Woody encroachment affects multiple dimensions of ant diversity in a neotropical savanna

Although savanna woody encroachment has become a global phenomenon, relatively little is known about its effects on multiple dimensions and levels of savanna biodiversity. Using a combination of field surveys, a species‐level phylogeny, and functional metrics drawn from a morphological dataset, we evaluated how the progressive increase in tree cover in a fire‐suppressed savanna landscape affects the taxonomic, functional, and phylogenetic diversity of neotropical ant communities, at both the alpha and beta levels. Ants were sampled along an extensive tree cover gradient, ranging from open savannas to forests established in former savanna areas. Variation in tree cover had a significant influence on all facets of diversity at the beta level, whereas at the alpha level tree cover variation affected the taxonomic and functional but not the phylogenetic diversity of the ant communities. In general, ant community responses to variation in tree cover were largely non‐linear as differences in taxonomic alpha diversity and in the taxonomic, functional, and phylogenetic composition of the sampled communities were often much stronger at the savanna/forest transition than at any other part of the gradient. This indicates that savanna ant communities switch rapidly to an alternative state once the savanna turns into forest. Ant communities in the newly formed forest areas lacked many of the species typical of the savanna habitats, suggesting that the maintenance of a fire suppression policy is likely to result in a decrease in ant diversity and in the homogenisation of the ant fauna at the landscape scale.


INTRODUCTION
Tropical savannas are one of the most biodiverse and most threatened terrestrial biomes (Bond & Parr, 2010;Murphy et al., 2016), as the expansion of agriculture and cattle ranching has rapidly and substantially reduced the extent of tropical savannas in many parts of the world (Strassburg et al., 2017). Changes in fire regimes are also a threat to tropical savannas, but as opposed to tropical forests, this threat comes not only from frequent, severe, and/or uncontrolled fires but also from fire suppression (Bond & Parr, 2010;Rosan et al., 2019). Fire was a key element in the evolution of tropical savannas and as such, many species of its flora have adaptations to or are dependent on fire (Simon & Pennington, 2012). In the absence of fire, the tree cover gradually increases and, as a result, savannas can be replaced by forests, notably in high-resource environments or in transitional regions where forest remnants share the landscape with grassland and savanna patches (Hoffmann et al., 2012). In higherrainfall savannas (also known as mesic savannas), for instance, this process can take place within a matter of only a few decades (Abreu et al., 2017;Durigan, 2020). The biome shift from savanna to forest involves a major switch in the structure and composition of plant communities-with the replacement of fire-adapted species by species that are both fire-sensitive and shade-tolerant-so that forests and savannas are regarded as alternative stable states mediated by plantfire feedbacks (Dantas et al., 2013;Hoffmann et al., 2012).
While the patterns and mechanisms involved in the development of the vegetation in fire-suppressed savanna areas are relatively well documented (e.g., Flake et al., 2021;Stevens et al., 2017), relatively little is known about how insect communities respond to woody encroachment. The few existing studies have compared communities from areas with highly contrasting woody-plant cover (such as those that have been protected from fire for a relatively long time versus those that have been burned) and therefore it is unclear whether the observed changes (e.g., such as those on termite activity or in ant community structure; Andersen et al., 2006;Parr et al., 2012;Leitner et al., 2018), are gradual or rather become more abrupt at certain phases of the woody encroachment process.
Furthermore, there is limited information on the extent to which woody encroachment affects the functional and phylogenetic structure of savanna insect communities (see Parr et al., 2012). Measures of functional and phylogenetic diversity can complement those provided by the traditional taxonomic diversity metrics. Functional diversity is a key component of biodiversity as it provides a link between species traits and the role they play in the environment (Swenson & Weiser, 2014). Phylogenetic diversity, on its turn, helps to address whether species in a community are drawn from the same clades or from distantly related lineages, while also providing insights into the evolutionary history of these species (Cavender-Bares et al., 2006;Swenson & Weiser, 2014;Webb et al., 2002).
Ants are one of the most diverse, dominant, and functionally important groups of insects in tropical savannas, and their responses to habitat disturbance are thought to be mediated mainly by changes in habitat structure (Andersen, 2019). This suggests that woody encroachment may have pervasive effects on savanna ant communities given its effects on tree cover, and thus habitat openness (Andersen, 2019). Herein we evaluated the extent to which variation in tree cover in a fire-suppressed savanna landscape affects the Neotropical ant communities. More specifically we attempted to answer the following questions: (a) What is the relationship between tree cover and the taxonomic, functional, and phylogenetic diversity of ant communities at the alpha and beta levels? (b) Are changes in ant diversity along the tree cover gradient gradual or are there evidence of a rapid switch from one community state to another at a particular point of the gradient? Since woody encroachment in fire-suppressed savannas often favours the occurrence of species adapted to forest habitats, at the expense of open-habitat savanna ant specialists (Abreu et al., 2017;Andersen et al., 2006;Maravalhas & Vasconcelos, 2014), we expected to find marked changes in the taxonomic richness and composition of ant communities along the tree cover gradient, notably at the savanna/forest transition. In addition, we expected that changes in tree cover (resulting from woody encroachment) would affect the functional and phylogenetic diversity of the ant communities, as the structure of the habitat can act as an environmental filter favouring the occurrence of certain ant traits and lineages (Blaimer et al., 2015;Gibb & Parr, 2013).

Study area
The study was performed at Santa Barbara Ecological Station (SBES), a biological reserve with an area of 2715 ha located inÁguas de Santa Barbara, São Paulo, Brazil (22 48 0 S, 49 14 0 W), at an elevation 600-680 m a.s.l. Climate in the region is classified as Koppen's Cwa, characterised by a warm summer and a dry winter (Alvares et al., 2013), annual rainfall varies from 1000 to 1300 mm, and the mean temperature of the coldest month is 18 C, while the hottest month exceeds 22 C of temperature (Meira-Neto et al., 2007).
At SBES, a fire suppression policy was in force for several decades, and an analysis of the vegetation development over a 30-year period (1985-2015) shows a steady temporal increase in tree cover (Abreu et al., 2017). At the time of our study, several of the former grassland and open savanna areas had turned into dense savanna or even forest (Abreu et al., 2017; Figure S4). In 2015, a prescribed fire experiment was initiated at SBES  and the data analysed here represents pre-fire data collected in December 2014 in all 30 plots (each 20 Â 50 m) designated for the fire experiment ( Figure S1). The experiment was replicated in three sites within SBES (hereafter blocks), located at least 1.7 km from each other. Plots located within each block encompassed the entire variation in tree cover found in the study area and included plots established in open savanna (campo sujo), dense savanna (cerrado sensu stricto), and forest (cerradão) (Figures S1 and S4).

Tree cover
As a measure of tree cover in each plot at the time of our sampling, we used the leaf area index (LAI), a satellite-derived metric that is strongly and positively correlated with tree basal area (Abreu et al., 2017). LAI data for the study plots were obtained in April and May 2015. Each sampling plot was subdivided into 10 subplots of 10 Â 10m. Each subplot had four collection points of LAI with hemispherical photographs. Photos were taken before sunrise, after sunset, or under homogeneous overcast skies. A tripod was used to position the camera (Canon EF 8-15 mm fisheye lens) at a height of 1 m, and the top of the camera was oriented relative to the north. Photos were taken with an underexposure of on f stop (Macfarlane et al., 2014), and the colour images we converted to black and white using Hemisfer 2.12 (Schleppi et al., 2007;Thimonier et al., 2010) and using maximum blue contrast (Nobis & Hunziker, 2005). The images were then analysed with Hemisfer 2.12 using an automatic threshold for closed-canopy vegetation and with a supervised manual threshold under open canopies. The LAI values were averaged over the 40 subplots to obtain a single value for each plot (for more details see Abreu et al., 2017).

Ant sampling
We used pitfall traps as the only method to sample ground-foraging ants, as the large number of traps installed (600 in total) and the amazing abundance and diversity of ants at SBES precluded the use of other sampling methods. We acknowledge, however, that this may have underestimated the number of species sampled, notably in the more structurally complex habitat, where pitfall trapping can be less effective than other methods to sample ants (Melbourne, 1999).
A total of 20 pitfall traps were installed in each plot, arranged in five grids of approximately 2.5 Â 2.5 m (keeping a minimum distance of 20 m between any two grids), with four pitfall traps installed at the corners of each grid. Each trap consisted of a small plastic cup (250 mL, 8.5 cm high and 7.8 cm in diameter) partially filled with water and detergent. The traps remained in operation for 48 h. Upon collection, the contents from the four traps set within the same grid were combined to make a composite sample. In the lab, ant workers were sorted into morphospecies and a representative specimen from each sample was dry-mounted for subsequent identification using available taxonomic keys (Bolton et al., 2007;Fernández, 2003;Fernández & Ortiz-Sepúlveda, 2019) or by comparison with specimens previously identified by ant taxonomists deposited at the Zoological Collection of the Federal University of Uberlândia and the Entomological Collection from the Federal University of Paraná, where the specimens collected were also deposited. Specimens for which a species-level identification was not possible received a morphospecies code.

Ant phylogeny
To compute the phylogeny-based diversity metrics (as detailed below), we used a maximum-likelihood tree created using DNA sequences of ultraconserved elements (UCEs) for one representative specimen of 166 of the 179 species/morphospecies collected ( Figure S2). Laboratory methods for sequence generation followed well-established protocols for library preparation, targeted capture, and sequencing of UCE data (see Supporting Information for a detailed overview). For the enrichment of UCE loci, we used an ant-customised bait set (myBaits UCE Hymenoptera 2.5Kv2A; ArborBiosciences, MI) that includes 9898 baits targeting 2524 UCE loci conserved in Hymenoptera (Branstetter et al., 2017).
The UCE loci were extracted to separate FASTA files, and each locus was aligned using MAFFT v. 7.130b (Katoh et al., 2002) and trimmed with GBLOCKS v. 0.91b (Castresana, 2000). The alignment was filtered to include only loci available for at least 50% of the taxa, resulting in an alignment including 1628 loci. The concatenated dataset was partitioned by partition schemes defined by PartitionFinder2 (Lanfear et al., 2016), incorporating also the SWSC-EN algorithm (Tagliacollo & Lanfear, 2018; see Supporting Information for a detailed overview).
The maximum-likelihood analyses were run with IQTREE v. 1.6.12 (Nguyen et al., 2015). Raw sequence reads in FASTQ format are deposited in the NCBI Sequence Read Archive (accession PRJNA927311).

Ant morphological data
We measured five morphological traits that are commonly used in studies of ant functional diversity (Martello et al., 2018;Parr et al., 2017;Weiser & Kaspari, 2006). Measurements were made of at least five specimens from each species; when fewer than five specimens were available, we measured all specimens. Whenever possible, the measured individuals were from different sampling plots. A

Diversity metrics
A dataset was built using information about the number of occurrences of each species in each plot (number of grids in which the species was recorded, for a maximum of five grids in each plot), the functional traits of each species, and the phylogenetic relationship between species. To estimate the taxonomic, functional, and phylogenetic alpha diversity of ant communities in each plot and the dissimilarity (beta diversity) between communities in different plots, we used Rao's quadratic entropy index (Pavoine et al., 2005;Rao, 1982). The Rao index is known to reflect community assembly processes well (Mouchet et al., 2010) and allows for comparing the alpha and beta components of taxonomic, functional, and phylogenetic diversity within the same mathematical framework (De Bello et al., 2010). All diversity metrics were calculated using the Rao function and applying Jost's correction as recommended by De Bello et al. (2010). We estimated the pairwise functional and phylogenetic Euclidean distances between species using the functions 'daisy' and 'cophenetic.dist'. The functions are implemented by 'cluster' and 'picante' packages, respectively (Maechler et al., 2019;Paradis & Schliep, 2018).
The standardised effect size (SES) of the functional and phylogenetic, alpha and beta diversity metrics were also calculated. Standardised metrics are less sensitive to species richness differences and, in addition, allow the test of hypotheses regarding the functional and phylogenetic structure of the communities (Webb, 2000). Negative values of SES (<À1.96) are indicative of a clumped structure (i.e., the greater importance of environmental filtering in determining the traits or lineages that are found in a given community), whereas positive values (>1.96) are indicative of a dispersed, functional or phylogenetic structure (i.e., the greater importance of interspecific competition and niche partitioning) (Gotelli & Rohde, 2002). We generated null communities by randomising (999 randomizations) the community data matrix using the 'independent swap' algorithm. This algorithm maintains the row and column totals of the original matrix, meaning that the species richness and the total number of ant records in each plot are not altered (Gotelli, 2000). The null models were built using the 'picante' package (Kembel et al., 2010). Then, we determined the difference between the observed diversity value and the mean of the null values and divided this difference by the standard deviation of the null values. All analyses were performed using the R environment (R Core Team, 2021).

Statistical analyses
To evaluate the influence of tree cover on ant alpha diversity we built mixed models, having blocks as a random factor. Analyses were done separately for each dimension of diversity (taxonomic, functional, or phylogenetic). We built simple linear regression models, quadratic models, and generalised additive models (GAM) and compared the adequacy of each of these models using the Akaike information criterion (AIC). Whenever possible, conditional and marginal r 2 values were calculated using the method of Nakagawa and Schielzeth (2013). We assumed a Gaussian error distribution in all models and checked model assumptions through visual examinations of the residual plots. Moran's I correlogram (Dray, 2011), based on six distance classes, was built to assess the presence of spatial autocorrelation in model residuals ( Figure S3).
To determine the relative influences of tree cover and geographic distance (inter-plot distances) on the taxonomic, functional, or phylogenetic dissimilarities (beta diversity) between ant communities, we employed multiple regression on distance matrices (MRM). The MRM was based on the Spearman correlation between the distance (dissimilarity) matrices, in order to test for a monotonic but not necessarily linear relationship between the matrices (Goslee & Urban, 2007). We used the Euclidean distance to build the matrix of dissimilarity in tree cover (using the LAI of each plot) and the geographic distance matrix (using the coordinates of each plot). Statistical significance of the regression coefficients and of the full model was assessed through permutation tests (1000 permutations).
We also used the matrices of pairwise dissimilarities generated by the 'Rao' function to perform non-metric multidimensional ordination analyses. We then compared the resulting ordination scores between plots previously classified (Abreu et al., 2017) as forest (LAI > 2), dense (LAI <2 and >0.5) or open savanna (LAI < 0.5) (Figure 1) using a multivariate analysis of variance (MANOVA), which also took into account the block (random) factor. A posteriori, pairwise multiple comparisons were performed using Hotelling's T-square test.

RESULTS
In total, we had 2201 species records of 179 ant species from 53 genera (Table S1). The most diverse genus was Pheidole with 33 species, followed by Camponotus (16 species) and Solenopsis (16 species

Tree cover and alpha diversity
Both taxonomic (TD) and functional alpha diversity (FD and ses.FD) varied significantly with tree cover (LAI), and in both cases, the GAM and quadratic models presented a better fit to the data than did the linear models (as indicated by the AIC values of each model; Table S2).
TD tended to decrease, whereas FD and ses.FD tended to increase as tree cover increased (Figure 1). Phylogenetic alpha diversity (PD and ses.PD), in contrast, was not significantly related to tree cover ( Figure 1; Table S2). Looking at all the 30 ant communities sampled along the tree cover gradient we found that neither mean ses.FD or mean ses.PD differed significantly from zero (one sample t-test: ses.FD, t = 0.14, p = 0.88; ses.PD, t = 0.69, p = 0.49), indicating that these communities present a random functional and phylogenetic structure.
Overall, we did not detect any sign of spatial autocorrelation in the residuals of the models that evaluated the influence of tree cover on ant alpha diversity ( Figure S3).

Tree cover and beta diversity
Our MRM models revealed that dissimilarity in tree cover (based on the LAI of each plot) was a much stronger determinant of the observed variation in ant community composition than was the geographic distance between the survey plots (Table 1). Community dissimilarity increased as the dissimilarity in tree cover increased, and this was true in the analyses involving the taxonomic, functional, and phylogenetic beta diversity (Figure 2). The amount of variation explained by our models, however, was much greater for models involving beta TD, than those involving beta FD or PD. Furthermore, the standardised regression coefficients suggest that TD increases at a faster rate with tree cover dissimilarity than FD or PD (Table 1).
The ordination analyses reinforced the view that variation in tree cover was a major determinant of the observed dissimilarities in taxonomic composition (Figure 3). We found strong and significant differ-

DISCUSSION
Our study is the first to evaluate how variation in tree cover within a fire-suppressed savanna-dominated landscape affects different facets of ant diversity at both the alpha and beta levels. As expected, we found significant variation in ant species richness (alpha TD) along the tree cover gradient. In contrast with the pattern detected in studies of secondary succession in tropical forests, during which ant species richness changes gradually as the forest grows older (Bihn et al., 2010;Rocha-Ortega et al., 2018), we found relatively little variation in alpha TD along the first half of the tree cover gradient (i.e., from open to dense savanna), but at plots that became forests as a result of the fire suppression policy at our study site, alpha TD declined markedly. We cannot discard the possibility that this result reflects, at least in part, the limitations of our sampling protocol which employed a single sampling method whose effectiveness may vary with the structure of the sampling habitat (Melbourne, 1999). A similar study in Australia that used a combination of methods for sampling ants also found a sharp decline in ant species richness with woody encroachment resulting from the long-term exclusion of savanna fires (Andersen et al., 2006).
As also expected, we found marked differences in ant taxonomic community composition along the tree cover gradient, with the dissimilarity in taxonomic composition (beta TD) between any two communities increasing as the difference in tree cover increased. These dissimilarities were much stronger when comparing the forest against the open and dense savanna communities than when comparing the two types of savanna. Many of the species found in the savannas (and even some genera) occurred exclusively or predominantly in this type of habitat, whereas others were much more characteristic of the forest habitat (Figure 4; Table S1). These findings reinforce those of earlier studies by showing that fire suppression in mesic savannas results in the local loss of ant species adapted to open-habitat conditions F I G U R E 2 Taxonomic, functional, and phylogenetic dissimilarities between ant communities (beta diversity) in relation to the Euclidean distance in leaf area index (LAI) between sampling plots. (a) Taxonomic beta diversity, (b, c) functional beta diversity (absolute and standardised, respectively), (d, e) Phylogenetic beta diversity (absolute and standardised). Lines represent a locally weighted scatterplot smoothing of the data (LOWESS). (Abreu et al., 2017;Andersen et al., 2006;Maravalhas & Vasconcelos, 2014). Furthermore, the fact that changes in TD along the tree cover gradient were largely non-linear (being much greater at the savanna/forest transition than at any other part of the gradient) suggests that ant communities switch rapidly to an alternative state once the savanna turns into forest.
Savanna woody encroachment affected not only the taxonomic but also the functional and phylogenetic diversity of the ground-  Table S1 for the full name of each species.
dwelling ant communities. The influence of tree cover on FD and PD was, however, weaker than its influence on TD, notably at the alpha level. In fact, variation in tree cover did not have a significant effect on alpha PD. Furthermore, in none of the ant communities we sampled did we find greater or lower than expected alpha ses.PD values, suggesting that these communities have a random phylogenetic structure along the entire tree cover gradient. Similarly, all the communities sampled presented non-significant alpha ses.FD values. Despite having a relatively low TD, however, forest communities tended to present a higher alpha FD (as well as a higher alpha ses.FD), especially in comparison with communities in open savanna. This is indicative of greater pairwise functional differences between species of the forest community than between the open savanna species. It also suggests that the relative importance of interspecific competition and niche partitioning on the functional structure of ant communities potentially increases as tree cover increases (Gibb & Parr, 2013;Swenson & Weiser, 2014;Webb, 2000).
At the beta level, in contrast, we found a much greater congruence between patterns of taxonomic diversity and those of functional, and phylogenetic diversity. This indicates that the changes in species composition observed along the tree cover gradient, notably those seen at savanna/forest transition, were accompanied by concomitant changes in the occurrence of certain ant traits and lineages. We did not record any species of Carebara or Hypoponera in the savanna plots. Conversely, we did not find any Dorymyrmex, Forelius, Gracilidris, or Linepithema in the forest. Interestingly, while both Carebara and Hypoponera have minute (or even vestigial) eyes, Dorymyrmex, Forelius, Gracilidris, and Linepithema have relatively large-sized eyes which are positioned more dorsally than laterally (Guerrero, 2019). These findings, thus, lend additional support to previous studies claiming that habitat structure is a strong determinant of the success (or not) of certain species and morphological traits in ant communities (Gibb & Parr, 2013).

Conclusions and implications
Overall, our results give support to the idea that habitat openness is a key driver of variation in ant communities and that ant responses to disturbance are strongly linked to their responses to habitat openness (Andersen, 2019). In addition, we found strong support for the idea that fire mediates alternative states of ant communities in tropical savannas (Andersen et al., 2012). Thirty years of a fire suppression policy in our study area resulted in a substantial increase in tree cover, with former grassland and savanna areas becoming forests in several cases (Abreu et al., 2017). Where this occurred, the composition of