a previously determined likelihood that a fire will spread (Scheiter & Higgins, 2009).
In contrast to many other DGVMs, simulating individual trees in aDGVM allows accounting for fire effects on individual plants and vegetation structure as a function of individual plant height (Scheiter & Higgins, 2009). Height influences if trees survive grass fires because only tall enough trees can escape the flame or topkill zone of a fire (Higgins et al., 2000). In addition, in contrast to forest trees, grasses and savanna trees in aDGVM are able to resprout after fire damage (Bond & Midgley, 2001). Population composition evolves dynamically in the model as a result of interactions of, for example, fire and [CO 2 ]. Random events in demography subroutines for tree populations add to the stochasticity in model results.

| Regional climate model simulations
Climate input data consisted of an ensemble of six downscaled GCM projections under two mitigation scenarios generated with the variable-resolution conformal-cubic atmospheric model (CCAM; McGregor, 2005). The simulations were performed at the Council for Scientific and Industrial Research in South Africa (Archer et al., 2018;Davis-Reddy et al., 2017;Engelbrecht et al., 2015). The GCM projections formed part of the Coupled Model Intercomparison Project Phase 5 (CMIP5 ; Table S1; Stocker et al., 2013). The downscaling procedure involved CCAM being integrated globally at a quasi-uniform resolution of about 50 km in the horizontal, forced at its lower boundary by sea-ice concentrations and bias-corrected sea-surface temperatures from the host GCMs (Engelbrecht et al., 2015). The CCAM simulations were performed for the period 1961-2099 and for the low mitigation scenario RCP8.5 and modest-high mitigation scenario RCP4.5.
The downscaled climate data sets were bias-corrected to the monthly climatologies of temperature and rainfall from CRU TS3.1 data for the period 1961-1990(Engelbrecht & Engelbrecht, 2016Engelbrecht et al., 2015). Previously, the CCAM downscalings have been shown to realistically represent present-day climate over southern Africa (e.g. Engelbrecht et al., 2009Engelbrecht et al., , 2013Engelbrecht et al., , 2015. CCAM output is available on a latitude-longitude grid of 0.5° resolution and at a daily time step. We used daily precipitation, daily minimum and maximum temperature, daily wind speed and daily relative humidity from the CCAM data set to force aDGVM (see Figure S1 for mean annual precipitation (MAP) and temperature maps). As projected radiation data were not available, we derived present-day radiation from sunshine percentage (Allen et al., 1998) from New et al. (2002). Thus, our vegetation simulations are based on an ensemble of climate data providing a range of GCM-projected climate change under two emission scenarios at high spatial and temporal resolution.

| Simulation design
We simulated vegetation dynamics in Africa from 1971 to 2099 at 0.5° resolution by forcing the aDGVM with the climate ensemble described above and soil data from the Global Soil Data Task Group (2000). A model spin-up of 210 years was simulated to allow vegetation to reach equilibrium with environmental forcing.
To that end, we used a random series of climate data from the period 1971-1979. In our experimental setup, we combined changes in climatic conditions with increases in [CO 2 ] (eCO 2 ) for two greenhouse gas emission scenarios, RCPs 4.5 and 8.5. To estimate the extent of the eCO 2 effect and its uncertainties, we repeated the same simulations with climate conditions following the two RCP scenarios, but with [CO 2 ] rising only to 400 ppm and then keeping [CO 2 ] fixed (fCO 2 ).
Stochastic effects within DGVMs can be factored in when conducting replicate runs for each ensemble member. Due to high computing times for continental-scale high-resolution simulations, for each GCM only one simulation was conducted per RCP-CO 2 scenario. In a previous study with the aDGVM at regional scale, more than 60% of replicate simulations agreed with respect to biome projections for the year 2100 for large parts of the simulated area, the Limpopo province in South Africa . As opposed to , the use of daily climate input data for our simulations helped to avoid the generation of daily climate time series with the aDGVM and thus removed the associated stochasticity. In addition, the high spatial resolution of our simulations effectively acts as a replication in space with smooth simulation patterns in space indicating low aDGVM-caused variability in the simulation.

| Analysis of simulation results
Aboveground carbon dynamics and plant-physiological effects of eCO 2 were analysed in more detail. We defined stand WUE as the ratio of NPP and plant transpiration, following the definition by De Kauwe et al. (2013) for FACE results. We used NPP, transpiration and WUE to isolate which processes-CO 2 fertilization of photosynthesis or increased WUE via reduced transpiration through reduced g s -are driving plant-physiological responses to eCO 2 . (ω 2 ) measure to quantify the effect size of explanatory variables (see Supporting Information S1 for R code; R Core Team, 2015).
CO 2 scenario, RCP scenario, and GCM and their two-way interactions were used to explain differences of the dependent variables.
Using ω 2 to estimate the proportion of variance explained (Olejnik & Algina, 2003), we quantitatively evaluate the magnitude of the differences between the ensemble members (White et al., 2014).
As computational limitations restricted us from implementing replicate runs, analysing for three-way interactions between the three explanatory variables was not possible.
To study potential future biome distributions and biome changes, we classified the simulated vegetation into seven biome types: desert, C 3 grassland, C 4 grassland, C 3 savanna, C 4 savanna, woodland and forest. The classification scheme is based on simulated tree cover, grass biomass and composition of tree and grass types  with minimum grass biomass levels for grassland classifications see Table S2). In this definition scheme, the main difference between savannas and woodlands is the predominance of savanna or forest tree types, which is in turn an indicator of fire activity.
Based on this biome classification scheme, we identified biomes for the above-listed time periods for Africa for each RCP-CO 2scenario (RCP4.5/8.5 with eCO 2 /fCO 2 ). The biomes identified for all GCM simulations per scenario were used to derive consensus biome maps for each scenario. For each grid cell, the biome type that was most common in the ensemble member maps was used to derive each scenario's consensus biome. According to the binomial distribution, two ensemble members agreeing in the simulated biome do not have a higher probability than an outcome by chance. Such grid cells therefore were denominated as having 'No consensus'. In addition, we plotted the number of ensemble members that simulated the consensus vegetation type for each time period as an indicator of agreement within the simulated ensemble.
We used the number of ensemble members per RCP-CO 2scenario combination that showed a biome change from 2000-2019 to 2080-2099 (maximum six ensemble members) as an indicator for the probability of biome change under future climatic conditions. We therefore derived a measure of uncertainty from the number of simulated biome changes within a scenario. The type of biome change, that is, whether ensemble members simulate the same or different biome transitions, was not considered in this uncertainty assessment.
To illustrate the overall potential consequences (i.e. areal increase or decrease) of biome changes for each biome type, we determined   Table 1).

| Current aboveground biomass and future projections
Both RCP-eCO 2 scenarios showed small biomass decreases in a narrow belt along the equator. North and south of the equator both RCP scenarios showed a general pattern towards woody encroachment, with higher biomass increases for RCP8.5.
Spatial patterns of biomass change in the fCO 2 scenarios were more variable. Increases prevailed over decreases under RCP4.5 and decreases prevailed under RCP8.5 (Table 1; Figure 1c-f). The Ethiopian Highlands were the only larger area that showed mean biomass increases for all four RCP-CO 2 scenarios. However, these increases were weaker under fCO 2 than under eCO 2 . This biomass increase is in line with precipitation increases predicted for this region ( Figure S1).
Both RCP-eCO 2 scenarios showed an appreciable carbon sink for Africa by the end of the 21st century in all GCM simulations ( Figure 2a; Figure S3). The increase under RCP8.5 was more pronounced, whereas under RCP4.5 mean total aboveground carbon started to saturate towards the end of the century (81.3 and 70.0 Pg C in 2080-2099, respectively, see Table 2 for % change). Saturating biomass increases under RCP4.5-eCO 2 followed the projected mid-century peak in greenhouse gas emissions for RCP4.5 (Stocker F I G U R E 1 Simulated and observationderived aboveground biomass in Africa and simulated change in biomass. Current aboveground biomass in Africa (t/ha) is derived from the ensemble mean across all 24 ensemble members of aDGVM-simulated biomass (2000-2019, a) and Avitabile et al. (2016, b). Change in aboveground biomass between 2000-2019 and 2080-2099 under RCP4.5 and 8.5 with eCO 2 (c, d) and with fCO 2 (e, f) is based on the mean of all six ensemble members of the respective scenario. aDGVM, adaptive Dynamic Global Vegetation Model; eCO 2 , elevated CO 2 ; fCO 2 , CO 2 fixed at 400 ppm; RCP, representative concentration pathway , 2013). Variability within each RCP ensemble due to differences between the six GCMs (±SD, 2 × SD = 9.7 and 11.2 Pg C, respectively for RCPs 4.5 and 8.5) was similar to the difference of the means of the two eCO 2 RCP scenarios (11.2 Pg C, see Table S3).  Table 2 for % change) was less than that in 2000. For RCP4.5-fCO 2 , the ensemble mean biomass in 2080-2099 was 55.2 Pg C. It depended on the downscaled GCM whether a small source or a small sink was projected for RCP4.5 under fCO 2 ( Figure S3). The ω 2 metric indicated that variation in total carbon between all 24 ensemble members was mainly explained by the CO 2 scenarios, followed by interaction effects of CO 2 and RCP scenarios (Table 3; see Table S5 for corresponding F-values from ANOVA).

| Future projections of WUE and underlying plant-physiological processes
Changes in WUE, here defined as the ratio of NPP and transpiration, until 2080-2099 were similar to the trends simulated for aboveground  RCP8.5, fCO 2 -6 to -22% -9 to -16% AGB, aboveground biomass; eCO 2 , elevated CO 2 ; fCO 2 , CO 2 fixed at 400 ppm; RCP, representative concentration pathway; WUE, water use efficiency biomass for all four RCP-CO 2 scenarios ( Figure 2b; Table 2). Yet, under RCP4.5-eCO 2 , WUE levelled off earlier than carbon in aboveground biomass. Both fCO 2 scenarios showed a decrease in mean WUE. For NPP and WUE, the shape of the time series was almost identical for each individual scenario (Figure 2b; Figure S4a). For transpiration on the other hand, all four scenarios showed increases ( Figure S4b).
Within an RCP, transpiration was lower for fCO 2 than for eCO 2 .

WUE increased in most of the simulated grid cells between 2000-
2019 and 2080-2099 in eCO 2 simulations, and relative change ratios were higher than in fCO 2 simulations (Figure 3a). With a stronger decrease under eCO 2 than fCO 2 and for RCP8.5 than for RCP4.5, g s showed the opposite trend ( Figure 3b). Although g s decreased with increasing [CO 2 ], transpiration increased for all four RCP-CO 2 scenarios ( Figure S4b). NPP was more sensitive to MAP changes under RCP8.5 than RCP4.5, with a steeper slope for eCO 2 (Figure 3c). We found increased NPP in both eCO 2 scenarios even for grid cells where MAP decreased. Variation in WUE was largely explained by the CO 2 scenario. The interaction effect between the CO 2 scenario and the RCP scenario, as well as the RCP scenario on its own also explain some of the variation (Table 3). The effect of RCP scenario on WUE was stronger than its effect on carbon in aboveground biomass.   aDGVM simulated that larger areas changed biome state than under RCP4.5 (Table 1). Large areas showed transitions from C 4 savanna to forest, especially under RCP8.5 ( Figure 4c).

| Projected biome changes and population dynamics
As opposed to the eCO 2 scenarios, a decrease in forests dominated biome cover changes in the fCO 2 scenarios ( Figure 5). In total, under fCO 2 , a smaller area was projected to experience biome transitions (Table 1). For RCP8.5, we found a strong decrease in forest cover (5.8 percentage points) and an increase in C 4 -dominated biomes. Forests mainly changed to woodlands in both RCP scenarios ( Figure 5c; Figure S7c). Under RCP8.5-fCO 2 , the core savanna area in 2080-2099 in southern Africa was smaller and located further north than in the other three scenarios (Figures S8e,f and S9e,f), with a pronounced transition from C 4 savanna to C 4 grasslands in the Kalahari region (Figure 5a,b). Under RCP4.5, the cover fractions of biome types did not change as much, with transitions in favour of C 4 savanna and woodlands.
Rates of tree biomass change were biome specific and varied between CO 2 scenarios (Figure 7). Under eCO 2 , C 4 savannas showed a stronger relative increase in tree biomass change rates than woodlands and forests for both RCP scenarios, hinting at population dynamic effects (Figure 7a,b). Tree biomass change in C 4 savannas was more sensitive to NPP change than in other woody biomes for both eCO 2 scenarios. With fCO 2 , an increase in NPP often did not lead to an increase in tree biomass, especially for RCP8.5 and forest biomes.

| Uncertainty of biome projections
The simulated biomes for the six ensemble members of all four RCP-CO 2 scenarios agreed well at the beginning of the simulations scenarios, under RCP8.5, a larger area in Africa had a medium to high probability for biome transitions than under RCP4.5 (Figure 9; Figure   S11).

| D ISCUSS I ON
The aim of this study was to provide projections of climate-and [CO 2 ]driven vegetation changes in Africa until the end of the century using a regionally adapted DGVM and regionally downscaled climate data.
We analysed plant-physiological and vegetation dynamic responses to eCO 2 and quantified the impact of three sources of uncertainty on vegetation projections: the influence of climate models (GCM implementations), future global socio-economic development (RCP scenarios) and strength of CO 2 effects on plant growth (CO 2 scenarios).
We found that large regions in Africa are likely to experience changes in biomes and carbon stocks, especially with eCO 2 response scenarios. The extent of the projected changes depended on the chosen scenario and patterns were less clear for biome changes. Projected change in aboveground biomass was driven primarily by CO 2 fertilization and showed that assumptions concerning CO 2 effects (eCO 2 or fCO 2 ) caused the strongest variability in future projections (Table 3; Figure 2; Figure S3). Even medium-impact scenarios (RCP4.5, irrespective of CO 2 scenario) suggested considerable ecosystem change ( Figure 6; Figure S6b,c).

| Plant-physiological effects of eCO 2 and drivers of woody encroachment
Elevated CO 2 led to strong woody encroachment in our projections and was accompanied by enhanced WUE. Increases in NPP clearly drove WUE enhancement, although both photosynthesis (increase in NPP via more efficient CO 2 fixation) and g s (decrease) are affected by eCO 2 in aDGVM. We therefore conclude that increased WUE, intensification of carbon sinks and woody encroachment are mainly driven by CO 2 fertilization of photosynthesis in aDGVM.
This finding seems to contradict observational studies that found that increased intrinsic WUE in mature tropical and subtropical trees did not translate into growth enhancement and often even resulted in declining tree growth rates (Peñuelas et al., 2011;Silva & Anand, 2013;van der Sleen et al., 2015). However, our simulated biomass increases in savannas are driven by a combination of demographic processes and CO 2 fertilization effects on tree seedlings and saplings, which are submature, unlike most empirical studies (Figure 7). availability. This limitation is driven by a combination of higher demand for water by more productive vegetation, higher transpiration ( Figure S4b), temperature increases and regionally decreasing MAP ( Figure S1). However, the g s model implemented in the aDGVM might not represent eCO 2 responses adequately in these increasingly water-limited ecosystems (Medlyn et al., 2001).
The ensemble modelling results are useful to estimate potential lower and upper bounds of potential eCO 2 effects. Disabled eCO 2 effect scenarios serve as a surrogate for low physiological sensitivity of plants to eCO 2 and represent an artificial lower bound of change. It is, however, unlikely that eCO 2 effects on plants will be negligible in systems composed of C 3 and C 4 plants, where feedbacks between CO 2 fertilization and fire disturbance affecting the competitive balance between C 3 and C 4 plants have been postulated (Midgley & Bond, 2015). However, DGVMs such as aDGVM might not adequately represent carbon sink-source processes and other non-photosynthetic growth limitations (Körner, 2015), though empirical work by Kgope et al. (2010) strongly implicate sink-source effects in controlling the demographic process under elevated CO 2 , as simulated by aDGVM.
The first submodels that consider long-term physiological acclimation effects in plants to eCO 2 are now being implemented in DGVMs (Haverd et al., 2018).
Nutrient dynamics in plants and ecosystems can stimulate or reduce eCO 2 effects on plants, but are not represented in aDGVM.
Low nutrient availability, for example, nitrogen and phosphorus deficits, limits the growth-enhancing effects of eCO 2 and therefore on aboveground biomass (e.g. Jiang, Caldararu, et al., 2020;Peñuelas et al., 2011). Vegetation models including nutrient cycling frequently predict smaller eCO 2 effects (Fleischer et al., 2019;Hickler et al., 2015). Field data suggest that tree encroachment might further reduce nitrogen availability in African savannas, which might lead to a negative feedback on CO 2 -induced tree encroachment . The ability of trees to benefit from eCO 2 under nutrient limitation depends on their association with different mycorrhizal types and a tree's ability to fix atmospheric nitrogen (Terrer et al., 2018) with nitrogen-fixing trees from the Fabaceae family having high abundances in African savannas (Stevens et al., 2017). Under nitrogen limitation, tree species associated with arbuscular mycorrhizae, the dominant type in the tropics, benefit less than trees associated with ectomycorrhizae or N 2 fixers (Terrer et al., 2016(Terrer et al., , 2018.

| Climate change impacts on biome patterns
Projected biome changes towards woody vegetation under eCO 2 are consistent with results from previous studies on regional (Doherty et al., 2010) and continental scale (Conradi et al., 2020;Gonzalez et al., 2010;Niang et al., 2014;Scholze et al., 2006;Sitch et al., 2008). The transition of grassland and savanna biomes, that is, C 4 -dominated biomes, to more woody biomes, that is, C 3 -dominated biomes, in eCO 2 simulations corroborates findings that savannas and grasslands are particularly vulnerable to biome changes under eCO 2 Osborne et al., 2018;. Woody encroachment into grasslands and savannas in Africa is a major threat to their biodiversity (Bond, 2016).
Climatic changes outweigh eCO 2 effects in a belt south of the equator with decreasing biomass in both CO 2 scenarios by 2080-2099, which is similar to DGVM results by Sato and Ise (2012) and projections of carbon sinks saturating in African tropical forests by 2040 (Hubau et al., 2020). In our simulations, this decrease in regional carbon stocks in forests is outweighed on the continental scale by increases in other tropical forest areas as well as savanna and grassland areas that were not included in the study by Hubau et al. (2020).  (Gonzalez et al., 2010;. Explanations for differences in biome change projections between different studies include the use of different modelling approaches (e.g. species distribution models, Conradi et al., 2020), different climate data sets (e.g. GCM data or interpolated data such as ISIMIP instead of RCM data), different DGVMs (Doherty et al., 2010;Gonzalez et al., 2010) and that precipitation changes until 2100 were not accounted for  are hampered by low soil water availability (Nowak et al., 2004;Reich et al., 2014). On the other hand, advantages of C 4 grasses over C 3 grasses in dry habitats may be reduced under drought conditions (Taylor et al., 2011).

| Uncertainties of vegetation projections
In our simulations, direct eCO 2 effects on plants are the primary determinants of the future carbon sink or source, WUE and biome states in Africa (Figures 2 and 6; Table 3). This is consistent with the results from Rammig et al. (2010) for the Amazon and Huntingford et al. (2013) for the global tropics. Direct eCO 2 effects are very likely, but considerable uncertainty remains with respect to the strength of the effect. Thus, the effect sizes in Table 3 only allow comparison of CO 2 effects and climate input data as simulated by the aDGVM.
Nevertheless, our results suggest that improving our understanding of plant-physiological CO 2 effects is crucial to achieve realistic future ecosystem projections.
Areas with high uncertainty in projected biome transitions only partially agree with patterns reported by Gonzalez et al. (2010).
Uncertainty is often higher in our study (Figure 9a). Both studies agree on high uncertainty for biome changes in the western parts of the transition zone between the Sahel and the Sahara. The uncertainty of biome change we inferred in this study was similar or smaller than that in Scholze et al.'s (2006) projections. Low agreement of projected biome changes in grassland, closed savanna and woodland areas in our ensemble can be explained by alternative vegetation states related to fire occurrence typical for Africa (Hoffmann et al., 2012;Moncrieff et al., 2014;Staver et al., 2011) and variability in climatic drivers derived from GCMs, in particular uncertainty regarding precipitation.
Uncertainty in future projections is increased by the interplay of eCO 2 , fire and feedbacks caused by population dynamics. Enhanced tree sapling growth under eCO 2 may eventually reduce grass biomass and fine fuel availability, which in turn reduces fire frequency and intensity. In combination with climate changes, this feedback mechanism can lead to a rapid woody encroachment in savanna ecosystems (Figure 7; Midgley & Bond, 2015). Fire disturbance is a factor shaping ecosystems (e.g. Bond & Midgley, 2012;Midgley & Bond, 2015;Scheiter et al., 2012) leading to multiple stable states for the same climate zones (e.g. Higgins Staver et al., 2011). Fire has been shown to influence and delay projected biome changes and can delay transitions to woody biomes and thus stabilization of ecosystem states after changes in environmental drivers stabilized (e.g. [CO 2 ], Scheiter et al., 2020). Even when ecophysiological processes have already saturated (e.g. here expressed via NPP), ecosystem states continue to change (e.g. biomass, Figure 2a; Figure  Despite the uncertainties in regional patterns, the projections of substantial biome changes across all scenarios of the ensemble suggest substantial future changes in habitat structure and biodiversity. These potentially large changes in climate and biomes sharply contrast with relative stability in past climates and disturbance regimes in Africa, which probably contributed to its high biodiversity (Midgley & Bond, 2015).

| Opportunities for future ensemble vegetation simulations
By downscaling climate data from a GCM with an RCM, regional change is represented consistent with global changes, but regional features such as topography that affect regional climate are accounted for (Morales et al., 2007). Future studies could increase the ensemble size by using downscaled climate from multiple RCMs. However, Morales et al. (2007) have shown for Europe that the choice of the GCM used as boundary condition for an RCM is more important than the choice of the RCM when applying downscaled climate data in an ecosystem model. We therefore considered the CCAM ensemble in combination with the CO 2 scenarios as sufficient for this study.
Extending the ensemble by applying the same simulation design to different DGVMs would allow assessing the impact of different implementations of vegetation processes in DGVMs and the associated variability in DGVM results (Nishina et al., 2015;Sitch et al., 2008;Warszawski et al., 2014). The physiological representation in many DGVMs is based on the biochemical model of photosynthetic assimilation in leaves of C 3 and C 4 plants developed by Farquhar et al. (1980) and Collatz et al. (1991Collatz et al. ( , 1992Sitch et al., 2008). Therefore, we expect similar eCO 2 responses in an ensemble of DGVMs, unless the DGVMs additionally include nutrient limitation, or different representations of carbon allocation or of stomatal control models.

| Socio-economic development and flexible adaptation
The stronger impacts under RCP8.5 compared to RCP4.5, irrespective whether eCO 2 or fCO 2 is assumed, entail that today's decisions from individual to policy level shape the degree to which climate, ecosystems and livelihoods will change. The strong carbon sink simulated here under RCP8.5-eCO 2 might suggest regional potential for land-based climate mitigation, but the associated conversion of grassy biomes to more encroached biomes is a threat to their biodiversity. Increased woody cover in grassy biomes might lead to losses of their ecosystem services, such as arable and range land (Bond et al., 2019;Parr et al., 2014).
The range of possible climate change impacts in Africa presented in our study can support policymakers and stakeholders in Africa in planning for alternative climate futures in climate change adaptation measures (Müller et al., 2014) even though the study was not specifically designed to address policy questions. Our PNV simulations, for example, could help to guide good practice in nature conservation and land-use planning (Loidi & Fernández-González, 2012) or to avoid afforestation activities in non-forest ecosystems (Bond et al., 2019;Brancalion et al., 2019;Veldman et al., 2015). Stakeholders such as As the large differences between the different scenarios imply large uncertainties, adaptation strategies must be highly adaptive and flexible. Over time, observations of climate change and impacts on ecosystems will provide an improved knowledge base (Fletcher et al., 2019). As opposed to using vegetation projections only, Bayesian statistics combined with a stochastic dynamic programming approach allow an upfront assessment of the opportunity of this increased knowledge base in the future (Fletcher et al., 2019;Yousefpour et al., 2012). Thus, trade-offs between investing in highly flexible measures compared to potentially simpler but less flexible measures can be weighed against each other (Fletcher et al., 2019). In combination with updated information on population developments and land-use change via continued monitoring, this method may support stakeholders in the development of adequate management plans. However, implementing this method requires close and iterative exchange between stakeholders and scientists, so that specific policy and planning questions and their uncertainties can be assessed using vegetation projections as an exploratory tool adjusted to the specific question (Weaver et al., 2013).
Combining DGVMs with land use, SSPs and population development would give deeper insights into the vulnerability of different regions in Africa to climate change. However, knowledge of PNV states is a basis for future conservation planning and provides an estimate of ecosystem and plant types that could persist and thrive under future climatic conditions. When considering human population development and future potential land use, information on PNV is essential to feed into development policies. Especially in regions with higher uncertainty regarding the impact of climate change on future PNV, a diverse set of climate change adaptation measures needs to be included in development strategies, such as diversification of income and flexibility in production methods, to account for this uncertainty (Müller et al., 2014). Policies need to embrace the opportunities of continuous knowledge gain about climate change impacts (Fletcher et al., 2019) based on latest monitoring results, empirical research and updated models with scientists and policymakers collaborating closely in the decision-making process.

ACK N OWLED G EM ENTS
This study and C.M. were funded under the SPACES programme

DATA AVA I L A B I L I T Y S TAT E M E N T
The aDGVM code that was used to produce the results in this publication is available on Github (https://github.com/aDGVM/ aDGVM1_CCAM