Abstract
The increasing worldwide interest on the conservation of tropical forests reflects the conversion of over 50% of their area into agricultural lands and other uses. Understanding the distribution of remaining biodiversity across agricultural landscapes is an essential task to guide future conservation strategies. To understand the long-term effects of fragmentation on biodiversity, we investigated whether forest fragments in southeastern Brazil are under a taxonomic homogenization or heterogenization process. We estimated pre-deforestation species richness and composition based on a Species Distribution Modelling approach, and compared them to the observed patterns of α- and β-diversity. In particular, we asked (i) if changes in β-diversity reveal convergence or divergence on species composition; (ii) if these changes are similar between forest fragments in Strictly Protected Areas (SPAs) (n=20) and within private lands (n=367) and in different regions of the state (West, Center, and Southeast). We detected steep reductions in observed local species richness in relation to our modeled predictions, and this was particularly true among forest fragments in non-protected private lands. The higher observed β diversity indicated an overall biotic heterogenization process, consistent with the idea that the originally diverse vegetation is now reduced to small and isolated patches, with unique disturbance histories and impoverished communities derived from a large regional species pool. Since conservation of biodiversity extends beyond the boundaries of strictly Protected Areas, we advocate forest fragments are valuable for conservation in agricultural landscapes, with particular relevance for private lands, which represent the most exposed and neglected share of what is left.
INTRODUCTION
Tropical forests are known for holding a substantial portion of the world’s terrestrial biodiversity (Myers et al. 2000; Chazdon, Harvey, et al. 2009; Slik et al. 2015), yet over 50% of its original area was converted into agricultural lands or other uses, with prospects for an increasing agricultural expansion in tropical developing countries (Hansen et al. 2013; Laurance et al. 2014). Understanding the distribution of remaining biodiversity across human-modified landscapes (HMLs) and what has changed after forest conversion are essential questions to guide future conservation strategies (Tabarelli et al. 2010; Melo, Arroyo-Rodríguez, et al. 2013; Laurance et al. 2014; Socolar et al. 2016). In HMLs, tropical forest fragments comprehend a variety of different sized habitats, including forests that never experienced clear cutting or severe impacts (i.e. primary forests), and the full spectrum of degraded forests that are regenerating after extraction, fire or abandonment of croplands and pastures, among other previous land-uses (i.e. secondary forests) (Melo, Arroyo-Rodríguez, et al. 2013; Malhi et al. 2014). A very narrow fraction of these forest fragments are under restrictive protected areas, where biodiversity conservation is tangible to a limited extent (Andam et al. 2008; Joppa et al. 2008; Coetzee et al. 2014). In this context, the variety of forest fragments located within private lands not only represent the largest share of what is left (Gardner et al. 2009; Sparovek et al. 2012; Soares-Filho et al. 2014; Mendenhall et al. 2016), but also the most neglected. These fragments are rarely explicitly targeted in conservation programs, as the focus is usually on avoiding deforestation (Chazdon, Harvey, et al. 2009; Barlow et al. 2016). Several studies have shown that secondary forests may play an important role in conservation (Santos et al. 2007, Chazdon et al. 2009; Dent & Wright 2009; Tabarelli et al. 2012), as they hold a depleted but relevant portion of biodiversity even within HMLs. Abundant evidence is available for birds (Karp et al. 2012; Emer et al. 2018), mammals (Galetti et al. 2009; Pardini et al. 2010; Beca et al. 2017) and plants (Arroyo-Rodríguez et al. 2008; Lima et al. 2015; Sfair et al. 2016; Farah et al. 2017).
The efects of habitat loss and fragmentation on biodiversity have been intensively studied over the past three decades, with a primary focus on more local scales (Karp et al. 2012; Vellend et al. 2013; Dornelas et al. 2014; Malhi et al. 2014; Newbold et al. 2015; Barlow et al. 2016). There has been an increase on studies focusing on broader extensions, based on the assumption that we cannot properly understand the consequences of deforestation if disregarding the influence of entire landscapes over local processes (Malhi et al. 2014). Also, beyond the intuitive interest on species richness loss, an emerging issue of interest is how community composition responds to fragmentation along spatial gradients and periods of time (Arroyo-Rodríguez et al. 2013; Solar et al. 2015; Morante-Filho, Arroyo-Rodríguez, et al. 2015; Collins et al. 2017; Olden et al. 2018). Measures of the variation in species composition among sites (β) can indicate if communities are converging or diverging in response to fragmentation, providing relevant information on the maintenance of regional diversity (Socolar et al. 2016).
Some studies have demonstrated that forest fragmentation and degradation result into biotic homogenization (Vellend et al. 2007; Lôbo et al. 2011; Karp et al. 2012; Marcelo Tabarelli et al. 2012; Püttker et al. 2015; Zwiener et al. 2017), i.e. the convergence of biotas in time and space, in which communities may suffer a simplification of their genetic, taxonomic and functional diversities (McKinney & Lockwood 1999; Olden & Rooney 2006). The rationale is that more ecologically specialized species (“losers”) are locally extinct, while a much narrower sub-set of generalists, with high dispersal abilities (“winners”), override them (Lôbo et al. 2011; Marcelo Tabarelli et al. 2012; Mendenhall et al. 2016). This process results in impoverished communities that represent sub-sets of a larger pool of species, translated by reduced β-diversity and high contribution of the nestedness component on β diversity. The predominance of nestedness suggests conservation efforts might focus on the richest sites, provided they are connected to support viable communities (Howe 2014; Socolar et al. 2016; Emer et al. 2018). An opposite consequence to fragmentation and degradation occurs when communities diverge on composition over time and space (i.e. enhanced β diversity) because they suffer different frequencies and levels of disturbances combined with dispersal limitations and environmental heterogeneity, resulting into biotic heterogenization (Dornelas et al. 2014; Solar et al. 2015; Sfair et al. 2016; Catano et al. 2017; Collins et al. 2017). In this scenario, widespread regional diversity conservation is only possible if targeting multiple sites.
Very few studies of plant community changes in response to fragmentation evaluate β-diversity patterns based on temporal replicates (Lôbo et al. 2011; Dornelas et al. 2014; Haddad et al. 2015; Collins et al. 2017); most of them adopt a space-for-time approach (e.g., disturbed × undisturbed) regardless of the fact that distinct sites could reflect distinct pre-disturbance conditions (Collins et al. 2017). In addition to the lack of temporal replicates, severely deforested landscapes may not be suitable for the space-for-time approach when in the absence of large forest remnants or high forest-covered regions to represent undisturbed ecosystems. For that matter, environmental niche modeling (ENM) and species distribution modeling (SDM) can be useful to provide species’ spatial occurrence disregarding the effects of anthropogenic disturbances, which is a major driver of community composition changes (Malhi et al. 2014, Catano et al. 2017). Since ENM is based on the niche concept and considers environmental conditions as the primarily influence over the establishment of a given species (De Marco Junior & Siqueira 2009), it results in maps representing the geographic space where the abiotic conditions are appropriate (Peterson & Soberon 2012). Distinctly from ENM, which disregards dispersal/colonization limitations and biotic interactions, SDM restrict the model calibration to accessible areas, incorporating dispersal issues into analyses and producing maps where a focal species may potentially occur, with varying degrees of suitability (De Marco Junior & Siqueira 2009; Peterson & Soberon 2012). For community-level modeling, further methodologies are available to adjust the over-prediction of the number of species coexisting at a given location (Guisan & Rahbek 2011; Calabrese et al. 2014; Gavish et al. 2017). Therefore, these approaches combined can be used to generate expectations of what communities would look like, in terms of species composition, if they were not disturbed by land use conversion.
In recent years, Brazil has stood out among tropical developing countries for its environmental engagement, which resulted on the exceptional decline in deforestation rates during 2000-2012 (Hansen et al. 2013; Loyola 2014), despite the fact that it already has 30% of its total area occupied by agricultural lands (Martinelli et al. 2010). In Sao Paulo state, southeastern Brazil, which includes two biodiversity hotspots (Atlantic Forest and Cerrado) (Myers et al. 2000; Laurance 2009), deforestation has taken place during the last three centuries (Metzger 2009), and surveys only became a common practice in the last 30 years (Haddad et al. 2015; Renato Augusto Ferreira de Lima et al. 2015). With a very long history of land conversion for agricultural purposes, most remaining vegetation is comprised by small forest fragments (i.e. <50ha) (Ribeiro et al. 2009), representing an unique opportunity to understand the long-term effects of fragmentation on biodiversity. The purpose of our study was to evaluate whether the woody assemblages on forest fragments are under a taxonomic homogenization or heterogenization process in response to habitat fragmentation. For that matter, we estimated pre-deforestation species richness and composition based on a Species Distribution Modelling approach, and compared them to the observed patterns of α- and β-diversity. In particular, we asked (i) if changes in β-diversity indicate convergent or divergent composition; (ii) if these changes are similar between forest fragments under strict protection or within private lands and in different regions of the state. In the hyper-fragmented landscapes of this study, we expected to find lower mean values of α-diversity within private lands relative to strictly Protected Areas. Additionally, because of intrinsic environmental heterogeneity strengthened by fragmentation disturbances, we expected an overall increase in β-diversity, indicating a taxonomic differentiation process. This pattern should be particularly more evident for unprotected forest fragments located within private lands, where fragments are more susceptible to a broader range of recurrent disturbances (Laurance et al. 2014; Malhi et al. 2014).
METHODS
Study region
The state of Sao Paulo is located within the range of two current global hotspots, the Atlantic Forest and Cerrado (tropical savannas) (Myers et al. 2000). With a long history of deforestation caused by timber extraction and agricultural cycles (coffee, pasture, orange, sugarcane) (Metzger 2009), the São Paulo state case-study may provide relevant insight about the long-term effects of habitat loss and fragmentation on biodiversity, which can be useful for other tropical regions facing the same threats (Laurance 2009). Both hotspots are poorly protected, with only 1.6% and 0.5% of Atlantic Forest and Cerrado’s original area protected as strictly Protected Areas (Durigan et al. 2006; Ribeiro et al. 2009; Carranza et al. 2014). The remaining vegetation cover in the interior plateau (i.e. excluding coastal areas) ranges from 1 to 30% (São Paulo State Forest Inventory 2011), mostly located within private rural properties (Gardner et al. 2009; Sparovek et al. 2012; Soares-Filho et al. 2014; Mendenhall et al. 2016).
In order to facilitate analyses’ interpretation along the extent region of this study, we adopted the ecological regions defined by Setzer (1966), which divide the state in 6 sub-regions based on climate, soil, topography and vegetation variables (Setzer 1966). We excluded the south and north coastal areas because their forest cover is well above the rest of the state. To meet a minimum of 30 localities per sub-region, we joined Southwest and Northwest into one single “West” sub-region (Figure 1).
Regarding vegetation, we focused on the predominant seasonal semi-deciduous forest (SSF), considering its transition to evergreen forests or forested savannas (Cerradão) (Oliveira-Filho & Fontes 2000; Durigan & Ratter 2006), and all other forest ecosystems included in this extension: swamp, alluvial and deciduous forests. Despite the fact that these forests are influenced and determined by soil, altitude and climatic conditions (Oliveira-Filho & Fontes 2000), their floristic composition are strongly influenced by the surrounding vegetation (e.g. SSF), creating complex transitional mosaics and continuum distribution of species (Kurtz et al. 2015; Oliveira-Filho & Fontes 2000).
Woody plant species occurrence data
We chose to use woody plant species (i.e. trees and shrubs) in this study because they represent a fundamental structure and functional component of forest ecosystems, as they support food webs and represent a substantial proportion of tropical diversity (Arroyo-Rodriguez et al. 2015; Slik et al. 2015). Based on a compilation of 367 floristic surveys in private lands (more details in Rodrigues et al. 2011 and Farah et al. 2017), we initially defined a species pool of 921 species. To enhance our sample of the environmental space occupied by these species and to improve modelling outcomes, we retrieved complementary national occurrence data for these species from SpeciesLink (http://splink.cria.org.br/) and NeoTropTree (http://prof.icb.ufmg.br/treeatlan/) (Oliveira-Filho 2017). Finally, we selected 691 species occurring in more than 10 localities along the Brazilian territory, which comprised 114,740 records.
We also gathered available checklists for 20 strictly Protected Areas within the study region (Figure 1) following the same criteria regarding forest types as those applied to select forests in private lands. We did not use the strictly Protected Areas localities to build species distribution models due to the lack of precision on their geographical coordinates, varying from a random point within their boundaries to the municipality centroid. We compiled the floristic surveys available at Fundação Florestal (http://fflorestal.sp.gov.br) and Instituto Florestal (http://iflorestal.sp.gov.br), both related to the São Paulo State Environmental Secretariat, and WWF Protected Areas’ Observatory (http://observatorio.wwf.org.br/). Species names were standardized using the Plantminer web tool (www.plantminer.com) (Carvalho et al. 2010), based on Flora do Brasil (www.floradobrasil.jbrj.gov.br) (Flora_do_Brasil_2020) and The Plant List (www.theplantlist.org/). Complementary queries were performed on The Missouri Botanical Gardens (www.tropicos.org). According to these databases, we excluded any exotic and unidentified species from final compilation.
Environmental data
We compiled 22 environmental predictors with spatial resolution of 1km2 and summarized them by using a Principal Component Analysis (PCA) considering 1,000 randomly distributed points within Sao Paulo state. Pairs of variables with scores > I1I (absolute value) were verified to avoid multicollinearity (correlation <0.7) finally selecting 6 variables with the highest PCA scores (Appendix 1).
Environmental Niche Modeling and Species Distribution Modeling
Environmental Niche Models (ENM) are statistical models that relate focal species occurrence to associated environmental conditions, generating correlative rules that allow extrapolation and prediction of occupancy patterns over wide geographic extents, representing a valuable tool for conservationist purposes (De Marco Junior & Siqueira 2009; Angelieri et al. 2016; Gavish et al. 2017; Guisan et al. 2013). We applied a Species Distribution Model (SDM) approach by restricting ENM to accessible areas, aiming to presume the distribution range of species if they were not affected by habitat loss, fragmentation and disturbance. In other words, if species’ distribution were primarily defined by abiotic conditions (i.e., environmental niche), in the lack of constraints imposed by altered habitat and landscape structure (e.g., fragmentation and patch isolation) (Peterson & Soberon 2012). As ENM and SDM can be based on the same sets of mathematical algorithms, occurrence data and environmental variables (Peterson & Soberon 2012), they are near-synonymous: the main difference is that SDM implies on some sort of restriction over ENM, which will be further detailed; for the purpose of this study, we hereafter will refer to our modeling approach only as SDM.
We built the SDM for each species using the Model-R framework (Sánchez-Tapia et al. 2018) with a three-fold cross validation procedure, meaning that two partitions were used for parameter estimation and algorithm training, and one to evaluate the model’s accuracy. Random pseudo-absence points (nback = 1000) were sorted within a mean distance buffer, where the radius of the buffer was the mean geographic distance between the occurrence points. If one species’ records were less than 20 km apart, they were rarefied to reduce effects of sampling bias and avoid modelling overfitting (Elith et al. 2006; Zwiener et al. 2017).
In the Model-R framework, for each partition and algorithm a model was built and its performance was tested by their True Skill Statistics (TSS) (Allouche et al. 2006). We previously tested several algorithms – BioClim, GLM, SVM, Random Forest, MaxEnt – and selected the last two based on their overall performance (Appendix 2), which was consonant to the results found by Diniz-filho et al. (2009). We then selected Random Forest and MaxEnt partitions with TSS>0.4, and applied a threshold that maximizes two error types: sensitivity (i.e. true presences) and specifity (i.e. true absences) (Sánchez-Tapia et al. 2018). The resulting binary models were averaged into a final model for each algorithm, an then combined into a final ensemble model with an average threshold that maximizes TSS values (Sánchez-Tapia et al. 2018), resulting in a final map indicating areas of probable presence. These maps were primarily generated for all the Brazilian territory and then cropped for São Paulo State.
Species richness and community composition
Our analyses were based on two different type of information: the observed species richness and community composition in each site and the species richness and community composition predicted by SDM. The observed metrics were adjusted to consider only modelled species, that is, species occurring in more than ten localities, at least 20km apart and with a final ensemble model derived from Random Forest and Maxent algorithms, using their partitions with TSS>0.4 (Figure 2a), as previously detailed. This adjustment was necessary in order to make proper comparisons between observed and predicted richness and compositions, and considered a final sub-set of 663 woody species.
Species richness based on SDM can be predicted either by stacking individual species-level models (Stacking SDM, S-SDM) or by modelling α-diversity itself (Macroecological Models, MEM) (Gavish et al. 2017; Calabrese et al. 2014; D’Amen et al. 2015; Guisan & Rahbek 2011). Since stacking binary presence/absence SDM tend to overpredict richness, as it does not account for biotic interactions or filters (Gavish et al. 2017; Guisan & Rahbek 2011), Calabrese et al. (2014) proposed S-SDM corrections to reduce these overpredictions and concluded that if stacked correctly, S-SDM are no worse than MEM. According to their findings, a corrected S-SDM approach involves summing-up the raw predicted suitabilities for each locality instead of summing-up their binary values (Calabrese et al. 2014; D’Amen et al. 2015) (Figure 2b). Hence, we inputted an average of the raw suitability values into the areas of probable presence defined on the final ensemble models; the predicted community species richness is the sum of these values considering any given location (i.e. pixel).
Following the estimation of potential richness, we predicted site-level composition by adopting the “top alpha” approach (Gavish et al. 2017), where we ranked the species’ suitabilities of occurrence per site from the highest to the lowest values – based on their individual ensemble SDM – and then selected the top number of species that equals to the predicted potential richness per site (D’Amen et al. 2015; Gavish et al. 2017) (Figure 2b).
β-diversity analyses
The compositional variation among communities from site-to-site (β-diversity) relates local diversity (α) to the regional species pool (γ) (Anderson et al. 2010). When evaluating the effects of habitat loss and fragmentation, changes on the organization of biodiversity over space and time can reveal if biological homogenization or heterogenization is taking place, an essential information to guide conservation planning over regional diversity (Socolar et al. 2016; Arroyo-Rodríguez et al. 2013; Püttker et al. 2015). Despite the valuable contribution from evaluating β-diversity, its interpretation must be very cautious, as there are several ways to measure and compare it (Koleff et al. 2003; Baselga et al. 2007; Jost 2007; Chao et al. 2012; Jost et al. 2010; Anderson et al. 2006, 2010; Tuomisto 2010).
There is extensive debate regarding the interrelationships among α, β and γ-diversity, in addition to measures for partitioning it (i.e. multiplicative or additive) and statistical approaches to properly analyze β-diversity (Anderson et al. 2010). A particular concern for our study is the fact that a great variety of metrics to estimate β-diversity depend on α and γ diversity – and therefore on scale and sample size. Considering that our samples represent a compilation of floristic surveys using distinct methods and sampling effort, we decided to use a β-diversity metric that weights on composition dissimilarities more than on richness differences (Koleff et al. 2003). For that matter, we calculated pairwise Sorensen (βSOR) and Simpson indices (βSIM) among sites, which indicate the overall variation on the species composition between pairs of sites (βSOR) and the variation related to its turnover component (βSIM), reflecting the replacement of species (Baselga 2010; Baselga et al. 2015; Socolar et al. 2016).
To evaluate differences between observed and predicted β-diversity, we: (i) ran a Principal Coordinate Analysis (PCoA) based on the Simpson (βSIM) observed dissimilarity matrix (as the turnover component was the main contribution to total beta diversity; see bellow); (ii) estimated the distance of each site (forest fragment) to the group centroid in the multivariate ordination space generated by the PCoA (βd, Figure 3); (iii) repeated steps (i) and (ii) using the SDM predicted species composition; and (iv) compared mean observed βd with mean predicted βd with permutational paired t-test. In this test, observed βd values were paired with predicted βd ones. βd is analogous to the local contribution to beta diversity (LCBD) proposed by Legendre and De Cáceres (2013), since higher values of βd represent higher distinctiveness of one site (forest fragment) within a group. This metric with standard effect size allowed us to test the null hypothesis that beta diversity does not differ among observed and predicted communities.
RESULTS
For the 663 woody species considered in this study, overall comparisons revealed that predicted richness at the site level (α-diversity) was 3.8 times higher than observed richness. This ratio was much lower when considering only strictly Protected Areas, ranging around 1.0. In fact, the few forest fragments that presented increased observed species richness in relation to predicted richness were mostly represented by Protected Areas (Figure 4) (Appendices 3 and 4).
Partition of total beta diversity (βSOR) indicated a consistent higher contribution of turnover (βSIM) to overall dissimilarity within regions and protection categories (Table 2) (Appendix 5).
We found that sites with lower predicted than observed βSOR (ratio < 1) also had higher predicted than observed species richness (ratio > 1) (Figure 5), which indicates a correlation between local species loss and heterogenization.
The mean observed distance to centroid (βd) was significantly higher than the predicted βd for all regions and for private lands, with strictly protected areas being the only exception (Table 3; Figure 6).
Individual βd values varied within regions and protection categories from almost zero to almost 0.8, especially in the central region (Figure 6a). Despite this variation, most observed-predicted pairs of sites showed an increase in βd from predicted to observed values (Figure 6b), indicating a clear trend for biotic heterogenization in the different regions. After splitting the data into protection categories (all regions pooled), we found that the higher beta-diversity in the observed data was due to an increase in βd values in private lands, as there was no difference between observed and predicted values in strictly protected areas. These results were confirmed by the paired t-test (Table 3).
DISCUSSION
We detected a general loss on observed local diversity in relation to our modeled predictions and this was particularly true among forest fragments in private lands, where we registered consistent reductions in species richness. The higher β diversity registered for the observed dataset imply an overall biotic heterogenization. However, our study also exposes the complexity of this process, with evidence indicating that both homogenization (positive observed-predicted pairwise slopes) and heterogenization (negative observed-predicted pairwise slopes) are taking place in these hyper-fragmented landscapes. Even though distinct regions of São Paulo state were gradually occupied and converted along time – from the eastern coast towards western countryside – all of them presented the same pattern: observed and predicted beta diversity were significantly different, with lower mean values for the latter. This is probably because of the results found among private lands, which represent almost 95% of our samples and drove the pattern registered for all state regions (i.e., heterogenization). As an exception, strictly Protected Areas had lower local species’ loss with no significant differences between observed and predicted β diversity, suggesting they may be fulfilling, to some extent, their protection purpose.
Biodiversity changes in tropical ecosystems are extremely complex to evaluate and understand, as they are scale and context dependent, differ among taxonomic groups and ecosystems, and often respond differently to similar environmental changes (Vellend et al. 2013; Dornelas et al. 2014; Newbold et al. 2015; McGill et al. 2015; Boesing et al. 2018; Catano et al. 2017; Magurran et al. 2018). For instance, recent studies found no evidence for systematic loss in local diversity (Vellend et al. 2013; Dornelas et al. 2014), while several others indicate this is not true for the tropics, where a variety of taxa experienced steep local species decreases in human modified landscapes (Haddad et al. 2015; Mendenhall et al. 2016; Beca et al. 2017; Ceballos et al. 2017; Farah et al. 2017; Galetti et al. 2017; Barlow et al. 2018; Bovendorp et al. 2018). Whereas the unprecedented level of forest degradation, fragmentation and intensive land use have an undeniable contribution to immediate and long-term local diversity loss (Haddad et al. 2015; Barlow et al. 2018), less is known regarding how these disturbances modify and drive the community composition along time and space, regardless of the recent growing interest and evidence on these compositional shifts (Dornelas et al. 2014; Haddad et al. 2015; McGill et al. 2015; Collins et al. 2017; Olden et al. 2018).
The comparison between observed and predicted species composition revealed the idiosyncratic responses of β-diversity (i.e. homogenization and heterogenization) in agricultural landscapes, which can be explained by several mechanisms suggested by the literature. Studies that observed biotic homogenization associate β diversity reduction to niche-selection processes, suggesting ecological filtering overrides environmental heterogeneity (Vellend et al. 2007; Lôbo et al. 2011; Marcelo Tabarelli et al. 2012; Arroyo-Rodríguez et al. 2013; Morante-Filho, Arroyo-Rodríguez, et al. 2015; Püttker et al. 2015; Zwiener et al. 2017). This statement assumes non-random local species extinctions occur because habitat fragmentation affects species differently, according to traits such as rarity, life span, dispersal, and reproductive mode (Haddad et al. 2015), supporting the proliferation of widespread, short-lived and small-seeded species (e.g. pioneer species, generalists or “winners” as defined by Tabarelli et al. 2012a) in detriment of rare and shade-tolerant species (e.g. specialists or “losers”) (Lôbo et al. 2011; Marcelo Tabarelli et al. 2012; Arroyo-Rodríguez et al. 2013; Morante-Filho, Arroyo-Rodríguez, et al. 2015; Zwiener et al. 2017). Additionally, non-random plant extinctions may also be related to selective logging, which overharvest valuable hardwood species, and to the disappearance of large and medium frugivores through overhunting and habitat loss, with cascading effects over plant-frugivore interactions, species persistence, ecosystem services and functioning in human-modified landscapes (Bello et al. 2015; Bovendorp et al. 2018). All of these mechanistic explanations may be related to the homogenization registered in our study region, where forest fragments are usually small and therefore exposed to edge effects, with depleted plant-animal interactions – especially large-sized species (Beca et al. 2017, Emer et al. 2018), and subject to recurrent fire and other disturbances (Farah et al. 2017).
Overall, however, our results showed that biotic heterogenization is the predominant process in our study region, accordingly to studies that found compositional shifts heading towards divergent communities (Smart et al. 2006; Dornelas et al. 2014; Solar et al. 2015; Sfair et al. 2016; Collins et al. 2017). In fact, a meta-analysis carried by Catano et al. (2017) found 21 cases of heterogenization among 22 studies evaluating herbaceous plants in disturbed/undisturbed grasslands and savannas. Assuming the long history of forest degradation and fragmentation in our study region as the main driver acting upon forest fragments, we refer to some mechanisms that may explain why the studied communities were more heterogeneous than compared to our modeled predictions. First, there is a combination of long-term disturbances that impose constant selection pressures (e.g., edge effects) with occasional and contingent perturbations (e.g., fires, windstorms etc.), resulting in unique disturbance histories and shifts in the physical environment (e.g. microhabitat conditions), that most likely enhance pre-disturbance compositional differences (Haddad et al. 2015; Catano et al. 2017). Second, forest fragments in private lands are more susceptible to disturbances due to the lack of formal and effective protection, proven by their altogether smaller patch sizes (Ribeiro et al. 2009). They also represent a greater variety of conditions that range from mature forests experiencing post-fragmentation changes to regenerating secondary forests (Laurance et al. 2014; Malhi et al. 2014; Farah et al. 2017). In common, they share reduced local species richness, which together with a large regional species pool (i.e., γ diversity) may create a sampling effect; i.e., a higher probability of more distinct composition between sites when a small portion of the species pool (i.e., low α diversity) is expected to occur in any random community, inflating β diversity (Karp et al. 2012; Newbold et al. 2015). The third explanation is particularly relevant in the studied hyper-fragmented landscapes, where dispersal limitation due to patch isolation might play a dominant role in making those communities such heterogeneous. This is supported by the strong positive relation between local species richness and seed arrival in plant communities (Myers & Harms 2009) and because dispersal limitation play a stronger role in determining community assembly in tropical forests (Myers et al. 2013). More specifically, Catano’s et al. (2017) findings on how disturbance and dispersal interact and alter community composition support that increased β-diversity in disturbed landscapes occurs when dispersal is limited, challenging the hypothesis that disturbances always homogenizes communities compositions through deterministic environmental filtering, that is, selecting those species best able to survive within HMLs (Vellend et al. 2007; Lôbo et al. 2011; Arroyo-Rodríguez et al. 2013; Püttker et al. 2015). Finally, other plausible mechanisms acting upon these forest fragments may be related to the reduction in the number of individuals and thus in community size, turning them more susceptible to ecological drift and other stochastic forces (Orrock & Watling 2010), and to competitive release arising from the removal of dominant species (Catano et al. 2017).
Given that, by definition, proper evaluation of biotic homogenization or heterogenization processes depend on quantifying changes in β diversity through space and time (McKinney & Lockwood 1999; Olden & Rooney 2006; Olden et al. 2018), the use of Species Distribution Model proved valuable for predicting community composition in the absence of habitat loss and fragmentation, serving as a temporal surrogate in our study. However, we must acknowledge that our modeling approach imposed some restrictions, notably the non-inclusion of rare or poorly-sampled species and biotic interactions. That said, we do not expect that the overall trend registered here – biotic heterogenization – would be affected by the absence of rare species because their inclusion would most likely increase the differences among communities, while biotic interactions were addressed by choosing a method to adjust or at least reduce an overprediction bias related to the lack of biotic interactions (Calabrese et al. 2014; Gavish et al. 2017; Guisan & Rahbek 2011; D’Amen et al. 2017). Another caveat is that the reduced local species richness among private lands may be related, to some extent, to sampling effort. Since the floristic assessments that compose most of the dataset used here aimed to quickly characterize the regional flora for restoration purposes (Rodrigues et al. 2011), we applied preliminary analysis of incidence-based estimated richness (e.g., Chao 2, Jackknife 1 and Jackknife 2 (Magurran 2013)) that indicated satisfactory sampling effort for both private lands and strictly protected areas. Furthermore, we chose β diversity metrics that focused on compositional changes to alleviate the contribution of α diversity and eventual uneven sampling effort (Koleff et al. 2003). With those considerations, we are confident that our results are consistent and would not be much different from what we have shown here.
Our study highlights the complexity and idiosyncrasies of community compositional shifts in hyper-fragmented landscapes, where both homogenization and heterogenization processes were detected, with the latter prevailing as an overall trend, especially in non-protected private lands. From an applied perspective, the implication of biotic homogenization or heterogenization alone is not sufficient to underpin conservation strategies, as its interpretation is not straightforward – human disturbances can cause β diversity to increase, decrease or remain unchanged (Socolar et al. 2016; Olden et al. 2018). However, the heterogenization process in our study is coupled with a scenario where (i) the originally diverse vegetation is now extremely reduced and fragmented, with small and isolated patches distributed within an intensive agricultural matrix; (ii) forest fragments, especially in private lands, represent unique disturbance histories that result in varying quality habitats, and often in reduced local diversity derived from a large regional species pool (γ diversity); (iii) community composition accumulate great variation among patches (high β diversity), predominantly from turnover (i.e. replacement of species). Bringing these facts together and recognizing that conservation of biodiversity extends far beyond the boundaries of strictly Protected Areas, we advocate that all forest fragments are valuable for conservation in HMLs, with particular relevance for private lands, which represent the most exposed and neglected share of what is left (Gardner et al. 2009; Mendenhall et al. 2016; Farah et al. 2017). Based on our results and supported by many other studies, we understand there is enough information to develop an evidence-based approach that should be considered in future management and conservation plans. To foster and sustain biodiversity conservation in HMLs, we thus recommend: (i) effective protection of strictly Protected Areas, which usually represent the largest regional core areas (Joppa et al. 2008) and where compositional shifts apparently are more stable; (ii) active restoration of forest fragments to enhance their alpha diversity, through the management of hyper abundant species (e.g., lianas) (César et al. 2016; Estrada-Villegas & Schnitzer 2018) and reintroduction of lacking groups of species (Garcia et al. 2014; Viani et al. 2015) (iii) active restoration of corridors where the vegetation is degraded and natural regeneration is unlikely, aiming to enhance forest cover and connectivity among forest fragments, allowing species to disperse and persist (Howe 2014; Emer et al. 2018). Finally, considering the growing development of more sustainable agricultural practices (Ferreira et al. 2012, Gonthier et al. 2014) and alternatives for ecological restoration with profitable purposes (Pedro H. S. Brancalion et al. 2012), we encourage the establishment of policies that foster a feasible production model, aligned with the conservation of the remaining biodiversity.
Author contributions
‘CYV conceived the idea with considerable inputs from RRR, MFS and TS. CYV, MFS and DSBR performed data compilation; MFS and DSBR modeled species distributions; CYV and TS analyzed the data and led the writing of the manuscript; all authors edited the drafts and gave final approval for publication.
ACKNOWLEDGEMENTS
This study was financed in part by the Coordenação de Aperfeiçoamento de Pessoal de Nível Superior – Brasil (CAPES) – Finance Code 001, by the National Council for Scientific and Technological Development (CNPq grant #140545/2016-6 and #561897/2010-7) and by The São Paulo Research Foundation (FAPESP grant # 2013/50718-5 and #1999/09635-0).