Not only trees: Grasses determine African tropical biome distributions via water limitation and fire

Aim: Although much tropical ecology generally focuses on trees, grasses are fundamental for characterizing the extensive tropical grassy biomes (TGBs) and, together with the tree functional types, for determining the contrasting functional patterns of TGBs and tropical forests (TFs). To study the factors that determine African biome distribution and the transitions between them, we performed the first continental analysis to include grass and tree functional types. Location: Sub-Saharan Africa. Time period: 2000 – 2010. Major taxa studied: Savanna and forest trees and C 4 grasses. Methods: We combined remote-sensing data with a land cover map, using tree functional types to identify TGBs and TFs. We analysed the relationships of grass and tree cover with fire interval, rainfall annual average and seasonality. Results: In TGBs experiencing < 630 mm annual rainfall, grass growth was water limited. Grass cover and fire recurrence were strongly and directly related over the entire subcontinent. Some TGBs and TFs with annual rainfall > 1,200 mm had the same rainfall seasonality but displayed strongly different fire regimes. Main conclusions: Water limitation to grass growth was fundamental in the driest TGBs, acting alongside the well-known limitation to tree growth. Marked differences in fire regimes across all biomes indicated that fire was especially relevant for maintaining mesic and humid TGBs. At high rainfall, our results support the hypothesis of TGBs and TFs being alternative stable states maintained by a vegetation – fire feedback for similar climatic conditions.

The interplay between biotic and abiotic variables that drives the TGB dynamics and determines their environmental limits and transition to tropical forests is not fully understood and is even controversial in some aspects (e.g., Lehmann & Parr, 2016;. The most important climatic factor determining the biome distributions is mean annual rainfall (MAR), followed by rainfall seasonality . At the driest end of the gradient, with low and highly seasonal rainfall, grass-dominated ecosystems predominate because trees are water limited (Higgins et al., 2012;Sankaran et al., 2005;Sarmiento, 1984) and suffer from grass competition, especially at their seedling stage (Baudena, D'Andrea, & Provenzale, 2010;D'Onofrio, Baudena, D'Andrea, Rietkerk, & Provenzale, 2015;February, Higgins, Bond, & Swemmer, 2013). The very seasonal rainfall affects the temporal distribution of soil water and probability of fire, thereby helping to maintain open canopies . In areas with high, year-round rainfall, tropical rain forests predominate (Walter, 1973), although there are no definite annual thresholds for precipitation that delimit the biome transitions.
Fire is also extremely important in TGBs (e.g., Bond, Woodward, & Midgley, 2005;Higgins et al., 2007;Scholes & Archer, 1997). As a result of their high productivity during the wet season, followed by rapid drying and high flammability in the dry season (Bond, 2008;Lehmann et al., 2011), C 4 grasses promote fires and maintain open canopies (Beckage, Gross, & Platt, 2011;Lehmann & Parr, 2016;Ratnam et al., 2011, and references therein), which in turn favours them, because they are shade intolerant. This positive feedback is reinforced by savanna trees also being adapted to fire (Bond, 2016;Hoffmann et al., 2012;Ratnam et al., 2011), thanks to a thick protective bark (Gignoux, Clobert, & Menaut, 1997;Hoffmann et al., 2009), and by a preference for open areas, because they too are generally shade intolerant (Bond, 2016;Ratnam et al., 2011). C 4 grasses have been present in Africa for 3-8 Myr, thus for this long time fire has been important in determining the dynamics of the TGBs (Edwards, Smith, & Thresholds, 2010;Lehmann & Parr, 2016). Tropical forests, on the contrary, are characterized mainly by shade-tolerant, fire-intolerant species (Ratnam et al., 2011 and references therein). Closed-canopy forests suppress fires because: (a) reduced light availability hinders C 4 grass growth (Hoffmann et al., 2009); (b) the presence of a humid understorey and lower temperatures limit flammability; and (c) reduced wind speeds limit fire spread (Cochrane, 2003).
They have not been included in the vast majority of the continental-scale studies, most of which have focused on woody variables, using remote sensing (Hirota et al., 2011;Staal et al., 2016;Staver et al., 2011b;Yin et al., 2016; but see Bertram & Dewar, 2013), or field sites (Dantas et al., 2016;Sankaran et al., 2005;Staal & Flores, 2015;. Using a recent moderate resolution imaging spectroradiometer (MODIS) product and the European space agency (ESA) Global Land Cover map, we set out to study the factors that determine the distribution of African biomes and the transitions between them, with the aim of conducting the first continental vegetation analysis to include information from grass cover. Furthermore, by using tree functional types to identify savanna and forest trees (generally classifiable in Africa as deciduous and evergreen, respectively; see the Materials and Methods section for details; Bowman & Prior, 2005;Scholes & Archer, 1997;Shorrocks, 2007), we could also avoid the problems associated with identifying the bimodality in MODIS tree cover data recently pointed out by Gerard et al. (2017).
We observed, for the first time, that water limitation at the driest end of the TGB distribution (for MAR 630 mm/year) affected not only trees but also grasses, and was the main factor determining the occurrence of dry TGBs. Mesic TGBs, however, were characterized by frequent fires associated with the presence of grass. Some humid TGBs occurred in similar climatic conditions (in terms of MAR and its seasonality) as TFs but had very different fire frequencies, thus supporting the hypothesis of alternative stable states maintained by vegetation-fire feedback.

| Satellite data
We analysed data for tree and herbaceous vegetation cover percentages, average fire interval, mean annual rainfall and seasonality, for sub-Saharan Africa from latitudes 358 S to 158 N at 0.58 resolution (c. 50 km). A sensitivity analysis performed with these satellite data with grain sizes ranging from 500 m to 100 km has shown that African relationships between climatic, vegetation and fire variables are insensitive to spatial resolution (Staver, Archibald, & Levin, 2011a).
We derived tree cover (T) and grass cover (G; hereafter also called woody and herbaceous cover, respectively), averaging in space and time the yearly percentage of tree and non-tree vegetation cover products of MODIS vegetation continuous fields (MOD44B VCF) version 051 for the period 2000-2010 (Townshend et al., 2011;Figure 1a,b).
We used the ESA global land cover map (ESA CCI-LC, v 1.6.1; 5-yearaveraged dataset centred in 2010) to remove pixels with more than one-third of the area affected by anthropogenic activities and covered by water (coastal and inland) and/or with more than one-half of the area occupied by shrubland (MOD44B underestimates tree cover in the presence of shrubs because it does not detect trees shorter than 5 m; Bucini & Hanan, 2007). Cultivations cover most of the area that we discarded in our study (see Supporting Information Appendix S1).

| 715
Using the same map, we classified pixels as TGB when they included 50% of the deciduous trees and grassland classes (ESA CCI-LC codes 60-62, 130) and as TF when they included 50% of evergreen and flooded tree classes (codes 50, 160, 170; Figure 1a). In Africa, unlike many Neotropical and especially Australian savannas, deciduous trees predominate in TGBs (Bowman & Prior, 2005;Scholes & Archer, 1997;Shorrocks, 2007; see also Supporting Information Table S4.1 for a compact summary of the literature), whereas evergreen trees predominate in TF (Bowman & Prior, 2005;Walter, 1973). Evergreen trees can also occur in African TGBs, but only locally (e.g., Scholes & Walker, 1993;Scholes et al., 2002), and would thus not be detected at the coarse grain of our analysis. Our distinction between TGB and TF based on land cover and tree functional types does not suffer from the drawbacks of previous analyses that separated them on the basis of tree cover (e.g., Hirota et al., 2011;Staver et al., 2011b), a method which could create problems owing to the uncertainties in the MODIS cover data (as recently pointed out by Gerard et al., 2017). We also used the same procedure to identify other land cover types using different ESA CCI-LC codes (see Supporting Information Appendix S1).
We used monthly rainfall measurements from the tropical rainfall measuring mission (TRMM 3B42) precipitation product to derive MAR and a rainfall seasonality index (SI; Walsh & Lawler, 1981) Lehsten, Harmand, Palumbo, and Arneth (2010). Given that AFI spans different orders of magnitude, in the analysis we used log 10 (AFI).
See Supporting Information Appendix S1 for further details.

| Identifying ranges of mean annual rainfall with different tree-grass dominance
We chose MAR as first independent variable because it has been repeatedly identified as the main explanatory variable for African woody vegetation Sankaran et al., 2005). We

MAR [mm/year]
FIGUR E 1 (a) Tree cover, (b) grass cover and (c) mean annual rainfall (MAR) for 0.58 cells across Africa (grey colour scale). In (a), red lines delimit TGBs (areas in which cells with 50% of their area is flagged on the ESA CCI-LC map as deciduous trees and grasslands), and green lines delimit TFs (areas in which cells with 50% of their area is flagged on the ESA CCI-LC map as evergreen and flooded trees). In (b), lines delimit areas with similar annual average fire intervals (AFI): yellow lines delimit AFI 10 years, green lines delimit AFI between 10 and 100 years, and violet lines delimit AFI > 100 years. In (c), lines delimit rainfall seasonality classes (Walsh & Lawler, 1981): equable with a definite wetter season [0.20 rainfall seasonality index (SI) 0.39; violet line]; rather seasonal with a short drier season (0.40 SI 0.59; light blue line); seasonal (0.60 SI 0.79; green line); markedly seasonal with a long drier season (0.80 SI 0.99; light green line); most rain in 3 months (1.00 SI 1.19; orange line); and extreme, almost all rain in 1-2 months (SI 1.20; red line). In each panel, dots represent pixels excluded from the analysis (see Materials and methods) analysed the cover of trees, grasses, total vegetation [T 1 G; which is connected to the bare soil fraction: 100% 2 (T 1 G)]. We also analysed the relative dominance of trees and grasses, which we defined by subtracting grass cover from tree cover (T 2 G). To identify the transition points where vegetation showed clearly different dependence on MAR, we used the marked changes in slopes and spread of the four vegetation variables along the MAR axis, because we considered these changes to indicate variations in the underlying ecosystem dynamics.
We especially focused on T 2 G, because it expressed the changes in dominance of the two vegetation types along the gradient.

| Multivariable statistical models
We used generalized linear models (GLMs; McCullagh & Nelder, 1989) to analyse the dependences of the four vegetation cover variables (T, G, T 1 G and T 2 G) with respect to three predictors: MAR, SI and log 10 (AFI). We performed the analysis within the different MAR intervals (identified as explained in the previous subsection). To select the models, we used Akaike's information criterion (AIC; Akaike, 1974), and we evaluated the goodness-of-fit with the explained variance, R 2 . See Supporting Information Appendix S1 for further details.
Finally, we applied the two-sided Wilcoxon rank sum test to check whether pairs of variables had significantly different statistical distributions at the p 5 .05 level.

| R ESU L TS
3.1 | Overall dependence of tree and grass cover on rainfall and fire Although tree cover strongly depended on MAR but also showed a large spread (see Figure 2b and Hirota et al., 2011;Staver et al., 2011b), total vegetation (T 1 G) varied with a very narrow spread for increasing MAR (Figure 2d). The T 1 G could easily be captured with a simple implicit-space logistic model for vegetation cover (Levins, 1969) Herbaceous cover and log 10 (AFI) were significantly correlated (R 2 5 .62; Figure 3). Fires with an average interval of < 10 years occurred only if grass cover was > 45-50%. Fires were most frequent at intermediate MAR (Supporting Information Figure S3.1), where grass was more common. In contrast, T was weakly related to log 10 (AFI) (Supporting Information Figure S3.2a), with low explained variance (R 2 5 .13).
3.2 | R1: Low mean annual rainfall (MAR 630 mm/year) The R1 range was mainly characterized by long dry seasons and rare fires ( Figure 4). Tropical grassy biomes represented 54% of the pixels (mainly grasslands), with the remainder mostly being sparse vegetation or bare soil (see Supporting Information Table S4.2). Mean annual rainfall was the best predictor for grass cover; it increased monotonically with annual rainfall (R 2 5 .55). Tree cover also increased with MAR, its best predictor, but with a lower explanatory power (R 2 5 .26; Figure Information Table S4.2), and on the right, TGB and tropical forests (TF) in R3. Outliers are not shown. The distribution of rainfall seasonality differed significantly between R1 and R2, whereas the distribution of log 10 (AFI) in R3 and R1 was indistinguishable 2b, Table 1). Trees were also negatively correlated with seasonal variability in precipitation (R 2 5 .19), and grasses and trees were positively correlated with fire frequency (R 2 5 .37 and .24, respectively); see Supporting Information Table S4.4. The positive correlation between trees and fire frequency can be understood by considering that in this range both trees and fuel availability and continuity (linked to grasses) increased with MAR (Table 1).
With respect to the other biomes in R1, TGBs experienced more annual rainfall and less seasonal regimes. These TGBs had more grass cover (and slightly more tree cover), with more frequent fires (although still fairly rare; Figure 4; Supporting Information Figure S3.3).
3.3 | R2: Intermediate mean annual rainfall (630 mm/year < MAR < 1,200 mm/year) In this range, most pixels were TGBs (86%), with markedly seasonal, although not extreme, rainfall and frequent fires (Figure 4). The selected GLMs included only one explanatory variable (Table 1); grasses decreased with log 10 (AFI) (R 2 5 .23; Figure 3), whereas trees increased with MAR and stabilized at c. 20% cover at the end of the range, but with largely scattered values (R 2 5 .14; Figure 2b). In general, in R2 the best GLMs explained a smaller fraction of variance of the vegetation variables than in R1 and R3.
3.4 | R3: High mean annual rainfall (MAR 1,200 mm/year) The wettest areas (R3) had mostly mild rainfall seasonality and rare fires (see Figure 4). Total vegetation cover decreased with increasing rainfall seasonality (Supporting Information Figure S3.2b; Table 1). The other variables were strongly related to fire interval; G decreased with log 10 (AFI), and T and T 2 G increased with log 10 (AFI) (Figure 3, Table 1; Supporting Information Figure S3.2a). Additionally, SI explained a large fraction of variance of these variables, which is not surprising given that SI and log 10 (AFI) were highly correlated (Supporting Information   Table S4.3). In this range, MAR explained the lowest variance of G and T and was not significantly correlated with total vegetation (Figure 2).
In R3, 61% of the pixels were TFs, with a high tree cover on average (66%). Tropical grassy biomes occupied 23% of the pixels, with Performing the GLM analysis over the two R3 biomes separately, we found that in R3 forests, trees and grasses were significantly correlated with seasonality; trees decreased with SI (R 2 5 .28; see Supporting Information Figure S3.5), and grasses increased (R 2 5 .25). In TGBs, however, neither T nor G was found to be significantly correlated with  Figure 4; Supporting Information Figure S3.3).

| DI SCUS SION
Across sub-Saharan Africa, we observed marked changes in tree and grass cover and in their relative dominance at different MAR; the changes were generally also associated with different rainfall seasonalities (Figure 4) was controlled by different water-fire-driven dynamics (as we summarize and discuss below), and in some humid areas we observed bistability of TGBs and TFs for similar climatic conditions ( Figure 5). Despite these different mechanisms and relationships determining the emergence of the various biomes along the MAR gradient, we found that over the entire subcontinent, grass and fire recurrence were strongly related, with low fire frequency corresponding to low grass cover (Figure 3), and that overall the total vegetation cover was controlled mainly and strongly by mean annual rainfall (Figure 2d).
At low precipitation (< 630 mm/year), grasses were dominant though water limited, and fires were rare (Table 1). Within this range, more availability of water, owing to higher MAR and lower seasonality, promoted large increases in grass (similar to what was observed for the Kalahari by Scholes et al., 2002) and, to a much smaller extent, in tree cover. We thus showed here, for the first time at a broad scale, how water limitation acted strongly on the herbaceous component, not only on the woody component ( Figure 2). In conditions of low and seasonal precipitation, grasses can prevail over trees, for different reasons (Sankaran, Ratnam, & Hanan, 2004), including higher photosynthetic efficiency in water use (Lloyd et al., 2008), lower costs for plant structure and maintenance (Orians & Solbrig, 1977), better adaptation to (clay) soils (Axelsson & Hanan, 2017;Fensham, Butler, & Foley, 2015;Sankaran et al., 2005), and overlapping rooting depths of grasses and trees (Holdo & Brocato, 2015;Kulmatiski & Beard, 2013) that allow grasses to suppress growth and establishment of tree seedlings (Baudena et al., 2010;D'Onofrio et al., 2015;February et al., 2013). In addition, in dry areas herbivores have a larger impact than fires, although the effects on vegetation structure differ, depending on the type of herbivory (Archibald & Hempson, 2016 (2011) and Staver et al. (2011b). The highly seasonal regimes in R2 increase the probability of droughts, which negatively affect both juvenile and adult trees by increasing the probability of mortality and reducing their growth rate . In the present study, one key finding was that at intermediate rainfall, grasses were no longer water limited. A second key finding was that overall grass cover increased with fire recurrence, thus providing evidence, for the first time at a continental scale, in favour of the grass-fire feedback hypothesis (Bond, 2008;Dantas et al., 2013;van Langevelde et al., 2003). The marked, albeit not extreme, rainfall seasonality in R2 was also likely to favour fire occurrence by enhancing grass productivity in the wet season and the availability of fuel in the dry season .
The relevance of the nonlinear dynamics of a system driven by finescale feedbacks between vegetation composition and fires (Pausas & Dantas, 2017) might also explain why only minor parts of the variance of the vegetation variables were explained in R2 (Table 1). Other factors that we did not consider in this research might be responsible for the unexplained variance. Herbivory plays a similar role to fire in shaping the vegetation states and transitions, especially at local scales, although with less effect than fire in terms of consumed biomass (Archibald & Hempson, 2016;Dantas et al., 2016;Hempson, Archibald, & Bond, 2015). Additional variability could be attributable to soil texture influencing the water balance (Staver, Botha, & Hedin, 2017) Table S4.2). Some TGBs and TFs were found for similar climatic conditions, as determined by overlapping rainfall seasonality in R3. Biomes are classified based on land cover information from the ESA CCI-LC map, on mean annual rainfall ranges and on bimodality information from our analysis. Dots represents pixels excluded from the analysis (see Materials and Methods) or pixels that were not represented here because they were parts of R2 or R3 but not classified as TGB or TF (< 14% of the pixels analysed) (Aleman, Blarquez, Gourlet-Fleury, Bremond, & Favier, 2017), which might still be having an impact even though we excluded areas subjected to strong human influence.
At high precipitation (MAR > 1,200 mm/year), both savanna and forest states were observed, with very distinct fire frequency distributions ( Figure 4). Most of the TGBs and TFs corresponded to distinct rainfall seasonal patterns, with forests growing in areas with the least seasonal regimes, thus revealing that seasonality plays a greater role in savanna-forest transitions than previously reported (Hirota et al., 2011;Staver et al., 2011b). In fact, seasonality and fire return time were highly correlated in this MAR range (Supporting Information Table   S4.3), indicating that the biome distributions are possibly also mediated by the connection between seasonality and fire occurrence. The different seasonalities of TGB and TF in humid areas at the continental scale might also be connected to land-atmosphere coupling (e.g., Baudena, D'Andrea, & Provenzale, 2008;Rietkerk et al., 2011;Van Nes, Hirota, Holmgren, & Scheffer, 2014), which is especially strong in tropical Africa (Green et al., 2017;Koster et al., 2004). Savanna and forest biomes with equal annual rainfall have different evapotranspiration and radiative fluxes, which can also affect rainfall seasonality by determining large-scale atmospheric circulation (Yin et al., 2016). Projected changes in seasonal distribution of precipitation, which may occur locally in the tropics (Arnell & Liu, 2001), would trigger transitions between TGB and TF.
Remarkably, however, although 20% of the humid TGBs and forests shared similar climatic constraints, including rainfall seasonality, they maintained different fire frequencies (Supporting Information Figure S3.4). This finding has not been influenced by the uncertainties that seem to have affected previous studies reporting analogous results from remote-sensing analyses, which identified the two biomes by their different typical tree cover values (Hirota et al., 2011;Staver et al., 2011b). That approach has recently been questioned because the algorithm that produces the MODIS cover values includes consistent patterns of under-and overestimation (see also Gerard et al., 2017;Hanan, Tredennick, Prihodko, Bucini, & Dohn, 2014;Staver & Hansen, 2015). Such data uncertainty was much less relevant in our analysis, as we identified TGBs and forests by using different tree functional types, based on their phenology. This finding thus reinforces the view that savanna and forest may be alternative stable states maintained by fires (as proposed by e.g. Dantas et al., 2016;Hirota et al., 2011;Staal & Flores, 2015;Staver et al., 2011b). Thanks to the availability of sufficient water, shade-tolerant forest trees can close their canopies, but the positive vegetation-fire feedback can maintain savannas as an alternative stable state (van Langevelde et al., 2003). A decrease in fire frequency (e.g., as a consequence of management strategies, see Andela et al., 2017), could thus lead to savanna transitioning to forest (see Bond, 2008, and references therein). Differences in soil nutrients are an alternative explanation for forest and savanna occurring in similar climatic conditions, as forest soils display higher nutrient content (Lloyd et al., 2008;. This explanation is still controversial (e.g., Staal & Flores, 2015), because observations suggest that deep savanna soil contains enough nutrients to sustain forests (Bond, 2010), and because of the existence of feedbacks between soil nutrient levels and plant community composition (Veldhuis, Hulshof, Fokkema, Berg, & Olff, 2016) and fires (Pivello et al., 2010). Finally, the distribution of TGBs and TFs ( Figure 5) showed that the bimodal areas tended to occur at the boundaries between the two biomes, indicating that spatial structure was also important (see also Pausas & Dantas, 2017;Staal et al., 2016;Wuyts, Champneys, & House, 2017).
The prominent correlation between grass cover and fire intervals across the entire subcontinent ( Figure 3) clearly demonstrates the pervasiveness of the connection between grasses and fire at a broad scale.
Correspondingly, a hump-shaped relationship of fire intervals with MAR was observed (Supporting Information Figure S3.1). We could generally confirm the intermediate fire-productivity/aridity hypothesis (Krawchuk, Moritz, Parisien, Van Dorn, & Hayhoe, 2009;Pausas & Ribeiro, 2013) for sub-Saharan Africa, given that plant cover grew monotonically with MAR (Figure 2d), and assuming cover as a proxy for productivity (similar to what was observed by, e.g., Dantas et al., 2016;Lehmann et al., 2011). According to the hypothesis, fires are limited by fuel availability and discontinuity in unproductive, dry areas, and by fuel moisture in productive, wet regions. However, the data mostly differed from the hypothetical one-to-one fire-productivity relationship at high rainfall, where TGBs and TFs could both occur as alternatively stable states.
Tree cover increased with MAR in the entire dataset but showed a Interestingly, this relationship could be captured very well with a simple implicit-space model for vegetation cover (Levins, 1969), in which plants colonize new space proportionally to the MAR. This shows that water limitation acted most strongly on vegetation as a whole. It is much easier to predict total vegetation than the partitioning between trees and grasses, which involves other types of dynamics and feedbacks (Scholes & Walker, 1993). Such excellent performance by an extremely simple model in predicting large-scale vegetation cover trend seems to support the idea that increasing ecological details at small scales does not assure improved predictions at a larger scale (Levin, 1998).
We found that some areas had low forest-tree cover (< 60%), yet fire was rare or absent. Thanks to the use of tree phenologies, we could identify these as 'degraded forests' that lack the C 4 grasses that permit fire to spread, and we could distinguish them from old-growth savannas (Ratnam et al., 2011;Veldman, 2016;Zaloumis & Bond, 2015). In TFs, the low tree cover was connected to higher seasonality (Supporting Information Figure S3.5), which can be interpreted as evidence of forest retreat during past dry periods, thereby showing that African rain forests are more sensitive to small variations in rainfall seasonality D'ONOFRIO ET AL.
than in MAR (Malhi, Adu-Bredu, Asare, Lewis, & Mayaux, 2013). However, the absence of fires in these tropical forest sites may be attributable, in part, to inaccurate fire data for humid regions (Favier et al., 2012). Furthermore, in tropical forests, anthropogenic deforestation is known to be very important (Achard et al., 2014;Hansen et al., 2013).
The similarity in structure of TGBs (grass layer with shadeintolerant, fire-tolerant, deciduous trees) has led to assumptions in the literature that they are all regulated by the same processes (Lehmann & Parr, 2016), fostering much debate on their origin. In our study across sub-Saharan Africa and with 0.58 grain size, we have been able to highlight the importance of the overlooked grass layer in characterizing TGBs, alongside the well-known role of woody vegetation. We found that water limitation to grass growth is fundamentally characterizing dry TGBs and acts alongside the well-known water limitation to tree growth. The role of fires was more evident at intermediate and high rainfall values. Tropical grassy biomes where grasses were not water limited were associated with similar tree and grass cover values and frequent fires and experienced a similar marked rainfall seasonality.
Despite these similarities, in mesic savannas the trees were still water limited, whereas in humid areas, by distinguishing between forest and savanna trees we found that some TGBs and TFs occurred as alternative states in a similar climate. It is possible that these humid TGBs are

DATA ACCESSIBILITY
The observational datasets used in this study are all freely available.