Identifying climatic drivers of tropical forest dynamics

Introduction Conclusions References Tables Figures


Introduction
Tropical forests are characterized by high annual precipitation and high evapotranspiration.Nevertheless, strong seasonal variations in rainfall inputs, partly driven by atmospheric movements related to the monsoon or latitudinal changes in the inter-tropical conversion zone, occur in most tropical regions around the world (Feng et al., 2013).Such seasonality implies various changes of the availability of resources, such as water and light, necessary to tree development and to forest functioning.The seasonality of tree growth and tree mortality is increasingly studied in tropical forests, with some studies having succeeded in linking sea-sonal tree demography to climate seasonality (Wagner et al., 2012;Grogan and Schulze, 2012;Brando et al., 2010).Tree growth is mainly related to water availability, resulting in growth during the wet months and static or even contracted states during the dry season months (Grogan and Schulze,35 2012).The use of a convenient water availability proxy like the relative extractable water (REW) (Wagner et al., 2011) shows that low levels of REW rather than lack of rainfall per se are the key drivers of the decrease in growth rate, or even of the stop, at a seasonal time step.(Wagner et al., 2012).

40
At another time scale, long-term forest dynamic changes may also be related to exceptional climate events.Effects of unusual dry periods on tree growth and mortality may enlighten us about the long-term processes linking water availability and tree dynamics.After the intense 2005 45 drought in Amazonia, the forest suffered an additional mortality, leading to a huge loss of alive tree biomass (Phillips et al., 2009).Similar major mortality events were observed in Panama (Condit, 1995), in chinese rainforests (Tan et al., 2013) or in South-East Asia (Slik, 2004).Water 50 exclusion experiments in Brazil provide results in line with a deep impact of drought on tree mortality (Nepstad et al., 2007;da Costa et al., 2010;Brando et al., 2008).Between the time scale of exceptional events and the time-scale of intra-annual seasonal rhythmicity, there is a gap in our 55 knowledge on the inter-annual scale.This gap is partly due to the weak magnitude of variation of the demographic rates when compared to what is observed from a seasonal point of view or to some spectacular events.This gap is also due to the lack of sites in tropical forests where annual 60 regular inventories of tree growth and death are performed and where precise climatic data on the same time-scale are available.Moreover, the potential links between inter-annual climate variations and tropical forest dynamics should be studied from a multi-decadal long-term perspective in order to be representative of the climatic variability and of the variability of forest dynamic responses (Clark et al., 2010).Some climatic variables (mainly water stress, water saturation and temperature) are expected to play a role in forest dynamics regarding the tree's physiological processes.Water stress due to drought is well documented (Phillips et al., 2009;Allen et al., 2010).Water insufficiency leads generally to higher mortality rates and lower growth (Choat et al., 2012).Water stress needs to be estimated, and diverse estimators may be found in the literature (Wagner et al., 2011;Toledo et al., 2011;Aragão et al., 2007;Malhi et al., 2009).The length of the dry season seems to be the simplest estimator.The relative extractable water (REW) described in the study of Wagner et al. (2011) estimates the quantity of water available for tree development and has been proved to be highly performant to predict intra-annual forest dynamics in Wagner et al. (2012).Although water availability is expected to reduce growth and increase mortality, these impacts have to be investigated on an inter-annual timescale.Rain may also be responsible for water saturation, a phenomenon that is far less studied but that can have an effect on tree mortality or growth.For instance, Ferry et al. (2010) underlined a higher mortality rate in waterlogged areas.Inter-annual variations of rain quantities can lead to more or less waterlogged soils, independent of their topographical location, implying instability that can cause cascading tree-falls.The effects of temperature are less consensual; some studies suggested that tropical forests can be near a high temperature threshold and that these systems may be more vulnerable to climate change than previously believed (Clark et al., 2003).For instance, Clark et al. (2003) showed a negative correlation between 16-year diameter increments and annual means of daily minimal temperature in La Selva, Costa Rica, while Toledo et al. (2011) found a positive correlation between annual diameter growth and temperature in Bolivia.An explanation for such apparently conflicting results was proposed by Dong et al. (2012), that the effects of variability in solar radiation and daily minimum temperature on tree growth appear to be largely independent.
In this study, we use a modelling approach in order to mechanistically link climate conditions and functional plant traits to tree growth and survival (Zuidema et al., 2013).
Functional traits have been recently used to include functional diversity in models of tree growth (Hérault et al., 2011;Rüger et al., 2012;Wagner et al., 2014) and tree mortality (Aubry-Kientz et al., 2013).We first question the potential relationships existing between climate variables computed on two-year time step and forest dynamics.We identify independent variable responsible for the inter-annual variation of growth and mortality rates.These variables are then included in a coupled growth-mortality model to test their multivariate effects.Finally, we include in the model some interac-120 tions between functional traits (wood density and tree size) and climate predictors to test for a potential differentiated response depending on the individual functional identity.First, tree species having high wood density have been reported to better resist drought events as compared to lower density 125 ones (Phillips et al., 2010).Part of these differences is related to differences in hydraulic failure, as wood density is linked to xylem structure.Second, the current tree size also influence resistance to drought events or other climatic perturbations (Nepstad et al., 2007;Condit et al., 2004).Two 130 main hypotheses are debated.First, small, young trees that are not well established and that do not have deep roots may be more sensitive and may suffer under stressful water conditions.Second, large, older trees may feel water stress because they must maintain their photosynthesis activities and carry 135 sap to a higher altitude.

Data Collection
Three datasets were used in this study.The study site is located in Paracou, French Guiana(5 18'N, 52 55'W).The 140 forest is typical of Guianan rain forests and the dominant tree families are Fabaceae, Chrysobalanaceae, Lecythidaceae, and Sapotaceae.More than 700 species of trees ≥10 cm DBH (diameter at breas height) have been described at the site.

145
Mean annual precipitation averages 2980 mm (30-year period), and the site receives nearly two-thirds of its annual precipitation during the long rainy season between mid-March and mid-June (Wagner et al., 2011), and less than 100 mm per month from August to November (Figure 11).

Tree dynamic
The first data set is an inventory of trees >10 cm DBH in the six natural forest plots of 6.25 ha in Paracou.Mortality and diameter growth have been censused every two years between 1991 and 2011.DBH was calculated from 155 circumference measures made to a precision of 0.5 cm.We excluded individuals with buttresses or other problems that required an increase in measurement height because we were unsure about the height of the initial points of measurement for these trees.The data set contained 20,340 160 trees from 642 species.For each tree and every two years, we know the location, DBH, vernacular name, status (dead or alive), and the mode of death for dead trees (tree-fall or standing death).Vernacular names are the common names used by local tree spotters.As botanical identification of 165 the trees species was completed in 2012, a large part of the trees that died during the study period  have only a vernacular name and no botanical determination.The method of Aubry-Kientz et al. (2013) is used to handle this uncertainty and to integrate the information on botanical determination contained in the vernacular names of trees that were not identified.

Functional traits
The second data set was a collection of five functional traits of 335 Guianan tree species that occur at the Paracou site (Table 11).These 335 species represent 79% of the total number of individual trees included in this study.We used the procedure described in Aubry-Kientz et al. (2013) to assign functional trait values to trees for which (i) the species is known but trait values were not available, (ii) the species was not determined at the species level and (iii) the tree was dead before being identified.Traits are related to leaf economics, stem economics and life history and are extracted from a large database (Baraloto et al., 2010a, b).
The leaf economics reflects a trade-off between investments in productive leaves with rapid turnover versus costly physical leaf structure with a longer payback.The stem economics defines a similar trade-off at the stem level: dense wood versus high wood water content and thick bark (Baraloto et al., 2010b).Life-history strategies describe how trees allocate resources to different organs and how these allocation translate into a species' ability to compete for resources and finally to grow, survive, reproduce and disperse (Rüger et al., 2012).Some of these functional traits are accurate proxies of growth trajectories (Hérault et al., 2010(Hérault et al., , 2011) ) and mortality rates (Aubry-Kientz et al., 2013).

Climate
The third data set consists of climate data (Table 12).Six variables were provided by the Climatic Research Unit (CRU) at the University of East Anglia (Mitchell and Jones, 2005), consisting in month-by-month variations in climate over the last century calculated on high-resolution grids (0.5*0.5 degree) (Mitchell and Jones, 2005).We used the aggregated variables (mean or sum, depending of the nature of the observed process) for two years, from July to July, to include the dry season (mid-August to mid-November).Selected variables that may have an impact on forest dynamics are the cloud cover (Cld), the potential evapo-transpiration (P et), the precipitation (P re), the daily mean temperature (T mp), the vapour pressure (V ap) and the wet day frequency (W et).
Three other climate variables were computed using the relative extractable water (REW) computed with a water balance model developed by Wagner et al. (2011) calibrated at our study site and taking the daily precipitation from the CRU into account; this REW index takes values between 0 and 1 at our study site, corresponding to the available water for trees.This REW index is used to compute N b under , the number of days under a REW threshold of 0.4, which is the 220 threshold recommended in Wagner et al. (2011); A under , the area over the REW curve and under the threshold of 0.4; and A over , the area situated under the REW curve and over the threshold of 0.95.N b under and A under are built to be indicators of drought, while A over is related to soil water sat-225 uration.All climate variables are centred to allow an easier interpretation of the results.

Model
The model used in this study consists of a model coupling growth and mortality processes at the whole community 230 scale.The model is build taking advantage of two preliminary studies where the growth (Hérault et al., 2011) and the mortality (Aubry-Kientz et al., 2013) sub-models were developped.The likelihood is computed using the distribution probability of mortality (equations 3 and 4) and the computed 235 growth rate (equations 5 and 6).A vigour index is added into the mortality process, taking the past growth of the two previous year into account.We added the climate variables into the two processes to highlight the links between some climate drivers and one particular process.Because the final forest 240 dynamic model was not linear, we build a MCMC algorithm under a bayesian framework to infer the parameter posterior distributions.Growth and mortality processes were linked through tree vigour and are parametrized simultaneously.If tree i stays alive, it grows at a growth rate AGR i,s,t , and 245 its diameter DBH i,t−1 becomes DBH i,t .The joint model likelihood is then if tree i stays alive during the length of the studied period, and with εi ∼ N (0, θ13).
where p i,s,t is the probability of dying of tree i of species s between time t−1 and t; AGR i,s,t−1 is the predicted growth between time t − 2 and time t − 1. AGR i,s,t−1 is the observed growth between time t−2 and time t−1; DBHmax s , Hmax s , W D s , T ough s and δ13C s are functional traits of species s to which tree i belongs (Table 11); θ 1 , θ 2 , • • • θ 13 are parameters to be estimated, and ε i is an individual error term following a normal distribution; γ 1 and γ 2 are the parameter vectors linking the climate predictors with the processes of mortality and growth respectively; clim 1 and clim 2 are the vectors of climate predictors included in the processes of mortality and growth, respectively.

Variable selection
To identify the different axes of variation of our climate data set and avoid including collinear variables in the model, we realized a principal component analysis (PCA) on the climate variables.We included all climate variables one by one in each process of the model and computed the partial likelihood for each sub-model of growth or mortality we obtained.This provides a first result about the importance of each climate variable.Depending on these results and on their degree of collinearity from the PCA, we selected some climate variables and included them in the growth model and in the logit function of mortality.

Model inference
We implemented a Markov Chain Monte Carlo algorithm to estimate the model parameters (Robert and Casella, 2004).

305
A random walk was used as a proposal distribution to sample new values of parameters that were or were not selected, using the ratio of Metropolis-Hasting.Only standard deviation was sampled in an inverse-gamma posterior distribution with a Gibbs sampler.The functional traits used as demo-310 graphical predictors were uncertain because botanical determination was incomplete for the older censuses, and not all values of functional traits were available for all species.We used the method developed in Aubry- Kientz et al. (2013) to handle these uncertainties.All the algorithms and statistical 315 treatments were implemented with R software (core Team, 2014).

Functional trait and forest dynamic responses
Functional traits were introduced in the final model with an interaction term by multiplying a climatic variable with a 320 functional trait.We did not test all possible interactions but, based on results from a literature survey, we we investigated biological-meaningful interactions only (Table 13).We included in the model an interaction between wood density and the drought estimator A under , an interaction between DBH 325 and A under , and an interaction between DBH and precipitation P re.
Species vary over one order of magnitude in their wood density (W D), ranging from 0.08 to 1.39 g.cm 3 (Iida et al., 2012), and the encountered range of wood density is partic-330 ularly large in species-rich tropical rainforests (Chave et al., 2006(Chave et al., , 2009)).Wood density is a key functional trait because of its importance for mechanical stability, defence against herbivores, hydraulic conductivity, photosynthetic carbon gain and diameter growth rates of plants (Poorter et al., 2008).

335
High wood density implies thin and short xylem vessels with small pit-pores, which decrease the risk of embolism and cavitation.Trees with high wood density are then expected to be less sensitive to drought.The term A under multiplied by (W D max − W D) accounts for the effect of drought on 340 trees with low wood density.This term is added in growth and mortality to test this effect (Table 13).

Variable selection
The variable selection was realised using the literature, the 345 PCA results, and the results of the univariate analysis.

PCA
The PCA underlines one principal axis, explaining 46% of the inertia and strongly negatively correlating with variables T mp and P et.The variables W et and Cld are positively correlated with this axis, while V ap, A under and N b under are negatively correlated with this axis (Figure 12).The second axis (20%) is strongly negatively correlated with P re and Area over .The third axis (12%) is essentially negatively correlated with A under .

Univariate analyses
When the climate variables are included one by one in each model, all climate variables but precipitation (P re) had an effect in the growth process, while only few had an effect in the mortality process (Table 14).The climates variables associated with the mortality process are P re, N b under and A under .In the growth model, A under is the best predictor according to the likelihood.In the mortality process, the best value of likelihood is obtained when N b under is included.

Variable selection
The Pet and temperature are indicators of the energy that the system receives and are expected to play a role in tree growth (Clark et al., 2003;Dong et al., 2012).These variables are strongly correlated (r=0.8) and negatively correlated with the first axis of the PCA (P et, C=-0.45 and T mp, C=-0.44).This is not surprising, as Pet is computed using the temperature (Allen et al., 1998).As these two variables are strongly correlated, we finally included only temperature, which had a better likelihood score than P et when it is included in the growth model.Neither P et nor the temperature had an effect if included in the mortality process.The second axis of the PCA is related to water saturation and is correlated with P re (C=-0.68)and A over (C=-0.61).
A over only had an effect when included in the growth process.However, both the effect size and the likelihood (Table 14) were the worst score obtained so that we did not include this variable in the final model.Concerning mortality, P re had a clear effect (Figure 13) and is thus included as a proxy of water saturation in the final mortality model.The third axis of the PCA is strongly correlated with the drought estimator A under , which is the better climate driver of growth regarding the likelihood and the effect size.A under also had an effect on the mortality process, and is finally included in the two processes in the final model.

Full model inference
The growth trajectory was adjusted by a size-dependent diameter growth model (Figure 14).Parameters linking the maximal growth to the functional traits DBHmax, W D, Hmax and δ13C have have similar values and interpretations to Hérault et al. (2011), i.e. maximum growth rates increase with increasing DBHmax, and decreasing W D, 400 Hmax and δ13C (Table 15).Maximum growth rate is attained for a tree diameter equal to 0.794*DBHmax.The parameters linking the probability of mortality to Hmax, W D and T ough converged around negative values, meaning that the probability of dying is lower when the tree is high, has a 405 high wood density and/or high laminar toughness.The drought estimator (A under ) converged to negative values in the growth and mortality processes; thus growth and mortality computed at our biannual time-scale are lower when the drought estimator A under is higher.The parameter linking mortality with precipitation (P re) is positive.This finding implies that mortality rate is higher during two-year timescale with high precipitation.In our dataset, the highest total precipitation was, albeit non-significantly, rather related to the highest proportion tree-fall deaths (Figure 15).The pa-415 rameter linking temperature (tmp) and growth takes negative values; thus growth values are lower during the warmest periods.

Functional variability of responses
In the growth process, interaction between (W D max −W D) 420 and drought is negative (Table 13), implying that trees with lower W D are more sensitive to drought and reduce their growth more.Moreover, interactions linking the current diameter and drought are also negative; thus larger trees are more sensitive to drought and reduce their growth more com-425 pared to smaller trees.None of the interaction terms included in the mortality process had an effect (Table 13).

Discussion
In this study, we questioned the importance of the climate drivers of tropical forest dynamics by using a community 430 growth-mortality modeling framework.First, one can note that few climate variables had an univariate effect when included in the mortality process, while almost all had an univariate effect in the growth process.However, the magnitude of the impact of climate variables is stronger 435 in the mortality process (observed mortality rate varying between 1.6 and 2.5% of mortality/2 years, while observed growth rates vary between 1.9 and 2.5 mm/2 years, Figure 13).Next, we developed bayesian algorithms to infer the multivariate nonlinear model and select the best predictors 440 with a great flexibility.We found that drought decreased annual growth and mortality rates, high precipitation through soil water saturation increased mortality rates and high temperature decreased growth (Figure 14).We confirmed that the vigour index is negatively related to mortality, i.e., 445 trees that grow more than expected have a lower probability of dying, and trees with lower-than-expected growth have a higher probability of dying.Moreover, the posterior values for obtained the functional trait parameters are coherent with results of Hérault et al. (2011) and Aubry-Kientz et al. (2013), increasing our confidence in (i) the developed algorythm and (ii) the biological determinisms of the ecological processes we want to model.This confirms that the functional trait-based approach could be successfully used to predict climate-induced tree dynamics in highly diverse tropical forests for which taxonomic data may be lacking but functional trait data are available.A limited number of interactions between climate variables and functional traits was tested because of our selection of three climate predictors.One can argue that some climatic variables that were disregarded in the first selection step would increase the likelihood if included in interactions with a functional trait.This pathological case is very improbable (Wagner et al., 2014) and will necessitate an impractical amount of computational time to be tested.

Water stress
The water stress during the dry season, estimated with A under , negatively impacts the growth and mortality processes.Trees will thus grow less quickly and have a lower probability of dying during two-year periods with the most intense dry seasons.The reduction of growth is expected, and has many ecophysiological causes.Indeed, water is essential for sap fluxes and for photosynthesis efficiency.The reduction of growth is furthermore linked with the current DBH and the species' wood density (Table 13).Big trees are more sensitive to water stress than small trees.This was expected in light of the results obtained after rainfall exclusion (da Costa et al., 2010).Indeed, maintenance costs are higher for big trees, making these trees more vulnerable to the driest periods.Regarding the wood density, species with high values are more resistant to drought.This is consistent with our hypothesis that high wood density implies thin and short xylem vessels and thus decreases the risk of embolism and cavitation.As the ability of trees to recover from periods of sustained drought is strongly related to their embolism resistance (Choat et al., 2012), a tree with high wood density will be more able to maintain growth during dry years.For similar reasons, we expected a positive impact of A under on mortality rates.Experimental trough-fall exclusions conducted in Tapajos and Caxiuaña indeed demonstrated that 50% rainfall exclusion led to very high mortality rates (Nepstad, 2002).Our results show no positive effect of drought intensity on mortality rates (Table 14) and look contradictory to Nepstad (2002).However, the natural variability of the drought intensity (total rainfall from 5486 to 6207 mm) in our dataset is hardly comparable to the experimental 50% reduction in total rainfall.Moreover, our modeling framework prevented us from seeing long-term effects induced by repeated drought events because the drought variable values depend only on the last 2-year climate.One may also expect that standing death is more frequent during the driest periods but, when plotting tree mode of death against drought estimator (A under ), no evidence was observed for a potential trend (Figure 15).To 505 conclude, our results confirmed that the relationship between drought and mortality may be challenging to estimate and to link with their underlying causes at an inter-annual time scale.

Water saturation
Water saturation P re had a strong effect on mortality; mortality rate varied between 1.5 and 2% per 2-years with increasing total precipitation.This is consistent with the hypothesis that trees are more vulnerable when the soil is water 515 saturated.In the Paracou forest, about half of tree deaths are due to standing death and half to tree-fall.This ratio looks, albeit non-significantly (R 2 =0.61,P=0.08) because of the low number of observations (n=9)), linked with total precipitation.The highest total precipitation led to the highest pro-520 portion tree-fall deaths (Figure 15).This confirms the observation of Ferry et al. (2010) and the hypothesis that waterlogged soils in space or in time are risky for trees.Moreover, during the rainy season, strong rainfall events often come with strong winds that may accelerate this process (Toledo 525 et al., 2011).Studies observing a relationship between tree mortality and excess of water in the soil primarily focus on geographical variation (Ferry et al., 2010;de Toledo et al., 2012) and conclude that excess water in the soil restricts root establishment because productivity of fine roots and rooting 530 depth are generally low in sandy soils and soils with high moisture content.Our results highlight that the time variation in soil water saturation is also very important and should be reassessed.

535
Temperature is identified as predictor of trees' decreasing growth.As the temperature rises, the velocity of reacting molecules increases, leading to more rapid reaction rates but also to damage of the tertiary structures of the enzymes and reduced enzyme activity and reaction rates (Fitter and Hay,540 2001; Lloyd and Farquhar, 2008).These two processes are responsible for a bell-shaped curve of growth response to temperature (Fitter and Hay, 2001).Temperature can affect photosynthesis through modulation of the rates of activity of photosynthetic enzymes and the electron transport chain, and 545 in a more indirect manner, through leaf-temperatures defining the magnitude of the leaf-to-air vapour pressure difference, a key factor influencing stomatal conductances (Lloyd and Farquhar, 2008).In tropical forests, as temperatures are already high, rising temperatures may imply lower growth, 550 consistent with results from Clark et al. (2003).This temperature effect may become the most problematic for tropical forest dynamics, considering the rising tempera-tures that are predicted, with a great degree of certainty, by climate models for the next century (Stocker et al., 2013).Indeed, as temperature was identified as a strong predictor of growth, all else being equal, averaged community growth and forest productivity may consequently decrease in time.This decline in productivity in time is perhaps what we are starting to see throughout the Amazon (Brienen et al., 2015).
As no consensus has been reached yet, additional studies using regular inventories are urgently needed (Reed et al., 2012;Corlett, 2011) to explain the conflicting patterns of the temperature effect found in the extant literature (Dong et al., 2012).Finally, we need to acknowledge that we do not know much about how forest dynamics will behave in the next century under temperature conditions that will be so different from what is actually observed.In this context, manipulative warming experiments are increasingly vital to better predict the future of tropical forest dynamics (Cavaleri et al., in press).

Conclusions
Global climate models converge to simulate, at least for the Amazonian region, a change in precipitation regime and temperature conditions over the coming decades (Malhi et al., 2009).Drought is expected to become longer and stronger in the future (Joetzjer et al., 2013) and the temperature will continue rising drastically during the next century (Stocker et al., 2013).Our modelling framework allows us to study interannual variations of climatic variables and identify which of these climatic variables are the key drivers of tropical forest dynamics.Drought, precipitation and temperature were highlighted as strong drivers of tree growth and/or mortality.Drought decreased annual growth and mortality rates, high temperature decreased growth and high precipitation events increased mortality rates.Moreover, we demonstrated best resistance to drought for trees with high wood density and for trees with small current diameters, giving us some possible indications on the future composition of a tropical forest where droughts are becoming more frequent.In light of these results, raising awareness of the current impacts of climate changes on tropical forest dynamics is urgent.
Appendix A: Growth and mortality simulations for Oxandra an Hevea Simulations presented in Figure 14 are realized using median values for tree functional traits.These median values do not have any ecological meaning, and the figure was realized only to show how climatic drivers impact the tree growth and mortality in reality (Figure 13) and in our model (Figure 14).To show more realistic simulations, the same patterns are plotted for two species that differ in their ecological strategies in Figure 16.The first column shows the simulated dynamics of Oxandra asbeckii, a relatively small tree.
The second column shows the simulated dynamics of Hevea guianensis, which is a canopy tree reaching heights of 50 me-605 ters and which has a low wood density.These two strongly contrasting species show two different growth and mortality rates, although the effects of climatic drivers stay the same.
Project (European structural funding, PO-feder).The funders had no role in study design, data collection and analysis, decision to publish or preparation of the manuscript.This work also benefited from an 'Investissement d'Avenir' grant managed by the Agence Nationale de la Recherche (CEBA, ref ANR-10-LABX-0025) and from a grant from the Centre de Coopération Internationale en Recherche Agronomique pour le Développement.Table 11.The five functional traits used in the growth-mortality model.Descriptions of the traits, abbreviations used in this study and ranges observed in our data set.Growth rises with reduced temperature and reduced water stress.This is more noticeable for large values of DBH/DBHmax, which means large, old trees.Mortality (% per 2 years) is computed with varying precipitation (third line) and with varying water stress (fourth line) and is plotted against the ontogeny (DBH/DBHmax).Mortality rate rises with rising precipitation and reduced water stress. 150 510

Figure 11 .
Figure 11.Ombrothermic diagram of the Paracou forest, data from the 2001-2014 time period (precipitation in m) on the left, temperature in ˚C on the right

Figure 13 .
Figure 13.Climatic drivers of tree dynamics.Observed mean growth (mm / 2 years) is plotted against temperature (a) and against the water stress (b).Observed mortality rate (proportion / 2 years) is plotted in abscissa against precipitation (c) and against the water stress (d).

Figure 14 .
Figure14.Climatic drivers of tree dynamics.Simulations are made using median values for tree functional traits.Growth (in mm / 2 years) is computed with varying temperature (a) and with varying water stress (b) and is plotted against the ontogeny (DBH/DBHmax).Growth rises with reduced temperature and reduced water stress.This is more noticeable for large values of DBH/DBHmax, which means for large, old trees.Mortality (% per 2 years) is computed with varying precipitation (c) and with varying water stress (d) and is plotted against the ontogeny (DBH/DBHmax).Mortality rate rises with rising precipitation and reduced water stress.This illustration clearly shows the effects of climate variables and ontogeny on tree growth and mortality, but the median functional traits used do not represent a real 'mean' tree.To evaluate more precisely the dynamics for two different species, we plotted the same curves for Oxandra Asbeckii and Hevea guianensis in Appendix A.

Figure 15 .
Figure15.Proportion of dead trees caused by tree-fall plotted against the climate variable P re (a) and proportion of dead trees caused by standing death plotted against the climate variable A under (b).About 50% of tree deaths are tree-fall; this proportion is quite higher but not significant (F-statistic test, P=0.079) during 2-yr periods with high precipitation.No significant correlation (Fstatistic test, P=0.814) between the mode of death and the drought intensity A under was noted.

Figure 16 .
Figure16.Predictions of growth and mortality depending of climatic drivers for Oxandra asbeckii and Hevea guianensis.Simulations are made using the functional traits values of the species Oxandra asbeckii (left) and Hevea guianensis (right).Growth (in mm / 2 years) is computed with varying temperature (first line) and with varying water stress (second line), and is plotted against the ontogeny (DBH/DBHmax).Growth rises with reduced temperature and reduced water stress.This is more noticeable for large values of DBH/DBHmax, which means large, old trees.Mortality (% per 2 years) is computed with varying precipitation (third line) and with varying water stress (fourth line) and is plotted against the ontogeny (DBH/DBHmax).Mortality rate rises with rising precipitation and reduced water stress.