Carbon allocation and flow through ecosystems regulates land
surface–atmosphere CO
Accurate projection of the changing global climate, given a particular scenario of future greenhouse gas emissions or concentrations, is largely determined by adequate representation of mechanistic processes in Earth system models (ESMs) (Taylor et al., 2012). Land surface models (LSMs) and their associated biogeophysical and biogeochemical parameterizations are key determinants of the ESMs' fidelity in characterizing and quantifying complex feedbacks in the Earth system (Arora et al., 2013; Friedlingstein et al., 2006; Pitman, 2003). Modeling studies have increasingly used observational data and mechanistic knowledge of processes to advance the development of LSMs (Best et al., 2011; Dai et al., 2003; Krinner et al., 2005; Oleson et al., 2013; Wang et al., 2011). Global and regional observations of land surface fluxes, states, and dynamic vegetation change offer insights into the large-scale interactions between the land surface and atmosphere and hence facilitate model improvements at relevant scales in space and time (Beer et al., 2010; Huntzinger et al., 2012; Luo et al., 2012; Randerson et al., 2009). However, to better quantify and reduce uncertainties arising from deficiencies in model process representation, parameters, driver data sets, and initial conditions, there has been significant effort to evaluate and to calibrate LSMs against site-scale observations and experimental manipulations (Baldocchi et al., 2001; De Kauwe et al., 2014; Hanson et al., 2004; Ostle et al., 2009; Raczka et al., 2013; Richardson et al., 2012; Schaefer et al., 2012; Schwalm et al., 2010; Stoy et al., 2013; Walker et al., 2014; Williams et al., 2009; Zaehle et al., 2014). Further, model development from these focused site-scale studies, especially in close collaboration with experimentalists, can inform and prioritize new experiments and observations that are specifically designed to advance understanding of critical terrestrial ecosystems and processes (Shi et al., 2015).
The Community Land Model (CLM) is an advanced LSM with a comprehensive mechanistic parameterization of carbon (C), water, and energy budgets for diverse land types that can be applied across multiple temporal scales (Oleson et al., 2010). CLM has been evaluated against observations from a wide range of sources, and these evaluations have resulted in improved model performance (Bauerle et al., 2012; Bonan et al., 2011, 2012; Koven et al., 2013; Lawrence et al., 2011; Mao et al., 2012a, b, 2013; Oleson et al., 2008; Randerson et al., 2009; Riley et al., 2011; Shi et al., 2011, 2013, 2015; Thornton et al., 2007). Nevertheless, little attention has been paid to CLM's ability to replicate short-term manipulative experiments, which provide an avenue for exploring and validating model response to sudden, large changes in environmental drivers that control physiological and ecological responses (Amthor et al., 2001; Bonan et al., 2013; Shi et al., 2015). Processes operating over short timescales can have long-lived ecosystem consequences through indirect effects; e.g., stomatal conductance varies on timescales of hours or shorter, but indirect effects on site-level water balance through controls on transpiration can extend to annual timescales and beyond. Combined model–experiment projects can focus efforts on specific mechanistic processes whose representation in the model may be neither adequate nor appropriate for specific sites (Walker et al., 2014; Zaehle et al., 2014). Extending these model–experiment evaluations and ensuing model refinements to additional sites of the same and different ecosystem types improves confidence in the regional- and global-scale adequacy of the LSM's mechanistic process representation and parameterization.
Photosynthetic carbon (C) assimilation, the allocation of photosynthetic products
into tissues with different turnover rates, and the respiration of C back
into the atmosphere are important determinants of CO
Carbon isotopes provide important constraints on specific processes and can
be used in labeling experiments to track pulses of carbon through plant and
soil components. Both diffusion through stomata and enzyme activity during
photosynthesis discriminate against the accumulation of
We evaluated the integrated response of a simulated tree–soil system to an
imposed alteration of shortwave radiation, the main environmental driver for
photosynthesis, and compared the observed trajectory of labeled carbon
pulses through that system with approximations of carbon allocation that are
typical of a global-scale model. We used a version of CLM4 that has been
modified to allow convenient application of the global-scale modeling
algorithms at single points (PTCLM, described in Oleson et al., 2013). We
evaluated the model against observations and experimental results from the
“Partitioning in Trees and Soils” (PiTS) experiment established in a young
loblolly pine stand in Oak Ridge, Tennessee, USA (Warren et al., 2013). The
project exposed a young loblolly pine (
The field component of the project was conducted in a young loblolly pine
stand at the University of Tennessee Forest Resources AgResearch and
Education Center in Oak Ridge, Tennessee. The soil is classified as a
silt–clay–loam (13.3 % sand; 35.7 % clay; 51.0 % silt), with bulk
density ranging from 1.2 to 1.4 g cm
In 2010, a subset of eight of the trees, adjacent to one another, and their
soils were instrumented with automated sensors to continuously measure soil
temperature, soil moisture vertically throughout the soil profile, soil
surface
Following several weeks of pretreatment measurements, the eight study trees
were enclosed with plastic film stretched over a frame surrounding the
trees, and then trees were exposed to 53 L of 99 atom %
To assess actual conditions under the shade cloth treatments, short-term
measurements of temperature, humidity, wind speed, and PAR were collected at
the canopy surface following shade cloth installation. Linear regressions
between meteorological data from under the shade cloth and from the open
field were used to estimate conditions at the canopy surface during the
experimental period. Temperature was
Non-destructive measurements of soil moisture, soil temperature, soil
respiration, sap flow, and stem growth were made prior to the labeling and
for the duration of the shade treatment. During the shade treatment,
destructive measurements of foliage, stem phloem tissue, roots, and soil were
collected to assess presence of the
We used CLM4 (Oleson et al., 2010), the land component of the Community Earth System Model (CESM) (Gent et al., 2011), to simulate the pretreatment and manipulated processes in the PiTS study. This CLM version includes fully prognostic carbon and nitrogen representations for its vegetation, litter, and soil biogeochemistry components (Oleson et al., 2010, 2013; Thornton et al., 2007; Thornton and Rosenbloom, 2005).
Carbon allocation in this version of CLM is simplistic. After maintenance
respiration demands are calculated and subtracted from gross primary
productivity (GPP), and following a step that downregulates GPP on the basis
of static allocation parameters, fixed tissue C : N stoichiometry, and plant
mineral N uptake, the available carbon is allocated to new growth, storage
for growth in subsequent growing seasons, and associated growth respiration.
The model includes pools for leaf, fine root, and several categories of stem
and coarse root, with over-season storage pools associated with each of
these “displayed” growth pools. The allocation ratio between stem and leaf
is a function of the previous year's net primary productivity (NPP; higher
fractional allocation to stem with higher annual NPP), while all other
allocation ratios are fixed throughout the simulation for a given vegetation
type. For
Several major developments of CLM performed specifically for this study
include (1) introducing the ability to represent the shade effect and
experimental labeling by driving the model with observed atmospheric
To perform simulations at the PiTS site, we used PTCLM, a scripting
framework to run site-level simulations of CLM efficiently with
site-specific forcing and initialization data (Oleson et al., 2013). We
performed the standard 600 years of accelerated decomposition spinup, in
which soil organic matter decomposition rates are increased (Thornton and
Rosenbloom, 2005), followed by 1000 years of normal spinup, in which the
decomposition rates are returned to their normal values, and a transient
simulation between 1850 and 2010 using historically varying CO
To simulate the treatment period, we replaced the meteorology from the eddy
covariance sites with observed data at the treatment sites starting at day
of
Default PFT-level, site-specific, and optimized parameters for the PiTS site used in CLM 4.0. PFT-level parameters are for the temperate evergreen needleleaf forest type. Optimized values were obtained using the pretreatment data (PRE_OPT) and for the transpiration data during the shading period (HS_MB). In the HS_MB optimization, only the mp and bp parameters were optimized, while other parameters retain their pretreatment optimization values.
Model evaluations are complicated by the co-occurrence of parametric and
structural uncertainty, which confounds the attribution of model errors
(Keenan et al., 2011). A model's performance might be negatively impacted by
misrepresentation of mechanistic processes, poor parameterization of
otherwise sound functional representations, or both. Parameter optimization,
however, can help to isolate structural deficiencies in the model. In this
study, we applied model calibration, by optimizing model parameters, as a
tool to highlight areas for model development rather than simply improving
predictive skill. We optimized selected CLM parameters against pretreatment
data. We then evaluated the performance of the calibrated CLM in the
pretreatment phase and again in the post-treatment phase without
recalibration following simulation of the canopy shading and
We first calibrated the model to simulate the pretreatment conditions using
observations and prior information about model parameters. Data constraints
for the calibration consisted of single pretreatment estimates for leaf,
stem, and root biomass from allometric relationships for similarly aged
loblolly pine (Baldwin and Feduccia, 1987; Naidu et al., 1998; Vanlear et al., 1986), a
pretreatment
Some model parameters were measured directly from observations (Table 1). Other parameters for which direct estimation was not possible were optimized to maximize fit between model results and the observed calibration data (Table 1). The selection of parameters for optimization was based on formal sensitivity analysis (Sargsyan et al., 2013) and prior experience with the model. We defined the sum of squared errors between simulation and observations weighted by data uncertainty as the cost function for the optimization. We used a genetic algorithm (Runarsson and Yao, 2000) to find a set of parameters that minimizes the cost function. Simulations were performed in parallel using two populations of 32 ensemble members in parallel over 100 iterations for a total of 6400 model simulations.
For the pretreatment (pre-labeling) period, we compared the standard “parameter” version of the model (PRE-STD) with the optimized “parameter” version (PRE-OPT). The model with optimized parameters was used in simulations for the shading treatment period for both the high shade and low shade treatments. Because of uncertainties associated with simulated stomatal conductance and transpiration in high shade conditions, we performed additional parameter calibrations for the parameters mp (slope of the Ball–Berry stomatal conductance formulation) and bp (intercept of the Ball–Berry stomatal conductance formulation) during the shade treatment period using the genetic algorithm with transpiration and stem growth data as constraints (HS_MB), with results discussed below.
Since we are interested in understanding the fate of photosynthetically fixed carbon as it is allocated to various tissues and fluxes, and how allocation dynamics respond to changes in photosynthesis as driven by changes in PAR, it is useful to evaluate model predictions of photosynthesis over a range of light levels. We used a functional unit testing framework (Wang et al., 2014) to evaluate CLM's representation of the photosynthetic light response at the scale of individual leaves against light-response curves obtained by Warren et al. (2012) for foliage in the upper canopy of trees at the PiTS experimental site prior to the shade treatment. This approach isolates the targeted model process to allow a direct comparison between instrumental data and simulation output, driving the model component with specified environmental conditions and parameter values.
Mean surface air temperature adjacent to the site decreased from days
The model predicted approximately exponential growth in all biomass pools
during the 8 years of pretreatment simulation, with some evidence of slowing
growth in the final years (Fig. 3a). Using default global-scale
ecophysiological parameters, the model significantly overestimated biomass
accumulation in leaf, stem, and root pools, by 85, 36, and 76 %,
respectively, on 1 September 2010 (PRE_STD curves,
Fig. 3a). Replacing default parameters with observed (lower) leaf N
concentration and with calibrated (higher) allocation ratios for stem : leaf
and root : leaf (complete set of parameter changes shown in Table 2) brought
the biomass accumulation curves in better agreement with observations (Fig. 3a). Using the PRE_OPT parameters, the bias for leaf, stem,
and root biomass accumulations was
Comparison of predicted vs. observed photosynthesis light-response curves
was used as an independent assessment of the model performance before and
after calibration across a range of PAR values characteristic of midday
values in the open field and under the LS and HS treatments (Fig. 3b). In
the range of PAR from 750 to 1588
Soil temperature predicted by the optimized model at 0–5 cm depth had a
consistent overestimation bias of 1–2
Observed transpiration during the pretreatment period was higher for HS than LS plots, likely a consequence of the higher biomass and leaf area of the HS trees (Warren et al., 2012) and perhaps also higher soil water content (Fig. 4b). We used the pretreatment transpiration data to calibrate CLM, and the model simulated the pretreatment observations well in terms of both magnitude and temporal variations (Fig. 4c). After the treatment initiation, decreased transpiration was seen in both observations and model simulations for the HS and LS trees. For the LS case, CLM captured the observed transpiration well. However in the HS case, CLM predicted a sharp reduction in transpiration, whereas the observations differed relatively little from the LS case. To investigate this difference further, we performed a second optimization for the Ball–Berry stomatal conductance slope and intercept terms (HS_MB). However, despite increasing these parameters to near the maximum acceptable values (Table 1), the HS_MB optimization failed to reproduce the measured transpiration.
Pretreatment state variables included in the optimization. Simulated
values were obtained using the default parameters (PRE_STD) and the
optimized parameters (PRE_OPT). The bias reduction (%) caused by the
optimization is listed in the last column. In the case of leaf, root, and
aboveground biomass, we use allometric equations from multiple sources
(Baldwin and Feduccia, 1987; Naidu et al., 1998; Vanlear et al., 1986) that went into
producing a range. The bias calculation uses the mean of the range. For sap
flow and soil respiration, daily observations were made, but the values
represent a mean over the 25 pretreatment days over both LS and HS periods.
Both HS and LS trees showed increasing trend in stem carbon during the pretreatment period, as inferred from stem thickness measurements. While the LS stems continued to grow during the treatment period, the observed HS stem size declined (Fig. 5a). Modeled relative increase in stem carbon was more rapid during the pretreatment period than observed; additionally, while the modeled LS trees continued to accumulate carbon during the treatment period (at a somewhat reduced rate), the modeled HS tree growth essentially stopped. The observed shorter-term (3–5 day) variation in stem carbon (based on diameter change) under shading (Fig. 5a) was attributed primarily to precipitation events and changing soil moisture (Figs. 2a and 4b) and the accompanying swelling and shrinkage of stem diameter, which translates through the allometric functions to apparent changes in stem biomass. Apart from whole-plant mortality and fire, the model has no physiological mechanisms allowing for negative growth of stems.
Both observed and simulated soil respiration tended to decline over the
study period (after day
Observations of foliar
The model reproduced observed pretreatment values for foliar, phloem, and
root tissue
The model predicted a steady dilution of labeled C in leaf, root, and
storage pools for the LS trees, compared to their HS counterparts. With a
severe reduction in PAR, GPP was greatly reduced in the modeled HS
treatment, and what little photosynthate produced was prioritized for
maintenance respiration, so the label appeared quickly in tissues and
remained relatively constant for that treatment. For the LS treatment GPP
remained relatively high following the labeling and initiation of the shade
treatment. In this case unlabeled C continued to accumulate as new growth,
causing a steady decline in the label concentration for LS trees over the
course of the experimental period (Fig. 6a, b, c, insets). In contrast to the
plant pools, modeled soil surface CO
Toward the end of the experimental period, the observed multi-day pulses of
labeled C in phloem and soil surface CO
Default model physiological parameters most appropriate to our site are
based on averages taken across numerous data sets collected in evergreen
needleleaf forests. There is considerable variation within that broad type
classification for all of the measured parameters (White et al., 2000), and
any time a site-level evaluation is used to assess model behavior (as here)
it is helpful to constrain within this range according to the local species
or species mixture. We used measurements taken directly from the site where
available and constrained the optimization of other parameters based on the
observed ranges for loblolly pine, when available. The fine-root-to-leaf
allocation ratio increased from 1.0 to 1.24, which is well within the range
of reported values (White et al., 2000). The fraction of leaf nitrogen in
RuBisCO was 70 % higher than the model default value and, while on the
high end, is consistent with measurements of other loblolly pine trees
(Tissue et al., 1995). The temperature sensitivity of maintenance
respiration (
Observed (black) and CLM simulated (blue) change in
The optimized model delivered very reasonable simulations of pretreatment
tree biomass, transpiration, and leaf
Independent evaluation of model results at the leaf scale demonstrated that the optimized parameters either reduced biases (LS and open-field light levels) or gave mixed results (HS light levels) at this scale. This provides additional confirmation that the optimization approach was reasonable and was not generating unrealistic parameter values to compensate for gross structural deficiencies in the model. This is further confirmed by the fact that optimized parameters (Table 1) controlling stomatal conductance changed only modestly from default values.
Independent evaluation of model against pretreatment
Though several changes in the canopy photosynthesis scheme were made in the
version 4.5 of CLM (Bonan et al., 2011; Oleson et al., 2013), in this work
the canopy photosynthesis process of CLM4 did a reasonably good job
against our evaluation metrics, including the leaf-level light-response
data. The ability of our optimized model to reproduce pretreatment biomass,
transpiration,
We did not attempt to optimize model predictions for soil temperature or soil moisture content. The model overestimation of soil temperature while faithfully reproducing the multi-day excursions in temperature is consistent through the pretreatment and treatment periods. Soil surface temperatures were not measured, so it is not clear whether the overestimation bias is related to a surface energy balance bias, to a bias in the overlying air temperature, or to parameterization error in thermal diffusivity and its relationship to soil texture and surface layer properties.
The overestimation bias in modeled soil moisture during the treatment period (there were no pretreatment observations) suggests a parameterization error for soil texture or variation in texture with depth. Small differences in the clay fraction, for example, could cause the observed offset in mean soil water content, and clearly there is variability in soil moisture states across the site, both within and between the shade treatments (Fig. 4b). We used a single estimate of sand, silt, and clay fractions from the site, and we were satisfied that the model was able to capture pretreatment transpiration with that soil parameterization and that the multi-day excursions of soil moisture were of similar magnitude in the model compared to observations during the treatment period. We also note that modeled stomatal conductance was not impacted by lack of soil water in these simulations. Periodic rainfall kept soils relatively wet throughout the pre-treatment and treatment periods, minimizing effects of bias in soil moisture on simulated photosynthesis or transpiration.
The very large difference between modeled and measured transpiration for the HS treatment is the most confounding result from our study. The model carbon and water dynamics are well behaved for the pretreatment period, and the model also captures the influence of light shading on transpiration accurately. Stem growth results indicate that reduced growth of LS trees and the cessation of growth for HS trees, is captured properly by the model. Through the Ball–Berry approximation linking stomatal conductance to photosynthetic rate, the model is forced into a state of reduced transpiration for the HS treatment, even with additional optimization that placed Ball–Berry parameters at their outer observational limits. It is possible that the sap-flow measurements in the HS treatment are biased, and that the actual tree-scale transpiration is not as high as suggested by these measurements, but if true we would expect that bias to occur for both pretreatment and treatment periods and not only, as observed, to appear in the treatment period. Connected to that hypothesis, it is possible that while actual leaf stomatal conductance shut down during the HS treatment, water continued to accumulate in the stem, moving past the sap-flow sensors and filling a capacitance in the xylem tissue. However, the sustained sap flow over the long duration of the treatment period and the negative observed trend in stem diameter for HS trees argue against that interpretation.
Alternatively, if we assume that the sap-flow measurements reflect actual high levels of transpiration in the HS trees, then we are forced to conclude that the Ball–Berry relationship as implemented in CLM (De Kauwe et al., 2013; Oleson et al., 2010, 2013) breaks down under these rather extreme experimental conditions. Under that hypothesis, it would seem that there is some “memory” of the expected range of light levels in the tree and, even when photosynthesis is nearly extinguished due to experimentally forced reduction in PAR, stomatal conductance remains at a relatively high level. Another possibility is that these trees exhibit a strong nonlinearity in the relationship between stomatal conductance and net photosynthesis, which has been observed at low light levels and strongly impacts estimated transpiration (Barnard and Bauerle, 2013). This type of nocturnal transpiration may indeed have been greater for the HS trees if the vapor pressure deficit were larger (Domec et al., 2012). Errors in modeled leaf temperature and leaf boundary layer vapor pressure deficit may also contribute to the discrepancy with observations. Conductance may have been maintained to some extent by vapor pressure differences between the foliage and the shade cloth – indeed, dew was observed on unshaded trees in early morning, yet not on the shaded trees. This hypothesis could be tested in future studies with additional leaf-level measurements under HS treatments, sampling both the diurnal cycle and the multi-day behavior of leaf physiology in trees subjected to high levels of shading. While the HS conditions are unlikely to be realized for extended periods under natural conditions, understanding this failure of the commonly used Ball–Berry parameterization may be helpful in understanding and predicting the broader case of adaptation of stomatal behavior to environmental change, which is known to influence water and carbon cycle predictions under future climates (Damour et al., 2010).
Stem diameter can shrink or swell based on changes in stem xylem water content, bark water content, and cambial growth and is dependent on xylem water potential, vapor pressure deficit, C availability, non-structural carbohydrate concentrations, and C allocation (Vandegehuchte et al., 2014). C allocation to stem growth is revealed by a step-wise increase in stem diameter that occurs in response to favorable conditions and maintained under less favorable conditions. The LS treatment clearly displayed the step-wise increases in stem diameter, while the HS treatment displayed a reduction in stem diameter. The shrinking stem diameter of HS trees indicates a decline in xylem and phloem water content likely linked to phloem sugar concentration. The HS treatment certainly reduced foliar C uptake and C available for phloem loading and allocation to cambial growth (Warren et al., 2012).
The modeled difference between LS and HS in biomass accumulation in stems is
in good agreement with observations based on stem diameter, with increases
of 1.9 and 1.6 % by treatment day 19 for model and observations,
respectively (Fig. 5a). Given the previously discussed pretreatment results
for biomass accumulation and leaf-scale photosynthesis, we are confident in
the optimized model's ability to capture carbon dynamics at the plant scale
on timescales of years to tens of days. It is reassuring to see that the
model prediction of soil respiration falls in the observed range, although
this could be the result of good luck as much as good performance. While
soil respiration on an annual basis is closely related to litter inputs and
belowground plant respiration, it is possible for compensating errors
between decomposition rates and litter inputs, or between litter inputs and
root respiration, to result in good model–observation agreement for the
approximately monthly timescale examined here. We note a potential bias in
the model relationship between soil respiration and soil moisture: while the
observed soil respiration is depressed after large precipitation events, the
model estimates an increase. Neither CLM4's carbon allocation to roots nor
its predicted root respiration is dependent on soil water conditions. CLM4's
heterotrophic contribution to soil respiration may also have too little
sensitivity and the timing of soil respiration response to soil water
variation may also be too simplistic. A more mechanistic treatment of
water–air–microbe interactions at the scale of soil pore space might help to
eliminate these differences. Resolved vertical transport of respired
CO
Beyond noting the obvious discrepancy in observed vs. modeled
Given that LS and HS leaves seem to have photosynthesized the pre-shading
labeled pulse of CO
Conceptual model of label transport, assuming a constant velocity
(
Plant storage pools in the form of non-structural carbohydrates are known to
play an important role in regulating allocation to structural pools, and they may
make up a significant portion of total biomass (e.g., Hoch et al., 2003).
Simple models that account for non-structural carbohydrates better compare
with observed
Representing the existence and dynamics of short-term photosynthate storage
pools in a model like CLM could also help to resolve the mechanisms relating
nutrient mineralization and availability in soils with plant–microbe
competition for available nutrients and the influence of nutrient uptake on
leaf-scale photosynthesis. In addition to the shading treatments described
here, other manipulations that would be useful to explore include elevated
CO
Limitations identified in this first PiTS model–experiment interaction have already led to improvements in follow-on experiments. For new experiments in a nearby dogwood stand, additional observations include multiple treatments in different seasons, a collection of absolute destructive tree biomass at the end of the study (rather than highly uncertain estimates based on allometric relationships), seasonal leaf-level photosynthetic measurements, assessment of mycorrhizal C flux, and improved meteorological measurements. Although model parameters can be improved through optimization as in this study, model parameters are being measured where possible. These additional observational data are necessary for more detailed model evaluation and improvement of model routines of C and allocation patterns at various timescales. Additional effort is being devoted to characterizing the system prior to manipulation, including measurements of biomass, soil physical, and soil biogeochemical states.
The point version of CLM4 was implemented, calibrated, and evaluated against
carbon and hydrology observations from a shading and labeling experiment in
a stand of young loblolly pines. We found that a combination of parameters
measured on-site and calibration targeting biomass, transpiration, and
Although the model lacks short-term photosynthate storage and transport mechanisms that are clearly present in the real plants, first-order monthly timescale dynamics for carbon allocation and growth do not seem to suffer greatly. We used observations from the experiment to develop a conceptual model (hypothesis) of short-term photosynthate storage and transport and suggested further studies that could be carried out to evaluate the generality of the hypothesized mechanisms. We suggest several research problems, which, if the proposed mechanism turns out to be generally valid, would benefit from model–experimental study in which the new mechanisms are incorporated into the model structure.
This work is supported by the US Department of Energy (DOE), Office of Science, Biological and Environmental Research. Oak Ridge National Laboratory is managed by UT-BATTELLE for DOE under contract DE-AC05-00OR22725. Edited by: T. Keenan