What drives the spatial distribution and dynamics of local species richness in tropical forest?

Downloaded from http://rspb.royalsocietypublishing.org/ on September 20, 2017

rspb.royalsocietypublishing.org

Research Cite this article: Wiegand T, May F, Kazmierczak M, Huth A. 2017 What drives the spatial distribution and dynamics of local species richness in tropical forest? Proc. R. Soc. B 284: 20171503. http://dx.doi.org/10.1098/rspb.2017.1503

Received: 5 July 2017 Accepted: 22 August 2017

Subject Category: Ecology Subject Areas: ecology, theoretical biology, computational biology Keywords: demographic rates, mortality, recruitment, null model, spatially explicit neutral model, species diversity

Author for correspondence: Felix May e-mail: [email protected]

Electronic supplementary material is available online at https://dx.doi.org/10.6084/m9. figshare.c.3868213.

What drives the spatial distribution and dynamics of local species richness in tropical forest? Thorsten Wiegand1,2, Felix May2,3, Martin Kazmierczak1 and Andreas Huth1,2,4 1

Department of Ecological Modelling, Helmholtz-Centre for Environmental Research—UFZ, Permoserstraße 15, 04318 Leipzig, Germany 2 German Centre for Integrative Biodiversity Research (iDiv) Halle-Jena-Leipzig, Biodiversity Synthesis, Deutscher Platz 5e, 04103 Leipzig, Germany 3 Institute of Computer Science, Martin-Luther University Halle-Wittenberg, 06099 Halle (Saale), Germany 4 Institute of Environmental Systems Research, University of Osnabru¨ck, 49069 Osnabru¨ck, Germany TW, 0000-0002-3721-2248; FM, 0000-0002-1106-8188 Understanding the structure and dynamics of highly diverse tropical forests is challenging. Here we investigate the factors that drive the spatio-temporal variation of local tree numbers and species richness in a tropical forest (including 1250 plots of 20 20 m2). To this end, we use a series of dynamic models that are built around the local spatial variation of mortality and recruitment rates, and ask which combination of processes can explain the observed spatial and temporal variation in tree and species numbers. We find that processes not included in classical neutral theory are needed to explain these fundamental patterns of the observed local forest dynamics. We identified a large spatio-temporal variability in the local number of recruits as the main missing mechanism, whereas variability of mortality rates contributed to a lesser extent. We also found that local tree numbers stabilize at typical values which can be explained by a simple analytical model. Our study emphasized the importance of spatio-temporal variability in recruitment beyond demographic stochasticity for explaining the local heterogeneity of tropical forests.

1. Background The high diversity of tropical forests is one of the enduring puzzles in ecology [1,2] that challenges ecological theories. Previous modelling studies seeking to understand the basic dynamics of tropical forests focused on non-spatial patterns (e.g. [3–5]) and explored equilibrium properties such as the species abundance distribution or the species –area relationship (e.g. [6,7]). Prominent examples are neutral models that consider dispersal limitation and ecological drift, and assume that species are functionally equivalent [3,8]. They provide realistic predictions of different biodiversity patterns [4,7,9]. Extended models have also been used to explain spatio-temporal patterns of tropical forests such as species extinction and temporal shifts in species clusters, or to assess environmental variation and temporal storage effects [5,10 –12]. Here we aim to investigate the mechanisms behind two spatio-temporal patterns that have rarely been studied with dynamic models: the variation in local tree numbers and in local species richness found in tropical forests [13 –15]. For example, tree species richness in the 50 ha forest dynamics plot on Barro Colorado Island (BCI; Panama) ranged from approximately 20 to 80 species at the local scale (20 20 m2; figure 1a,b). We argue that these patterns can provide important additional information on the spatial and temporal dynamics of species rich communities that has not been investigated until now. We propose here a novel approach to explore the spatial and temporal dynamics of species rich forests using simple simulation models that are built around the local spatial variation of demographic rates. A specific feature of our

& 2017 The Author(s) Published by the Royal Society. All rights reserved.


(b)

y

300 200 100 0 0

250

500 x

750

1000

2 300

200 100

200 100 0

0 20

40 60 80 no. species

100 200 300 no. trees

approach is inclusion of three spatial scales: the 20 20 m2 quadrat scale (where forest state and dynamics are analysed), the 50 ha community scale (the extent of our spatially explicit simulation models), and the regional scale represented by a simple metacommunity. Our approach allows us to test different biological hypotheses. We accomplish this by combining field data analysis with a dynamic null model approach that simulates recruitment and mortality in each quadrat by maintaining certain aspects of the observed distributions (of the number of recruits per quadrat and the mortality rates per quadrat, respectively) while randomizing other aspects. In total, we test six simulation models that make alternative assumptions on local mortality and recruitment in their ability to explain the observed spatio-temporal variation in local tree and species numbers. We underpin our simulation results by an analytical model for local tree numbers and additional data analyses. Our simplest hypothesis is that variations in local tree numbers and in local species richness emerge solely as a consequence of random birth and death processes [3,6] (i.e. the demographic stochasticity hypothesis). However, additional mechanisms not considered in classical neutral theory may contribute to the local dynamics of tropical forests. We focus here on the additional contribution of spatio-temporal variability in demographic rates. One possible mechanism is that local clusters of high tree density (and species richness) may be related to locally favourable environmental conditions for recruitment and/or survival (habitat hypothesis) [16]. Alternatively, heterogeneity in recruitment (or mortality) may be caused by spatially and temporally varying biotic habitat conditions at the local scale (e.g. temporal occurrence of tree fall gaps [15,17,18]) that might result in randomly moving local clusters of high tree density and species richness (moving cluster hypothesis) [11]. A similar effect for recruitment would be expected in the case of strong spatio-temporal variation in seed production [19]. Within the framework of our approach we can analyse such different hypothesis. To test the performance of the alternative model versions we assess how well they reproduced several spatial and temporal patterns observed at the 50 ha plot of neo-tropical forest BCI in Panama [20]. Our primary focus is on the observed quadrat-scale distributions of tree abundance N and species richness S per quadrat. As additional spatiotemporal patterns we test, for example, the changes in the number of trees and species per quadrat and time (DN versus N0 and DS versus S0), and to explain the empirical pattern of DN versus N0 we use an analytical model. We show that the dynamics of local tree density were stabilized mainly

by local recruitment being largely independent of the local tree density. The observed large spatio-temporal variability in local recruitment was the key factor for emergence of the observed variation in local tree density, and limited dispersal was important for reproduction of the observed pattern of local species richness.

2. Material and methods (a) Study site We used data from the 50 ha (1000 500 m2) forest dynamics plot on BCI [15,20,21]. The plot is located on an island in the Panama Canal (98150 N, 798860 W) and the vegetation on BCI represents seasonally moist tropical forest with a pronounced dry season and mean annual rainfall of 2600 mm. The plot comprises old growth lowland moist forest and contains up to 305 tree and shrub species. Elevation in the plot ranges from 120 to 155 m.a.s.l. All trees with a stem diameter at breast height (dbh) of 1 cm or more have been mapped and had their diameters measured in 1982 and 1985, and from then on every five years until 2010, resulting in seven censuses.

(b) Data We used the six censuses from 1985 to 2010. As recruits we classified all trees which appeared for the first time in a given census (i.e. grown during the last 5 years to a dbh 1 cm). As dead trees we classified those trees that were alive in a given census and recorded as dead or disappeared during the next one. For data analysis we divided the forest area into 1250 non-overlapping contagious quadrats with a size of 20 20 m2. However, to test the generality of our approach we also conducted supplementary analyses using quadrat sizes of 10 10 m2 and 50 50 m2. For every quadrat and census we counted the number of living trees and the number of species, and determined their distribution (figure 1). We also determined the distributions of the numbers of recruits and of dying trees per quadrat and census interval (figure 2c,d). Finally, we determined the distribution of the per capita mortality rates within quadrats by dividing in each quadrat the number of dying trees between census t and t þ 1 by the number of trees at census t. Electronic supplementary material, figure C1 shows all distributions and pairwise correlations among these variables (using rank-correlation analysis). To assess the temporal variability in several quadrat-scale variables (e.g. number of recruits or per capita mortality rates) we estimated the coefficients of variation (CV) of these variables for each quadrat across the available censuses and then determined the distribution of the CV values of all 1250 quadrats. We also determined the analogous distribution for spatial variability.

Proc. R. Soc. B 284: 20171503

Figure 1. Distributions of the number of species and the number of trees (above 1 cm dbh) in the BCI forest plot for census 2010. (a) Raster map that indicates the number of species in all 20 20 m2 quadrats. (b) Distribution of the number of species per quadrat for the 2010 census. (c) The same as (b), but for the number of individuals per quadrat.


no. species 80 70 60 50 40 30

400

(c) no. quadrats

BCI 2010

500

no. quadrats

(a)


(b)

(c)

(d)

0.020

0.10

0.08

0.08

0.015

0.08

0.06

0.06

0.04

0.04

0.02

0.02

0.06

0.010

0.04

0.005

0.02

0

0 0

100 200 300 400

100

0 0 20

60

100

0 20

(f)

(g)

(h)

0.020

0.10

0.08

0.08

0.015

0.08

0.06

0.06

0.04

0.04

0.02

0.02

0.06

0.010

0.04

0.005

0.02

0

0 0

100 200 300 400

(i) model 5: R2/M2

60

0 0 20

60

100

60

100

( j)

(k)

(l)

0.08

0.08

0.015

0.08

0.06

0.06

0.04

0.04

0.02

0.02

0.04

0.005

0.02 0

100 200 300 400 no. trees

0 20

60 100 no. species

60

100

0

0

0

0

mean dispersal distance (m) 10 20 30 40 50 LDD 0 20

0.10

0.010

100

0 0 20

0.020

0.06

60

0 20

60 100 no. death

0 20

60 100 no. recruits

Figure 2. Comparison between field data from BCI and model predictions of selected scenarios. The histograms show the distribution of the field data for the censuses 1985 – 2010 in 1250 20 20 m2 quadrats. The coloured lines show the corresponding model predictions. Each line summarizes 30 model replicates and the different colours indicate the different dispersal kernels. The different colours indicate the different dispersal kernels (LDD, long-distance dispersal). The corresponding results of all models are shown as electronic supplementary material, figure C2.

3. Simulation models (a) General structure of the simulation model at the 50 ha community scale We developed several simple, but spatially explicit simulation models describing the community dynamics at the 50 ha scale based on individual trees. All models include stochastic representations of mortality, dispersal, recruitment and immigration of trees with dbh 1 cm in a 50 ha (1000 500 m2) forest plot. Each tree is characterized by its position in space (continuous x, y coordinates) and its species identity. For simplicity, tree size is not considered. In each time step mortality, recruitment and immigration events are calculated. As simple reference model we used a spatially explicit simulation model, which corresponds to a spatially explicit implementation of neutral theory [3]. The number of trees that die per time step (i.e. the 5-year census interval) is calculated based on a per capita mortality rate, which was directly calculated from the data (m ¼ 0.15 per 5 years). When a tree dies, a new tree (recruit) is added to the community (i.e. the zero-sum assumption [3]). For determining the position of the recruit, we first randomly select a parent tree among all trees at the 50 ha community scale and then apply a dispersal kernel. In the more complex model versions we make alternative assumptions on mortality and recruitment (see ‘Submodels for hypothesis testing’). To investigate the consequence of dispersal, we use an isotropic Gaussian dispersal kernel [22,23] with specified mean dispersal distance (10, 20, 30, 40 and 50 m). We use periodic boundary conditions to avoid edge effects. In addition, we simulate a long-distance dispersal (LDD) scenario, where

we select a random position for the recruit and a random parent anywhere in the plot. Similar to previous models (e.g. [4]), we also considered a low probability that a recruit is an immigrant of a new species. The probability was set to 1.1 1024 per recruitment event. This parameter reflects the observed rate of species immigrations of approximately four species per 5 years [24] and represents a simplified metacommunity. The model is written in Cþþ and source code for all model versions is published on the online repository GitHub (https:// github.com/FelixMay/confettiSquares).

(b) Submodels for hypothesis testing We use three alternative submodels for recruitment (R1, R2 and R3) and two alternative submodels for mortality (M1 and M2), summarized in table 1. Other possible submodels could be excluded a priori by analysis of our data. We investigate all six combinations together with the six alternative assumptions on dispersal, resulting in a total of 3 2 6 ¼ 36 alternative simulation models of spatial and temporal forest dynamics to explain the spatial variation in local tree density and species richness. To test the demographic stochasticity hypothesis, we use submodels for recruitment (R1) and mortality (M1) based on constant per capita mortality and recruitment rates m and r, respectively. This model version corresponds to spatially explicit neutral models [7,25,26]. To assess the moving cluster hypothesis, we test if the observed spatio-temporal variability in recruitment contributed to the emergence of our focal patterns. We determine for each quadrat q the absolute number Rq of recruits by taking

Proc. R. Soc. B 284: 20171503

model 2: R2/M1

(e)

0 0 20

3


model 1: R1/M1

(a)


100 0 −100 −200 50

100 200 300 initial no. trees (N0)

40 20 0 −20 −40 20 30 40 50 60 70 80 initial no. species (S0)

(c)

rank-corr. = –0.53

200 100 0 −100 −200 50 100 200 300 initial no. trees (N0)

(d ) change in no. species in 25 years (DS)

200


change in no. trees in 25 years (DN)

(b) change in no. species in 25 years (DS)



40 20 0 −20 −40 20 30 40 50 60 70 80 initial no. species (S0)

Table 1. Summary of the recruitment and mortality submodels. We analysed all six combinations of the three recruitment and two mortality submodels. submodel

name

biological background

R1

constant per capita recruitment rate

each tree has the same probability to be mother tree of a new recruit; the observed mean per capita recruitment rate r is maintained

R2

numbers of recruits from data

the observed distribution prec(R) of the number R of recruits per quadrat is maintained

R3

local density dependence as observed

the observed distribution of recruits per quadrat and its covariation with tree abundance is maintained

M1

constant per capita mortality rate

each tree has the same probability to die; the observed mean per capita mortality rate m is maintained

M2

local per capita mortality rates from data

the observed distribution pmort (m) of the per capita mortality rates mq per quadrat is maintained

randomly a value from the distribution prec(R) observed in the field for all six censuses together (submodel R2). Because we draw each time step a new value of Rq from prec(R), we assume here equivalence of spatial and temporal variability in Rq. Because in this model version the number of recruits in a given quadrat is set, we determine the species identity of the recruit by ‘backward’ application of the dispersal kernel (see electronic supplementary material, appendix A). In this submodel we expect emergence of higher variability in the recruit numbers per quadrat (compared with submodel R1). To assess the relative contribution of the observed spatial variability in local mortality rates we draw in submodel M2 for each quadrat a local per capita mortality rate mq from the observed distribution [pmort (m)] (similar approach as for R2). To test for possible density dependence in recruitment we extend submodel R2 by additionally maintaining the observed covariation between the number Rq of recruits and the number Nq of trees in quadrat q (submodel R3). A detailed description of the simulation model is provided in the electronic supplementary material, appendix A. We also test an extreme case of the habitat hypothesis for local tree numbers where Rq and mq are temporally constant but vary in space following the observed distributions prec(R) and pmort (mq), respectively. This model can be solved without simulation (electronic supplementary material, appendix B).

(c) Model initialization, scheduling and simulation scenarios All model simulations were initialized with the observed BCI census of 1995 (in total 228 819 trees, stem diameter . 1 cm).

For each parameter set we conducted 30 replicate model runs. Each simulation was run for 5025 years (1005 time steps, one time step corresponds to 5 years), and the last six simulated time steps were analysed and compared with the field data. We checked that this simulation time was sufficient to reach a dynamic equilibrium for the patterns investigated in this study.

(d) Comparison of field data and model predictions We compared the output of our alternative models with a number of observed patterns. They included (i) the observed distributions of the number of species or trees per quadrat (figure 1b,c), (ii) the correlations between the 25-year change in the number of species (and trees) per quadrat and the initial numbers (i.e. the 1985 and 2010 census from the data and time steps 1000 and 1005 from the simulations) (figure 3), (iii) the correlations between the number of trees and dead trees, between the number of trees and the mortality rate, between the number of trees and recruits and between the number of dead trees and recruits, and (iv) the correlation between the number N of trees and the number S of species per quadrat (electronic supplementary material, figure C1). Because our simulation models are rather simple (they can be seen as null models) we do not expect perfect fits of these patterns, but comparison of results allows us to infer on the aspects of recruitment and mortality that are critical for tropical forest dynamics. We compared the mean and the standard deviations between the simulated and observed distributions and applied Kolmogorov–Smirnov tests to quantitatively check their match.

Proc. R. Soc. B 284: 20171503

Figure 3. Correlations between changes in tree and species numbers over 25 years and the corresponding tree and species numbers in the beginning of this period. Blue dots represent observed quadrats and red dots simulated quadrats. (a) Correlation between the number of trees in quadrats observed in 1985 and the change in tree number over 25 years. (b) Same as (a), for the number of species in quadrats. (c) Same as (a), but for the simulation results of one replicate of the most complex model R2 þ M2 with a mean dispersal distance of 20 m. (d) Same as (c), but for the number of species. Panels (a) and (c) show also the DN line of the analytical model of equation (4.2) and the stable equilibrium point (black dot). (Online version in colour.)

4


change in no. trees in 25 years (DN)

(a)


DN ¼ Ntþ1 Nt ¼ rNt mNt :

ð4:1Þ

This model assumes that all recruits stem from parents in the same quadrat (i.e. strong dispersal limitation) and shows unstable dynamics. However, in our spatially explicit simulations only a certain proportion c of the offspring may remain in the home quadrat of their parents (i.e. a term þ crNt with c , 1) whereas the rest is dispersed into other quadrats. As a consequence, each quadrat will receive an additional input of recruits R from other quadrats that is independent on the local number Nt of trees. We assume here for simplicity that R does not vary in time. A generalization of model (4.1) is therefore given by DN ¼ R þ crNt mNt :

ð4:2Þ

The equilibrium of equation (4.2) is given by N ¼ R=ðm crÞ. Additionally, at equilibrium the total number of recruits must equal the mean number of recruits taken over all quadrats at the 50 ha community per quadrat R scale, such that R þ crN ¼ R:

ð4:3Þ

RÞ=rN that can This results in an estimate of c ¼ ðR be compared with the proportion of recruits remaining in the parent quadrat for different mean dispersal distances (electronic supplementary material, figure B1). The analytical model of equations (4.2) and (4.3) shows a stable equilibrium if c , 1. This is a consequence of the recruitment input R that operates independently on local tree numbers and ‘rescues’ quadrats with low tree numbers from local extinction in a way known from metapopulation theory [28].

5. Results (a) Demographic stochasticity hypothesis Our reference model (R1 þ M1), based only on demographic stochasticity and dispersal limitation, substantially underestimated the variation in the local number of trees and species (figure 2a,b; electronic supplementary material, figure C3). A mean dispersal distance of 20 m produced the best agreement in the spatial distribution of local species richness (figure 2b; electronic supplementary material, figure C3). The assumption of constant per capita mortality and recruitment rates led to a substantial underestimation of the local variability in recruitment (figure 2d) and to a moderate underestimation of the local number of dead trees (figure 2c). Thus, we find strong spatial variability in local recruitment and a moderate spatial variability in mortality.

(b) Spatio-temporal variability in recruitment Recruitment submodel R2 reproduces the observed variation in the local number of trees and species well when combined with a per capita mortality rate (i.e. submodel M1 for a mean dispersal distance of 20 m; figure 2e,f; electronic supplementary material, figure C3). Thus, spatio-temporal variability in the number of recruits beyond demographic stochasticity is a key factor in generating the observed variation in the local number of trees and species. Further data analysis of the spatial and temporal variability in the local number of recruits provides hints on the underlying mechanism. We found very high temporal variability in local recruitment among the five censuses with a mean CV of 0.63 and 2.5%, and 97.5% percentiles of 0.23 and 1.27, respectively (electronic supplementary material, figure C5a). The temporal variability was basically identical to the analogous spatial variability (electronic supplementary material, figure C5b), thus making the moving cluster hypothesis more likely for recruitment than the habitat hypothesis. Additionally, implementation of the habitat hypothesis without temporal variability produced too wide a variation in local tree numbers (electronic supplementary material, figure B2b).

(c) Spatio-temporal variability in mortality We find a positive correlation between the number of dead trees and the number of trees in a quadrat (electronic supplementary material, figure C1; r ¼ 0.62), which suggests that mortality can be described in first approximation by a constant per capita rate. Consequently, we find that model R2 þ M1 with a constant per capita mortality rate m approximates the distribution in the number of dead trees per quadrat fairly well (figure 2g). However, model R2 þ M1 produced too high a correlation between the observed numbers of trees and dead trees per quadrat (electronic supplementary material, figure C6c), suggesting that the spatial variability in local mortality cannot be ignored. Indeed, mortality submodel M2 that determines mq by random draws from the observed distribution pmort (m) improved the (already reasonable) results when combined with recruitment submodel R2 (cf. figure 2g and figure 2k), and also produced a good match of the correlation between the observed numbers of trees and dead trees per quadrat (figure C6b). However, model R2 þ M2 does not improve the prediction of the distribution of the number of trees and species in quadrats (figure 2e,f versus i,j) and the correlations between

5

Proc. R. Soc. B 284: 20171503

We develop an analytical model to explain the empirical pattern of the change in the number of trees per quadrat and time step relative to the initial number of trees (i.e. DN versus N0) (figure 3a). In classical population ecology [27] this type of relationship is frequently used and it allows us to determine the stability properties of equilibriums N* at the local quadrat scale (see electronic supplementary material, appendix B). First, we assume similar to our simulation model M1 þ R1 constant per capita mortality and reproduction rates r and m, respectively, and obtain

Our analytical model (equations (4.2) and (4.3)) approximates the relationship between DN (the 25 year change in the number of trees per quadrat) and N0 (the initial number of trees in 1985) well (figure 3a; R 2 ¼ 0.45). It allows us also to assess the proportion c of recruits that remain in the parent quadrat (^c ¼ 0:05; electronic supplementary material, appendix B), which corresponds to a mean dispersal distance of the Gaussian kernel of approximately 40 m (figure B1). Thus, recruitment at the local quadrat scale is largely independent of the local number of trees. The analytical model (equations (4.2) and (4.3)) predicts an increase in the correlation between DN and N0 with increasing mean dispersal distance. The reference simulation model (R1 þ M1) reproduced this pattern (electronic supplementary material, figure C4), but the field data show stronger correlation between DN and N0 than predicted by R1 þ M1 (figure 3a,c).


4. Analytical model for the dynamics of tree numbers at the local scale


(d) The density dependence hypothesis

(e) Results with different quadrat sizes We repeated all simulation analyses with quadrat sizes of 10 10 m2 and 50 50 m2 (see electronic supplementary material, figures C7 and C8) and found that all main findings are robust to these changes. There was only a slight change of the optimal dispersal distance that is needed to predict the variation in local species richness (between 10 m and 20 m for 10 10 m2 and approx. 40 m for 50 50 m2 quadrats). Our findings with respect to the recruitment and mortality submodels were not affected by quadrat size.

6. Discussion In this study, we provide insights into the mechanisms that drive the spatial and temporal variation of local tree density and species richness in a neo-tropical forest. For this purpose, we tested a number of biological hypotheses by using dynamic models that are built around the local spatial variation of mortality and recruitment (extracted from the census data by using 1250 non-overlapping 20 20 m2 quadrats). We then asked which of the hypotheses (i.e. model versions) are able to generate the observed spatial and temporal variation of local tree numbers, species richness and other spatio-temporal patterns extracted from the data.

(a) Biological hypotheses We find that our two main patterns, the variation in local tree numbers and species richness, do not emerge solely as a consequence of stochastic birth and dead processes and dispersal limitation as assumed by our simple reference model R1 þ M1 (the demographic stochasticity hypothesis). Thus, additional processes not included in classical neutral theory are operating at the BCI forest in generating these patterns. We identified a large spatio-temporal variability in the local number of recruits (that operated independently on local tree density) as the main missing mechanism to reproduce our main patterns, whereas variability in local per capita mortality rates contributed to a lesser extent. There are three general possibilities for the origin of the excessive variability in local recruitment: it can be spatial,

6

Proc. R. Soc. B 284: 20171503

The observed correlation between the local number of trees and the local number of recruits was low (electronic supplementary material, figure C1), and maintaining the observed covariation between number of trees and number of recruits in a quadrat (i.e. recruitment submodel R3) did not lead to any improvements with respect to submodel R2 (electronic supplementary material, figure C2). Similarly, the correlation between the per capita mortality rate and the number of trees in a quadrat was low (electronic supplementary material, figure C1; r ¼ 0.16). Thus, density dependence of demographic rates at the community level (without regard to species) had no influence on the emergence of our focal patterns.

temporal or spatio-temporal. Environmental stochasticity alone (e.g. [5,12]) (i.e. the second possibility) can be excluded because spatial variability in recruitment was large for individual censuses (electronic supplementary material, figure C9). One obvious hypothesis that corresponds to the first possibility is that differences in local habitat conditions would generate spatial variability in local recruitment (the habitat hypothesis). Indeed, several studies at BCI found associations of the distribution of species with habitat features such as topography or soil attributes (e.g. [16,29,30]). However, the habitat hypothesis did not fit the observed distribution of local tree numbers (electronic supplementary material, figure B2b), and we found that spatial and temporal variations in local recruit numbers show similar magnitudes (electronic supplementary material, figure C5a,b). This rather points to spatio-temporally varying biotic conditions as the underlying mechanism (i.e. the moving cluster hypothesis). Tree fall gaps, for example, are able to generate locally high recruit numbers by making light available that is necessary for seedling establishment and sapling growth [15,17]. Variability in seed production [31] can also contribute to the locally high recruit numbers. The lack of support for the habitat hypothesis for recruitment agrees with results of a study by Kanagaraj et al. [32] showed that local recruit assemblages at the BCI plot showed weak and temporally inconsistent habitat structuring with respect to topographic variables. Additionally, several studies have attempted to explain the variability in local species richness of large plots of tropical forests by environmental variables [13,33 –35] and found that an overwhelming proportion of this variability remained unexplained by local environmental variables. Seri et al. [11] also found very little effect of environmental heterogeneity on the dynamics of tree clusters over time. Recent modelling studies outlined the importance of environmental stochasticity in explaining observed patterns in tropical forests (e.g. [5,12]). They found considerable temporal variability in tree population abundances beyond what is expected based on neutral theory. Our study, focusing on the spatial pattern of tree and species numbers at the local scale, reaches a similar conclusion, but with respect to spatiotemporal variability in local recruitment (and, to a lesser extent, mortality). Our study calls for a better understanding of the causes of the high spatio-temporal variability in local recruitment. For example, this variability can facilitate coexistence of competing species through a storage effect by producing resource partitioning through space or time [19,36]. Usinowicz et al. [19] suggested that the strong temporal variability and asynchrony in seed and seedling production observed at BCI is able to produce storage effects. However, it is not clear if our results can be interpreted as evidence for a storage effect because submodels R2 and R3 represent strong spatio-temporal variability in the total number of recruits without regard to species, but we did not explicitly include mechanisms that desynchronize recruitment of different species as required for the storage effect. Nevertheless, local clustering of species in combination with occasional ‘recruitment pulses’ at the quadrat scale may result in spatiotemporally desynchronized recruitment of species required for the storage effect. The observed strong spatio-temporal variability in recruitment (with 21% of the quadrats having fewer than 10 recruits and 55% fewer than 19) generates for many species recruitment limitation (i.e. failure to have viable juveniles at all available


DN and N0 and DS and S0 (electronic supplementary material, figure C4) when compared with model R2 þ M1. The spatial and temporal variability in local mortality rates was substantially smaller than that of the number of recruits (electronic supplementary material, figure C5a –d). We also found that the spatial variability was very similar to the corresponding temporal variability in mortality (electronic supplementary material, figure C5c,d).


(b) Dynamics of local tree and species numbers

(c) Link with forest gap models Ecosystems—even when they can be assumed to be in a state of overall equilibrium like an old-growth forest—are constantly changing, and different parts of these systems represent

7. Conclusion Hubbell’s neutral theory [3] has been highly successful in explaining important characteristics of tropical forests—like species abundance curves. Here we showed that focusing on local dynamics can provide important additional insights on demographic mechanisms that drive the spatial and temporal variation of local species richness in a neo-tropical forest. We detected strong spatio-temporal variation in the local number of recruits, much beyond that expected by demographic stochasticity alone. This variability, when combined with a constant per capita mortality rate, was critical for predicting the variation in local tree and species numbers. Our research reveals that, despite the observed variability in mortality and recruitment, local tree numbers and local species richness seems to stabilize at typical values. However, more research is needed to investigate the causes. Investigating a high numbers of replicate plots bears high potential to improve our understanding of species-rich plant community dynamics. Data accessibility. The census data for the BCI plot are publicly available at the website of the Center for Tropical Forest Science: http://dx.doi. org/10.5479/data.bci.20130603. Use of the data has been agreed upon with the principal investigators of the BCI plot: Stephen Hubbell, Richard Condit and Robin Foster. Authors’ contributions. M.K., A.H., F.M. and T.W. conceived of the study. M.K., F.M. and T.W. analysed the data. M.K. and F.M. implemented and analysed the null models. T.W. and M.K. wrote the draft of the manuscript including contributions of A.H. and F.M. All authors discussed the results and revised the paper.

Competing interests. We have no competing interests. Funding. All authors acknowledge support from the ERC advanced grant 233066. A.H. was also supported by the Helmholtz-Alliance Remote Sensing and Earth System Dynamics.

Acknowledgments. The authors thank Friedrich Bohn, Rico Fischer, David Alonso, Jose´ Capita´n, Dominique Gravel, R. Condit, S.P. Hubbell, Fangliang He and an anonymous reviewer for comments on earlier versions of this manuscript. The BCI forest dynamics research project was founded by S. P. Hubbell and R. B. Foster, and is now managed by R. Condit, S. Lao and R. Perez under the Center for Tropical Forest Science and the Smithsonian Tropical Research in Panama. Numerous organizations have provided funding, principally the US National Science Foundation, and hundreds of field workers have contributed.

7

Proc. R. Soc. B 284: 20171503

In the field data, we found surprisingly strong negative correlations between the change in the local number of trees and species over 25 years and the corresponding initial numbers (figure 3a,b). Quadrats with a high tree or species number tend to lose trees or species, respectively, while quadrats with few trees or species tend to gain trees or species. This finding also indicates that local clusters of high tree density and high species richness may move stochastically through the plot. Our analytical model (equations (4.2) and (4.3)) for the dynamics of local tree numbers quantitatively explains the observed relationship between DN and N0. It emerges if the number of recruits at quadrat scale is only weakly associated with (or independent on) the corresponding number of trees. In the case of a density-independent constant recruit input R, this model leads to a stable equilibrium at the local quadrat scale. A recruit input R independent of the local tree number can emerge in simulation model R1 þ M1 as a consequence of a wide enough kernel that locates most recruits outside the quadrat of their parents. However, R1 þ M1 severely underestimated the variability in local recruitment numbers. Instead, the good fit of models R2 þ M1 and R2 þ M2 (figure 2) suggested that the local dynamics at BCI is driven by spatio-temporally varying local conditions that allow occasionally at some sites establishment of a high number of recruits. This puts more emphasis on the spatial and temporal pattern of ‘safe sites’ that provide just the right conditions for recruitment, a perspective shared by many gap forest models [39]. Additionally, independence between the number of recruits in a quadrat and the actual number of trees will be fostered by long time delays between seed dispersal and actual recruitment (e.g. due to seed, seedling or sapling banks). Furthermore, models R2 þ M1 and R2 þ M2 approximate the observed correlations between DN and N0 and between DS and S0 well, although the observed correlations were stronger. Previous work let us expect strong correlations between the local number of trees and species [40]. However, our best models (R2 þ M1 and R2 þ M2) with mean dispersal distances of 20–30 m produced too large correlations in the number of species and number of trees in the quadrats, and stronger species clustering due to shorter mean dispersal distance produced weaker correlations (electronic supplementary material, figure C6a). The data show that only one-third of the variation in local quadrat-scale species richness is explained by the number of individuals compared with two-thirds in our best models. This result suggests existence of additional smallscale spatial structures, beyond those created by the dispersal kernel, which make this relationship more stochastic. A candidate mechanism for this is seed dispersal by animals that may cause conspecific clumps of recruits (e.g. [41,42]).

different possible states. This idea has a long history in ecology. It was first proposed by Aubreville [43] and made popular by Richards’ ‘mosaic theory’ for forests [44]. This notion also underlies forest-gap models [39,45,46], which exhibit smallscale disturbance through gap-creation as a major driving force of forest dynamics, or individual-based Markov models of forest dynamics (e.g. [47]). Common to these approaches is the idea that different parts of a forest can be seen as partially linked and that it is the small-scale heterogeneity of these parts that drives the patterns and dynamics of forests as a whole. We used this approach to gain insight into its dynamics over time as well as the spatial distribution of local species richness and density. The actual processes governing the dynamics of tropical forests are highly stochastic, but when one uses a high number of replicates (e.g. 1250 plots) over sufficiently long time periods (25 years in our case), patterns in the fundamental characteristics of the forest emerge. Our results confirm our basic premise that the patterns of local species richness can provide additional information on the spatial and temporal dynamics of species rich communities.


sites [15,37]). Recruitment limitation can be a powerful force for maintaining diversity in species-rich communities, because it allows many sites to be won by ‘forfeit’ by species that are not the absolutely best competitor for the site [37,38].


References

2. 3.

4.

6.

7.

8. 9.

10.

11.

12.

13.

14.

15.

16.

17.

18. Runkle JR. 1989 Synchrony of regeneration, gaps, and latitudinal differences in tree species diversity. Ecology 70, 546 –547. (doi:10.2307/1940199) 19. Usinowicz J, Wright SJ, Ives AR. 2012 Coexistence in tropical forests through asynchronous variation in annual seed production. Ecology 93, 2073–2084. (doi:10.1890/11-1935.1) 20. Condit R, Lao S, Pe´rez R, Dolins SB, Foster R, Hubbell S. 2012 Barro Colorado forest census plot data (version 2012). See https://repository.si.edu/handle/10088/ 20925. (doi:10.5479/data.bci.20130603) 21. Condit R. 1998 Tropical forest census plots. Berlin, Germany: Springer. 22. Chave J, Muller-Landau HC, Levin SA. 2002 Comparing classical community models: theoretical consequences for patterns of diversity. Am. Nat. 159, 1 –23. (doi:10.1086/324112) 23. Rosindell J, Cornell SJ. 2007 Species–area relationships from a spatially explicit neutral model in an infinite landscape. Ecol. Lett. 10, 586–595. (doi:10.1111/j.1461-0248.2007.01050.x) 24. Condit R, Chisholm RA, Hubbell SP. 2012 Thirty years of forest census at Barro Colorado and the importance of immigration in maintaining diversity. PLoS ONE 7, e49826. (doi:10.1371/journal.pone.0049826) 25. Brown C, Law R, Illian JB, Burslem DFRP. 2011 Linking ecological processes with spatial and non-spatial patterns in plant communities. J. Ecol. 99, 1402– 1414. (doi:10.1111/j.1365-2745.2011.01877.x) 26. May F, Wiegand T, Lehmann S, Huth A. 2016 Do abundance distributions and species aggregation correctly predict macroecological biodiversity patterns in tropical forests? Glob. Ecol. Biogeogr. 25, 575 –585. (doi:10.1111/geb.12438) 27. Vandermeer JH, Goldberg DE. 2013 Population ecology: first principles, 2nd edn. Princeton, NJ: Princeton University Press. 28. Gotelli NJ. 1991 Metapopulation models: the rescue effect, the propagule rain, and the core-satellite hypothesis. Am. Nat. 138, 768–776. (doi:10.1086/ 285249) 29. Comita LS, Condit R, Hubbell SP. 2007 Developmental changes in habitat associations of tropical trees. J. Ecol. 95, 482 –492. (doi:10.1111/j. 1365-2745.2007.01229.x) 30. John R et al. 2007 Soil nutrients influence spatial distributions of tropical tree species. Proc. Natl. Acad. Sci. USA 104, 864– 869. (doi:10.1073/pnas. 0604666104) 31. Wright SJ, Muller-Landau HC, Calderon O, Hernandez A. 2005 Annual and spatial variation in seedfall and seedling recruitment in a neotropical forest. Ecology 86, 848 –860. (doi:10.1890/03-0750) 32. Kanagaraj R, Wiegand T, Comita LS, Huth A. 2011 Tropical tree species assemblages in topographical habitats change in time and with life stage. J. Ecol. 99, 1441–1452. (doi:10.1111/j.1365-2745.2011.01878.x) 33. Legendre P, Mi X, Ren H, Ma K, Yu M, Sun I-F, He F. 2009 Partitioning beta diversity in a subtropical

34.

35.

36.

37.

38.

39.

40.

41.

42.

43.

44.

45.

46.

47.

broad-leaved forest of China. Ecology 90, 663 –674. (doi:10.1890/07-1880.1) Baldeck CA et al. 2013 Habitat filtering across tree life stages in tropical forest communities. Proc. R. Soc. B 280, 20130548. (doi:10.1098/rspb. 2013.0548) Punchi-Manage R, Wiegand T, Wiegand K, Getzin S, Gunatilleke CVS, Gunatilleke IAUN. 2014 Effect of spatial processes and topography on structuring species assemblages in a Sri Lankan dipterocarp forest. Ecology 95, 376–386. (doi:10.1890/12-2102.1) Chesson PL, Warner RR. 1981 Environmental variability promotes coexistence in lottery competitive systems. Am. Nat. 117, 923– 943. (doi:10.1086/283778) Hurtt GC, Pacala SW. 1995 The consequences of recruitment limitation: reconciling chance, history and competitive differences between plants. J. Theor. Biol. 176, 1–12. (doi:10.1006/jtbi.1995.0170) Tilman D. 1994 Competition and biodiversity in spatially structured habitats. Ecology 75, 2–16. (doi:10.2307/1939377) Shugart HH. 1998 Terrestrial ecosystems in changing environments. Cambridge, UK: Cambridge University Press. Condit R, Hubbell SP, Lafrankie JV, Sukumar R, Manokaran N, Foster RB, Ashton PS. 1996 Speciesarea and species-individual relationships for tropical trees: a comparison of three 50-ha plots. J. Ecol. 84, 549–562. (doi:10.2307/2261477) Howe HF. 1989 Scatter-and clump-dispersal and seedling demography: hypothesis and implications. Oecologia 79, 417– 426. (doi:10.1007/BF00384323) Kwit C, Levey DJ, Turner SA, Clark CJ, Poulsen JR. 2007 Out of one shadow and into another: causes and consequences of spatially contagious seed dispersal by frugivores. In Seed dispersal: theory and Its application in a changing world (eds AJ Dennis, EW Schupp, RJ Green, DA Westcott), pp. 427 –444. Wallingford, UK: CABI. Aubreville A. 1937 Contribution a l’e´tude des re´serves naturelles et des parcs nationaux. Paris, France: Paul Lechevalier. Richards PW. 1945 The floristic composition of primary tropical rain forest. Biol. Rev. 20, 1 –13. (doi:10.1111/j.1469-185X.1945.tb00310.x) Shugart HH, Smith TM. 1996 A review of forest patch models and their application to global change research. Clim. Change 34, 131 –153. (doi:10.1007/ BF00224626) Fischer R et al. 2016 Lessons learned from applying a forest gap model to understand ecosystem and carbon dynamics of complex tropical forests. Ecol. Model. 326, 124– 133. (doi:10.1016/j.ecolmodel. 2015.11.018) Brisson J, Bergeron Y, Bouchard A, Leduc A. 1994 Beech-maple dynamics in an old-growth forest in southern Que´bec, Canada. E´coscience 1, 40 –46. (doi:10.1080/11956860.1994.11682226)

Proc. R. Soc. B 284: 20171503

5.

Brown JH. 1981 Two decades of homage to Santa Rosalia: toward a general theory of diversity. Am. Zool. 21, 877–888. (doi:10.1093/icb/21.4.877) Carson W, Schnitzer S. 2011 Tropical forest community ecology. New York, NY: John Wiley & Sons. Hubbell SP. 2001 The unified neutral theory of biodiversity and biogeography. Princeton, NJ: Princeton University Press. Volkov I, Banavar JR, Hubbell SP, Maritan A. 2003 Neutral theory and relative species abundance in ecology. Nature 424, 1035 –1037. (doi:10.1038/ nature01883) Chisholm RA et al. 2014 Temporal variability of forest communities: empirical estimates, of population change in 4000 tree species. Ecol. Lett. 17, 855–865. (doi:10.1111/ele.12296) Rosindell J, Cornell SJ. 2009 Species-area curves, neutral models, and long-distance dispersal. Ecology 90, 1743 –1750. (doi:10.1890/08-0661.1) May F, Huth A, Wiegand T. 2015 Moving beyond abundance distributions: neutral theory and spatial patterns in a tropical forest. Proc. R. Soc. B 282, 20141657. (doi:10.1098/rspb.2014.1657) Bell G. 2001 Neutral macroecology. Science 293, 2413–2418. (doi:10.1126/science.293.5539.2413) Etienne RS. 2005 A new sampling formula for neutral biodiversity. Ecol. Lett. 8, 253 –260. (doi:10. 1111/j.1461-0248.2004.00717.x) Azaele S, Pigolotti S, Banavar JR, Maritan A. 2006 Dynamical evolution of ecosystems. Nature 444, 926–928. (doi:10.1038/nature05320) Seri E, Maruvka YE, Shnerb NM. 2012 Neutral dynamics and cluster statistics in a tropical forest. Am. Nat. 180, E161 –E173. (doi:10.1086/668125) Kalyuzhny M, Kadmon R, Shnerb NM. 2015 A neutral theory with environmental stochasticity explains static and dynamic properties of ecological communities. Ecol. Lett. 18, 572–580. (doi:10.1111/ele.12439) He F, Legendre P, LaFrankie J. 1996 Spatial pattern of diversity in a tropical rain forest in Malaysia. J. Biogeogr. 23, 57 –74. (doi:10.1046/j.1365-2699. 1996.00976.x) Wiegand T, Gunatilleke C, Gunatilleke I, Huth A. 2007 How individual species structure diversity in tropical forests. Proc. Natl. Acad. Sci. USA 104, 19 029–19 033. (doi:10.1073/pnas.0705621104) Hubbell SP, Foster RB, O’Brien ST, Harms KE, Condit R, Wechsler B, Wright SJ, de Lao SL. 1999 Light-gap disturbances, recruitment limitation, and tree diversity in a neotropical forest. Science 283, 554–557. (doi:10.1126/science.283.5401.554) Harms KE, Condit R, Hubbell SP, Foster RB. 2001 Habitat associations of trees and shrubs in a 50-ha neotropical forest plot. J. Ecol. 89, 947 –959. (doi:10.1111/j.1365-2745.2001.00615.x) Augspurger CK. 1984 Seedling survival of tropical tree species: interactions of dispersal distance, lightgaps, and pathogens. Ecology 65, 1705 –1712. (doi:10.2307/1937766)


1.

8

Seed dispersal increases local species richness and reduces spatial turnover of tropical tree seedlings.

Estimating species richness and modelling habitat preferences of tropical forest mammals from camera trap data.

Historical dynamics and current environmental effects explain the spatial distribution of species richness patterns of New World monkeys.

Species richness and biomass explain spatial turnover in ecosystem functioning across tropical and temperate ecosystems.

Estimating the spatial and temporal distribution of species richness within Sequoia and Kings Canyon National Parks.

Spatial genetic structure in four understory Psychotria species (Rubiaceae) and implications for tropical forest diversity.

Human influence on the temporal dynamics and spatial distribution of forest biomass carbon in China.

Structural Dynamics of Tropical Moist Forest Gaps.

Controls of Soil Spatial Variability in a Dry Tropical Forest.

Sandfly distribution and abundance in a tropical rain forest.

Patterns of sandfly distribution in tropical forest: a causal hypothesis.

Mixed-forest species establishment in a monodominant forest in central Africa: implications for tropical forest invasibility.

Abundance of common species, not species richness, drives delivery of a real-world ecosystem service.

Historical and ecological drivers of the spatial pattern of Chondrichthyes species richness in the Mediterranean Sea.

Otolith microchemistry of tropical diadromous fishes: spatial and migratory dynamics.

Spatial congruence in language and species richness but not threat in the world's top linguistic hotspot.

Revisiting spatial scale in the productivity-species richness relationship: fundamental issues and global change implications.

Does species richness affect fine root biomass and production in young forest plantations?

Hydrological networks and associated topographic variation as templates for the spatial organization of tropical forest vegetation.

Impact of Forest Management on Species Richness: Global Meta-Analysis and Economic Trade-Offs.

Spatial and environmental correlates of species richness and turnover patterns in European cryptocephaline and chrysomeline beetles.

Limited sampling hampers "big data" estimation of species richness in a tropical biodiversity hotspot.

Relative importance of current and past landscape structure and local habitat conditions for plant species richness in dry grassland-like forest openings.

Data gaps in anthropogenically driven local-scale species richness change studies across the Earth's terrestrial biomes.