Most of the remaining forests in the tropical region is restricted to steep areas, hindered by human occupation, agriculture, and represented by small isolated remnants (Metzger et al., 2009, Ribeiro et al., 2009), which are highly disturbed (Ribeiro et al., 2009) and embedded in agricultural landscapes (Melo et al., 2013). Typically, such landscapes are an agricultural mosaic, containing forest fragments with different degradation histories and drivers, that usually include (i) old-growth forest remnants that have never experienced clear-cutting, but are still degraded by anthropogenic disturbances, (ii) secondary forests that are regenerating after a clear-cut, fire or the abandonment of croplands and pasturelands (Gardner et al., 2009; Molin et al., 2017), and (iii) secondary forests under recurrent events of disturbance (Molin et al., 2018; Reid et al., 2018).

Recent studies have reported the importance of such secondary forests to conserve biodiversity, supply ecosystem services, provide stepping stones, corridors, fauna and flora refuges, and other key conservation elements (Arroyo-Rodríguez et al., 2009; Brancalion et al., 2012; Chazdon et al., 2009; Viani et al., 2015). Nonetheless, different recurrent disturbances play important roles in forest succession when defining the plant community (Arroyo-Rodríguez et al., 2015). Disturbances such as wind, fire, temperature increases and exotic species invasion are important components of many ecosystems, and variations in the disturbance regime can affect an ecosystem's structure and functioning (Hobbs and Huenneke, 1992). The effects of disturbances can be expressed at different levels. Within fragments, environmental (soil conditions and light availability) and ecological filtering (demographic stochasticity, seed dispersal limitations, competition, recruitment, etc.) can alter the species composition and the spatial patterns of plant regeneration (Wright, 2010), by selecting species and defining future biodiversity (Arroyo-Rodríguez et al., 2015; Wills et al., 2006; Wright, 2010). For established woody plants, time since disturbance can play an important role on defining species composition, considering time-lag response that can lead to an extinction debt (Tilman et al., 1994). Such effects on both strata can lead to decreasing local diversity or to biotic homogenization at the local level (Püttker et al., 2015). From a regional perspective, disturbances have been observed to increase the compositional and functional differentiation among forest remnants, especially in highly deforested landscapes with long land use histories (Arroyo-Rodríguez et al., 2015). On the other hand, tropical forests under severe landscape disturbances can also be susceptible to floristic convergence (i.e., species homogenization) (Lôbo et al., 2011) driven by the proliferation and dominance of some fast-growing plant species, such as pioneers and vines (Arroyo-Rodríguez et al., 2015; Liebsch et al., 2008; Lôbo et al., 2011; Melo et al., 2007). Both floristic differentiation and homogenization are dependent on the landscape configuration and spatial scale of analysis (Arroyo-Rodríguez et al., 2013).

Landscape characteristics (e.g., amount of vegetation and matrix context) are known to heavily influence population and metapopulation dynamics (Püttker et al., 2011; Tscharntke and Brandl, 2004). Fragmentation, habitat loss or the integration of both mechanisms has led to several changes in landscape structure and configuration (Fahrig, 2003), driving plant and animal populations to reductions or even to extinctions at the landscape scale in tropical forests (Benchimol et al., 2017; Lima and Mariano-Neto, 2014; Morante-Filho et al., 2015). Species diversity in several tropical plant communities has been strongly and positively related to forest cover and connectivity.

The degree of change in species diversity along with a range of different scales (e.g., samples, habitats or communities) is measured by beta-diversity (β), since it provides a direct link between biodiversity at local (α diversity) and regional scales (γ diversity) (Magurran, 1988). In the Atlantic Forest, some studies have addressed patterns of beta diversity in native vegetation (Carneiro et al., 2016; Machado et al., 2016; Sfair et al., 2016; Bergamin et al., 2017; Benchimol et al., 2017), which remains a crucial field in understanding the patterns related to plant community assembly and its maintenance in highly degraded landscapes. The landscape-divergence hypothesis, postulated by Laurance et al. (2007), states that differential human disturbances among forest remnants, such as the different degrees of reduction of local richness (patches) after habitat loss, can lead to differential trajectories of vegetation succession. Those disturbances may increase the compositional and functional differentiation of communities in fragmented landscapes, leading to high beta diversity between patches in a landscape (Benchimol et al., 2017; Sfair et al., 2016). Along with understanding communities' structures along landscapes, it is also important to understand the historical transformations that habitat patches and biodiversity have been submitted to (Molin et al., 2017). This is particularly important in dynamic landscapes, where deforestation, regeneration, harvesting, grazing, urbanization, and other actions have shaped the distribution of these patches throughout the landscape, playing an important role in selecting species (Arroyo-Rodríguez et al., 2015).

Even with the increase in the number of studies about the Brazilian Atlantic Forest, little is still known about its biodiversity. Less than 1% of the remnant area has been sampled, and most of the current knowledge about this domain comes from private lands and areas outside of forest reserves (Lima et al., 2015). In addition, our understanding of the biological dynamics of such human-modified ecosystems (Arroyo-Rodríguez et al., 2013) has been inadequately studied. Therefore, the combination of beta-diversity analyses with different landscape contexts can inform us about the fundamental mechanisms shaping the remaining diversity. Given such importance, this approach must be investigated in the most varied landscapes possible.

Here, we investigated several ecological processes interacting collectively at multiple landscape contexts and scales of analysis. To discuss these issues, we assessed the floristic composition and diversity of forest remnants embedded in human-modified landscapes which are highly fragmented with a dynamic history of occupation. Our specific goals were (1) to evaluate the general patterns of α-plant diversity of the superior and inferior strata among forest remnants in different periods of regeneration, and among sugarcane and pastureland matrices; (2) to analyze the β-plant diversity patterns at different scales (among forest remnants, among landscapes, and among agricultural matrices); and (3) to assess the influence of landscape structure and dynamics (present and past events) on plant species diversity. Considering the studied forest remnants are embedded in different matrices, isolated from each other, and have been experiencing different disturbance histories over time, we would expect a plant community differentiation among forest remnants, among landscapes and among agricultural matrices resulting in higher β-plant diversity. In contrast, we would expect a lower α-plant diversity within forest remnants (among plots) as a result of the limited seed dispersal inter-forests imposed by the landscape structure.

Previous studies have shown, and our results support, that there is an important plant diversity and richness that remains in the agricultural and deforested landscapes with recurrent impacts on biodiversity (Magnago et al., 2014; Machado et al., 2016; Farah et al., 2017). Plant diversity differed among agricultural matrices, probably as a response to the different disturbance regimes of pasturelands and sugarcane fields. Forest remnants in pasturelands were overall more dynamic, where deforestation and regeneration occurred more frequently over the past few years (Ferraz et al., 2014), and therefore, the highest habitat heterogeneity could also explain the highest diversity (Ricklefs, 1977). The inferior stratum was richer and more diverse than the superior, due to the presence of plants from different growth forms on the understory (shrubs and small trees). Furthermore, in the inferior stratum, diversity varied among agricultural matrices, being higher in the pastureland matrix. Thus, the matrix is likely to be important to forest dynamics. For example, the matrix will often act as a selective filter for species’ movement across the landscape. The type of vegetation, height, or management (e.g., use of herbicides, pesticides, fire and fencing to restrict livestock transit) in the matrix will determine the permeability level for movements of individuals (da Silva et al., 2015). Characteristics of the surrounding matrix can strongly influence the plant regeneration patterns in forest remnants (Grau, 2004). The presence of isolated tress and the scattered distribution of forest remnants, common in pasturelands, are proven to facilitate seed dispersal (Chazdon and Guariguata, 2016; Holl et al., 2016; Toh et al., 1999). Our results support that pastureland matrices seem to provide better conditions for plant regeneration in similar forest edges when compared to sugarcane matrices.

Although individuals from superior stratum are present in lower densities when compared to other fragments of Tropical Semi-deciduous Forests (Ivanauskas et al., 1999), this species pool and high individual density per hectare suggests that these forest remnants still present resilience and therefore may be able to persist in the landscape. The high number of animal-dispersed plant species, as adults and regenerating, suggest that seed dispersal processes are still happening in such anthropic landscapes. However, the local population persistence of such tree species is firmly related to the effective conservation of their frugivorous bird communities responsible for complementary functional processes (Rother et al., 2016). The process of defaunation of large-bodied dispersers observed in human-modified landscapes remains one of the main issues to be addressed and limits the potential persistence of large-seeded plant species in the study sites and other tropical regions (Cramer et al., 2007; Dirzo et al., 2014; Melo et al., 2006).

Previous studies have also shown a great influence of agricultural uses on plant diversity, ecological processes such as seed dispersal and recruitment, and ultimately, forest resilience (Herrera et al., 2010; Jakovac et al., 2015). It is important to recognize that our plots were located near forest edges, where these impacts are stronger, and diversity is expected to be low (Murcia, 1995). Our results showed that diversity is low at the plot level, which might be explained by a myriad of factors that influence plant diversity at different scales. At the plot level, there are natural processes such as environmental filtering (soil conditions and light availability) and ecological filtering (demographic stochasticity, seed dispersal limitations, competition, recruitment, etc.) that have a considerable influence on selecting species and defining biodiversity (Arroyo-Rodríguez et al., 2015; Wills et al., 2006). In degraded environments, these processes are strengthened by anthropogenic disturbances, such as edge effects, competition with hyper abundant native species, mainly vines in our case (Schnitzer and Carson, 2010; Tabarelli et al., 2012), and regenerating species selection due to cattle pressure (Griscom et al., 2009). These disturbances have been shown to decrease diversity, thus leading to biotic homogenization at the local level (Püttker et al., 2015). Disturbances at plot level are even more important as our results showed that plots with eight years of regeneration are more diverse than plots with 13 or 30 years of abandonment. Although previous studies show an increase in forest cover in recent years for other tropical regions along the Corumbataí River basin (Ferraz et al., 2014; Molin et al., 2017), this coincides with a decrease in forest quality, because old-growth forests are being replaced by secondary forests in many stages of succession. This result can be explained by two observations. Firstly, there is an urgent need for the conservation and restoration of old-growth forest patches in human-modified landscapes. These patches are irreplaceable in biodiversity conservation (Gibson et al., 2011), particularly in the protection of groups of species sensitive to fragmentation and anthropogenic disturbances, such as large-seeded species (Melo et al., 2007), endemic species (Liebsch et al., 2008) and rare species (Mouillot et al., 2013). The removal of degradation drivers and restoration of these patches could reverse the trajectory of degradation and increase their potential for housing such sensitive groups (Vidal et al., 2016). Secondly, this result reinforces the importance of forest regeneration for conserving biodiversity at the landscape scale (Latawiec et al., 2016). In our study sites, young forest regeneration accounted for the greatest diversity within communities at the plot level. Natural regeneration has also been included as one of the main methods to restore degraded and open areas in large-scale projects, which would increase forest cover in highly fragmented landscapes (Chazdon and Guariguata, 2016). Together with old-growth forests (and not by replacing them), young forest patches and disturbed patches increase habitat cover for native species, increase habitat heterogeneity at the landscape level, and enhance landscape connectivity, which enhances the potential for biodiversity conservation of human-modified landscapes (Tabarelli et al., 2010).

It is interesting to point out that different matrix types may influence the forest fragment edge(Laurance and Yensen, 1991; Mesquita et al., 1999). The higher diversity observed in eight year plots in the pastureland matrices shows that matrix quality clearly affects plant dynamics and dispersal across the landscape. Pasturelands are widely considered to be a hostile mode of land use (Peres et al., 2010). However, sugarcane is often considered even more detrimental due to the high level of degradation caused by management practices, such as burning the adjacent forests, since sugarcane leaf biomass was typically burned before harvesting in the dry season up until recently (Martinelli and Filoso, 2008). This could explain the differences we observed for species diversity in eight year plots compared to 13 and 30 year plots occurring mainly in sugarcane matrices. Betadiversity was higher than expected by chance for total diversity among plots, among time of regeneration and among matrices, and in most cases for species richness (except among plots). This is consistent with previous studies in other tropical regions (Arroyo-Rodríguez et al., 2013; Machado et al., 2016). In our study area, the impermanence of secondary forests together with recurrent human pressure, implies different successional pathways for such plant communities. Each trajectory implies different forest assemblages, which allows the accumulation of a higher number of species at the landscape level (Arroyo-Rodríguez et al., 2015; Püttker et al., 2015). Essentially, as proposed by Tscharntke et al. (2012), the landscape-moderated dissimilarity of local communities (alpha) determines landscape-wide biodiversity (beta) and overrides any negative local effects of habitat fragmentation on biodiversity.

This understanding of biodiversity in fragmented habitats has a substantial influence on management strategies and the conservation role of forests under the distinct age of regeneration in human-modified landscapes. Previous studies have shown, and our results also support, that these disturbed forests have low plant diversity and richness, especially when compared to large tracts of pristine forests (Santo-Silva et al., 2015). On the other hand, all remaining and regenerating forests together still preserve high diversity at the landscape level (i.e., high beta diversity) even when compared to large conservation units (Farah et al., 2017), and not only for plants but also for other biological groups, such as large-bodied mammals (Beca et al., 2017). In other words, in human-modified landscapes, each forest regardless of its size has an immense value to the conservation of overall biodiversity, and the discussion regarding the conservation role of secondary forests should be migrated from the local to the landscape level (Arroyo-Rodríguez et al., 2015, 2009; Bergamin et al., 2017). Our results reinforce the need for secondary forests to be prioritized in future conservation and restoration efforts at the landscape scale to improve their persistence (Reid et al., 2018).

Notwithstanding, it is not often clear what the best strategy is to conserve remnant biodiversity in highly modified landscapes. Previous studies have reported that in landscapes with a low amount of forest patches, increasing connectivity is ineffective, whereas at low connectivity increasing high patch numbers is useless (Johst et al., 2011). Others have found that 30% of habitat cover is the turning point for species occurrence and the persistence of functional groups sensitive to fragmentation (Estavillo et al., 2013; Pardini et al., 2010). Therefore, landscapes below this threshold of 30% should not be targeted as priority areas for restoration, due to the low cost-effectiveness of these projects (Banks-Leite et al., 2014; Tambosi and Metzger, 2013). Nonetheless, each biological group has a different response to fragmentation thresholds and landscape structure, and for plants, regenerating individuals tend to reflect the current quality and configuration of the landscape, while adults are able to remain in the landscape for a very long time following habitat loss (Rigueira et al., 2013).

In the Atlantic Forest, 88% of forest cover is present in landscapes below the theoretical biodiversity thresholds (Banks-Leite et al., 2014), and therefore would be in danger of not being targeted for restoration programs. In addition, landscapes with different anthropic occupations present different distributions of patches and different historical dynamics. In our study sites, the influence of landscape structure and dynamic metrics on plant diversity differed among agricultural matrices, suggesting that, in highly fragmented landscapes, only the percentage of forest cover is not enough to define restoration and conservation strategies. In our study region, sugarcane landscapes present larger and less connected forest remnants, whereas, in pastureland landscapes, forest remnants are mostly connected but usually smaller and linear (Ferraz et al., 2014). Therefore, different strategies should be applied to each landscape to conserve biodiversity. In landscapes with low forest cover and highly dynamic histories, it might be more effective to increase biodiversity persistence by decreasing patch destruction than by increasing patch creation (Johst et al., 2011), so it is crucial that the remaining biodiversity should be maintained through conservation and restoration of current forests. This strategy must be considered to implement new protected areas, therefore increasing the number of conservation units, along with the assistance of financial incentive programs and specific legislation since biodiversity is spread in forest remnants in the landscape (Bergamin et al., 2017). Although nowadays there are legal instruments that can increase landscape connectivity through the restoration of riparian areas, very few require an increase in quality of the actual remaining forests and prevention of within-forest disturbances (Brancalion et al., 2012; Viani et al., 2015; Barlow et al., 2016; Rother et al., 2018).