Forest regeneration may reduce the negative impacts of climate change on the biodiversity of a tropical hotspot

2.1 Natural forest regeneration models and scenarios of landscape restoration

We first employed a model of potential natural forest regeneration based on the seed dispersal from forest patches into the matrix promoted by the movement of frugivorous animals (Niebuhr, 2018; Figures 1 and 2). After that, we projected the potential distribution of a set of selected species of forest birds and arboreal mammals (primates and sloths) in four different scenarios that considered both current and potential future conditions of climate and landscape. The two climate conditions were (1) the current and (2) a future climate scenario based on the expected climate change, while the two landscapes were (a) the current configuration of the BAF native forest remnants and (b) our scenario of BAF regeneration. We choose the taxonomic groups for which we predicted distributions because they are considered good indicators of habitat quality and are of conservation concern in fragmented environments (Arroyo-Rodríguez & Fahrig, 2014; Pizo & Tonetti, 2020), thus serving as environmental indicators of priority areas for biodiversity conservation.

The theoretical models of potential natural forest regeneration proposed here are based on three main ecologically rooted assumptions extracted from reviews and recent large-scale studies relating landscape ecology, seed dispersal, and natural regeneration. Our main assumption is that, since most of the seed dispersal is led by animals in the tropics (Corlett, 2017; Howe & Smallwood, 1982), places which receive a larger load of seed rain have the higher probability of natural regeneration, mainly degraded and open areas that are abandoned or subject to less intense human management (Molin et al., 2017; Silva et al., 2017). Second, we assume that the richness and abundance of the main frugivorous animals (e.g. bats, birds, non-flying mammals, reptiles and even invertebrates) increase within larger forest patches (Böhning-Gaese, 1997; Bovo et al., 2018; Kirika et al., 2008; Magioli et al., 2015; Figure 2). Accordingly, the higher the richness and abundance of frugivorous animals, the higher the probability that they spread seeds around forest patches and for long distances (McConkey et al., 2012). Third, we assume that seed dispersal decreases as the distance of forest patches increases (Bullock et al., 2018; Camargo et al., 2020; Nathan & Muller-Landau, 2000; Figures 1 and 2). In combination, these assumptions mean that sites closer to larger forest patches have higher seed deposition, which leads to a higher potential for natural regeneration.

To operationalize these premises, we first classified BAF forest patches into five size classes, corresponding theoretically to different functional assemblages of seed dispersers (Table S1). The size classes were (1) up to 10 ha (<10), (2) between 11 and 25 ha, (3) between 26 and 50 ha, (4) between 51 and 250 ha, and (5) higher than 250 ha (Figure 2). Species abundance and diversity thresholds in tropical forests are generally assessed in terms of habitat amount in the landscape (e.g. Muylaert et al., 2016; Tambosi et al., 2014), but hardly using patch size and connectivity. In BAF, there is evidence that forest patches separated by size thresholds of 60 and 2050 ha present different functional assemblages of mammals (Magioli et al., 2015).

________________________________________________________________________

Read Also : Forests the size of France have regenerated over past 20 years

________________________________________________________________________

We selected the above-mentioned patch size classes to allow for variation within the seed dispersal process derived from the smaller patches (<250 ha), which comprehend 95% of the forest patches in BAF (Ribeiro et al., 2009), and considering that a great part of the seed dispersal that leads to regeneration is driven by small bodied species (Galetti et al., 2013). We then created curves of potential natural forest regeneration around these patches using an exponential function which assigns a decreasing chance of forest regeneration as one walks away from the edges of forest patches. The exponential distribution is defined by the scale parameter (λ), which determines the decreasing rate of the curves in relation to the distance of forest patches, varied among the different size classes of forest patches, according to an expected scale of mobility of the disperser animals (Table S1). Even though the decrease of these curves alone is monotonic, when the curves of potential regeneration around multiple patches in different landscape arrangements are combined, complex spatial patterns of potential natural regeneration are created by our models (Figure 2).

Continuous values for the chance of seed dispersal (and consequently forest regeneration) were estimated for each pixel in the landscape, and they decrease as forest fragment distance increases (Figure 2). Although these values may be estimated for virtually any type of matrix, we established scenarios of forest regeneration only for pastures because BAF regeneration is more likely to occur in these areas (Latawiec et al., 2015). Pastures are more likely to be abandoned than other types of land use, allowing BAF and other tropical forests to regenerate (Rezende et al., 2015). Pasture map used in our analysis was developed by the Processing and Geoprocessing Lab (LAPIG; https://www.lapig.iesa.ufg.br/lapig/; Table S2; available in https://pastagem.org). Since this is an initiative to map Brazilian pastures, we considered only the Brazilian portion of Atlantic Forest. We established a threshold of 0.5 in the continuous values of forest regeneration to create a scenario of landscape restoration, that is, we assumed that all pixels with a potential natural regeneration of >= 0.5 would be converted to forests (Figures 1 and 2). Although arbitrary, with this threshold value our amount of forest regeneration is similar to the one proposed by the Atlantic Forest Restoration Pact. A chance of natural regeneration below 0.5 was considered too low for the analysis (see below; Crouzeilles et al., 2019; Melo et al., 2013).

2.2 Study area

Extending from the Brazilian coast (from north-eastern to southern Brazil) to the inner southern parts of the continent, the Atlantic Forest is the second largest tropical forest domain in South America, and is mostly distributed in Brazil (99% of its area; Muylaert et al., 2018). Despite its high species richness and endemism, the Atlantic Forest was severely devastated in the last centuries, and only 28% of its native vegetation persist (Rezende et al., 2018; Figure 3). Moreover, most (80%) native forest patches are small (<50 ha) and distant from each other (1440 m in average; Ribeiro et al., 2009). The high latitudinal and altitudinal gradient of the Atlantic Forest causes great variation in climate and vegetation and its domain can be divided into different biogeographic regions (Figure 3). Due to limited availability of pasture data (see above), only the Brazilian portion of the Atlantic Forest was considered in our analyses (BAF; Figure 3).

High species richness and endemism associated with high devastation rates make the domain a priority region for biodiversity conservation (Myers et al., 2000; Pizo & Tonetti, 2020). Because forest remnants are fragmented and surrounded by different land cover types, BAF is an appropriate region for evaluating how natural forest regeneration may promote biodiversity in human-dominated environments. BAF delimitation and forest remnants considered in this study (Figure 3) were processed by the Spatial Ecology and Conservation Lab at the São Paulo State University, Brazil (LEEC – UNESP) in an association with Agroicone, a Brazilian non-governmental organization (https://www.agroicone.com.br/) and are detailed in Table S2.

2.3 Species distribution models

We projected the potential distribution of forest birds, primates, and sloths in BAF via SDM approach. SDM identifies areas with different suitability for the occurrence of species based on predictor variables (Guisan et al., 2017). We employed two categories of predictor variables – (1) climatic and (2) landscape. Species distribution at biogeographic scales (e.g. within a domain) is usually determined by climatic variables, while species occurrence at mesoscales, within a smaller regions, such as watersheds, are influenced by landscape ones (Peterson et al., 2011). We modelled species distribution separately based on climatic and landscape predictors and identified areas in which both predictors indicated suitability to species occurrence (Sobral-Souza et al., 2021; Figure 4).

Several SDM studies that predict species occurrence under future scenarios assume static conditions of landscapes, which may be unrealistic as land cover is usually dynamic (Martin, Van Dyck, et al., 2013). Some investigations have shown that considering different landscape scenarios in SDM may result in more accurate predictions (Regos et al., 2018; Visconti et al., 2016), probably because the different predictors (climate and landscape) that influence the distribution of species operate at different scales (Sobral-Souza et al., 2021).

To build the climate SDM we used the 19 WorldClim 1.4 variables with 30-arc seconds resolution (~1 km). These variables are also known as ‘bioclimatic’ as they are biologically meaningful and derived from temperature and precipitation data collected in weather stations around the world (1960–1990 temporal range, https://www.worldclim.org/data/v1.4/worldclim14.html). We considered bioclimatic variables within BAF delimitation and, to avoid correlation among then, we applied a Variance Inflation Test (VIF) using the USDM package in the R environment (Naimi et al., 2014). From the 19 variables, we eliminated those with VIF values ≥5 (Guisan et al., 2017), and used the remaining five, namely: mean diurnal range (bio 2); mean diurnal range/temperature annual range (bio 3); mean temperature of the wettest quarter (bio 8); precipitation of the wettest month (bio 13) and precipitation of the coldest quarter (bio 19).

In addition to the current climate, we projected the potential distribution of species for 2070 using the 6.0 representative concentration pathway (RCP). The emission of greenhouse gases is simulated by RCP scenarios according to different trends of human activity (IPCC, 2014). We opted for using an intermediated RCP scenario over the scenarios of highest (RCP = 8.5) and lowest gas emission (RCP = 2.6), as both are unlikely to occur and should not be used in environmental planning studies (Hausfather & Peters, 2020). To reduce the effect of uncertainty in the different general circulation models of the RCP scenario, we projected species distributions for three different models – MIROC ESM, HadGEM2-AO, and GFDL-ESM2G and averaged environmental suitability of these three models. We selected these models following Sanderson et al. (2015), that suggest choosing the models least related to each other.

To build the landscape SDM, we calculated the following landscape variables based on the BAF remnant map (30-m resolution): habitat amount, edge amount (edge depth = 30 m), functionally connected area and patch size. The functionally connected area is the sum of the area of all patches closer than a gap-crossing distance – here we considered the value of 60 m. We used these variables because they are widely known to influence the occurrence of vertebrates (Baguette & Van Dyck, 2007; Fahrig, 2013; Pfeifer et al., 2017), and the value of edge depth and gap crossing were chosen based on the literature (Graham & Socha, 2019; Uezu et al., 2005). With the increase of forest amount in the regeneration scenario, it is expected that the values of the four landscape variables will also increase. Despite the relative value of edge amount in relation to habitat amount may diminish, its absolute value is also expected to increase (Figure S1). These variables were calculated for each landscape scenario (current and potential regeneration) with the aid of LSMetrics (Niebuhr, 2018; Niebuhr et al., 2020; https://github.com/LEEClab/LS_METRICS), a package designed to calculate landscape metrics in the GRASS GIS environment (Neteler et al., 2012).

For both habitat and edge amount, we considered a moving window of 2 km (i.e. for each central pixel we considered an average value of all pixels within a 2 km-distance). This distance was based on the scale at which the landscape may affect the occurrence of the studied species (Barros et al., 2019; Gestich et al., 2019). As elevation has proven to be an important variable in predicting species distribution (Guisan et al., 2017), we added this layer to the landscape variables and, after applying a VIF test (threshold value = 5), we removed edge amount to avoid high correlation among variables. After calculating landscape variables using 30-m pixel size, we resampled the data to the same resolution of bioclimatic variables (~1 km), a suitable resolution for SDM. We calibrated the SDM for the current landscape layers, in which species are currently distributed, and projected for both the current landscape and the potential regeneration scenario.

________________________________________________________________________

Read Also : Yoga Vs Gym – Which you Should Opt for and Why?

________________________________________________________________________

We downloaded species occurrence data with the aid of the spocc package in the R environment (Chamberlain, 2020) from the following databases: Global Biodiversity Information Facility (GBIF; https://www.gbif.org); iNaturalist (https://www.inaturalist.org); VertNet (http://vertnet.org); eBird (https://ebird.org) and Integrated Digitized Biocollections (iDigBio; https://www.idigbio.org). We also considered primates and sloths occurrences in datasets of the Atlantic Forest region (Culot et al., 2019; Santos, Bocchiglieri, et al., 2019). We only considered primate and sloth data locations assigned as ‘precise’ or with a geographic accuracy of 1 km according to the authors (Culot et al., 2019; Santos, Bocchiglieri, et al., 2019). We excluded occurrence data recorded prior to 1980 to avoid mismatching between species presence and landscape features. The literature suggests that BAF forest remnants suffered no substantial changes in most of the domain from the 1980s onward (Ribeiro et al., 2009). Endemic birds considered in this study are listed in Vale, Souza, et al. (2018) and Vale, Tourinho, et al. (2018), non-endemic ones in Hasui et al. (2018), primates (endemic and non-endemic) in Reis et al. (2015), and sloths in Santos, Bocchiglieri, et al. (2019). We only considered species that have forest as main habitat type according to IUCN (https://www.iucnredlist.org) and included taxonomical synonyms in our searches.

To avoid spatial correlation and model overfitting, we spatially rarefied points keeping a distance of at least 10 km between them with the aid of the spThin package in R (Aiello-Lammens et al., 2015; Guisan et al., 2017; Peterson et al., 2011). To avoid potential location errors, we discarded species occurrence near biodiversity institutions, urban areas, state capitals and with zero decimal values for latitude and longitude with the aid of the package CoordinateCleaner in R (Zizka et al., 2019). Moreover, we plotted species occurrence and searched for potential outliers and records that failed to correspond to the expected distribution of the species based on our knowledge and on databases such as IUCN and Wiki Aves (https://www.wikiaves.com/). After these filtering procedures, we only modelled species with at least 10 occurrence points. We assumed this to be the minimum number of occurrences to sample a gradient of environmental variability adequate to estimate areas environmentally suitable for the species studied (Peterson et al., 2011).

As the study area in SDM should ideally consider the area accessible for the organisms (Barve et al., 2011), for each species we used a background region based on a minimum convex polygon of the occurrence points (comprising geographical extreme occurrences) and summed by a buffer of two coordinate degrees (~220 km). We then cropped variables for each species using the convex hull and sampled 5000 background points. We considered these types of backgrounds more adequate than using the entire BAF for all species, which could result in model overfitting and the overestimation of model evaluation (Barve et al., 2011; Guisan et al., 2017; Peterson et al., 2011). This approach is similar to other analyses employing SDM in BAF (Sales et al., 2021).

As there is no single best algorithm for use in SDM (Qiao et al., 2015), we combined the prediction of four different algorithms of presence-only data: bioclim, domain, random forest and maxent via the sdm package in R (Naimi & Araújo, 2016). For each species we ran 10 replicates of each algorithm assigning 70% of the occurrence points to training the models, and 30% to evaluate each run via bootstrapping (Peterson et al., 2011), resulting in 40 models per species for climate variables (four algorithms, 10 replicates) plus 40 models for landscape variables. Ensemble projections for the 10 replicates were performed via a weighted mean for the individual models based on a threshold that maximizes specificity +sensitivity considering the True Skill Statistic test (Allouche et al., 2006). We used the lowest presence-threshold in the raster of continuous values from the ensemble of the 40 individual models to create binary maps for each species (presence/absence; Liu et al., 2016). Then, we combined maps of climatic and landscape suitability and analysed regions where both landscape and climate were suitable for species occurrence (Figure 4).

After training the models for the current scenarios of climate and landscape, we projected the SDM for both current and future climate, current landscape configuration and future potential landscape regeneration. We added the binary maps of the potential distribution of species and identified areas with different potential species richness and suitable areas for the occurrence of species in the following scenarios: (1) current climate and current landscape, (2) current climate and future regenerated landscape, (3) future climate and current landscape and (4) future climate and future regenerated landscape. To calculate the area of occurrence of species in each scenario, we considered suitable areas only within forest patches. As forest patches may be small, to facilitate the visualization of species richness in different regions, we calculate the mean values of potential species richness considering only forest patches in hexagonal grids (~1000 km² hexagons, bottom line in Figure 5; see the potential species richness in forest fragments in Figure S2).

2.4 Assessment of extinction risk

Based on the relation between geographical range and probability of extinction, we evaluated the percentage of species that would become extinct in the future in the two different BAF landscape scenarios (current and with regeneration). We evaluated extinction risk assessment based on an equation developed by Thomas et al. (2004), which takes into account the average proportional loss of distribution area for each species: E = 1 – {(1/n) [∑(A_fut/A_original)]}^z in which E is the proportion of extinct species, n, the number of species, A_fut, the future occupied area, A_original, the original occupied area and z, a constant that reflects the proportion of species overlap between areas (Tjørve & Tjørve, 2008). Based on the literature, we set z = 0.25 (Brooks & Balmford, 1996; Thomas et al., 2004).

________________________________________________________________________

Read Also : Climate Change Driving Forests To Smaller And Younger Trees

________________________________________________________________________

4.1 Natural forest regeneration and biodiversity

Despite the possible negative impacts of climate change on the distribution of 67% of the studied species, our analysis indicated that, as expected, promoting natural forest regeneration at large scales may not only prevent biodiversity loss but also increase the potential distribution area of species that may be impacted by climate change in future scenarios. The 252 species that might be negatively impacted by climate change may suffer a reduction of 74.4% (SD = 28.7%) of their potential distribution area when considering future climate and current landscape scenarios. However, if our natural regeneration scenario is considered, these same species might increase their potential distribution area by 75.3% (SD = 9.3%) in the future climate scenario (Figure 6). From these 252 species, except for six species that may zero their potential distribution areas with climate change regardless of the landscape scenario, the remaining species might increase their area of occurrence with forest regeneration. Several tropical forest species of birds and arboreal mammals, including endemic species, tolerate fragmented and secondary forests (Develey & Phalan, 2021; Pizo & Tonetti, 2020), which suggests that the regenerated environments could be used for many of the studied taxa (but see Santos, Bailey, et al., 2019).

Thus, as expected, with the increase of forest amount in the regeneration scenario as modelled here, it may be possible to reduce the proportion of species that would become extinct because of future climate change. In an assessment of the risk of extinction of Atlantic Forest birds due to deforestation, endemic species were the most vulnerable (Brooks & Balmford, 1996). Indeed, the potential loss of area of distribution due to climate change may also increase the extinction risk in different environments around the world, including other tropical forests (Thomas et al., 2004). Loss of species may have deleterious cascading effects on biodiversity. Some studies have identified that endemic large-bodied birds (Vidal et al., 2014) and primates (Bufalo et al., 2016) are key elements involved in seed dispersal in tropical forests. Therefore, the loss of such large-bodied frugivores in BAF because of climate change may pledge seed dispersal systems with consequences for the dynamics of forest ecosystems (Sales et al., 2021).

Our models indicated that regions with higher forest amounts (such as Serra do Mar, Araucaria, and northern Bahia; Figures 3 and 5) have higher chances of natural regeneration and, thus, sustain a higher number of species, both in the present and future scenarios. Other studies found similar results (de Souza et al., 2011; Santos, Feltran-Barbieri, et al., 2020; Santos, Sobral-Souza, et al., 2020; Vale, Souza, et al., 2018; Vale, Tourinho, et al., 2018; Zwiener et al., 2017). When considering models based on landscape predictors only, species richness was higher in regions where landscapes have higher amounts of forest (Figure 4; Figure S4). These results were expected as the studied taxa are forest-dwelling species (Pizo & Tonetti, 2020; Reis et al., 2015; Santos, Bocchiglieri, et al., 2019).

Contrary to areas with high forest amount, regions where forest patches are small and distant to each other may show lower chance of forest regeneration and sustain lower potential species richness according to our models. Although some may presume that the higher the species richness the higher is the importance of an area for conservation, this must be interpreted with caution. In our study, some severely devastated regions that presented comparatively fewer potential species richness in all scenarios are relevant for the conservation of tropical biodiversity, as the Pernambuco Endemism Center in BAF (PEC; Figure 3). PEC is remarkable for its endemism of birds, mammals, and other groups, harbouring several species on the brink of extinction (Develey & Phalan, 2021; Garbino et al., 2018; Pizo & Tonetti, 2020). Therefore, when planning natural forest regeneration in large scales based on seed dispersal, it might be important to use restoration techniques beyond passive restoration to increase the amount of forest in landscapes with low amount of remaining forest (such as the PEC). These techniques may comprise the mixed restoration approach – that is, when less human input than used in active regeneration is employed. This has lower costs when compared to active regeneration and may increase the chances of regeneration and decrease the time needed for new forest formations to be established (Brancalion et al., 2016).

4.2 Forest restoration as an auxiliary measure of biodiversity conservation

Although secondary forests are relevant for the maintenance of species and ecological processes in human-dominated landscapes, older forests are crucial to preserve most of the original biodiversity and as sources of individuals that can establish populations in regenerated areas, especially in fragmented landscapes (Gibson et al., 2011). As in other tropical environments (e.g. Meyfroidt & Lambin, 2008; Smith et al., 2021), evidence shows a process of ‘rejuvenation’ of forests in BAF. Young forests increased at a rate of ~150,000 ha annually since 2000, however the loss of older forests occurred at rate of ~100,000–200,000 ha annually in the same period (Rosa et al., 2021). It is thus important that forest monitoring is carried out properly so that the allocation of areas for natural regeneration does not cause forest loss in other regions.

Regarding climate change, some of the imposed challenges to biodiversity may not be properly mitigated by increasing forest amount only. Population and community responses to altered climate may also drive complex changes on species interactions with consequences difficult to predict (Oliver & Morecroft, 2014), as for example, changes in phenology that can alter frugivory, seed dispersal, and pollination (Scranton & Amarasekare, 2017). With a more direct impact on forest regeneration, global warming is also expected to increase the frequency and intensity of fires in tropical forests, resulting in negative consequences to species persistence and to the growth of early forests (Herawati & Santoso, 2011). Added to the effects of fire, prolonged droughts caused by climate change may hinder seedling recruitment and survival, which may further compromise natural regeneration (Anderson-Teixeira et al., 2013). In sum, despite having several benefits, forest restoration should be seen as an auxiliary measure in the mitigation of the effects of climate change and landscape conversion and should be associated with policies that aim to drastically reduce carbon emissions and deforestation (Holl & Brancalion, 2020). Restoration policies that promote actions to enrich and accelerate the succession rate of secondary and regenerating forests might also be important, to account for the already increased loss of mature forests.

4.3 Tropical forest regeneration and model limitations

The summed area of native forest remnants in our BAF map (~34 Mha; Figure 2) is similar to previous spatial analysis, which identified ~28–32 Mha of native forests in the domain (Rezende et al., 2018; Rosa et al., 2021). Our regeneration scenario indicated an addition of 13 Mha of forests in BAF (Figure 3), an area close to the 15 Mha of native forests that the Atlantic Forest Restoration Pact aims to restore (Crouzeilles et al., 2019; Melo et al., 2013). The total area of pastures in BAF sum ~38 Mha. Therefore, for the amount of regenerated forests indicated in our study to be recovered (~13 Mha), 34% of current pastures need to be set aside for this purpose. Despite the area of regenerated forests in pastures indicated in our models seem excessive, studies suggest that the area of unproductive or underused pastures in BAF is also high – with the consequence that degraded pastures are often the land use class more prone to regenerate into forest (Silva et al., 2017). The cattle stocking rates in BAF are on average 1.17 animals per hectare, considered an extremely low production efficiency (Latawiec et al., 2015). It is estimated that 47% of pastures in Brazil could increase production by twice or even more (Santos, Feltran-Barbieri, et al., 2020; Santos, Sobral-Souza, et al., 2020). Thus, areas that could be allocated for conservation indicated in our regeneration model can be viable if productivity is increased in pastures that have low potential for forest regeneration.

________________________________________________________________________

Read Also : Sustainable Farming and its Relationship with Soil Biodiversity

________________________________________________________________________

Not all pastures that we identified as suitable to be converted into forests are unproductive and, thus, will not be destined for this purpose. Furthermore, although seed dispersal is one of the most important factors to promote forest regeneration, some other aspects may limit natural regeneration such as topography (Rezende et al., 2015; Silva et al., 2017), the past intensity of land use, and the amount of soil nutrient available (Crouzeilles et al., 2016; Mendes et al., 2019; but see Rozendaal et al., 2019). Socioeconomic factors such as the size of rural villages and the possibilities of farm credit might also influence how prone are landowners to convert their lands into crops, what might affect forest regeneration (Silva et al., 2017). Our model of forest regeneration is based on theoretical seed dispersal curves and their response to forest patch size and proximity and, although rooted in the literature, has limitations on its prediction of areas with higher potential of seed deposition and natural regeneration. Data-driven models that use seed dispersal data sampled at landscape-scale or natural regeneration data and take these aspects into account can generate more accurate results. For instance, Crouzeilles et al. (2020) developed a modelling approach based on machine learning algorithms to fit natural regeneration data from remote sensing to environmental and socioeconomic variables. Even though they used more complex models, they concluded that the most important variable to predict natural regeneration at large scales was the proximity to forest (Crouzeilles et al., 2020), which is the main assumption of our models. Therefore, we believe that data-driven approaches might have more reliable predictions locally and may be useful for driving location restoration planning, but do not change from our approach on a biogeographical scale.

Overall, most tropical forests have high potential for natural regeneration (Wright, 2010), which is also the case for the Atlantic Forest (Crouzeilles et al., 2019). There are evidences showing that even tropical regions where past land use intensity were high and forest cover low (~10%), the area of naturally regenerated forests was substantial (Baptista & Rudel, 2006; Brancalion et al., 2016; Crouzeilles et al., 2019; Lira et al., 2012; Mendes et al., 2019; Rezende et al., 2015). Moreover, the amount of degraded area is also large in other tropical regions (Lamb et al., 2005). This suggests that our natural regeneration model may be a viable strategy to passively restore degraded areas in several tropical forest regions and pinpoint regions where natural regeneration may be more successful.

Regarding species distribution, although SDM is an adequate approach to predict species occurrence at large scales, some limitations should be considered. In addition to the intrinsic errors of models to detect all areas potentially suitable for a species indicating otherwise unsuitable areas (Guisan et al., 2017; Peterson et al., 2011), these models may not be adequate in predicting population parameters (Lee-Yaw et al., 2021). The risk of species extinction could be better estimated if variations of population size were assessed (https://www.iucnredlist.org/assessment/process). Moreover, despite considering the effects of landscape and climate on SDM may potentially improve predictions, some limitations need to be recognized, as the fact that climate and landscape might affect species disproportionally depending on the taxonomic group and the studied region (Betts et al., 2019; Newbold, 2018; Vermaat et al., 2017).

Finally, to be successful a landscape regeneration program must include multiple stakeholders, involve local communities, use democratic and inclusive approaches, and consider political issues, so that restoration is economically and socially viable (Díaz-García et al., 2020; Holl, 2017b). Summed to all factors expressed favouring tropical forest regeneration in our study area, the Atlantic Forest Law establishes that rural landowners are not allowed to suppress forest patches at intermediate and advanced stages of regeneration (http://www.planalto.gov.br/ccivil_03/_ato2004-2006/2006/lei/l11428.htm). This provides some protection to the recovering areas, increasing the chances of success of a large-scale regeneration program in the domain. Therefore, in sum, our study reinforces that the regeneration of a tropical forest domain can have high benefits for biodiversity, is feasible in economical and legislation terms, pending, however, the political will to effectively occur.