Unraveling the relative role of light and water competition between lianas and trees in tropical forests: A vegetation model analysis

Abstract Despite their low contribution to forest carbon stocks, lianas (woody vines) play an important role in the carbon dynamics of tropical forests. As structural parasites, they hinder tree survival, growth and fecundity; hence, they negatively impact net ecosystem productivity and long‐term carbon sequestration. Competition (for water and light) drives various forest processes and depends on the local abundance of resources over time. However, evaluating the relative role of resource availability on the interactions between lianas and trees from empirical observations is particularly challenging. Previous approaches have used labour‐intensive and ecosystem‐scale manipulation experiments, which are infeasible in most situations. We propose to circumvent this challenge by evaluating the uncertainty of water and light capture processes of a process‐based vegetation model (ED2) including the liana growth form. We further developed the liana plant functional type in ED2 to mechanistically simulate water uptake and transport from roots to leaves, and start the model from prescribed initial conditions. We then used the PEcAn bioinformatics platform to constrain liana parameters and run uncertainty analyses. Baseline runs successfully reproduced ecosystem gas exchange fluxes (gross primary productivity and latent heat) and forest structural features (leaf area index, aboveground biomass) in two sites (Barro Colorado Island, Panama and Paracou, French Guiana) characterized by different rainfall regimes and levels of liana abundance. Model uncertainty analyses revealed that water limitation was the factor driving the competition between trees and lianas at the drier site (BCI), and during the relatively short dry season of the wetter site (Paracou). In young patches, light competition dominated in Paracou but alternated with water competition between the wet and the dry season on BCI according to the model simulations. The modelling workflow also identified key liana traits (photosynthetic quantum efficiency, stomatal regulation parameters, allometric relationships) and processes (water use, respiration, climbing) driving the model uncertainty. They should be considered as priorities for future data acquisition and model development to improve predictions of the carbon dynamics of liana‐infested forests. Synthesis. Competition for water plays a larger role in the interaction between lianas and trees than previously hypothesized, as demonstrated by simulations from a process‐based vegetation model.


| INTRODUC TI ON
The terrestrial biosphere is a critical component of the Earth system, responsible for the uptake of up to 30% of anthropogenic carbon emissions (Le Quéré et al., 2018). Globally, forests hold more than 80% of the terrestrial above-ground carbon (Sedjo, 1993), about 50% of which can be found in tropical ecosystems (Pan et al., 2011).
Furthermore, tropical forests account for about one-third of terrestrial photosynthesis (Beer et al., 2010), and thus play a key role in global carbon dynamics (Wieder et al., 2015).
Lianas are woody vines which are especially abundant in tropical forests (Schnitzer & Bongers, 2011) where they comprise up to 40% of all woody stems and substantially contribute to ecosystem leaf (Schnitzer, 2002) and root (Collins et al., 2016;Smith-Martin et al., 2019) biomass. However, their comprehensive contribution to the global carbon cycle remains poorly understood (Schnitzer, 2018).
A better understanding of the role of lianas is urgently needed as current estimates of the carbon balance of tropical ecosystems are highly uncertain (Avitabile et al., 2016;Pan et al., 2011). The widespread increase in liana abundance observed in the Neotropics (Phillips, 2002) might be among one of the multiple causes of the long-term transition of the tropical forests from carbon sinks to net sources (Baccini et al., 2017), after decades of carbon sink strength decline (Hubau et al., 2020).
As lianas allocate less carbon to woody biomass compared to trees, they are poor contributors to long-term forest carbon storage (van der Heijden et al., 2013) and strong competitors for resources (Alvarez-Cansino et al., 2015;Schnitzer et al., 2005). Trees are negatively impacted by the interactions with lianas in many different ways: reduced growth (Schnitzer & Carson, 2010), increased mortality (Ingwell et al., 2010) and increased turnover (Durán & Gianoli, 2013). At the ecosystem level, a liana removal experiment in Panama revealed that tree competition with lianas was responsible for a reduction of 76% in the forest net above-ground biomass accumulation (van der Heijden et al., 2015). Furthermore, many liana species can thrive in degraded and early successional forests, where they could slow forest regeneration and hence further strengthen the negative impact of forest disturbance on the long term.
As of today, it is unclear what (if any) mechanism dominates the competition between lianas and trees. The most limiting resource for which plant communities compete varies depending on forest site (Schnitzer, 2005), stand age (Barry et al., 2015) and season (Alvarez-Cansino et al., 2015). Yet, light is often thought to be the main limiting factor for plant growth and development in very dense closedcanopy ecosystems (Bongers and Sterck, 1998;Poorter et al., 2003).
However, liana-tree competition was driven by below-ground resource acquisition (water and nutrients) in at least one tropical site of Ivory Coast . Furthermore, water seems to play a key role in the interactions between lianas and trees as liana density is negatively correlated with mean annual precipitation and positively correlated with dry season length and site seasonality (DeWalt, 2010). Experimentally determining the relative magnitude of the different competition strengths is challenging as it requires establishing replicated manipulated field experiments, followed over time . Process-based vegetation models therefore have a key role to play in disentangling the different forms of competition between growth forms across sites.
Vegetation models are numerical tools that track pools and fluxes of carbon, water and energy in ecosystems. They have been routinely used for projecting ecosystem dynamics under contrasting climatic and land use scenarios (Dietze & Latimer, 2011). Despite their relevance in tropical forest dynamics, lianas have been largely 5. Model uncertainty analyses revealed that water limitation was the factor driving the competition between trees and lianas at the drier site (BCI), and during the relatively short dry season of the wetter site (Paracou). In young patches, light competition dominated in Paracou but alternated with water competition between the wet and the dry season on BCI according to the model simulations.
6. The modelling workflow also identified key liana traits (photosynthetic quantum efficiency, stomatal regulation parameters, allometric relationships) and processes (water use, respiration, climbing) driving the model uncertainty. They should be considered as priorities for future data acquisition and model development to im-ignored by dynamic vegetation models (Verbeeck & Kearsley, 2015). was therefore limited in its potential to disentangle above-and below-ground competition between lianas and trees. As a consequence, no clear signature emerged from the model simulations for sites with different hydrological drivers while liana abundance is expected to be sensitive to rainfall regime (Schnitzer & Bongers, 2011). Locally observed drought stress episodes (Alvarez-Cansino et al., 2015) were also not reproduced by model runs.
To accurately simulate competition for resources between lianas and trees, vegetation models need to comprehensively integrate the functional differences between the two growth forms.
Numerous in situ studies have indeed revealed functional and structural differences in leaf-level gas exchange (Slot & Winter, 2017;Slot et al., 2013), hydraulic properties (De Guzman et al., 2016), rooting depth (De Deurwaerder et al., 2018;Smith-Martin et al., 2019), root and stem vessel diameters Gartner et al., 1990), or leaf properties and allocation (Wyka et al., 2013). Those contrasts in the hydraulic architecture and functioning of lianas and trees need to be accounted for in vegetation models to determine the exact role and impact of lianas in the forest biogeochemical cycles. This was not the case in the previous model version in which several liana trait values were directly copy pasted from the list of ED2 pioneer tree parameters.
The objective of this study is to estimate the relative contribution of below-and above-ground competition between lianas and trees in order to better predict the dynamics of tropical forests as affected by lianas. In particular, we aim to (a) determine how lianas contribute to tropical forest ecosystem fluxes and plant community competition, (b) identify the liana physiological/ecological parameters that contribute the most to liana-tree competition and (c) assess the relative strengths of above-and below-ground competition between lianas and trees over time and across sites and forest stand ages.
To do so, we first updated the ED2 liana plant functional type to include the plant hydraulics module recently implemented in ED2 . We then used the Predictive Ecosystem Analyser (PEcAn, LeBauer et al., 2013) to (a) exhaustively parameterize the liana PFT according to the most recent available observational data and (b) run an uncertainty analysis in order to identify where and when light (or water) was the most limiting resource in two sites (Paracou, French Guiana and Barro Colorado Island [BCI], Panama).
These two sites are relatively wet (yearly rainfall of about 3,100 and 2,650 mm, respectively) but differ in the length and intensity of their dry season. In particular, we wanted to investigate if the relatively short and weak dry season in Paracou was sufficient to trigger a strong water competition between lianas and trees. We hypothesized that the quest for water would drive liana-tree competition on BCI where the dry season is longer and stronger. Contrastingly, under the wet conditions prevailing in Paracou, we expected competition to be primarily for light. In both sites, we assumed that young patch dynamics (where light is abundant, and root systems not fully developed) would be mainly driven by water competition. The comprehensive liana PFT meta-analysis allowed us to better constrain the model parameters, update the original implementation of di Porcia e Brugnera et al. (2019) and estimate the reduction of parameter uncertainty gained thanks to such a literature review.

| The ecosystem demography model
The Ecosystem Demography model version 2 (ED2) is a terrestrial biosphere model that accounts for horizontal and vertical heterogeneity across the landscape as well as plant diversity (Medvigy et al., 2009). ED2 is a size-and age-structured approximation of an individual vegetation model that is able to represent the stochastic nature of mortality, reproduction and dispersal processes . ED2 simulates both the short-term response of the ecosystem to changes in atmospheric conditions as well as the long-term dynamics of ecosystem composition driven by resource limitations (Raczka et al., 2018), which makes it a suitable tool to investigate competition between growth forms or functional groups.
In ED2, the energy, carbon and water cycles are solved separately for each single group of plants belonging to the same functional type and sharing a similar diameter at breast height (DBH), that is, the plant cohorts (Moorcroft et al., 2001). The cohorts belong to patches, which are defined as areas of the forest with a certain age, that is, time since last disturbance. Each patch represents the collection of similar canopy gap-sized areas within a given site (Moorcroft et al., 2001). Patch area corresponds to the relative chance of finding a forest portion sharing the same disturbance history. Plant cohorts and patches are spatially implicit: the horizontal position of each plant in a patch and the position of patches relative to one another are not simulated . Instead, the model computes the plant density of each cohort within each patch and its dynamics.
Previous studies have demonstrated the capacity of the ED2 model to realistically simulate important aspects of carbon and water dynamics in different types of ecosystems: temperate (Medvigy & Moorcroft, 2012;Medvigy et al., 2009), boreal (Ise & Moorcroft, 2010) and tropical . Importantly, ED2 could reproduce reductions in above-ground biomass of Amazon forests subjected to drought experiments (Powell et al., 2013), capture multiple benchmarks (e.g. mortality rates, above-ground biomass stocks) on Barro Colorado Island, Panama (Powell, Kueppers, et al., 2017), and represent leaf and biomass spatial and temporal variability in tropical dry forests .

| Model relevant processes and parameter description
Among other biological and physical processes, ED2 simulates soil hydrology (Walko et al., 2000), biogeochemistry (Bolker et al., 1998), leaf phenology (Botta et al., 2000), photosynthesis (Farquhar et al., 1980) and plant hydraulics , which all in turn impact the energy, carbon and water balances of the ecosystem. For further details about the model structure, we refer the readers to the latest model description  as we only briefly describe a subset of the model parameters and the underlying processes relevant to this study.

| The plant functional types
Plant functional types (PFTs) reflect an ensemble of morphological, physiological and life-history traits that mimic the plant strategy for resource acquisition and use (Fisher et al., 2010). In this study, we simulated the competition for light and water between one liana PFT and three tree (early-, mid-and late-successional tropical evergreen trees) PFTs. We used the tree PFT definitions of , in which self-supporting plants are represented by a discrete approximation of the continuous distribution of life strategies, ranging from fast-growing, resourceacquisitive (early-successional PFT) to conservative, slow-growing (late-successional PFT).
For this analysis, we focused on lianas and selected 32 parameters related to various aspects of their ecophysiology, competition and demography (Table 1). These specific plant parameters were chosen based on previous sensitivity and uncertainty analyses (Dietze et al., 2014;LeBauer et al., 2013;Wang et al., 2013)  First, the liana PFT was integrated in the most recent model version of ED2 that includes plant hydraulics and a process-based description of water uptake and transport (Powell et al., 2018;Xu et al., 2016). Second, as simulations were initialized with observed inventory data rather than started from bare ground, we introduced height restrictions for liana cohorts based on the patch tree height distribution rather than the height of a tracked cohort.
Liana heights were allowed to deviate from the prescribed height allometry so that large lianas can overtop the tallest tree cohort in each forest patch by no more than a small offset (Figure 1).
Without the structural support of the host tree, they can indeed not grow any higher.
In the explanatory schematics of Figure 1, the forest is composed of a liana cohort overtopping a tree cohort by a fixed maximal h offset .
Because of carbon allocation to growth, both tree and liana cohorts increase in diameter and hence height (tree/liana structural growth).
However, the resulting liana height is larger than the updated tree cohort height and is therefore reduced to just overtop it (height restriction), which results in a deviation from the prescribed allometry.
Liana initial height was similarly restricted for all lianas larger than a threshold DBH fixed to 3 cm. Details of implementation are given in Appendix B. In addition, we give an overview of the liana plant functional type functioning as well as the details of the differences with the original implementation of di Porcia e Brugnera et al. (2019) in Appendix C.

| Model predictive uncertainty and parameter sensitivity
To quantify the model uncertainty with respect to liana-tree competition, we used the automated workflow in PEcAn, which consists of three main steps : (a) a meta-analysis to constrain PFT functional, physiological and morphological parameters from observational trait data, (b) a model sensitivity analysis of the selected parameters and (c) an uncertainty analysis that combines the results of the first two steps to estimate the relative importance of each parameter on the overall parametric uncertainty. In this study, we kept tree parameters constant while letting liana parameters vary. Lianas being in competition with a range of competitors (slow-to fast-acquisitive tropical trees), we assumed that incorporating tree parameters in the sensitivity analysis would only increase the number of parameters without clarifying the picture.

| Meta-analysis
First, the meta-analysis aims to generate a posterior distributions ( 0 p ) for each parameter p from a prior distribution and the existing trait observational data. Prior distributions represent the a priori knowledge of the model parameters and define the widest range of variation as well as the probabilistic distribution of each single trait. Liana priors were adapted from tree distributions (Dietze et al., 2014;LeBauer et al., 2013;Raczka et al., 2018) to encompass the original parameterization of the liana PFT and reflect the allegedly differences between growth forms according to 'liana/ED2 expert' opinion, see Table 2. Prior distributions were also chosen to TA B L E 1 List of model parameters for the liana PFT analysed in this study alongside with their description, units and classification into organs, competition type and model processes. A more detailed description of the underlying processes that those parameters affect can be found in Appendix A

| Sensitivity analysis
Parameters were varied one-at-the-time around their median values (±1, 2, 3 SD) and several model responses (GPP, NPP and evapotranspiration, as well as the liana contribution to these fluxes) were fitted using a Hermite cubic spline function g p , which allowed us to estimate the model sensitivity to each parameter. Model sensitivity was estimated as the slope of the spline function

| Uncertainty analysis
The outputs of the two previous steps, the parameter posterior distribution 0 p and the model response function g p were further used to estimate the contribution of each parameter to the model parametric uncertainty. The total parametric uncertainty was calculated as the model output variances generated by each single parameter summed up over the total number of parameters N, as shown in Equation 3. In this study, total and parametric uncertainty are synonymous as we only account for the latter type of uncertainty.
Parameter contribution to the total parametric uncertainty rel.var p was then computed as the fraction of variance explained by each parameter (Equation 4).
The entire workflow is illustrated in Figure 2 for a specific input parameter (liana stem conductivity K max ) and two model responses (ecosystem GPP and its liana contribution). The hierarchical Bayesian meta-analytic model shifts the parameter median tot.var F I G U R E 1 Liana model and initial forest composition. The figure illustrates the DBH-height (h) allometry for both trees (dashed green line) and lianas (dashed blue line) as well as the liana initial distribution on BCI as derived from forest inventory and allometric equation (blue dots). Initially, all liana cohorts larger than DBH threshold are assumed to have reached the canopy (i.e. to be slightly taller than the tallest tree within that specific forest patch). For each single liana cohort (each blue dot in the graph), an initial DBH-offset (ΔDBH) is calculated as the DBH-difference between the allometric equation and the actual allometric position and is used to shift the cohort DBH-height allometric relationship. As opposed to trees, the liana growth is a two-step process: the available carbon is spent by lianas to grow in diameter and compute a potential height which is further restricted by the tallest tree height within that patch incremented by a small offset (h offset ). ΔDBH is then updated with its new value and reduces the uncertainty by ingesting observational trait data.
Model univariate sensitivity analyses (star and circle symbols) are then fitted with the spline functions to estimate the predictive uncertainties corresponding to the parameter prior (light) and posterior (dark colours) distributions. Figure 2 further illustrates how ecosystem-scale variables can be more constrained than their individual components due to PFT compensation effects.
The benefit of the liana parameter constraining and meta-analysis was assessed by the ratio of posterior to prior ensemble run spreads.
Model output spreads were generated from ensemble simulation TA B L E 2 Parameter distributions for the liana PFT as used in the sensitivity analysis alongside with the prior and posterior medians and the ED2 default parameters (di Porcia e Brugnera et al., 2019). The values a and b define the constants of the prior distribution function  for each parameter analysed in this study. The sample size (N) is the number of mean trait observations collected for the meta-analysis. Parameter units can be found in Table 1 Parameter These parameters are actually negative and were multiplied by (−1) after sampling. f Indicates when the 95% CI interval of the posterior did not include the ED2 default parameter. runs (n = 250), using either the prior or the posterior distributions sampled using Monte Carlo techniques (Raczka et al., 2018).

| Simulated sites
The model uncertainty analysis was performed for two sites: Barro Colorado Island, Panama and Paracou, French Guiana. These two specific sites were selected based on the local abundance of liana and ecosystem empirical data, their difference in liana contribution to forest biomass and rainfall regimes (Table 3; Supporting Information Figure E1).
The forest of BCI is an old-growth seasonally moist lowland tropical forest with an average annual rainfall of about 2,640 mm (Detto et al., 2018) and a well-marked dry season (total rainfall between late-December and mid-April is about 175 mm on average). Located on the coastal part of French Guiana, the Paracou research station is classified as a lowland moist primary forest (Aguilos et al., 2018;Bonal et al., 2008;Malhi, 2012) which, compared to BCI, experiences higher precipitation rates (recorded mean annual precipitation is almost 3,100 mm), and a weaker and shorter dry season spanning from mid-August to mid-November (total rainfall during this period is 238 mm). Both sites support tropical evergreen moist forests and we therefore imposed an evergreen phenology to all plant functional types of this study, following Powell et al. (2018)

| Prescription of atmospheric forcings
For both sites, we used the meteorological data from the local flux tower measurements as atmospheric forcings (see Table 3 for respective spanning periods) and used the observed carbon and energy exchange fluxes obtained with the eddy-covariance method to benchmark the modelled productivity and evapotranspiration (Aguilos et al., 2018;Bonal et al., 2008;Powell, Kueppers, et al., 2017). Meteorological data of the simulated years were readily available at hourly resolution for air temperature, wind speed, specific humidity, precipitation rate, short-and long-wave radiation and were hence used as ecosystem upper boundary condition. To exclude CO 2 fertilization effects and keep the same meteorological drivers as in our previous study (di Porcia e Brugnera et al., 2019), the atmospheric concentration of CO 2 was fixed at a constant value of 370 ppm, which corresponds to initial concentrations measured by the flux towers. Because censuses were not available for trees <10 cm DBH at Paracou, we extrapolated the number of tree individuals in the 1-10 cm DBH class range using a linear model applied to the log-log transforms of the DBH size class versus the plant density. We filled the missing class of trees by generating the estimated number of plants from the three tropical tree PFTs based on their relative frequency in the inventory.

| Vegetation initial conditions
From liana inventories, it appeared that liana density was much higher on BCI and not only because the inventory in Paracou did not include smaller lianas comprised between 1 and 2 cm (Table 3). Paracou counted a few more large (DBH > 14 cm) lianas (four individuals ha −1 ) as compared to BCI (two individuals ha −1 ).
Similarly, areas with different levels of liana infestation co-existed in Paracou: Large liana density ranged from 0 to 210 individuals ha −1 .
Here and everywhere in the manuscript, we refer to liana stems (ramets) as liana individuals while they are not always individuals in the genetic sense (i.e. genets).

| Competition and model scenarios
To determine the driving force of competition between lianas and trees, we classified each liana parameter according to its relevance for below-ground (water) or above-ground (light) competition (Table 1). We also classified them by plant organ (leaf, stem, root, seed or entire plant for parameters that could not be primarily related to a single organ) and ecophysiological process (allocation, water use, photosynthesis, respiration, mortality, tissue turnover or structural parameters), see Table 1.
Tissue turnover represents the maintenance costs of leaves and roots. Summing up the relative contribution of all parameters belonging to each group (of process, organ, competition types) allowed us to determine the most critical parameter categories for uncertainty.
All model simulations were run for 5 years. ED2 was run as standard and all patches and cohorts were allowed to age, grow or disappear. To investigate competition shifts over time, we assessed the model uncertainty both over the full simulation duration and during dry periods only. We defined as dry the months during which rainfall did not exceed 100 mm. To account for competition changes across forest stand ages, the uncertainty analyses were run starting either from the full set of initial conditions or from young forest

TA B L E 3
Main features of the two simulated forest sites patches only. Because forest inventories did not provide any information on age, we assumed patch age based on three criteria: the initial liana density, the initial abundance of late successional trees and the initial patch height. Thresholds of these criteria were progressively modified from the most extreme values to include a minimum of five patches on BCI (of 1,250) and one (of 10) in Paracou.
By doing so, we ended up selecting six patches on BCI and one in Paracou in which the liana initial density was among the highest in the respective sites, alongside with a disproportionately low initial representation of late successional trees in patches that were initially shorter than the average. Distributions of these criteria are represented for both sites and the selection of young patches from the inventory highlighted in the 50-ha plot of BCI in Supporting Information Figure E2.

| Parameter distributions and meta-analysis
We were able to collect data for 19 of the 32 liana parameters we selected in this study ( Figure 3; Table 2): six hydraulic, six allometric, four photosynthetic and two structural parameters as well as the density-independent (i.e. ageing) mortality rates of lianas that were extracted from Phillips et al. (2005). The priors of the remaining 13 parameters could not be constrained by data. In Table 2 Figures F1 and F2).

F I G U R E 3
Liana PFT parameter distributions. The prior distributions (grey) are relatively broad and were established to encompass natural variability of parameter values and cover all field observations (black vertical lines smoothed into the green distributions). A Bayesian metaanalysis was performed to combine the prior distributions with trait data (whenever available) to create posterior distributions (blue), which were further used to estimate both the model and parameter uncertainties. The figure only illustrates the parameter distributions for which data were available. Note that we did not use the Bayesian meta-analysis for the allometric parameters (b1Rd, b2Rd, B1Bl, B2Bl, b1Bs and b2Bs). Instead, a posterior distribution constrained to data was directly built for each of those. The units of each parameter are given in Table 1 In  (Christoffersen et al., 2016).
Liana rooting depths estimated from posterior distributions were considerably shallower than the ones that used default allometric coefficients ( Figure 3; Table 2 In Paracou, the ratio of the parameter coefficients of variation after and before meta-analysis was always lower than 1 except for the rooting depth slope allometric coefficient (b2Rd). The standard deviation of b2Rd was also reduced after data ingestion (i.e. stronger constraints), but this effect was overcompensated by a large decrease of the distribution median (Table 2). This indicates that the posterior distributions were systematically more constrained than the a priori distributions of the model parameters. The posterior to prior CV p ratio varied between 0.07 (b1Bl, rho) and 1.31 (b2Rd) with a mean of 0.35. These results were essentially the same for BCI.

| Ensemble runs and liana impacts on forest
In both sites, the model could capture many of the structural characteristics of the ecosystems. On BCI, the simulated total leaf area (LAI, 4.6 ± 0.3 for the posterior ensemble runs) and the above- On average, simulated lianas accounted for about one-fourth (24%) and one-eighth (12.5%) of the landscape average leaf area on BCI and in Paracou, respectively, while accounting for less than 3% of the above-ground biomass in both ecosystems (2.8% and 1.6%, respectively). Those numbers are averages across ensemble runs and over the duration of simulation. However, liana abundance did not dramatically change over time and hence neither did liana contribution to forest biogeochemical cycles (Supporting Information Figure F5). After 5 years, liana density on BCI remained higher than in Paracou (0.14 liana m −2 vs. 0.013 liana m −2 ), which is in agreement with observations/initial conditions (Table 3). All confidence intervals (CI) of the ensemble runs were reduced after meta-analysis, especially for the landscape average variables. This indicates a successful parametric constraining through the meta-analysis. Ecosystem and liana LAI CI spread decreased by more than 45% in both sites over the entire duration of the simulation (Supporting Information Figure F5) and ecosystem AGB CI decreased about 75%. In addition, the reduction in ecosystem flux CI was on average about 70% in Paracou and between 30% and 50% on BCI (GPP and latent heat respectively). The reduction in liana flux uncertainty was around 20% yearly in both sites, and reached 60% on BCI and 40% in Paracou during the dry season ( Figure 4).   and less negative tree leaf water potentials (Pérez-Salicrup & Barker, 2000), observed right after liana removal.

| Uncertainty analysis and competition factors
In this section, we explore the outputs of the uncertainty analysis of the liana PFT. We mainly illustrate these results using the liana contribution to ecosystem GPP as model output since this represents the capacity of lianas to maintain, grow or thrive through competition with the tree PFTs. As detailed below, results for other fluxes such as evapotranspiration and NPP are very similar.
After integration of the available data during the meta-analysis, liana photosynthetic quantum efficiency (with a relative contribution of 37%) and the stomatal closure regulation parameter stoma_ psi_b (31%) were the strongest drivers of liana GPP uncertainty on BCI, with all other parameters contributing <10% to the overall uncertainty ( Figure 5). Not only did the relative contribution of these For the liana contribution to the ecosystem NPP, the growth respiration parameter played a very important role (mean partial variance of 37% for the uncertainty analysis of posterior runs across sites and seasons, see also below). This was similarly found for other tree PFTs F I G U R E 6 Comparison of the contribution of the liana parameters to model uncertainty (here, the liana GPP) during the dry season (red) or yearly (blue) on BCI, Panama (left) and Paracou, French Guiana (right). The parameters are sorted by their contribution to uncertainty on BCI over the entire year, which is the posterior shown in the last panel of Figure 4. The total standard deviation (the square root of the sum of the variances) is also given for each single scenario F I G U R E 7 Relative contribution of liana parameters to liana GPP model uncertainty (relative variances, rel. par p ) in both BCI, Panama (left) and Paracou, French Guiana (right) for the prior, posterior and posterior in young patches only distributions. Parameters were ordered by their partial variance contribution summed up over both sites and all three scenarios. Only the contributions superior to 1% (in at least one of the sites or one of the scenarios) are shown in the figure. In addition, the left column shows the classification into water-related (blue underlines) and lightrelated (green underlines) parameters as determined in Table 1. Finally, parameter relative contributions to both competition types (water and light) are presented considering entire simulations (black, which is basically the sum of the relative contributions presented just above) and during the dry season only (red, single relative parameter contributions not shown). The competition type dominating for each particular scenario and site is presented in bold in ED2 and relates to the model structure (see Section 4). This parameter aside, partial variances of liana parameters for the contribution of lianas to GPP and NPP were also very well correlated (r 2 = 0.84, slope = 1.12). Therefore, the conclusions drawn above for liana GPP remain valid for modelled fluxes of liana NPP and evapotranspiration.
While the partial variances of liana parameters varied over time and between forest sites and stand ages, the contribution of the different plant organs and processes remained relatively consistent for the different model outputs (Figures 8 and 9). On BCI, the leaf-related parameters (60% on average Figure 8) and water use-related parameters (43%, Figure 9) overall dominated model uncertainties, even though respiration-related parameters (driven by the growth respiration parameter) became almost as important as water use for liana NPP (32% vs. 35%, Figure 9). During the dry F I G U R E 8 The relative contribution of liana parameters to liana evapotranspiration, GPP and NPP model uncertainty (partial variances) on BCI, Panama (left) and Paracou, French Guiana (right) over the whole year or during the dry season as aggregated by organ according to the classification of Table 1 F I G U R E 9 The relative contribution of liana parameters to liana evapotranspiration, GPP and NPP model uncertainty (partial variances) on BCI, Panama (left) and Paracou, French Guiana (right) over the entire year or during the dry season as aggregated by process according to the classification of Table 1. Turnover indicates the living tissue maintenance costs season, allocation parameter contribution (driven by rooting depth allometric coefficients) increased (+17% on average) while water use-related parameters either remained constant (liana GPP and NPP) or decreased (−13%, liana evapotranspiration), so that in total, water-related parameter contribution always increased during the dry season.
In Paracou, leaf organ importance systematically decreased in favour of entire plant-scale parameters (especially b1Ht and b2Ht): on average, leaf-related parameters contributed to 49% of the total variance and plant-scale parameters to 26%. On BCI, the contribution of these leaf-related parameters and plant-scale parameters reached 60% and 16%, respectively. Similarly, Paracou was characterized by a higher contribution of allocation parameters as compared to BCI (+6% on average) at the expense of water-use parameters (−11%), driven by growth for light competition and the height allometric coefficients.

| Liana impact and competition across simulated sites and forest stand ages
This study is an important step towards realistically representing lianas in vegetation models. Our approach completes the first attempt to include the lianescent growth form in ED2, as it fills several gaps in the previous study (di Porcia e Brugnera et al., 2019). Primarily, it mechanistically accounts for the hydraulic architecture differences between lianas and trees, as observed by many studies (Ewers et al., 1990;Johnson et al., 2013;Maréchaux et al., 2017;Tyree & Ewers, 1996;van der Sande et al., 2013van der Sande et al., , 2019Zhu & Cao, 2009) and therefore allows us to extend the use of such a model to drier sites or to more extreme (i.e. future) climatic conditions. It also targets shorter time-scales as compared to the original publication (years vs. centuries) to focus on the mechanistic processes driving intergrowth form competition.
Moreover, we extended the use of the liana PFT to prescribed initial conditions in addition to near-bare ground initialization. It is worth noting that the liana PFT is slightly different to the one used for the production runs in the original publication (di Porcia e Brugnera et al., 2019) as liana height limitation was here applied at the patch level rather than at the cohort level (see Appendix C for more details). Furthermore, the new version of the liana PFT was parameterized using the most up-todate observational data as opposed to the default pioneer tree parameters that were used before (see Figure 3; Table 2).
The model simulations presented in this study captured many features of two tropical forests characterized by contrasting amounts and seasonality of rainfall, as well as liana abundance. Both forest structural properties (total LAI and AGB, Figure E5) and flux measurements derived from eddy-covariance observations (GPP and latent heat,  Figure E5) and reduced the overall model uncertainties ( Figure 5).
The impact of lianas on forest dynamics was also reproduced by model simulations. By strengthening competition for below-ground resources, lianas increased the simulated drought stress experienced by trees, especially during the dry season, as experimentally observed in liana removal experiments (Alvarez-Cansino et al., 2015).
Liana removal triggered tree drought-stress relief in the simulations, as suggested by experimental data (Pérez-Salicrup & Barker, 2000;Reid et al., 2015). Overall tree growth considerably increased in both sites when removing the liana PFT from the simulations (+30%-40%), which is in line with observed tree growth increases after liana removal (van der Heijden et al., 2015). Similarly, the predicted increase in tree mortality (+30% on BCI) relative to liana-free simulations is confirmed by experimental observations (van der Heijden et al., 2015). In the model, the increase in mortality was caused by a reduction of carbon gains for trees when lianas were added to the runs. This, in turn, was due to a combination of decreased tree stomatal conductance due to drought stress (below-ground competition) and declined light interception by tree PFTs (above-ground competition) caused by lianas. Liana removal in the simulations led to forest recovery, enhanced forest productivity and recovered sink strength, just like in the experimental plots (van der Heijden et al., 2015). According to the model, the effect of lianas on the forest does not differ between seasons: The strengths of water and light competition compensate each other over time, as observed experimentally (van der Heijden et al., 2019).
The vegetation model also enabled disentangling the contrasting impact of lianas on the forest composition: abundance and productivity decreased more in early successional trees than in the other tree PFTs because the former shared more similar ecological niches (fast acquisitive, low wood density, high mortality rates).
Water competition played a more important role than hypothesized. In silico, the competition between growth forms was dominated by water acquisition all year long on BCI and during the dry season in Paracou (Figure 7), even though the two selected sites were quite wet (Table 3). Several seminal studies investigating growth forms competition already indicated that water is critical for determining the impact of lianas on forest dynamics (Andrade et al., 2005;De Deurwaerder et al., 2018;Tobin et al., 2012). Our numerical findings reinforce the idea that below-ground competition is crucial in liana tree relationship as water acquisition dominated the competition even during the relatively short and weak dry season in Paracou. In sites characterized by lower yearly rainfall, and hence higher liana densities (DeWalt, 2010;Schnitzer, 2005;Schnitzer & Bongers, 2011), the relative importance of below-ground competition is expected to increase even more than we found for BCI and Paracou (Figure 7). The relative contribution of light competition that we observed (during the wet season and in young patches in both sites) decreased with decreasing water availability ( Figure 7) and will probably keep doing so in drier conditions. Therefore, the simulated relative contributions of water to the liana-tree competition (35% and 50% in Paracou and on BCI) are likely to be on the low side, and could increase if stronger seasonality or decreased precipitation is expected in the future.
Despite the fact that we only included two sites in this analysis, the modelling workflow and the new model development allow expanding simulations over a larger rainfall gradient in the future. Next steps should focus on the ability of ED2 to reproduce trends of liana abundance with dry season length and mean annual precipitation (Schnitzer & Bongers, 2011) or seasonality (Parolari et al., 2020) over a larger number of sites. This expansion to drier conditions should confirm the observed trend in this study of water dominating the liana versus tree competition.

| Uncertainty analysis and key parameters
The Bayesian workflow that we applied here and that was devel- Additional observations that would feed the meta-analysis could inform us if multiple liana functional types need to be accounted for according to their natural variability and the respective role that they have on forests.
Some specific liana parameters were systematically the largest contributors to model output uncertainty (growth respiration, quantum efficiency, plant hydraulics) and the list of these parameters largely overlaps with the ones of tree-PFT parameters from previous uncertainty analyses. Table 4 compares the uncertainty analysis results for ecosystem NPP from this study and from Raczka et al. (2018) and Dietze et al. (2014). Except for the height allometric coefficients (not considered in previous studies and quite specific to lianas, see Figure 1), all parameters identified as critical for the liana-tree competition were previously identified as crucial for trees as well (note that soil-plant water conductance was replaced by a set of mechanistic parameters in this study, e.g. stoma_psi_b, stoma_psi_c and K max ). It would be interesting to extend the uncertainty analysis accounting for both tree and liana parameters. While it would increase the number of parameters to constrain, it would also allow refining the mechanisms behind which lianas compete more with pioneering trees.
The uncertainty analysis also highlighted processes that lack a sufficiently mechanistic approach to be constrained with existing trait data, and therefore contributed the most to the overall variances. As high uncertainties were similarly found in growth respira-

| Study limitations and perspectives
As a vegetation modelling study, this research has important intrinsic limitations, the most critical of which is probably its ecophysiological boundaries. Lianas and trees interact more than through resource competition. It has been demonstrated that lianas can damage their hosts directly by mechanical abrasion and passive strangulation or indirectly by increasing the hosts' susceptibility to wind damages and likelihood of treefall (Putz, 1984). Hence, in reality, lianas might affect tree productivity and mortality rates in more ways than those that can be determined physiologically while we only focused on the latter in this study. Furthermore, many of the model predictions presented here are preliminary and should be validated using new and relevant datasets. Such a model-data fusion loop approach would help keep improving and refining model accuracy.
Similarly, liana abundance is not only driven by a combination of competition with self-supporting plants but also by fundamental limits on their capability to exist under different abiotic conditions, for example, freezing air temperature (Ewers, 1985;Schnitzer, 2005). In addition, several putative mechanisms of increasing liana abundance in the Neotropics (Schnitzer & Bongers, 2011) were not considered at all in the vegetation model. In addition, there are several plant processes that could be taken into account in the future for a more accurate representation of lianas in ED2. Currently, the model assumes an infinite ability for xylem refilling for both lianas and trees, while lianas, in some cases, may be better than trees at cavitation recovery (Ewers & Fisher, 1991;Fisher et al., 1997). As lianas cavitated more during our simulations, they also refilled their cavitated vessels more than trees. Yet, this could be simulated more explicitly in the future.
Differences in leaf phenology were also not accounted for while lianas produce leaves over a greater fraction of the year than trees whatever their successional status (Putz & Windsor, 1987). Actually, all plants were simulated as evergreen in our model runs while few lianas and nearly half of the canopy trees on BCI are brevi-or facultatively deciduous and hence lose a fraction of their leaves during parts of the dry season (Putz, 1990). In the future, contrasting seasonal phenology strategies should be considered to reproduce the seasonal differences in liana and tree growth (Schnitzer, 2005).
It must also be emphasized that the positive impacts of liana removal on forest productivity and carbon sequestration as observed in experimental plots (van der Heijden et al., 2015) and confirmed in our model simulations might be temporary. The substantial benefit of tree growth after liana cutting (Mills et al., 2019) presumably diminish with time, even if some seminal studies suggest that they could persist as long as 6-10 years after removal (Kainer, 2014

| CON CLUS IONS
We presented in this study the first vegetation model able to disentangle the contribution of water and light in the competition for resources between lianas and trees. While being critical for the fundamental understanding of forest dynamics, it is a question that is extremely difficult to answer as isolating below-and above-ground competition between lianas and trees requires heavy manipulations and measurements. Vegetation models therefore have an important role to play to unravel interactions between plant functional types. By further developing a liana PFT in the ecosystem demography model (ED2), and analysing it with the bioinformatics toolbox PEcAn, we identified that liana quantum efficiency and stomatal regulation parameters were the most critical parameters controlling liana productivity and hence the liana versus tree competition. Model simulations with parameters constrained by data successfully reproduced the magnitude and seasonality of GPP and ET, and the magnitude of aggregated properties such as LAI and AGB. Competition with lianas was predicted to negatively impact tree growth (between −30% and −40%) and reduce forest net productivity in both sites. Uncertainty analyses suggested that water competition was more critical in the relationship between lianas and trees than expected. Indeed, water acquisition dominated the yearly growth-form competition on BCI and was even important in a relatively wet site as Paracou. This workflow can now serve to predict the impacts of lianas on tropical forest carbon sink strength or storage at large scale or in a climate change context where decreased rainfall, increased disturbance and stronger seasonality are expected to promote lianas.

PE E R R E V I E W
The peer review history for this article is available at https://publo ns.com/publo n/10.1111/1365-2745.13540.

DATA AVA I L A B I L I T Y S TAT E M E N T
All data necessary to reproduce the results presented in this study, including ED2 source code, PEcAn source code, initial vegetation conditions, MET drivers and parameter distributions and traits, can be found on https://zenodo.org/recor d/41152 06#.X5Ayj HUzYig (https://doi.org/ 10.5281/zenodo.4115206) (Meunier et al., 2020).