Big trees drive forest structure patterns across a lowland Amazon regrowth gradient

Degraded Amazonian forests can take decades to recover and the ecological results of natural regeneration are still uncertain. Here we use field data collected across 15 lowland Amazon smallholder properties to examine the relationships between forest structure, mammal diversity, regrowth type, regrowth age, topography and hydrology. Forest structure was quantified together with mammal diversity in 30 paired regrowth-control plots. Forest regrowth stage was classified into three groups: late second-regrowth, early second-regrowth and abandoned pasture. Basal area in regrowth plots remained less than half that recorded in control plots even after 20–25 years. Although basal area did increase in sequence from pasture, early to late-regrowth plots, there was a significant decline in basal area of late-regrowth control plots associated with a decline in the proportion of large trees. Variation in different forest structure responses was explained by contrasting variables, with the proportion of small trees (DBH < 20 cm) most strongly explained by topography (altitude and slope) whereas the proportion of large trees (DBH > 60 cm) was explained by plot type (control vs. regrowth) and regrowth class. These findings support calls for increased efforts to actively conserve large trees to avoid retrogressive succession around edges of degraded Amazon forests.

day 31 and disperse seeds of more than 70 tree species 32 . The loss of mid-to large-bodied mammals may release some plant species from herbivory and increase their dominance, which subsequently decreases tropical forest biodiversity 33,34 .
Given the need to understand the patterns of forest structure in Amazonian forests, here we aim to identify how biotic and abiotic factors (Table 1) can explain patterns in forest structure across a successional gradient in 15 lowland (90-120 masl) smallholder properties in the eastern Brazilian Amzon. These sites were surrounded by a matrix of continuous closed canopy forest cover (Table 2), a configuration far different from more widely studied fragmented landscapes. We compare 5 forest structure attributes across late second-regrowth, early second-regrowth and abandoned pasture plots to examine how forest structure changes in relation to topography, hydrology, regrowth class, time and mammal diversity.

Results
Variation in stand structure variables. There were clear differences in forest structure between control and regrowth plots (Fig. 1). On average control plots had increased basal area and increased proportion of large trees (Fig. 1). In contrast regrowth plots tended to have increased proportion of small trees (< 20 cm DBH).
The number and basal area of living trees tended to increase with altitude and this relationship was stronger in regrowth areas (Fig. 2). The relationship with altitude was strongly affected by low lying (90 masl) pasture plots with no trees that generated significant leverage on the linear relationship (Fig. 2).
Basal area ranged from 0 to 76.4 m 2 /ha across the 30 survey plots (Table 2), with control plots showing an average fourfold increase in basal area compared with regrowth plots (mean basal area 44.1 and 11.5 m 2 /ha, control and regrowth respectively, Fig. 2). The patterns in plot basal area also differed between regrowth classes (Fig. 3, Supplementary Table S1). There was a significant interaction between plot type (control/regrowth) and regrowth stage, with basal area increasing across pasture, early and late regrowth plots but control plots showing the opposite trend, with basal area decreasing significantly in late-regrowth control plots (Fig. 3).
There was a highly significant positive linear relationship between overall basal area and large tree basal area (F 1,28 = 127.5, R 2 = 0.82, P < 0.0001). The basal area of large trees decreased significantly in regrowth compared  www.nature.com/scientificreports/ with control plots (Fig. 3). On average large trees accounted for 42% of the basal area in control plots compared with only 4% in regrowth plots ( Table 2). Indeed a single large tree (> 60 cm DBH) was recorded only once in a late-regrowth plot. This relationship was also reflected in the decline in basal area of late-regrowth control plots (Fig. 3), which was associated with a decline on the proportion of large trees that accounted for a reduced 31% of the basal area in late-regrowth control plots (Table 2).
Relationships between forest structure, mammal diversity and environmental variables. Mammal diversity varied considerably across the survey plots (Fig. 4). There appeared to be a tendency for basal area to increase with mammal diversity in Late-regrowth plots, yet basal area was only weakly associated with mammal diversity within the different regrowth classes (Fig. 4). Indeed, the diversity of mammals was found to be only weakly informative for explaining the basal area of trees across the 30 sample plots ( Table 3).
Comparison of models showed that plot type (control v regrowth) and topography were the most important first ranked variables for the five forest structure attributes ( Table 3). The most simple model including only  www.nature.com/scientificreports/ plot type explained more than 50% of model deviance for all forest structure attributes except for the number of trees (DBH > 10 cm). Plot type and regrowth class were both included in the 95% confidence set of models for the basal area of large trees (Table 3). In contrast Topography was the most important (first ranked) model for the number of trees, proportion of small trees and tree basal area (Table 3). Mammal diversity, Time and Hydrology models were not well supported and were not included in the 95% confidence set of models for any of the forest structure attributes (Table 3).

Discussion
We integrate field and remotely sensed data to establish support for the relative importance of different variables explaining patterns in forest structure across a lowland Amazon regrowth gradient. We establish that different variables are supported for different structure attributes. Here we discuss these findings in terms of prospects for the passive restoration of degraded Amazon forests. The mean basal area value from our 15 control plots (44.1 m 2 /ha) was close to the mean from 42 Guyana Shield forest plots (43.4 m 2 /ha, range 10-65 m 2 /ha) in French Guiana 35 . The results from Molto et al. 35 were obtained from an extensive survey of 0.5-1 ha plots. Although our plot size was smaller compared to Molto et al. 35 , the similarity in mean values suggests that our plots do provide a representative sample of forest structure in the regrowth areas. The basal areas obtained from our regrowth plots followed a similar trajectory to those reported from abandoned pasture in Costa Rica 36 , where the most recently abandoned pasture plots (< 14 year) had mean basal area of 13.5 m 2 /ha, with basal area increasing to 26.1 m 2 /ha after 21-30 year 36 .compared with 11.1 and 17.6 m 2 /ha respectively in our Early (1-5 year) and Late (20-25 year) regrowth plots. This also follows a similar pattern to values reported from 370 successional forest plots in the Brazilian Amazon, with basal area values typically < 10 m 2 /ha in early stages (< 5 year) and reaching 25 m 2 /ha after 15 years 45 .
Although results from lowland forest sites in Costa Rica suggest rapid recovery of pasture areas 36 this could be related to the substantially lower basal area in the seven old growth reference plots (26.1 m 2 /ha, range 19.3-32.2 m 2 /ha) compared with those in our study area. Our results are similar to those reported from the central Amazon, where 25 y of regrowth restored half of the mature-forest biomass 42 . A recent analysis of 45 Neotropical secondary forest study sites found that secondary forests in the lowland tropics reach 90 percent of old growth biomass in a median time of 66 yr 13 . Our findings do suggest nuanced difference in successional trajectories. Basal area increased rapidly in early regrowth stages and this could be explained by the less intensive land use (i.e. lack of pasture) and the proximity to large areas of intact forest. In contrast basal area of late-regrowth areas was less than those reported from other areas 36,45 . This could be related to soil productivity, as previous studies www.nature.com/scientificreports/ show that highly diverse Guyana Shield wet forests can take longer to establish 13 . With basal area of control plots dominated by large trees it seems likely that many decades will be necessary for forest structure (total basal area, proportion of large trees) to return to pre-disturbance values. The success of active and passive restoration can depend on ecological conditions 46 . We found topography was the most informative model for explaining patterns in number of tress, tree basal area and proportion of small trees (Table 3, Fig. 2). Differences in altitude and slope have been shown to affect floristic structure of tropical forests from local to regional scales 16,35,43,[47][48][49] . Indeed, even relatively small variations in topography can generate changes in local-scale soil chemistry, hydrology and microclimate 47,50 . The effects of topography do not operate in isolation from hydrology and the increased numbers of small trees and tree biomass with increasing altitude (Fig. 3) agree with previous studies that show trees grow more slowly in more low lying (and often more waterlogged) terrain 43 . It remains uncertain how differences in soil type may affect regrowth trajectories and additional experimental studies are needed to examine the relative contributions of soil, hydrology and topography.
We found a weak association between mammal diversity and regrowth forest structure. Previous studies in a nearby protected area show that this group of mammals (mid-to large-bodied Artiodactyla, Perissodactyla and Rodentia) are more strongly associated with factors such as access to water 51 and altitude 51,52 . A recent study also showed that mammal abundances were more strongly associated with phenology (fruit fall) than basal area along 10 km of forest in the western Guyana Shied 53 . Additionally, regrowth class was found to be the primary driver of mammal species encountered independent of forest cover 54 . For example the number of species detected in control and regrowth plots (all with forest cover > 87%) varied between 1 and 6 ( Fig. 4). Mid-to large-bodied seed dispersers are a critical component of Amazon forests 18,19,25 and are also widespread and ubiquitous across myriad Amazonian forest types [55][56][57] . The eight species are therefore not strictly dependent on the quality of forest habitat compared with other more specialist groups such as primates 58 . The lack of a strong relationship between diversity of these eight mammal seed dispersers and forest structure attributes (i.e. overall basal area and proportion of small trees) is therefore to be expected.
Decades of research show that myriad edge effects can extend up to 150 m in fragmented Amazon forests 59,60 . Considering the range of expect edge-effects it is highly probable that the natural regeneration and/or restoration of regrowth habitats in Amazon smallholdings (typically < 100 ha) will strongly depend on species ecological responses to habitat edges 61 . Previous studies show that edge effects increase mortality of large trees, which in turn has major impacts on forest ecosystems 62 . In highly fragmented areas edge-effects can drive tree communities through a process of "retrogressive succession" 63 and toward an early successional state that may persist indefinitely. This early successional state can be characterized by functional and structural differences in that larger slower-growing tree species with high wood density tend to decline whereas faster-growing tree and liana species with lower wood density increase 63,64 . The decline in the number and basal area of large trees from our www.nature.com/scientificreports/ control plots along 20-25 year old edges suggest that retrogressive succession may establish even in relatively un-fragmented areas surrounded by extensive forest cover. More empirical evidence is however required to strengthen support for the retrogressive succession hypothesis in relatively unfragmented Amazon landscapes. Additional studies are needed that focus on quantifying the distribution of other structural and functional components of tropical forests such as lianas and short-lived pioneers, changes in forest stratification and convergence of tree assemblages [63][64][65] . There also remained uncertainty in our models that could be reduced through increased sampling effort. Studies show that within and among plot uncertainty can be reduced by multi-plot census 66 , with smaller plot sizes (e.g. < 0.25 ha) requiring more replicates 66,67 . An increased long-term monitoring effort is therefore necessary to enable a more robust understanding of the forest structure dynamics across our study sites.
Our findings provide an early warning that even under a best case scenario there is potential for "retrogressive succession". We found not only a lack of large trees in regrowth plots but also that large tree basal declined in older late-regrowth control plots. We suggest that this decline in large tress may be the primary driver of differences between regrowth and old growth forest and as such represent an unquantified component of resilience and time to recovery of Neotropical secondary forests. We also suggest that the continued presence of mid-and large bodied mammal seed dispersers in the study area are likely to be vital in order to avoid such "retrogressive succession". Table 3. Summary of the Generalized Linear Models created to explain forest structure in 30 plots (15 control and 15 regrowth). Models ordered by decreasing AIC (Akaike Information Criterion) values. a Models used to explain forest structure. All models contained plot type (control/regrowth) as categorical factor. Variables and associated estimates in the different models can be found as Supplementary Information Table S2 online. b Akaike weights (W i ) from largest to smallest.  The soil is predominantly low-fertility oxisols, including a mix of red, yellow and red-yellow latosols following the Brazilian soil classification system 72 . The regional climate is classified by Koppen-Geiger as Am (Equatorial monsoon) 73 , with annual rainfall greater than 2500 mm 74 . The driest months are September to November (total monthly rainfall < 150 mm) and the wettest months from February to April (total monthly rainfall > 300 mm) 74 . The State of Amapá has the lowest deforestation rate in Brazil and > 70% of the Amapá receives some form of legal protection. There is no large scale agricultural developments or monocultures along the waterways and properties retain typically small (< 1000 ha) areas of opened land, which are cleared for small scale family agriculture, which focuses on acai (Euterpe oleraceae 75 ), small scale production of fruits and vegetables for sustenance and limited commercial sale of regional produce (e.g. manioc flour) in local markets. There are some 54 properties upstream of the nearest town (Porto Grande 76 ). There has never been any expansive clearcutting in www.nature.com/scientificreports/ the region and there are no monocultures (e.g. soy) or cattle production. All sites were at least 26 km from the nearest town by river, and all sites are surrounded by matrix of continuous closed canopy forest cover (Table 2). Pesticides and/or herbicides had never been used at any of the sites. As described previously 54,68 the 15 smallholder properties were all lowland (90-120 masl) and selected based on differences in land-use histories and forest succession/regrowth stage. All sites were close (110-554 m, Table 2) to 100-200 m wide rivers that are navigable by motorized boats, but due to riverbank formation the sites are never flooded. These 15 sites were grouped into three regrowth classes based on the land-use history: late secondregrowth forest (N = 5, most recent human disturbance between 20 and 25 years), early second-regrowth (N = 5, most recent human disturbance between 1 and 5 years), and pasture (N = 5, recently cleared and abandoned pasture areas dominated by grasses/herbs but that had never been used to raise livestock, with the most recent disturbance between 1 and 17 years). Each of the 15 regrowth sites was paired with a nearby (60 to 150 m) control site i.e. 20-30 m tall terra-firme forest site without a history of mechanized timber extraction. We did not record any evidence of logging or hunting in the control areas. We therefore assume that there were minimal anthropogenic impacts in the control sites that if present did not generate any systematic bias in our comparisons between regrowth stages. To reduce the possible confounding influence of edge effects that are known to strongly influence the distribution of trees in Neotropical forests, all regrowth and control sites were established at a standardized distance (approximately 30 m) from the nearest control-regrowth habitat edge.
Forest structure. Data were collected from May to August 2016. Forest structure data (i.e., number of trees and basal area) were obtained from plots measuring 50 × 10 m (500 m 2 /0.05 ha), at each of the 30 points, totaling 1.5 hectare. This plot size was selected as it has been widely used to examine structural changes in tropical forests 19,42,43,77 and several of the regrowth areas were too small ( Table 2) to enable the establishment of larger spatially independent plots. We obtained five measures (responses) to characterize the forest structure in each plot. These were selected based on previous studies that show their appropriateness to distinguish attributes of regrowth/successional stages related to biodiversity of Amazon forests 13,45,78,79 . The number of all trees ≥ 10 cm DBH (diameter at breast height at a standard 1.3 m above ground, or above tallest root buttress) was used to quantify the tree density in each plot (m 2 ). This count included all trees which had at least half of their basal trunk inside the plot. The proportion of small (10-20 cm DBH) trees was calculated to represent the expected increase of younger trees in regrowth areas. The proportion of large (> 60 cm DBH) trees was calculated as this is known as an important characteristic of mature/late succession areas 45,79 . We also calculated the basal area of all and large trees as this is known to be strongly correlated with tropical tree biomass 36,80 . Explanatory variables. A total of 10 variables (Table 1) were used to build models to represent 5 working hypotheses (topography, hydrology, regrowth class, time and mammal diversity) that based on the findings from previous studies were likely to explain the observed patterns 13,28,36,41,42,81 . We chose to work with mainstream, widely available environmental variables. Four of these (the topographic and hydrographic model variables) were computed from remotely sensed digital terrain model (SRTM-DTM): altitude (masl), slope, TWI (Topographic wetness index), DND (Distance to Network Drainage) calculated from the interaction between HAND (Height above network drainage) and HDND (Horizontal distance to network drainage). The time model included years since the regrowth site was opened and years since last use, both of which were obtained from interviews with local landowners.
Mammal functional diversity was obtained from a camera-trap survey conducted at the same time (May to September 2016) and in the same plots as forest structure was sampled 54 . Camera traps equipped with infrared triggers (Bushnell Trophy Cam, 8MP, Overland Park, KS, USA) were installed in each of the 30 plots following standardized protocols 51,52,82 . This camera trap survey [full details provide in 54 ] including a sampling effort of 827 camera-trap days (450 and 377 camera-trap days, control and regrowth sites respectively) was used to estimate functional diversity of eight terrestrial mammal seed dispersers (Cuniculus paca, Dasyprocta leporina, Myoprocta acouchy, Mazama americana, M. nemorivaga, Pecari tajacu, Tayassu pecari and Tapirus terrestris). Data analysis. Tree Basal Area in each plot was obtained as the sum of the basal area value for each individual tree derived from the DBH of each tree following the formula BA (basal area in m 2 ) = 0.00007854 × DBH 2 (constant obtained by solving the following equation to obtain BA in m 2 from the DBH measured in cm 78 ): We calculated basal area of all and large (> 60 cm DBH) living trees 78,79,81 . We also calculated the proportion of small stems (10-20 cm DBH trees) as this has been shown to be an important measure of stand structure in forest regrowth areas 35,78 .
To represent diversity of terrestrial mammal seed dispersers we calculated a richness and functional diversity (FD) value for each of the 30 plots 54 . Richness was calculated as the observed number of species (hereafter "species richness") at each plot. Although there are many diversity metrics, we chose species richness as it is widely used and clearly interpretable 83,84 and with relatively few (eight) species and 30 plots there were strong correlations between species richness values and alternative diversity metrics such as Shannon and Simpson diversity (Spearman rho > 0.89). We used Functional Dispersion (FDis) 85 as an index of functional diversity as it is not strongly influenced by outliers, accounts for relative abundances, is unaffected by species richness and can be calculated from any distance/dissimilarity measure 85,86 . Functional Dispersion was estimated with the dbFD function 86 using default settings. www.nature.com/scientificreports/ To examine patterns in forest structure attributes we used Generalized Linear Models. We used an information theoretic model averaging framework 87 to compare five models representing the five non-mutually exclusive working hypotheses-topography, hydrology, regrowth class, time and mammal diversity (see Table 1 for variable description and ecological relevance). There were only weak insignificant correlations between regrowth classes and the topographic and hydrological variables (all pairwise Polyserial correlation coefficients 88 : rho < 0.45, P > 0.05), which supports a lack of systematic bias in the distribution of topographic and hydrologic variables across the three regrowth classes. We evaluated models based on their information content, as measured by AIC-Akaike Information Criterion. The relative importance of the models was measured by the models Akaike weights (Burnham and Anderson 87 , pp. 75-77, 167-172), which is a scaled measure of the likelihood ratio that ranges between 0 (least important) and 1 (most important). None of the unexplained variation (model residuals) was related to the geographic distance among plots so we did not need to control for spatial dependence. All analysis were conducted using the R language and environment for statistical computing 89 , with base functions and functions available in the following packages: vegan 90 , ggplot2 91 , MuMIn 92 , and tweedie 93 .

Data availability
The raw forest structure and environmental data used in the analysis of this study have been deposited in the OSF-Center for Open Science at https ://doi. www.nature.com/scientificreports/