Savannah trees buffer herbaceous plant biomass against wild and domestic herbivores

Questions: Given the growing abundance and dominance of domestic herbivores in savannah ecosystems, can trees maintain plant herbaceous standing biomass under increasing herbivore pressure? Are there differences in the capacity of leguminous and non-leguminous trees in sustaining understorey herbaceous biomass? And finally, to what extent does plant community composition underneath trees modulate the effects of herbivore assemblage and abundance on herbaceous biomass? Location: Pasturelands and protected areas along the borders of the Serengeti National Park, Tanzania, East Africa. Method: Monthly herbaceous biomass was monitored non-destructively using a calibrated pasture-disc. Measurements were taken within a network of paired exclosure and open plots, underneath large leguminous and non-leguminous tree canopies and outside canopies. Herbivore community assemblage composition and abundances were characterized using monthly dung counts, and herbaceous plant community composition was surveyed annually every wet season over two years. Results: Overall, we found that trees promote herbaceous standing biomass, particularly in the presence of moderate herbivory rather than under herbivore exclusion. Greater herbivore abundance and livestock dominance reduced herbaceous plant biomass, but trees, particularly leguminous trees, limited these negative effects. This capacity for trees to limit the effect of herbivores was related to herbaceous plant species composition. Understorey plant communities A cc ep te d A rt ic le


INTRODUCTION
Tropical savannahs are characterised by the continuous grass-dominated herbaceous layer interspersed by scattered trees or shrubs (Belsky 1993;Scholes & Archer 1997). Trees can act as small-scale biodiversity hotspots, provide structural complexity and spatial heterogeneity in savannah ecosystems (Belsky 1993;Munzbergova & Ward 2002). Most notably, trees play an important role in shaping spatial heterogeneity of herbaceous biomass. By competing for water, nutrients and light, trees can supress understorey plant biomass (Belsky 1993;Ludwig et al. 2004a;Smit 2005;Riginos et al. 2009). Yet, it is widely documented that savannah trees can also facilitate understorey biomass by improving water relations, soil nutrient cycling and/or providing protection from herbivory (Belsky 1993(Belsky , 1994Riginos & Grace 2008). The balance between competitive and facilitative effects of trees on the herbaceous understorey has been found to be contingent on wider environmental factors such as annual rainfall (Belsky 1993;Dohn et al. 2013), fire (Coetsee et al. 2010), tree density (Riginos et al. 2009), individual tree size (Treydte et al. 2009;Moustakas et al. 2013), functional type (i.e. leguminous with nitrogen fixing symbionts versus non-leguminous trees), and species (Scholes & Archer 1997;Blaser et al. 2013). Although insightful, the majority of these tree-herbaceous understorey generalities are from studies in wildlife-dominated savannahs with low herbivore densities and several meta-analyses and reviews on savannah tree-herbaceous understorey interactions exclude heavy grazing altogether (Blaser et al. 2013;Dohn et al. 2013;Treydte et al. 2013). Yet, the savannah landscape is rapidly changing with a widespread loss of wild herbivores and their replacement with higher densities of livestock

Accepted Article
This article is protected by copyright. All rights reserved (Hempson et al. 2015(Hempson et al. , 2017. A pertinent question is whether trees exert competitive or facilitative effects on herbaceous understorey productivity in savannah ecosystems which are increasingly grazed by livestock.
The competition-facilitation continuum predicts that with increasing stress, the net outcome of plant-plant interactions will be positive supporting growth and reproductive success.
Herbivores can add to stress, thus cancelling out facilitative interactions (Smit et al. 2009), for example, by consuming the facilitated growth of understorey plants. However, facilitative effects of savannah trees on understorey plants can be greater in the presence of herbivores rather than when excluding them (Augustine & McNaughton 2006;Treydte et al. 2009). Herbivores are attracted to nutrient-rich grasses found underneath trees, concentrating dung and urine deposits enhancing growth and nutrient enrichment of understorey vegetation (Ritchie et al. 1998;Treydte et al. 2007Treydte et al. , 2009Treydte et al. , 2010Treydte et al. , 2013. At low densities of wild herbivores, such tree facilitative effects have been found to be stronger underneath leguminous trees with nitrogen fixing symbionts or larger savannah trees due to already elevated soil and plant nutrient contents (Treydte et al. 2009;Blaser et al. 2013;Dohn et al. 2013;Moustakas et al. 2013). Yet, beneficial impacts of herbivores on tree understorey vegetation may decline at high herbivore densities, irrespective of tree functional type or stature. From a single study in Kenya, beneficial soil fertility and herbaceous biomass effects under large leguminous and non-leguminous trees were lost under heavy livestock grazing due to soil compaction . Nevertheless, much more work is required to derive generalities of how herbivore pressure and livestock dominance in savannahs interacts with the functional capacity of trees in tree-understorey interactions.
At low herbivore density, in wildlife-dominated savannahs it is well-known that plant communities found underneath trees can be compositionally distinct from outside tree canopies, supporting understorey specialists Munzbergova & Ward 2002;Smit 2005;Riginos et al. 2009;Treydte et al. 2010;Stahlheber et al. 2015). Less attention has been given to how understorey plant species composition potentially mediates the impact of wild and domestic herbivores on savannah tree-herbaceous vegetation interactions. Compared to plant communities outside tree canopies (see Porensky et al. 2013;Young et al. 2013;Young et al. 2018), understorey plant communities may be more resistant or resilient to changes in herbivore assemblage and intensity. To support growth in shaded conditions understorey plant species have a higher proportion of non-digestible fibres, tannins and other polyphenols that deter herbivores (Rohner & Ward 1997;Treydte et al. 2007). Yet, greater access to soil nutrients may support

Accepted Article
This article is protected by copyright. All rights reserved resilient understorey communities that are able to regrow and persist following herbivore damage (Ritchie et al. 1998;Hawkes & Sullivan 2001;Riginos et al. 2009). Given the desirability of understorey specialist plant species by herbivores (Smit 2004(Smit , 2005, we would expect many understorey specialist plant species to be reduced and eventually lost with increasing herbivore densities. Across the savannah biome, the highest rates of tree clearance and land conversion from wildlife to livestock-dominated savannah are concentrated at the borders of protected areas, e.g. national parks (Beale et al. 2013;Fynn et al. 2016;Veldhuis et al. 2019). Thus, we directed our attention to an increasingly common juxtaposition of wildlife-versus livestock-dominated savannah across the border of the Serengeti National Park, Tanzania, and its impact on treeherbaceous vegetation interactions. Utilising a network of small-scale exclosures underneath trees and outside tree canopies, we studied the following: (1) whether savannah trees maintain herbaceous standing biomass with increasing livestock dominance and abundance; (2) whether there are differences in the maintenance of herbaceous biomass beneath leguminous and nonleguminous trees across different herbivore assemblages and abundances; and lastly (3) whether understorey species composition modulates the effect of wild and livestock herbivory on savannah tree-herbaceous interactions. We hypothesized that savannah trees would facilitate herbaceous biomass from low to moderate herbivore pressure across herbivore assemblages with greater plant biomass underneath leguminous trees due to nitrogen-fixing symbionts. However, we predicted that facilitative effects of both tree functional types would be lost at high herbivore pressure due to a loss of productive understorey canopy specialists sensitive to intense herbivory.

STUDY SITES
The study was performed in grazed wooded savannahs south of the Serengeti National Park, Tanzania (latitude 2°00 to 2°30′′S and longitude 34°00 to 34°30′′E) with an elevation of 1200-1670 m a.s.l. Two rainfall regions were selected: a wet region (1287  191 mm annual rainfall, mean  standard deviation, between 2015 -2017) in the west close to Lake Victoria and a mesic region (699  105 mm annual rainfall) in the east within the rainfall shadow of the Ngorongoro crater (Huffman 2017) (Fig. 1). Rainfall is seasonal, with a dry season from June to October and a bimodal wet season between November and May, including a short and long wet season, although this seasonality is stronger in the mesic rather than wet region. Soils were less fertile in the wet

Accepted Article
This article is protected by copyright. All rights reserved region (carbon 6.0  1.5 kg C m -2 ; nitrogen 0.4  0.1 kg N m -2 in upper 0 -15 cm; means  sd) having sandy arenosols, compared to the mesic region (carbon 5.4  1.0 kg C m -2 ; nitrogen 0.6  0.1 kg N m -2 ) with gravelly leptosols and underlying calcareous tuff and granite/gneiss bedrock (ISRIC 2018). Within rainfall regions, sites encompassing a wildlife-livestock herbivore gradient from inside to outside protected areas were selected on similar soil types. Tree overstorey composition was similar across sites, with dominance of leguminous Vachellia tortilis (the genus Vachellia was formerly Acacia) and Vachellia robusta and non-leguminous Commiphora africana and Commiphora schimperi trees with the exception of pasturelands in the wet region where the overstorey comprised leguminous Vachellia drepanolobium and non-legume Balanites aegyptiaca trees. Herbaceous plant species richness was higher in the mesic region totalling 113 species in wildlife-protected areas and 101 in pastureland compared to with 43 and 74 species in the wet region wildlife-protected areas and pastureland. However, herbaceous species composition was not strongly influenced by rainfall region or herbivore regime but diverged primarily based on tree canopy type, fire history and tree height (see below).
Within wildlife-protected areas, apart from periodic episodes of migrating wildebeest

STUDY DESIGN
To capture a wide range of wildlife to livestock herbivore assemblages and intensities, we established sites inside protected areas, at the border of protected areas and outside protected areas on village pasturelands (Fig. 1). In each rainfall region we initially established four sites inside protected areas, at park boundaries and on pasturelands, with an extra four sites in an additional protected area in the mesic region (Fig. 1). Vandalism and robbery at the boundaries of protected

Accepted Article
This article is protected by copyright. All rights reserved areas reduced our original experimental set-up. Furthermore, a lack of livestock dung at remaining border sites suggested an absence of illegal grazing. As such, sites were classified as either wildlife-protected areas or pasturelands. The final design for the experiment consisted of 8 sites in the wet region (4 × protected areas and 4 × pasture) and 14 sites in the mesic region (10 × protected areas and 4 × pasture) (Fig. 1). Within rainfall regions, all sites were within 10 km of one another.
Outside the park, intense livestock grazing usually results in reduced fire frequencies due to insufficient fuel-loads in the dry season in contrast to regular burning undertaken by park At each site, we selected four replicate blocks based on the same co-dominating canopy species at a given site as well as soil characteristics (type, texture, organic matter and nutrient contents). Each block comprised a 50  50 m area and blocks were spaced approximately 500 m apart. At each block we established six permanent experimental plots, these included three open and three exclosed plots underneath leguminous trees, underneath non-leguminous trees and outside tree canopies. Due to the importance of tree size for herbaceous productivity we aimed to minimize variation in selected tree heights (Treydte et al. 2007;Blaser et al. 2013). Across the system, mean tree heights (± sd and median in parenthesis) for each species were as follows: the legumes were V. tortilis 7.3 ± 1.2 m (6.8 m), V. robusta 6.7 ± 1.1 m (7.0 m), V. drepanolobium 4.4 ± 0.5 m (4.2 m) and the non-legumes were B. aegyptiaca 4.9 ± 1.1 m (4.8 m) and Commiphora species 4.9 ± 1.0 m (4.6 m). Across our study system there remained a 2.5 fold variation in selected tree heights, so tree height was also included in our statistical analyses (outlined below).
Tree height was measured annually using a clinometer (Nikon, Forestry Pro 550). Plots outside of tree canopies were selected following random cardinal directions from the centre of each block and situated a minimum of distance of 1.5 times the canopy radius of the nearest tree (Treydte et

Accepted Article
This article is protected by copyright. All rights reserved al. 2007).
In each block, exclosure and open plots were paired underneath one leguminous and one non-leguminous tree. Exclosures underneath trees were triangular with the corner adjoining either north or south facing side of the tree with an approximate exclosed area of 2.3 m 2 . Outside tree canopies we constructed square 1.6 m 2 exclosures paired with marked open plots within 10 m of these exclosures. Exclosures were either constructed out of steel metal mesh or wooden poles with barbed wire and sisal depending on the likelihood of fire and theft at specific sites, respectively.
All exclosures were designed to exclude mammalian herbivores > 5 kg body mass. Due to significant human and animal damage to trees and exclosures later in experiment (after initial robbery outlined above) a further four plots were excluded from the study, namely, paired exclosure and open plots under a leguminous tree that was burnt down in the mesic pastureland, one exclosure under a leguminous tree and one outside tree canopy in different mesic protected area sites that were repeatedly damage by wildlife. Across all sites we had 128 vegetation plots.

MEASUREMENTS
Herbaceous biomass underneath trees and outside tree canopies was measured non-destructively using a pasture disc meter (Bransby & Tainton 1977;Trollope & Potgieter 1986;Moustakas et al. 2013), specifically calibrated to the Serengeti ecosystem. Herbaceous biomass was estimated from pasture disc meter readings (in cm) using the formula y = 0.0563x 0.68 where x is the disc reading in cm and y the grass biomass in kg m 2 . Pasture measurements were applied to herbaceous vegetation, namely graminoids and forbs, avoiding any large shrub or tree saplings. Ground cover in plots was primarily composed of herbaceous graminoids and forbs, averaging 57 ± 24 % cover (mean ± SD). Woody forbs and dwarf shrubs accounted for 0.5 ± 2.5 % cover and did not influence pasture disc estimates (R-INLA model, posterior mean 0.408 ± -1.025/1.840 with ± lower and upper quartile estimates). The pasture meter disc had an area of 0.16 m 2 and was used four times inside each experimental plot. Herbaceous biomass measurements estimated using pasture disc meters were taken at monthly intervals between May 2016 to May 2018.
Spatial and temporal variation in herbivore abundance was characterized by counting fresh dung piles at each site. All fresh dung (i.e. less than a few days old) was counted along a 4 × 50 m transect every month. The herbivore assemblage was characterized by grouping dung into four herbivore guilds separating wild or domestic herbivores and loosely based on dietary niche and body size (Riginos & Grace 2008;Young et al. 2018). Our herbivore guilds included: large

Accepted Article
This article is protected by copyright. All rights reserved livestock grazers (referring to cattle), small livestock grazers and mixed feeders (referring to sheep and goats), wild grazers (for example wildebeest, zebra, impala and gazelles and including mixed feeders) and wild browsers (mainly large herbivores such as giraffe and elephant). Cattle dung comprised on average 64% of dung deposited in pastureland and 4% inside wildlife-protected areas. Our livestock gradient encompassed an estimated cattle metabolic biomass range of 0 to 1522 kg per site on pastureland and 0 to 354 kg per site in wildlife-protected areas (408  418 for pastureland and 26  73 kg for wildlife-protected areas, means  sd) over the duration of the study. Dung counts were used as part of a multivariate analysis (below) to generate a score for herbivore assemblages as well as using total dung count as a coarse measure of total herbivore abundance (Riginos & Grace 2008;Young et al. 2013).
We estimated herbaceous species cover at peak aboveground biomass in all plots in May -June 2016, May 2017 and May 2018. Percent cover of each species was visually estimated within a 0.36 m 2 quadrat inside each plot. In total, 175 herbaceous plant species were recorded across all plots over the duration of the experiment.

RAINFALL DATA
As soil moisture can modulate the facilitative effect of trees on the herbaceous understorey in savannah ecosystems (Ludwig et al. 2004a;Ludwig et al. 2004b), we estimated soil moisture by proxy using daily precipitation. We obtained satellite-based daily rainfall from NASA's Goddard Earth Sciences Data and Information Services Center (Huffman 2017) and half-hourly measurements of cloud cover were taken using multi-satellite microwave data at 10  10 km spatial resolution. Between May 2016 and May 2018 we further measured soil moisture in all plots at a 5 cm depth using a hand-held probe (Theta probe ML2, Delta-T, UK) every three months. Both satellite-based rainfall estimates and plot-scale soil moisture measurements were significantly positively related with a correlation of 0.78 (Pearson's correlation test: t 780 = 34.9, p<0.001). Using the daily satellite rainfall estimates, we calculated cumulative rainfall for the interval preceding each monthly biomass and herbivore dung measurement measured in mm.

STATISTICAL ANALYSIS
To generate a single measure of herbivore assemblage and plant species composition we used multivariate techniques. Firstly, to generate a herbivore assemblage score that could differentiate herbivore assemblages through time we applied principal response curves (PRC) to herbivore

Accepted Article
This article is protected by copyright. All rights reserved dung counts. PRC is based on redundancy analysis, but designed to help interpret multivariate response within a repeated measure framework (Van den Brink & Braak 1999). Dung counts were log + 1 transformed prior to analysis. Our herbivore guilds were strongly correlated with dominance of wild and domestic herbivores, as such the analyses was simplified to wild herbivore or livestock dominance. The analysis associated wild herbivore dominance with positive scores and livestock dominance with negative scores. Plant community composition over the two years (2016 -2018) was analysed through non-metric multidimensional scaling (NMDS) using the annual survey of all experimental plots (Oksanen et al. 2017). We restricted the NMDS to two axes; however, the first axis was strongly collinear with tree canopy type and was not included in our final model. To identify whether specific species cover was enhanced under different tree canopy types as well as different tree canopies across land-uses (pastures and wildlife-protected areas), we used a method of differential analysis of count data. This analysis accounted for the low number of observations for the majority of species. P-values were generated contrasting estimated standard error of the log 2 change in species cover to zero (Love et al. 2014).
In total, 2916 plot pasture-disc measurements were analysed over a 22 month-period generating data from July 2016 to May 2018. Due to initial disturbances the first two months of the biomass data collection were omitted from the analysis. To test the potential effect of trees on herbaceous understorey standing biomass across a range of herbivore assemblages and intensities, we fitted a linear mixed model with an auto-correlated structure (accounting for repeated measures) using the Integrated Nested Laplace Approximation (INLA). INLA uses a Bayesian framework where the importance of factors can be determined via posterior parameter estimates (Rue et al. 2009). By using INLA, we were able to ensure positive estimates (i.e. above zero) of herbaceous biomass using a gamma distribution. Our primary global model included the following fixed factors: exclosure treatment (exclosed vs. open to herbivory), tree canopy type (leguminous, non-leguminous and outside tree canopy), herbivore assemblage (PRC scores -outlined above), rainfall and herbaceous species composition (NMDS axis 2 scores) and their two-way interactions.
In the global model we had two three-way interactions between treatment, tree canopy and herbivore assemblage as well as treatment, tree canopy and herbaceous species composition. For the random component of the model, a nested structure was used reflecting the experiment design with plot pair (i.e. paired open versus exclosed plots), nested within block and then site (n = 5, two in wet and three in mesic savannah). The potentially confounding variables tree height and fire history on herbaceous biomass (mentioned above) could not be included in our global model,

Accepted Article
because tree height covaried with tree canopy type, where leguminous trees were taller than nonleguminous trees and fire history covaried with NMDS axis 2. To consider the potential influence of tree height and fire history on our results we re-ran the model outlined above using tree height instead of tree canopy type and fire history instead of NMDS axis 2. Various assessment criteria for our global model suggested a good model fit, including: Wantanabe-Akaike Information

Herbaceous biomass under savannah trees with increasing livestock dominance and abundance
Herbaceous standing biomass was greater under savannah trees and when herbivores were excluded, independent of spatial and temporal rainfall patterns ( Fig. 2; Fig. 3). Over the duration of the study, mean herbaceous biomass was 0.19 ± 0.11 kg m -2 (mean ± standard deviation) under leguminous trees, 0.18 ± 0.11 kg m -2 under non-leguminous trees and 0.14 ± 0.1 kg m -2 outside tree canopies. On average, increasing livestock dominance reduced herbaceous biomass across canopy types ( Fig. 2; Fig. 3). Excluding herbivores had the greatest impact in pasturelands and livestock-dominated assemblages, leading to a 119% increase in herbaceous biomass in pasturelands compared to 54% increase in wildlife-protected areas across canopy types over the duration of the study ( Fig. 2; Fig 3 herbivore assemblage × exclosure treatment). The tree-induced (i.e. underneath tree compared to outside tree) increase in herbaceous understorey biomass was stronger in the presence of herbivores compared to excluding them ( Fig. 2; Fig.3 e.g. Tree canopy: Open vs. leguminous tree). Overall standing herbaceous biomass was 60% and 48% higher underneath leguminous and non-leguminous trees, respectively, compared to outside tree canopies when accessible to herbivores; the magnitude of these tree-induced increase in herbaceous biomass reduced to 39% and 40%, respectively, when herbivores were excluded (Fig. 2). The

Accepted Article
This article is protected by copyright. All rights reserved greatest positive effect of excluding herbivores was outside tree canopies with an increase of 159% in pasturelands and 78% in wildlife-protected areas when contrasting exclosed versus open plots.
Trees demonstrated a consistent facilitative effect in maintaining understorey herbaceous biomass across temporally variable wild herbivore and livestock abundances (estimated by dung counts) (Fig. 2; Fig. 3). Higher herbivore abundance of both wildlife-and livestock-dominated assemblages was negatively related to herbaceous biomass, this effect being more pronounced outside of the tree canopy (Fig. 4). Moreover, the model predicted differences between underneath trees and outside trees were stronger at high herbivore abundances than low wild herbivore and livestock abundances (Fig. 4). In part, this was due to higher variation in herbaceous biomass at low herbivore abundances. Despite a weak influence of herbivore assemblage on herbaceous biomass differences between canopies (Fig. 3), the relationship was strongly driven by the high livestock abundances ( Fig. 2; Fig. 4).
Herbaceous biomass was positively related to rainfall with little influence of canopy type or exclosure treatment on this relationship (Fig. 3). The increase in herbaceous biomass was greater in the mesic compared to the wet region during periods of high rainfall (Fig. 3). The negative influence of herbivore abundance on herbaceous biomass was strongest during periods of low rainfall (Fig. 2). This was due to a stronger negative influence of livestock-dominated assemblages during low rainfall periods compared to a positive influence of wild herbivore abundance during wetter periods ( Fig. 2; Fig. 3). In our system, we found that trees still exerted a positive effect on understorey herbaceous biomass even during periods of high livestock dominance in the dry periods.

Herbaceous biomass underneath leguminous vs. non-leguminous trees
Overall there was no differences in the quantity of herbaceous biomass underneath leguminous versus non-leguminous trees whether exclosed or not and across spatially and temporally variable rainfall ( Fig. 2; Fig. 3). However, herbaceous biomass underneath leguminous and nonleguminous trees differed in their response to herbivore assemblage and abundance. Higher herbivore abundance reduced herbaceous biomass underneath non-leguminous trees and outside tree canopies accessible to herbivores ( Fig. 3; Fig. 4). In contrast, biomass underneath leguminous trees was not strongly influenced by herbivore abundance, suggesting the capacity of the herbaceous vegetation to limit herbivore effects ( Fig. 3; Fig. 4). Surprisingly, greater livestock

Accepted Article
This article is protected by copyright. All rights reserved dominance negatively influenced herbaceous biomass underneath non-leguminous trees whether exclosed or not, whereas the negative impact of livestock was only observed under leguminous trees and open plots accessible to herbivory (Fig. 3). Likewise, herbivore abundance still exerted a negative influence on exclosed swards outside tree canopies (Fig. 3). Both of these findings may allude to a legacy of previous livestock-dominated assemblages and higher herbivore abundance (mainly livestock) indirectly influencing herbaceous biomass via species composition. Indeed, the influence of herbivory on herbaceous biomass depended upon herbaceous community species composition (Fig. 3).

Modulating effects of understorey species composition on the tree-herbaceous response to herbivory
Herbaceous species composition was primarily governed by whether the herbaceous plant community resided underneath a tree or outside of the tree canopy (R 2 = 0.25, p<0.001) ( Fig. 5; Appendix S1). However, herbaceous species composition was also strongly influenced by year since last fire (R 2 = 0.25, p<0.001) and study-tree height (R 2 = 0.21, p<0.001) (Appendix S1).
Tree canopy functional type (legume and non-legume tree or outside canopy) (R 2 = 0.13, p<0.001) significantly influenced the herbaceous community but less so than underneath versus outside tree canopies (Appendix S1). Plant community composition underneath trees was more dissimilar to outside tree canopies, particularly in wildlife-protected areas (Fig. 5). The composition of plant communities underneath trees in pasturelands varied from being similar to wildlife-protected tree understories (i.e. positive NMDS1 and negative NMDS2) to communities typically found outside tree canopies across land-use types (i.e. negative NMDS1 and positive NMDS2). Tree canopies and plant community composition were strong determinants of herbaceous biomass; however, these patterns were clearly modulated by herbivore abundance (Fig. 5). Taller study-trees were associated with understorey communities of greater herbaceous biomass in wildlife-protected areas (Appendix S1). However, tree height did not directly influence herbaceous standing biomass in this study (Appendix S3). Furthermore, longer periods without fire were negatively related to herbaceous standing biomass in our system (Appendix S4), but fire history did not significantly interact with tree canopy type in determining the quantity of standing biomass at our study sites (Appendix S4).
Understorey plant communities in wildlife-protected areas sustained greater herbaceous biomass across herbivore abundances (Fig. 5). In pasturelands, the relationship between

Accepted Article
This article is protected by copyright. All rights reserved herbaceous species composition, herbaceous biomass and herbivore abundance was more complex. Herbaceous plant communities underneath trees in pasturelands that were compositionally similar to understories in wildlife-protected areas, could sustain herbaceous biomass with increasing herbivore abundance (Fig. 5). However, the positive effect of trees diminished at high herbivore abundances (Fig. 5). Panicum maximum was the primary understorey specialist in our study system, with a significant 11-fold cover increase underneath leguminous and non-leguminous trees compared to outside tree plots (Appendix S2; Table 1). P. maximum was followed by Achyranthes aspera with a 5-fold cover increase underneath trees (Appendix S2; Table 1).
Themeda triandra and Panicum coloratum were abundant outside trees in wildlife-protected areas, but instead were found mainly underneath trees in pasturelands (Table 1; Appendix S2). From the species identified as associating with tree understories, P. maximum was the only species strongly positively related to greater herbaceous biomass inside exclosures (Pearson's correlation 0.47, n = 57, p<0.001), but at the same time negatively related to increasing herbivore abundance outside exclosures (Pearson's correlation -0.35, n = 59, p =0.006). In total, 10 out of 175 species observed over three years were strong candidates as tree leguminous and/or non-leguminous tree understorey specialists. Apart from the understorey specialists mentioned above, most understorey species occurred at low frequencies and percentage covers (Table 1; Appendix S2; Appendix S5).

DISCUSSION
Our results show that trees can maintain herbaceous understorey vegetation despite high herbivore pressure in both wildlife-and livestock-dominated savannahs. We observed these facilitative effects across spatially and temporally variable rainfall patterns. The facilitative effect of trees on the herbaceous layer was stronger in the presence of herbivores than without them. Although we found no difference in herbaceous biomass underneath leguminous and non-leguminous trees during our study, as is often the case in tropical savannahs (Belsky 1993(Belsky , 1994Blaser et al. 2013), plant biomass dynamics underneath leguminous trees were less strongly influenced by mammalian herbivory. Moreover, our study demonstrates that the three-way interaction between mature trees, understorey plant biomass and herbivores is mediated by the herbaceous community composition.
Plant communities found underneath trees in open wooded savannahs had a higher capacity to sustain herbaceous biomass with increasing herbivore abundance and livestock dominance. Therefore, to maintain herbaceous biomass in an increasingly livestock-dominated savannah Accepted Article ecosystem one must consider properties of the isolated trees (i.e. functional type and height) and their associated understorey herbaceous species composition.
Utilisation of tree understoreys by herbivores in savannah ecosystems is well-known and intensification of herbivory would be expected to negate positive effects of trees Smit et al. 2009). Yet, we were able to detect stronger differences (albeit lower biomass) underneath trees versus outside trees at higher wild herbivore and livestock abundances. We did not measure herbivore consumption rates beneath and away from trees that may have shaped observed plant biomass differences. Nevertheless, swards outside tree canopies, particularly in more livestock-dominated savannahs, showed higher standing plant biomass following release from herbivory (i.e. inside exclosures). This could be explained by either higher herbivore pressure away from trees (Augustine & McNaughton 2006) or growth constraints of plants beneath trees following herbivore release, such as light limitation and/or self-thinning processes (Belsky 1994;Hawkes & Sullivan 2001). Nonetheless, observed herbivore dung counts are suggestive of higher foraging rates in savannahs dominated by understorey specialist herbaceous species, for example P. maximum, compared to savannahs without these herbaceous species.
Consistent with our hypotheses, our study shows that mature trees can buffer the impact of increasing herbivore presence, especially livestock dominance. As herbivore abundances increase, an ecological warning sign of the potential loss of this tree buffering capacity in African savannahs is the loss of understorey specialist species such as P. maximum.
Soil nutrients and water availability have been shown to be tightly coupled with growth and herbaceous biomass production of tree understories (Ludwig et al. 2004b;Augustine & McNaughton 2006). We expected greater biomass underneath leguminous compared to nonleguminous trees due to the presence of nitrogen-fixing symbionts enhancing nitrogen availability (Blaser et al. 2013). Although we found no difference in herbaceous standing biomass underneath tree functional types, plant biomass under leguminous trees was less influenced by wild and livestock herbivory than under non-leguminous trees. This could still suggest that leguminous trees enhance the resilience of understorey communities due to higher soil fertility allowing plants to compensate for tissue consumed by herbivores (Belsky 1993(Belsky , 1994Riginos et al. 2009;Blaser et al. 2013). Cover of the understorey specialist P. maximum, strongly associated with greater biomass, was similarly negatively related to herbivore abundance. Thus, it is more likely that both leguminous and non-leguminous trees act as a refuge for herbaceous species sensitive to herbivory (Treydte et al. 2009), particularly in pasturelands. Supporting this, T. triandra and P. coloratum Accepted Article dominate herbaceous swards outside tree canopies in savannahs with low wild herbivore densities but were more often found underneath trees in livestock-dominated pastureland.
Optimising the spatial and temporal variability between mammalian wildlife and livestock foraging has increasingly been proposed as a way to facilitate herbivore coexistence in savannahs (Fynn et al. 2016;Tyrrell et al. 2017;Young et al. 2018). Here, we show that elevated herbaceous biomass underneath trees occurred during greater wildlife dominance in wetter periods but extended into drier periods and endured peaks in livestock dominance. Thus, pastoral grazing management plans should consider the temporal heterogeneity of herbaceous biomass provided by savannah trees (Smit 2004(Smit , 2005. For instance, greater gains in herbaceous plant biomass could be achieved by land managers that cease livestock grazing during peak growing season on pasturelands with low densities of mature trees compared to land owners that remove large trees or continuously graze livestock throughout the seasons. Additionally, our study suggests fire may also play a role in promoting heterogeneity of herbaceous biomass. Longer periods without fire contributed to homogenization of plant species composition underneath trees and outside tree canopies. Although our analysis suggested that fire did not significantly interact with tree canopies to influence herbaceous biomass, our measure of fire was coarse and unlikely to detect localised reduced fire prevalence underneath tree canopies (Coetsee et al. 2010). Long-term fire suppression is becoming more widespread in savannah ecosystems as agropastoralists seek to remove rank grass through livestock grazing rather than fire (Hempson et al. 2015(Hempson et al. , 2017Young et al. 2018).
However, if the absolute number of wild herbivores and/or livestock becomes too high, this will result in the loss of distinct understorey plant communities and a decline in herbaceous biomass.
Although we did not account for the impact of individual herbivore species in our study, these observations align with herbivore manipulations that identify a stronger negative influence of total savannah herbivore numbers on productivity rather than the functional attributes of livestock per se (Veblen et al. 2016;Charles et al. 2017;Young et al. 2018).
Pasturelands adjacent to wildlife-protected areas in tropical savannahs provide an important buffer area that can be occupied by wild herbivores for short durations of the year, as seen in our system (see also (Veldhuis et al. 2019). To ensure these pasturelands sustain herbaceous biomass as forage for both wildlife and livestock herbivores, it is necessary to preserve tall trees (Smit 2004(Smit , 2005, i.e. greater than 4 m, as in our study. Supporting a growing consensus of studies, we found taller trees maintain greater herbaceous biomass in the understorey (Treydte et al. 2009;Blaser et al. 2013;Moustakas et al. 2013), and our results show that this maintenance

Accepted Article
This article is protected by copyright. All rights reserved of biomass also occurs as an indirect effect, mediated through herbaceous species composition.
Moreover, in our nutrient-rich tropical savannahs, plant species turnover due to varying herbivory pressure is likely to be quicker than in non-tropical savannahs where understorey specialists can persist beside dead tree stumps for several decades (Stahlheber et al. 2015). The cover and distribution of understorey species, such as those identified in this study, are therefore important indicators of the resilience of the African tropical savannahs under increasing herbivore pressure. Table 1. Herbaceous species cover increasing underneath leguminous and non-leguminous trees. These were identified using differential analysis of count data (Love et al. 2014). Species functional type, occurrence (out of 375 quadrats surveyed between 2016 and 2018), plant cover when present and statistical summary from differential count analysis. P-values have been generated by contrasting estimated standard error of log 2 change in species cover to zero and have been back-calculated to x fold change. Mean cover  1 standard deviation.

Accepted Article
This article is protected by copyright. All rights reserved