Contribution of Vouacapoua americana fruit-fall to the release of biomass in a lowland Amazon forest

Fruit-fall provides the transfer of biomass and nutrients between forest strata and remains a poorly understood component of Amazon forest systems. Here we detail fruit-fall patterns including those of Vouacapoua americana a Critically Endangered timber species across 25 km2 of lowland Amazon forest in 2016. We use multi-model comparisons and an ensemble model to explain and interpolate fruit-fall data collected in 90 plots (totaling 4.42 ha). By comparing patterns in relation to observed and remotely sensed biomass estimates we establish the seasonal contribution of V. americana fruit-fall biomass. Overall fruit-fall biomass was 44.84 kg ha−1 month−1 from an average of 44.55 species per hectare, with V. americana dominating both the number and biomass of fallen fruits (43% and 64%, number and biomass respectively). Spatially explicit interpolations provided an estimate of 114 Mg dry biomass of V. americana fruit-fall across the 25 km2 area. This quantity represents the rapid transfer by a single species of between 0.01 and 0.02% of the overall above ground standing biomass in the area. These findings support calls for a more detailed understanding of the contribution of individual species to carbon and nutrient flows in tropical forest systems needed to evaluate the impacts of population declines predicted from short (< 65 year) logging cycles.


Dominance of
). This total included fruits of 86 species from 28 families and 51 genera (Supplementary Table S1). It was possible to identify 81 (94.1%) to species and five (5.9%) to genus (Supplementary Table S1). Fabaceae was the most species rich family with 24.4% of collected species, followed by Sapotaceae (12.7%) and Lecythidaceae (8.1%).
From May to June 2016 the number and biomass of fallen fruits was dominated by Vouacapoua americana (43% and 64%, number and biomass respectively, Fig. 2, Supplementary Table S1). Although V. americana was the most dominant and widespread of the fallen fruits (recorded in 62 of 90 plots), there was substantial variation in the number, taxonomic diversity and biomass of fallen fruit between the sampled plots ( Table 1, Supplementary  Table S1, Supplementary Fig. S1).
Meso-scale fruit-fall patterns. There was substantially more variation in biomass compared with occurrence of V. americana fallen fruits (Fig. 3). Considering the spatial scale of the samples (pairwise distances between plots ranged from 53 to 6642 m), spatial autocorrelation was detected only across relatively short distances (Fig. 3), with complete spatial randomness characterizing the variance in both responses beyond 200 m. The modelled variograms showed relatively high nugget values, representing 50 and 68% of sill values, for fruitfall occupancy and biomass respectively (Fig. 3).
A combination of spatial, topographic, hydrographic and vegetation cover variables were retained as important for explaining both fruit-fall presence and biomass ( Table 2). The spatial model was the most important for explaining patterns in meso-scale masting presence and biomass (Table 2). Topography, hydrography, and vegetation cover models only weakly explained meso-scale biomass patterns (Table 2). In contrast, a relationship with vegetation cover was strongly supported for the meso-scale presence of masting, but topography and hydrography were again only weakly supported. The RMSE of minimal model fitted values (Table 2) was well below the observed SD values for both responses (RMSE = 50.6, 0.35; SD = 74.9, 0.47; biomass, presence respectively).

Discussion
Despite the promotion of sustainable use of timber and non-timber resources in tropical forests, the current management criteria still typically lack inclusion of species-specific ecological features. We were able to model the meso-scale distribution of fallen fruit biomass from the critically endangered V. americana. The models enable us to describe, explain and predict meso-scale fruit-fall biomass patterns across 25 km 2 of lowland Amazon forest. We found that spatial effects most strongly explained variation in fruit-fall patterns and that the contribution of spatial, topographic, hydrographic and vegetation variables differed between responses. We discuss these findings in relation to what is known regarding fruit-fall patterns across lowland Amazonia and then consider the implications for understanding patterns in biomass below the forest canopy.
Our sample provides a representative snapshot of fruit-fall patterns across the 25 km 2 study area. The composition of families and species follows the general pattern found in nearby forest sites and across the Guiana shield [39][40][41][42][43][44] . We found fallen fruits from four (Crhysophyllum, Licania, Protium and Eschweilera) of the five most abundant genera that Pereira, et al. 43 identified in 1.9 ha of nearby (32 km distant) lowland terrra-firme forest. Fabaceae was also the dominant family in a recent inventory of large (> 40 cm DBH) trees, close (< 30 km) to our study area 39 .
Fruit-fall biomass was similar to values reported by studies from other regions of the Guianan Shield. In the most productive month in the rainy season, Sabatier 44 reported a mean of 50 kg ha −1 of fruit-fall production, with 86% of species producing fruits during this season. The less productive soils of the Guianan Shield result in lower values of arboreal species richness 2,45 . This pattern also appears to be reflected in fruit production 46 , with values from French Guyana (292 kg ha −1 annual fruit-fall dry biomass) less than half those reported from western Amazonia (e.g. Hanya, et al. 46 recorded 796 kg ha −1 of annual fruit-fall dry biomass in Cosha Cashu, Peru).
Clearly V. americana was the main source of fruit-fall, in our study site in 2016. From 20 fruits (representative sample of 10 mature fruits from 2 different trees), we obtained a mean dry mass of 18.2 g, the majority of which was the single seed (90%, mean dry weight from 20 seeds = 16.2 g). This provides an estimate of 72.8 kg (4000 × 18.2 g) of dry fruit per tree. Our dry fallen fruit biomass values therefore represent a range of 1 to 5 fruiting trees per Ha. Our aim was not to estimate tree density, but as these values fall within the expected range reported by previous studies they do reinforce the representativeness of our meso-scale sample. There is an urgent www.nature.com/scientificreports/ need for additional studies to establish the density and distribution of V. americana in the study area. This data could then enable our model predictions to be validated. Predicting across a continuous gird enables a variety of analyses that are not possible with sparsely sampled data. Several studies have used remote sensing data to create accurate models of predictions of tropical forest carbon or biomass in diverse scales 30,33,[47][48][49][50][51][52] , but such approaches have not been applied to fruit-fall biomass. The maps presented (Fig. 5) are useful for visualizing the environmental space in more than one dimension and for understanding the predicted responses in the 25 km 2 study area. These maps provide a baseline reference that enables evaluation of future management and/or silvicultural actions.
Our model predictions show that field data collection will be necessary to generate robust estimates of fruit-fall biomass and enable evaluation of future management and/or silvicultural actions. These estimates will improve our knowledge about what is being actually conserved and where we can find it within the protected area. However, comparison with fruit-fall patterns in other lowland sites is necessary to enable more rigorous model testing and evaluation. As the spatial structure described may also change significantly over time e.g. plots with low fruit-fall in 2016 may be plots with high fruiting in the following years, additional studies are also required to establish the multi-year patterns of fruit-fall dynamics in the study area.
We found that remotely sensed data provided useful environmental explanatory variables for modelling the distribution of fruit-fall dry biomass. More generally, this study shows remotely sensed variables have potential for predicting meso-scale fruit-fall biomass. More studies are necessary to improve predictive power of biomass models for understanding impacts of compositional changes driven by anthropogenic and climate changes.   Fig. 1). The regional climate is classified by Köppen-Geiger as Am (Equatorial monsoon) 54 , with an annual rainfall greater than 2000 mm (2240-2510 mm per year from 2010 to 2016) 55 . 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) 55 (Supplementary Fig. S2). The ANF is located within the Uatumã-Trombetas moist forest ecoregion (tropical and subtropical moist broadleaf forests biome) and consists of continuous tropical rainforest vegetation, predominantly never-flooded closed canopy "terra firme" forest 56 . Canopy trees within the ANF typically reach a height of 25-35 m interspersed with emergent trees reaching up to 50-m 56 . The soil is predominantly low-fertility oxisols, including a mix of red, yellow and red-yellow latosols following the Brazilian soil classification system 56,57 .

Study species. Vouacapoua americana is a Critically Endangered (A1cd + 2cd 23 ) endemic to the eastern
Guiana shield rainforests 24,58 . Guiana Shield forests are highly biodiverse, with a more complicated history than temperate and boreal forests, due to a mixture of gradual compositional changes and expansion from refugia 45,59 . V. americana may be a typical example of this situation, as it shows population contraction in its hypothesized range 24,58 . V. americana is a large canopy tree with a wide crown that can bear 3000-4000 large (32.6 ± 1.8 g 27 ) single-seeded fruits. It is highly valued for its durable hardwood timber (heartwood density 0.91 g/m 3 at 12% moisture content) and pharmacological potential 60,61 . The density of individuals over 10 cm diameter at breast height (d.b.h.) is of the order of 10 per hectare 59,62,63 , but varies widely with adults showing a locally clustered distribution 65 . The spatial distribution of adults depends on abiotic and biotic conditions, including topography and soil hydromorphy, light [63][64][65][66] and the availability of short and long range dispersal agents 27,63,64 .
Long-term studies from French Guiana show V. americana is a mast-seeding species [defined as bimodal seed production with no overlap between two tails 36 ] with an average interval between masting events ranging from 2.3 66 to 4 years 44 . Fruiting occurs during a relatively rapid (typically two month) masting event that is synchronized with the wet season 27,66 . Immature fruits are consumed by primates (including Ateles spp. and Alouatta spp., http://visio n.psych ol.cam.ac.uk/spect ra/guian a/fdiet .html). Fallen seeds may germinate underneath parent trees but this abiotic mode of recruitment is generally unsuccessful for V. americana 27,28,66 unless there is a canopy opening nearby 28 . Intense removal of V. americana seeds leads to high rates of seed dispersal compared with low predation throughout the season 27 . Most seeds are dispersed approximately 10 m away by rodents (Dasyprocta leporina and Myoprocta exilis) 27,63,65 , with tapir and peccaries thought to be responsible for longer range dispersal 26 .
Fruit-fall data collection. Fruit-fall ground surveys are a well-established, relatively efficient and straight forward method to assess fruit production in tropical forests and can reflect seasonal fruiting phenology well 5,67 . For example, fruit availability, estimated by fruit-fall, positively affected the biomass and the number of species among frugivorous primates 46,68 . As part of a wider study 69,70 fruit-fall surveys were conducted in May and June 2016 that coincided with the fruit-fall of V. americana. Although we did not examine phenological patterns, based on the findings from previous long-term studies 44,63,64 we assume that this was a masting fruit-fall event for V. americana.
Within the 25 km 2 PPBio grid, a total of 30 regularly spaced (1-km interval) points were sampled (Fig. 1) 71 . This regular arrangement and sample size of 30 has been shown to be adequate for capturing variation in mesoscale species diversity responses across lowland Amazonia 72 . At each of the 30 sample points, surveys were conducted along three plots (one permanent plot and two trail plots, Fig. 1). The permanent plots (250 m long) are nonlinear and follow altitudinal contours to minimize the internal variation in both altitude and correlated covariates such as soil type 53,71 . Additionally, we sampled two linear trail plots (250 m each), one before and the other after the start point of the permanent plots (Fig. 1). Fallen fruit were sampled 1 m to each side of the plot centerlines i.e. covering a total area of 500 m 2 (2 × 250 m) for the linear trail plots. This effort provided a total sampled area of 4.42 ha (plot area m 2 mean ± SD = 490.7 ± 46.9).
To obtain robust and reproducible estimates of fresh fallen fruit we established a number of inclusion and exclusion criteria 6 (Fig. 5). All fresh (i.e. not rotten or desiccated) fallen fruits were counted. Fruits considered unlikely to change in appearance between samples (i.e. Vouacpoua americana, Apeiba sp, Hevea brasiliensis) were removed in order to avoid counting the same fruit twice. Fresh fruits that had been partially eaten were also counted (Fig. 5).
The collected fruits were identified to the lowest taxonomic level (Supplementary Fig. S3) and named following APG III 73 by botanists from the Amapá State Scientific Research and Technology Institute (Instituto de Pesquisas Científicas e Tecnológicas do Estado do Amapá, IEPA). To obtain dry fruit biomass estimates the mean dry weight from a maximum of 30 mature fruits of all fallen fruit species was calculated (Supplementary Table S1). Fruits with seeds beginning to germinate were not weighed. The collected fruits were dried to constant Scientific Reports | (2021) 11:4302 | https://doi.org/10.1038/s41598-021-83803-y www.nature.com/scientificreports/ weight in an oven at 50 °C and then weighed with a precision balance (fruits < 10 g) or digital balance with error ± 0.01 g (fruits > 10 g). Values of fruit-fall dry biomass were expressed as kg ha −1 .

Environmental explanatory variables.
Remote sensing data represent continuous measurements of environmental variables that can be applied in ecological studies 69,72,74,75 . To explain and predict patterns in fallen fruit we used a total of 10 remotely sensed variables (Table 3). These remotely sensed variables were selected to represent meso-scale patterns in topography, hydrography and vegetation cover (Table 3, Supplementary Fig. S4).

Model development.
Complementary approaches were adopted to explain and predict patterns in fruitfall presence and biomass. Firstly, to examine spatial patterns in the responses we used variograms that enable quantification of different aspects of the spatial patterns, including: range (the limit of spatial dependence), nugget (portion of semi-variance attributable to random/environmental factors), and sill (distance at which the variogram becomes constant) 78 . Then, to examine the environmental factors explaining patterns of fruit-fall presence and dry biomass Generalized Additive Models (GAMs) were employed 79,80 . GAMs are a nonparametric extension of general linear models that provide the flexibility to model non-parametric and nonlinear relationships that are typical of many ecological patterns. Explanatory GAMs were generated using the "mgcv" package 81 using R 3.10 software 82 . Penalized cubic regression splines determined the shape of nonparametric functions, with the degree of smoothing selected automatically via maximum likelihood using the mgcv package defaults for all models.
To avoid subjective bias in model development we did not examine correlation among variables until after formulation of our a priori models. We applied a two stage approach to explain patterns in the fruit-fall responses. Firstly, separate models were developed to examine the effects of space, topography, hydrography and vegetation cover ( Table 2). Within each of our four models, explanatory variables were tested for collinearity. The explanatory power of the models and the level of support for the different hypotheses were examined within a maximum likelihood framework 83 , with models tested and compared using a combination of model deviance explained 81 and information criteria (AIC and BIC) 83 . The accuracy of model fits were examined using the root mean square error (RMSE) and correlations between observed and fitted values.
Secondly, we obtained the most parsimonious ("minimal") model, with the most important variables selected using a manual backward stepwise selection based on minimizing AIC values. Manual selection started from the full model without correlated variables, with stepwise exclusion of variables if they did not make a statistically significant (p ≥ 0.05) contribution.
After explaining variation in responses, we then used an ensemble model to predict the presence and dry biomass of masting fruit-fall across the 25 km 2 sample grid. The similarity in environmental variable values  (Table 3) means that we consider the predictions to be an interpolation within the range of environmental values. Ensemble approaches were used to decrease the predictive uncertainty of single-models by combining their projections. The ensemble model was obtained from the weighted mean of six single model methods. The six model algorithms were inverse distance weighted (IDW 84 ), universal kriging (UK 84 ), generalized linear model (GLM), generalized additive model (GAM 80 ), random forest (RF 85 ) and support vector machine (SV 86 ). Predictive models were developed for both presence and biomass with the weighted mean ensemble derived from model predictions weighted by the correlation between observed and predicted values. Finally, we compared the predicted values with overall live wood biomass. We estimated the range of overall live wood biomass in the 25 km 2 study area from two sources 31,38 . The 500 m resolution mapped values from Baccini et al. 31 were used to provide minimum value and the maximum value was estimated by extrapolating the mean (423.8 Mg ha) from Johnson et al. 38 across the 25 km 2 (2500 ha) area.

Data availability
The datasets generated during and/or analyzed during the current study are available from the corresponding author on reasonable request.