Accurate simulations of soil respiration and carbon
dioxide (
This paper has been authored by UT-Battelle, LLC under contract no. DE-AC05-00OR22725 with the U.S. Department of Energy.
The United States Government retains and the publisher, by accepting the article for publication, acknowledges that the United States Government retains a non-exclusive, paid-up,
irrevocable, worldwide license to publish or reproduce the published form of this paper, or allow others to do so, for United States Government purposes.
The Department of Energy will provide public access to these results of federally sponsored research in accordance with the DOE Public Access Plan
(
Globally, soils store over twice as much carbon (C) as the atmosphere (Chapin
III et al., 2011). Soil respiration (SR) is the second largest C flux between
terrestrial ecosystems and the atmosphere (Luo and Zhou, 2006). An accurate
simulation of SR is critical for projecting terrestrial C status, and
therefore climate change, in Earth system models (ESMs) (IPCC, 2013). Despite
significant experimental data accumulation and model development during the
past decades, simulations of soil
To assess the performance of ESMs, different types of data can be used. For
example, using atmospheric
Despite the significance of large global SR fluxes, SR has rarely been evaluated in ESMs using long-term observations. Among the factors that influence SR, soil water potential (SWP) provides a unified measure of the energy state of soil water that limits the growth and respiration of plants and microbes. Unlike soil temperature (ST) or soil volumetric water content (VWC), however, SWP is difficult to directly monitor in the field. Accurate estimation of SWP largely relies on the soil water retention curve (i.e., the relationship between VWC and SWP), which is highly specific to soil properties (Childs, 1940; Clapp and Hornberger, 1978; Cosby et al., 1984; Tuller and Or, 2004; Moyano et al., 2013). Site-level data have been used to evaluate model representations of other processes, such as phenology, net primary production (NPP), transpiration, leaf area index (LAI), water use efficiency, and nitrogen use efficiency (Richardson et al., 2012; De Kauwe et al., 2013; Walker et al., 2014; Zaehle et al., 2014; Mao et al., 2016; Duarte et al., 2017; Montané et al., 2017). In Powell et al. (2013), the only aspect influencing the modeling of SR was the sensitivity of SR to VWC in an Amazon forest, but the study resulted in no improvements to simulated SR. Here, we focus on improving simulations by using site-specific measurements to assess multiple factors influencing SR.
We will evaluate the simulation of SR step by step. We assessed underlying
mechanisms in the Energy Exascale Earth System Model (E3SM) land model
version 0 (ELMv0) by using intensive observations at the Missouri Ozark
AmeriFlux (MOFLUX) forest site in the central US. We first evaluated the
effects of two abiotic factors, ST and SWP, on the simulation of SR. Then we
evaluated the effects of biotic factors, such as GPP, LAI, and
The MOFLUX site is located in the University of Missouri's Thomas H. Baskett
Wildlife Research and Education Area (latitude
Ecosystem C, water and energy fluxes, SR, LAI, and supporting meteorological
measurements were initiated in June 2004 (Gu et al., 2016). Soil respiration
was measured within the ecosystem flux tower footprint using non-flow-through
non-steady-state auto-chambers. From 2004 through 2013, SR was measured using
eight automated, custom-built chambers (ED system; Edwards and Riggs, 2003;
Gu et al., 2008) coupled with an infrared gas analyzer (LI-820 LI-COR Inc.,
Lincoln, Nebraska). In 2013, this system was replaced with 16 auto-chambers
operated using the closed-path system (model LI-8100; LI-COR Inc., Lincoln,
Nebraska). The two systems (ED and LI-8100) were operated side by side for
several weeks in 2010 and found to produce comparable responses (Paul Hanson,
personal communication, 2017). Half-hourly SR time series were generated to
coincide with the ecosystem flux dataset by averaging those chambers sampled
in the corresponding averaging period. Net ecosystem
Flux-tower GPP was estimated from measured NEE. To reduce biases resulting
from individual methods, three NEE-partitioning approaches were employed. The
average and variation of the three methods were used to evaluate the
model-simulated GPP. In the first two methods, ecosystem respiration (ER) was
estimated from nighttime NEE and extrapolated to daytime, and daytime GPP was
calculated from NEE and the extrapolated ER (Reichstein et al., 2005). The
only difference between the two methods was whether they excluded nighttime
data under non-turbulent conditions. In the third method, GPP was estimated
by fitting the light-response curve between NEE and radiation (Lasslop et
al., 2010). All the partitioning calculations were conducted using the R
package
ELMv0 used in this study is structurally equivalent to the Community Land
Model 4.5 (CLM 4.5), which includes coupled carbon and nitrogen cycles
(Oleson et al., 2013). In ELMv0, the soil biogeochemistry can be simulated
with a one-layer or multi-layer converging trophic cascade (CTC, i.e., CLM-CN)
decomposition model. We used the vertically resolved CTC decomposition in
this study. In the model, SR was calculated by different
ELMv0 is a grid-based model. To assess it using site-level observations, we used a point-run framework which allows the model to simulate individual sites (Mao et al., 2016). Single-point runs forced with site-level measurements have a long history to evaluate model representations of phenology, NPP, transpiration, LAI, water use efficiency, and nitrogen use efficiency (Richardson et al., 2012; De Kauwe et al., 2013; Walker et al., 2014; Zaehle et al., 2014; Mao et al., 2016; Duarte et al., 2017; Montané et al., 2017). With site-specific forcing, a 200-year accelerated decomposition spin-up was performed, followed by a 200-year normal spin-up, before the transient simulation was performed from 1850 to 2013. The vegetation was set as 100 % temperate deciduous forest.
Soil water potential values for the Weller soils were estimated from observed VWC and soil water retention curves that were developed for the site. To derive the soil water retention curves, soil samples were collected in the area of the flux tower base at two depths: 0 to 30 cm and below 30 cm. Samples were evaluated periodically for soil water potential using a dew-point potentiometer (Decagon Devices, Model WP4C) as they dried over time (Hanson et al., 2003).
In ELMv0, the SWP was calculated from VWC based on the Clapp and
Hornberger model (Clapp and Hornberger, 1978), in which the SWP–VWC
relationship was expressed as
Root mean square error (RMSE) and Akaike information criterion (AIC) of different models in simulating the SWP–VWC relationship for the soil in the MOFLUX site at two depths: 0 to 30 cm and below 30 cm.
Observed (black dots) and simulated relationship between soil water
potential (SWP) and volumetric water content (VWC) by the different models at
two soil layers:
In the Brooks and Corey model, the SWP–VWC relationship was expressed as
In the Fredlund and Xing model, the SWP–VWC relationship was described as
In the Hanson model (Hanson et al., 2003), soil matric potential was
modeled by a double exponential function:
In the van Genuchten model, the SWP–VWC relationship was described as
In addition to the default SWP–VWC relationship in ELMv0, all five
empirical models were parameterized using non-linear fitting against measured
VWC and SWP data from the study site. For the calibration of the Clapp and
Hornberger model, instead of using the hard-coded parameters in
Eqs. (11)–(13), we calibrated the three parameters (i.e.,
The evaluation of SR was conducted step by step. We first compared
observations with the model default output of SR and related factors,
including ST, SWP, GPP, and LAI. Thereafter, we attempted to improve the
simulation of these factors in order to improve the overall SR simulation by
(i) implementing the best-fit SWP–VWC relationship and (ii) modifying model
parameters related to GPP, LAI, and SR. GPP-related parameters included the
specific leaf area (SLA) at the top of canopy and the fraction of leaf
nitrogen in the RuBisCO enzyme. LAI-related parameters included the number of
days to complete leaf fall during the end of growing season, the critical day
length for senescence (i.e., the length of the day when leaves start to
senesce), and a parameter
For the upper 30 cm of soil, the ELMv0 simulations using the default Clapp and Hornberger model tended to underestimate the SWP when VWC was less than 15 % (Fig. 1a), while SWP rapidly approached zero when VWC was greater than 25 % (Fig. 1a). For soil below 30 cm, ELMv0 showed a consistent overestimation of SWP (Fig. 1b). The default ELMv0 showed relatively high RMSE for both soil layers, indicating that the SWP–VWC relationship was not well simulated in ELMv0 (Table 1). Although the Clapp and Hornberger model performed better by using parameters from non-linear fitting, its performance was not as good as the Hanson and van Genuchten models (Table 1, Fig. 1). The Hanson model was the best-fit model for the MOFLUX site, showing the smallest RMSE and AIC values for both soil layers (Table 1, Fig. 1), and was therefore implemented in ELMv0 to calculate SWP from measured VWC.
The ELMv0 default run significantly underestimated both annual SR and GPP
(Fig. 2). In addition, the simulated SR had smaller interannual variability
compared to the observations. The model was not able to simulate the steep
drop of SR or GPP during the extreme drought in 2012. The simulations of ST
and SWP were isolated to analyze their contributions to model performance.
Whereas the model-simulated ST well at 10 cm depth (Fig. 3a), it tended to
underestimate SWP when water was limited and to overestimate SWP otherwise
(Fig. 3b). Implementing the data-constrained Hanson model significantly
improved the simulation of SWP, showing a greater
Annual SR and GPP.
Blue and red lines are model outputs before (MOD
Despite the improved simulation of SR, the model still underestimated SR and GPP during peak growing seasons when SR and GPP were high and overestimated them during non-growing seasons (Figs. 4, S4). In other words, though the improved simulation of SWP increased SR and GPP during peak growing seasons, the model still showed systematic errors. We attempted to improve the seasonal simulations of SR, GPP, and LAI by modifying several related parameters (Table 2). Using measurements of C and energy fluxes from the MOFLUX site, Lu et al. (2018) calibrated a polynomial surrogate model of ELMv0. Based on their results, we modified two parameters, i.e., the SLA at the canopy top from 0.03 to 0.01 and the fraction of leaf nitrogen in the RuBisCO enzyme from 0.1007 to 0.12.
Daily ST and SWP at 10 cm. Blue and red lines/dots are model outputs before
(MOD
Modified parameters to better simulate GPP and LAI at the MOFLUX site in ELMv0.
The annual mean cycles of LAI, GPP, and SR. OBS: observation;
MOD
Comparing the simulated LAI with the observations (Fig. 4), we found that the
parameter
In addition, we analyzed changes in simulated evapotranspiration (ET), runoff, photosynthesis, net primary production, C allocations to fine roots, leaf and woody tissue in response to the changes in the soil water scheme and parameters (Figs. S6, S7). The change in soil moisture scheme and parameter adjustments slightly increased ET and decreased runoff. Despite these slight changes, the model-simulated ET generally fell within the observed range, with or without changes in soil water scheme and parameters (Fig. S6). The improved SWP and parameter adjustments generally increased all photosynthesis, NPP, and carbon allocations to different tissues during the growing season (Fig. S7).
Constraining the SWP–VWC relationship with site-specific data and using the
Hanson model instead of the ELMv0 default model (Fig. 1) significantly
improved the model representation of SWP (Fig. 3) and annual SR (Fig. 2a).
The improvements in model fits could be due to the following reasons. First,
the changes in SWP with the Hanson model increased plant transpiration and
GPP in the model. The default ELMv0 underestimated GPP (Fig. 2b), similar to
a recent study where CLM4.5 significantly underestimated GPP at a coniferous
forest in northeastern United States (Duarte et al., 2017). GPP can directly
affect the magnitude of root respiration, as shown in many previous studies
(Craine et al., 1999; Högberg et al., 2001; Wan and Luo, 2003; Verburg et
al., 2004; Gu et al., 2008). Additionally, increased GPP can build a larger
SOC pool, which is the substrate for heterotrophic respiration (Fig. S3).
Second, the Hanson soil moisture model increased the moisture modifier
(
The simulation of SWP in the default ELMv0 was poor compared with that of ST
(Fig. 2), which may be a common issue in ESMs. For example, using a
reduced-complexity model, Todd-Brown et al. (2013) demonstrated that the
spatial variation in soil C in most ESMs is primarily dependent on C input
(i.e., NPP) and ST, showing
In this study, we derived a better SWP–VWC relationship by using non-linear fitting, primarily because of the availability of soil moisture retention curve data. It is an efficient method when site-level data are available, but it is not realistic to calibrate the water retention curve for every site. The SWP–VWC relationship is dependent on soil texture (Clapp and Hornberger, 1978; Cosby et al., 1984; Tuller and Or, 2004), so building relationships between model parameters and soil texture may allow efficient extrapolations of site-level measurements to regional and global scales.
Parameters in the default Clapp and Hornberger model used in ELMv0 were derived from synthesizing data across soil textural classes (Clapp and Hornberger, 1978; Cosby et al., 1984; Lawrence and Slater, 2008). The data were derived from over 1000 soil samples from 11 USDA soil textural classes (Holtan et al., 1968; Rawls et al., 1976). The dependence of model parameters on soil texture was derived from a regression of these 11 data points, i.e., the mean parameter values of 11 soil textural classes against the sand or clay fractions (Cosby et al., 1984). Because no actual sand or clay content of soil samples was reported in the original databases (i.e., only the soil textural classes were reported), the sand and clay fractions used for the regression were obtained from midpoint values of each textural class (Clapp and Hornberger, 1978; Cosby et al., 1984). One potential issue is that soil samples in the same textural classes can have different sand and clay content and SWP–VWC relationships, which may not be fully represented when they are grouped together. An updated SWP–VWC database with actual sand and clay content measurements could provide improved empirical relationships between model parameters and soil texture in the water retention model.
In addition, different empirical models have been developed to describe the SWP–VWC relationship (Brooks and Corey, 1964; Clapp and Hornberger, 1978; van Genuchten, 1980; Fredlund and Xing, 1994; Hanson et al., 2003). These models could be evaluated against data, and the selected best-fit model(s) could be used to calculate SWP in the field from continuously monitored VWC (e.g., from the AmeriFlux network) on different spatial and temporal scales. The database could also be used as a benchmark to evaluate simulations of soil water and biogeochemical processes in ESMs.
Moreover, we also explored whether the calibrated Clapp and Hornberger model
can lead to similar improvements with the Hanson model (Fig. S8). Generally,
both the Hanson model and the calibrated Clapp and Hornberger model improved
the simulation of GPP and SR in the ELM, in comparison with the default run
(Fig. S8). ELMv0 with the Hanson model consistently produced higher GPP
and SR than that with the calibrated Clapp and Hornberger model. In
comparison with the observations, the modeled SR generally fell within the
Although the SWP simulations using the Hanson model improved the representation of both annual SR and GPP, the model continued to overestimate SR during the non-growing season (Fig. 4), resulting in significant overestimations of the annual SR fluxes (Fig. S5). No matter which SWP simulations were used, ELMv0 had smaller interannual variability than the observations (Fig. 2). Specifically, the model was not able to capture the steep decreases in GPP and SR in the extreme drought year (i.e., 2012; Fig. S9). These results indicate that the current model structure is not sensitive enough to environmental changes. Several potential reasons may contribute to the underestimated seasonal and interannual variability. For example, field inventory data at the study site showed that the severe drought–pathogen interactions in 2012 resulted in a significant stem mortality of tree species (Wood et al., 2017). Thus, the observed steep decreases in GPP and SR could be due to mortality. The stem mortality could lead to lower evapotranspiration (Fig. S9), minimizing soil moisture losses (Fig. S10). However, ELMv0 simulated the moisture effect on biogeochemical cycles at the physiological level but not at the plant community level. In addition, the strong dependence of GPP and SR on the upper layer soil moisture could explain the model's difficulty in capturing interannual variability. Although better representation of SWP improved the mean annual simulation of biogeochemical processes, the model could not capture the mortality or the interannual variability of GPP and SR.
The calculation of the moisture scalars (e.g., btran and
Modeled contributions of autotrophic
(
In ELMv0, heterotrophic respiration contributed the majority (i.e., over
85 %) of total SR during non-growing seasons (Fig. 5), suggesting that the
overestimation of SR during these seasons was primarily due to the biased
heterotrophic respiration simulation. A potential reason for the biased
heterotrophic respiration simulation may be related to the temperature
sensitivity (
Another potential reason for the biased heterotrophic respiration simulation
may be that the seasonality of microbial organisms was not adequately
represented in the model. Like most ESMs, ELMv0 represents soil C
dynamics using linear differential equations and assumes that SR is a
substrate-limited process in the model. However, producers of
Additionally, the lack of representation of macroinvertebrates and other
forest floor and soil fauna in ELMv0 may be another reason. There is a
high density of earthworms at the MOFLUX site (Wenk et al., 2016). Earthworms
can shred and redistribute soil C and change soil aggregation structure,
which may alter soil C dynamics and
Our analyses also showed that the modeled SR was not able to reach the observed peak in many years during the peak growing season, even when the modeled GPP exceeded the observation. In addition, the parameter modification increased GPP during both peak and non-growing seasons, resulting in an even greater overestimation of SR during non-growing seasons. These results suggest that simply increasing GPP may not be adequate to increase the seasonal variability of the simulated SR. A potential reason may be that the current model does not include root exudates. Root exudates are labile C substrates that are important for SR (Kelting et al., 1998; Kuzyakov, 2002; Sun et al., 2017). The root exudate rate is primarily dependent on root growth, showing a seasonal cycle in temperate forests (Kelting et al., 1998; Kuzyakov, 2002). Thus, including root exudates in the model may further increase the model-simulated SR during the peak growing season without needing to increase GPP.
In this study, we used temporally extensive and spatially distributed site observations of SR to assess the capabilities of ELMv0. These results indicated that an improved representation of SWP within the model provided better simulations of annual SR. This underscores the need to calibrate SWP in ESMs for more accurate projections of coupled climate and biogeochemical cycles. Notwithstanding this improvement, however, ELMv0 still underestimated seasonal and interannual variabilities. It may be that inadequate model representation of vegetation dynamics, moisture function, and the dynamics of microbial organisms and soil macroinvertebrates could be explored as means to achieve better fit. Future incorporation of explicit microbial processes with relevant data collection activities may therefore enable improved model simulations.
The code for ELMv0 is available on GitHub (
The data for this paper are available upon request to the corresponding author.
The supplement related to this article is available online at:
JL, GW, and MAM designed the study. JL, GW, and DMR ran the model. LG, PJH, and JDW contributed to data collection. JL wrote the paper with input from all authors.
The authors declare that they have no conflict of interest.
The authors thank Dan Lu for sharing unpublished data, and William Wieder and two anonymous referees for constructive comments. This work is financially supported by the U.S. Department of Energy (DOE) Office of Biological and Environmental Research through the Terrestrial Ecosystem Science Scientific Focus Area (TES-SFA) at Oak Ridge National Laboratory (ORNL), the Climate Model Development and Validation (CMDV) project, and the Energy Exascale Earth System Model (E3SM) project. ORNL is managed by UT-Battelle, LLC, under contract DE-AC05-00OR22725 with the U.S. DOE.
This paper was edited by Tomomichi Kato and reviewed by William Wieder and two anonymous referees.