Introduction
The amplitude of the atmospheric CO2 seasonal cycle is largely controlled
by vegetation growth and decay in the Northern Hemisphere (NH)
(Bacastow et al., 1985; Graven et al., 2013; Hall et al., 1975; Heimann et al., 1998;
Pearman and Hyson, 1980; Randerson et al., 1997). Since 1958, atmospheric
CO2 measurements at Mauna Loa, Hawai'i, have tracked a 15 % rise in the
peak-to-trough amplitude of the detrended CO2 seasonal cycle
(Zeng et al., 2014), suggesting an enhanced ecosystem
activity due to changes in the strength of the ecosystem's production and
respiration and to a shift in the timing of their phases
(Randerson et al., 1997). In addition, some evidence
suggests a latitudinal gradient in CO2 amplitude increase in the NH,
with a larger increase at Pt. Barrow, Alaska (0.6 % yr-1) than at Mauna
Loa (0.32 % yr-1) (Graven et al.,
2013; Randerson et al., 1999). Previous studies have attempted to attribute
the long-term CO2 amplitude increase to stimulated vegetation growth
under rising CO2 and increasing nitrogen deposition
(Bacastow et al., 1985; Reich
and Hobbie, 2013; Sillen and Dieleman, 2012). Another possible explanation
offered is the effect of a warmer climate, especially in boreal and
temperate regions, on the lengthening of growing season, enhanced plant
growth (Keeling et al., 1996; Keenan et al.,
2014), vegetation phenology (Thompson, 2011), ecosystem
composition and structure (Graven et al., 2013). The
agricultural green revolution, due to widespread irrigation, increasing
management intensity and high-yield crop selection, could also contribute to
the dynamics of the CO2 seasonal amplitude
(Zeng et al., 2014; Gray et al., 2014). Even
though these studies are helpful in understanding the role of CO2,
climate and land use/cover changes, detailed knowledge of the relative
contribution of these factors is still lacking.
Dynamic vegetation models are useful tools not only to disentangle effects
of various mechanisms but also to offer insights on how terrestrial
ecosystems respond to external changes. Attribution of the role of CO2,
climate and land use has been attempted with a single model
(Zeng et al., 2014), but comprehensive multimodel
assessment efforts are still missing. Two important questions must be
addressed in such efforts, namely, whether the models can simulate observed
CO2 amplitude increase, and to what extent their factorial attributions
agree. For the first question, the Coupled Model Intercomparison Project
Phase 5 (CMIP5) Earth System Models seem to be able
to simulate the amplitude increase measured at the Mauna Loa and Point
Barrow surface stations (Zhao and Zeng, 2014); however, they
underestimate the amplitude increase compared to upper air
(3–6 km) observations significantly (Graven et al., 2013). It is
possible that uncertainty in vertical mixing in atmospheric transport models
(Yang et al., 2007), instead of biases in
dynamic vegetation models themselves, causes the severe underestimation of
upper air CO2 amplitude increase. For the second question, in a unique
modeling study conducted by
McGuire et al. (2001), both
CO2 fertilization and land use/cover changes were found to contribute
to CO2 amplitude increase at Mauna Loa, but the four models disagreed
on the role of climate and the relative importance of the factors they
studied. Since then, no published study has explored the reliability of
models' simulation of seasonal carbon cycle and quantified the relative
contribution of various factors affecting it.
An important trait of the three main factors (i.e., CO2, climate and
land use/cover change) we consider in this study is their different regional
influence. Rising CO2 would likely enhance productivity in all
ecosystems. Climate warming may affect high-latitude ecosystems more than
tropical and subtropical vegetation, and droughts would severely affect
plant growth in water-limited regions. Similarly, the effect of land
use/cover change may be largely confined to agricultural fields and places
with land conversion, mostly in midlatitude regions. Because of their
different spatial traits, it is possible to determine which factor is most
important with strategically placed observations. Forkel
et al. (2016) recently derived a latitudinal gradient of CO2 amplitude
increase based on CO2 observational data, which would provide strong
support that high-latitude warming is the most important factor. However,
with only two sites north of 60∘ N, the robustness of the result is limited.
In lieu of additional observational evidence, as a first step, it is
necessary to investigate how the models represent the regional patterns of
seasonal change of carbon flux.
A number of recent studies have addressed different aspects of the seasonal
amplitude topic. For example, the latitudinal gradient of CO2 seasonal
amplitude was used as benchmark in assessing the performance of JSBACH model
(Dalmonech and Zaehle, 2013; Dalmonech et al.,
2015). Based on a model intercomparison project – Multiscale Synthesis and Terrestrial Model Intercomparison Project (MsTMIP; Huntzinger
et al., 2013; Wei et al., 2014) – Ito et al. (2016)
focused on examining the relative contribution of CO2, climate and land
use/cover changes, but little model evaluation was performed. In order to
further explore and understand the seasonal fluctuation of carbon fluxes, a
more comprehensive study including both the model evaluation and factorial
analysis is needed. The TRENDY model intercomparison project provides a nice
platform for such analysis
(Sitch et al., 2015). Site-level model–data comparison of seasonal carbon fluxes has
been performed extensively in Peng et al. (2015) for the
first synthesis of TRENDY models. Using both the second synthesis of TRENDY
models simulations and observations, in this study we aim to achieve two main
goals. (1) Assess how well the models simulate the climatological seasonal
cycle and seasonal amplitude change of the carbon flux against a number of
observational-based datasets (CO2 observations and atmospheric
inversions). (2) Analyze the relative contribution from the three main factors
(CO2 fertilization, climate and land use/cover change) to the seasonal
amplitude increase, both at the global and regional level.
Method
Terrestrial ecosystem models and TRENDY experiment design
Monthly net biosphere production (NBP) simulations for 1961–2012 from nine
TRENDY models participating in the Global Carbon Project
(Le Quéré et al., 2014) were examined (Table 1). A set of three offline
experiments driven by either constant or varying climate data and other
input such as atmospheric CO2 and land use/cover forcing were designed
in the TRENDY project to differentiate the role of CO2, climate and
land use (Table 2). We primarily evaluated results from the S3 experiment,
where the models are driven by time-varying forcing data (Appendix A). In
addition, we also used results from the S1 and S2 experiments.
Basic information for the nine TRENDY models used in this study.
Model name
Abbreviation
Spatial resolution
Nitrogen
Fire
Harvest
Reference
cycle
simulation
flux
Community Land Model 4.5
CLM4.5BGC
1.25∘ × 0.94∘
yes
yes
no
Oleson et al. (2013)
ISAM
ISAM
0.5∘ × 0.5∘
yes
no
yes
Jain et al. (2013)
Joint UK Land Environment Simulator
JULES
1.875∘ × 1.25∘
no
no
no
Clark et al. (2011)
Lund-Potsdam-Jena
LPJ
0.5∘ × 0.5∘
no
yes
yes
Sitch et al. (2003)
LPX-Bern
LPX-Bern
0.5∘ × 0.5∘
yes
yes
yes
Stocker et al. (2014)
O-CN
OCN
0.5∘ × 0.5∘
yes
no
yes
Zaehle and Friend (2010)
ORCHIDEE
ORCHIDEE
2∘ × 2∘
no
no
yes
Krinner et al. (2005)
VEGAS
VEGAS
0.5∘ × 0.5∘
no
yes
yes
Zeng et al. (2005b)
VISIT
VISIT
0.5∘ × 0.5∘
no
yes
yes
Kato et al. (2013)
Observations and observational-based estimates
In light of the large difference in the Coupled Climate Carbon Cycle Model
Intercomparison Project (C4MIP) models' sensitivity to
CO2 change (Friedlingstein et al., 2013), it is essential to evaluate whether the terrestrial biosphere models
are able to capture important features of CO2 seasonal cycle. The
scarcity of observational constraints, especially the lack of long-term
continuous observational records, limits our capacity to fully evaluate the
dynamic processes in terrestrial ecosystem models. Nevertheless, in this
study we make a first-order approximation of the evolution of the global
CO2 seasonal cycle, using limited CO2 observation data. Following
Zeng et al. (2014), monthly Mauna Loa records from 1961
to 2012 and a global monthly CO2 index for the period of 1981–2012 were
retrieved from NOAA's Earth System Research Laboratory (ESRL; www.esrl.noaa.gov/gmd/ccgg/trends/).
Details on the data processing, choice of stations and quality control
procedures in deriving the global CO2 index (globally averaged CO2
concentration) can be found in Thoning et al. (1989) and Masarie and Tans (1995).
Fluxes from process-based models can be directly compared with monthly
gridded fluxes from atmospheric inversions, which combine measured
atmospheric CO2 concentration at multiple sites across the globe with
atmospheric transport driven by meteorological data. Two representative
inversions, Jena (Jena81 and Jena99, Rödenbeck et al., 2003) and the
CarbonTracker (Peters et al., 2007), are included for comparison (Appendix B).
For an exhaustive intercomparison of the atmospheric inversions, please
refer to Peylin et al. (2013).
Calculating the seasonal cycle and its amplitude change
All monthly NBP- and inversion-derived fluxes are first resampled
(box averaging, conserving mass) to a uniform 0.5∘ × 0.5∘
global grid in units of kg C m-2 yr-1. For the TRENDY
model simulations, we further define net carbon flux from the land to the
atmosphere (FTA), which simply reverses the sign of NBP, so that
positive FTA indicates net carbon release to the atmosphere, and
negative FTA indicates net carbon uptake. FTA represents the sum of residual
land sink and land use emission, including fluxes from ecosystem production
and respiration, fire, harvest, etc.; although some models may not simulate
all the processes. Changes in global atmospheric CO2 concentration are then
equal to FTA plus ocean–atmosphere flux and fossil fuel emission. For
inversion-derived fluxes, only terrestrial ecosystem fluxes are used
(optimized global biosphere fluxes plus fire fluxes in CarbonTracker), which are conceptually
similar to FTA, except that atmospheric transport is included.
Atmospheric transport can significantly affect local carbon fluxes
(Randerson et al., 1997); however, the impact is limited
on global and large zonal band totals.
The seasonal amplitudes of Mauna Loa Observatory CO2 growth rate,
global CO2 growth rate and fluxes from model simulations and
inversions are processed with a curve fitting package called CCGCRV from
NOAA/ESRL (http://www.esrl.noaa.gov/gmd/ccgg/mbl/crvfit/crvfit.html). This package
first filtered out the high-frequency signals with a series of internal
steps involving polynomial and harmonic fitting, detrending and band-pass
filtering, and then the amplitude is defined as the difference between each
year's maximum and minimum. For the latitudinal plots only, we simply use
maximum and minimum of each year to define the seasonal amplitude without
first filtering the data. Previous studies
(Graven et al., 2013; Randerson et al.,
1997) have established that FTA accounts for most of seasonal amplitude
change from atmospheric CO2, and the Mauna Loa CO2 record is
considered to represent the evolution of global mean CO2 well
(Kaminski et al., 1996). Therefore, similar to our earlier work
(Zeng et al., 2014), we evaluated the amplitude change of
modeled FTA with Mauna Loa CO2, ESRL's global CO2 and the
atmospheric inversions, to assess whether the models are able to capture
both the global trend and latitudinal patterns. For relative amplitude
changes, we compute the multimodel ensemble mean after deriving the time
series (relative to their 1961–1970 mean) from individual model simulations,
so that models with large amplitude change would not have a huge effect on
the ensemble mean. Additionally, global and regional mean seasonal cycles
over 2001–2010 between the models and inversions are compared. We further
compared the seasonal amplitude of zonally averaged FTA from TRENDY and
atmospheric inversions. To smooth out minor variations but ensure similar
phase in aggregation, we first resampled FTA into 2.5∘
resolution, then summed it over latitude bands for the 2001–2010 mean FTA
seasonal cycle.
Factorial analyses
Relative amplitudes for 1961–2012 (relative to 1961–1970 mean seasonal
amplitude) from the experiments S1, S2 and S3, respectively, are calculated
using the CCGCRV package for each model, and a linear trend (in % yr-1)
is determined for that period. We use relative amplitude for
percentage change to minimize impacts of some differing implementation
choices like climate data in S1 (CO2) among the models. The effect of
CO2 on the relative amplitude change is represented by a trend of S1
(CO2 only) results; the S2 (CO2 + climate) results show a trend
that is the sum of CO2 and climate effects, and the S3
(CO2 + climate + land use/cover) simulations include trends from
time-varying CO2, climate and land use/cover change (abbreviated as
land use for text and figures). For simplicity, the effect of “climate” as
used in this paper includes the synergy of CO2 and climate, and
similarly the effect of “land use/cover” also includes the synergy terms.
Therefore, the effects of CO2, climate and land use/cover are then
quantified as the trend for S1, the trend of S2 minus the S1 trend and the trend of S3
minus the S2 trend, respectively. Note that the synergy terms are likely small
in some of the current generation dynamic vegetation models, such as those
shown in previous sensitivity experiment results (Zeng et
al., 2014).
Spatial attribution
Spatial attribution of global FTA amplitude change can be difficult due
to the phase difference at various latitudes. For example, the two amplitude
peaks at northern and southern subtropics caused by monsoon movements are
largely out of phase, and the net contribution to global FTA amplitude
increase after their cancelation is small (Zeng et al.,
2014). To quantify latitudinal and spatial contributions for each model, a
unique quantity – FkAi – the difference between the
maximum month (i_max) and the minimum month
(i_min) of model i's global FTA, based
on model i's 2001–2010 mean seasonal cycle is defined as Eq. (1):
FkAi=FkA(i_max)i-FkA(i_min)i.
The subscript k denotes the index of each latitudinal band or spatial
grid, and A is the index of the year, ranging from 1961 to 2012.
FkAi could be quite different for each model: for
VEGAS, FkAi is FTA in November
(i_max = 11) minus FTA in July
(i_min = 7) in year A, and for
LPJ, FkAi is FTA in March
(i_max = 3) minus FTA in June
(i_min = 6) in year A. The
indexes i_max and i_min are
fixed for each model, as summarized in Table 3. For all three experiments,
FkAi is computed each year in 1961–2012 and at every
latitude band or spatial grid (k), and then the trends of
FkAi are calculated. The spatial aggregation of the
resulted latitudinal-dependent trends would then approximately be equal to the trend
of global FTA maximum-minus-minimum seasonal amplitude.
Results
Mean seasonal cycle of <inline-formula><mml:math display="inline"><mml:mrow><mml:msub><mml:mi>F</mml:mi><mml:mtext>TA</mml:mtext></mml:msub></mml:mrow></mml:math></inline-formula>
Four of the nine models (CLM4.5BGC, LPX-Bern, ORCHIDEE and VEGAS) simulate a
mean global FTA seasonal cycle of similar amplitude and phase compared
with the Jena99 and CarbonTracker inversions (Fig. 1, Table 3). The other
five models have much smaller seasonal amplitude than inversions, and the
shape of the seasonal cycle is also notably different. As a result, the models'
ensemble global FTA has seasonal amplitude of 26.1 Pg C yr-1 during
2001–2010, about 40 % smaller than the inversions (Fig. 4 inset, Table 3).
The model ensemble annual mean FTA (residual land sink plus land
use emission) is -1.1 Pg C yr-1 for 2001–2010, 30 % smaller than
the inversions (Table 3). In some models (ISAM, JULES and LPJ for the
northern temperature region in Fig. 2) FTA rebounds back quickly,
resulting in a late summer FTA maximum. The midsummer rebound is
unlikely a model response to pronounced seasonal drought after 2000, as it
is persistent in the mean seasonal cycle over every decade since 1961. A
probable cause is the strong exponential response of soil respiration to
temperature increase, which may lead to heterotopic respiration higher than
net primary production (NPP) in summer. For example, HadGEM2-ES and HadCM3LC, which employ a forerunner
of JULES3.2 used in this study, are found to have a comparatively better
simulation of the seasonal cycle (Collins et
al., 2011), due to a combination of a more sensitive temperature rate
modifier combined with a larger seasonal soil temperature that is used in
the later version of JULES. Alexandrov (2014) shows that
both the amplitude underestimation and phase shift of FTA seasonal
cycle can be improved by increasing water use efficiency, decreasing
temperature dependence of heterotrophic respiration and increasing the
share of quickly decaying litterfall. Another probable factor is the
simulation of plant phenology. With the help of remote sensing data, better
phenology in model simulation has been shown to improve seasonal cycle
simulation of carbon flux (Forkel et al., 2014).
Additionally, the effect of carbon release from crop harvest is considered.
If harvested carbon is the main cause for the midsummer rebound in some
models, the rebound should be much less pronounced for the S2 (constant 1860
land use/cover) experiment, given that cropland area in 1860 is less than
half of the 2000 level. However, based on the comparison between the S2 and
S3 experiments over the global and northern temperate (major crop belts)
FTA seasonal cycle (Figs. S1 and S2 in the Supplement), the impact of harvested carbon
flux is unlikely to explain the midsummer rebound. This is probably due to
modeling efforts to prevent the sudden release of harvested carbon. Instead,
carbon release of harvested products and/or their residuals is usually
either spread over 12 months (i.e., LPJ, LPX-Bern, OCN, ORCHIDEE), or it enters
soil litter carbon pool (i.e., ISAM) for subsequent decomposition over time.
Mean seasonal cycle of global net carbon flux from nine TRENDY
models (S3 experiment) and two inversions, Jena99 and CarbonTracker,
averaged over 2001–2010.
![]()
Experimental design of TRENDY simulations.
Name
Time period
Atmospheric CO2
Climate forcing
Land-use history**
S1
1901–2012
Time-varying
Constant*
Constant (1860)
S2
Time-varying
S3
Time-varying
* Constant climate state is achieved by repeated or randomized or fixed
climate cycles depending on each model. ** Only the crop,
pasture and wood harvest information is included, so land use in this
study refers specifically to the related agricultural and forestry
processes.
Global mean net land carbon flux, seasonal amplitude, the maximum
and minimum months of FTA for the nine TRENDY models and their ensemble
mean during 1961–1970 and 2001–2010 periods. For the later period,
characteristics of the atmosphere inversions Jena99 and CarbonTracker are
also listed.
Name
Net carbon flux
Seasonal amplitude
FTA
FTA
(Pg C yr-1)
(Pg C yr-1)
minimum
maximum
1961–1970
2001–2010
1961–1970
2001–2010
2001–2010
2001–2010
CLM4.5BGC
0.1
-2.4
38.4
44.3
Jun
Nov
ISAM
0.7
0.0
17.6
19.1
Jun
Oct
JULES
-0.2
-1.7
15.1
19.0
May
Aug
LPJ
1.3
-0.6
18.6
23.4
Jun
Mar
LPX-Bern
0.6
0.0
33.0
37.9
Jun
Jan
OCN
0.9
-1.8
16.1
21.6
Jun
Nov
ORCHIDEE
0.1
-0.7
35.7
39.9
Jul
Mar
VEGAS
-0.4
-1.5
40.7
46.7
Jul
Nov
VISIT
0.2
-1.4
25.3
28.9
Jun
Nov
Ensemble
0.4
-1.1
22.4
26.1
Jun
Nov
Jena99
-1.7
46.8
Jul
Oct
CarbonTracker
-1.6
39.9
Jul
Nov
TRENDY models and inversions agree best over the boreal region (Fig. 2a).
While underestimating the global seasonal cycle, LPJ and VISIT both simulate
similar boreal FTA amplitude as inversions. In addition to ORCHIDEE and
Mean seasonal cycle of net carbon flux totals over boreal
(50–90∘ N), northern temperate (23.5–50∘ N), northern tropics (0–23.5∘ N),
southern tropics (0–23.5∘ S) and southern extratropics (23.5–90∘ S) from nine
TRENDY models and two inversions, Jena99 and CarbonTracker, averaged over
2001–2010.
![]()
Latitudinal dependence of the seasonal amplitude of
land–atmosphere carbon flux from the TRENDY multimodel median (red line,
and the pink shading indicates the 10 to 90 percentile range of model spread), two
atmospheric CO2 inversions, Jena99 (black dashed line) and CarbonTracker
(gray dashed line), and each individual model (thin line). Fluxes are first
resampled to 2.5∘ × 2.5∘, then summed over
each 2.5∘ latitude band (Pg C yr-1 per 2.5∘
latitude) for the TRENDY ensemble and inversions.
![]()
VEGAS, LPJ and LPX-Bern also simulate maximum CO2 drawdown in July for
the boreal region, same as the inversions, while the other five models have
the FTA minimum in June. Large model spread is present for the northern
temperate region, especially in summer. Both inversions and models agree
marginally over the phase of the FTA seasonal cycle in the tropics. The
northern and southern tropics show seasonal cycles that are largely out of
phase except for LPJ (Fig. 2c, d), due to the seasonal movement of the
tropical rain belt in the intertropical convergence zone (ITCZ). The
southern extratropics exhibit even smaller FTA amplitude due to
the relatively small biomass of the southern extratropics, and most models and inversions indicate a maximum FTA
around July, opposite in phase to its NH counterpart.
The latitudinal pattern of the multimodel median FTA amplitude is
remarkably similar to the inversions (Fig. 3). A notable feature is the
large seasonality over the NH midlatitude to high-latitude region driven by temperature
contrast between winter and summer. The model median also captures the two
subtropical maxima around 10∘ N and 15∘ S that are caused by tropical monsoon
movement. The main difference between the TRENDY models and the two
inversions is in the tropics and SH, where several models (JULES, LPJ, OCN
and especially ORCHIDEE) show much higher amplitude than the inversions.
Seasonal amplitude over 37–45 and 53–60∘ N is also larger from TRENDY models
than the inversions. A majority of the models display larger amplitude in
the tropics and northern temperate regions. Only three models (ISAM, JULES
and OCN) exhibit an underestimation of seasonal amplitudes north of 45∘ N.
Because of phase difference among the models and at different latitudinal
bands, for spatial and cross-model aggregated carbon fluxes, the seasonal
amplitude is reduced. Similarly, analyses by Peng et al. (2015) with an earlier set of TRENDY models
(Sitch et al., 2015) show an approximately equal number of models overestimating and
underestimating carbon flux compared to flux sites north of 35∘ N. However,
once the carbon fluxes of different phases are transported and mixed, seven
out of nine models underestimate the CO2 seasonal amplitude compared to
CO2 site measurements (Peng et al., 2015). Note that even
at the same latitude band, factors like monsoons, droughts and spring snowmelt, etc. could lead to longitudinal difference in the phase of seasonal
cycle (Figs. S3 and S4).
Trends for seasonal amplitude of TRENDY simulated multimodel
ensemble mean land–atmosphere carbon flux FTA (black), of the
Mauna Loa Observatory (MLO) CO2 mixing ratio (CO2MLO, green) and global CO2 mixing
ratio (CO2GLOBAL, purple), and of FTA from atmospheric inversions
of Jena81 (red), Jena99 (orange) and CarbonTracker (blue). The trends are
relative to the 1961–1970 mean for the TRENDY ensemble and Mauna Loa CO2,
and the other time series are offset to have the same mean as the TRENDY
ensemble for the last 10 years (2003–2012). A 9-year Gaussian smoothing
(Harris, 1978) removes interannual variability for all
time series, and its 1σ standard deviation is shown for CO2MLO
(green shading). Note that the gray shading here instead indicates 1σ
models' spread, which is generally larger than the standard deviation of the
TRENDY ensemble's decadal variability. Inset: average seasonal cycles of
models' ensemble mean FTA (Pg C yr-1) for the two periods: 1961–1970
(dashed line; lighter gray shading indicates 1σ model spread) and
2001–2010 (solid; darker gray shading indicates 1σ model spread),
revealing enhanced CO2 uptake during spring/summer growing season. Mean
seasonal cycles global FTA from the atmospheric inversions for
2001–2010 are also shown (same color as the main figure) for comparison.
![]()
Temporal evolution of <inline-formula><mml:math display="inline"><mml:mrow><mml:msub><mml:mi>F</mml:mi><mml:mtext>TA</mml:mtext></mml:msub></mml:mrow></mml:math></inline-formula> seasonal amplitude
The seasonal amplitude of global total FTA from the TRENDY model
ensemble for 1961–2012 shows a long-term rise of 19 ± 8 %, with large
decadal variability (Fig. 4). Similarly, the seasonal amplitude of
CO2 at Mauna Loa increases by 15 ± 3 % (0.85 ± 0.18 ppm)
for the same period. This amplitude increase appears mostly as an earlier
and deeper drawdown during the spring and summer growing season, mostly in
June and July (Table 3, Fig. 4 inset). Changes in trend of yearly minima
(indicating peak carbon uptake) and yearly maxima (dominated by respiration)
contribute 91 ± 10 and 9 ± 10 % to the FTA amplitude
increase, respectively. Gurney and Eckels (2011) suggest
trend in respiration increase is more important, but they averaged all
months instead of using maxima and minima in their amplitude definition. The
multimodel ensemble mean tracks some characteristics of the decadal
variability reflected by the Mauna Loa record: stable in the 1960s, rise in
the 1970–1980s, rapid rise in the early 2000s and decrease in the most recent
10 years. Strictly speaking, Mauna Loa CO2 data are not directly
comparable with simulated global FTA because this single station is
also influenced by atmospheric circulation as well as fossil fuel emissions
and ocean–atmosphere fluxes. Nevertheless, the comparison of the long-term
amplitude trend is still valuable because the Mauna Loa Observatory data
constitute the only long-term record, and it is generally considered
representative of global mean CO2 (Heimann, 1986;
Kaminski et al., 1996). The global total CO2 index (CO2GLOBAL) and
FTA from three atmospheric inversions are also included in the
comparison. All data (Jena81, CO2MLO, CO2GLOBAL) show a decrease
in seasonal amplitude in the late 1990s, possibly related to drought in the
Northern Hemisphere midlatitude regions (Buermann et
al., 2007; Zeng et al., 2005a), and about half of the models (LPJ, OCN,
ORCHIDEE, VEGAS) also exhibit similar change (Fig. 7). Details on models'
FTA global and regional changes in 2001–2010 compared to 1961–1970 are
listed in Table 4.
The seasonal amplitude (maximum minus minimum, in Pg C yr-1) of
mean net carbon flux for 2001–2010 relative to the 1961–1970 period,
according to the nine TRENDY models (values are listed as percentage change
in brackets, for both regions and the entire globe). The four large
latitudinal regions are the same as in Fig. 3: boreal (50–90∘ N), temperate
(23.5–50∘ N), northern tropics (0–23.5∘ N), southern tropics (0–23.5∘ S) and
southern extratropics (23.5–90∘ S). Values from the two inversions, Jena99 and
CarbonTracker, are also listed for comparison.
Name
Global
Boreal
Northern
Northern
Southern
Southern
temperate
tropics
tropics
extratropics
CLM4.5BGC
44.3 (15 %)
31.9 (17 %)
19.2 (15 %)
7.2 (22 %)
6.5 (-2 %)
4.9 (4 %)
ISAM
19.1 (9 %)
12.1 (11 %)
7.4 (13 %)
6.0 (1 %)
6.9 (-8 %)
0.4 (4 %)
JULES
19.0 (26 %)
12.2 (24 %)
14.3 (9 %)
11.6 (0 %)
11.3 (11 %)
2.2 (-24 %)
LPJ
23.4 (26 %)
23.0 (18 %)
14.7 (11 %)
10.5 (9 %)
11.8 (16 %)
2.0 (-12 %)
LPX-Bern
37.9 (15 %)
26.9 (10 %)
19.3 (6 %)
8.3 (9 %)
4.6 (-6 %)
4.2 (15 %)
OCN
21.6 (34 %)
12.3 (33 %)
11.1 (23 %)
9.7 (17 %)
8.3 (3 %)
2.0 (14 %)
ORCHIDEE
39.9 (12 %)
23.4 (14 %)
19.1 (5 %)
22.7 (9 %)
18.7 (2 %)
1.4 (37 %)
VEGAS
46.7 (15 %)
22.3 (17 %)
24.7 (10 %)
4.0 (11 %)
3.4 (12 %)
2.1 (6 %)
VISIT
28.9 (14 %)
22.9 (12 %)
15.6 (8 %)
3.4 (9 %)
3.2 (1 %)
3.1 (18 %)
Ensemble
26.1 (17 %)
18.0 (19 %)
12.4 (15 %)
8.0 (8 %)
4.9 (-3 %)
2.1 (13 %)
Jena99
46.8
23.3
21
8.2
8.5
1.5
CarbonTracker
39.9
26.5
16.3
5.3
5.8
2.4
Attribution of the seasonal amplitude trend of global net land
carbon flux for the period 1961–2012 to three key factors of CO2,
climate and land use/cover. The red dots represent models' global amplitude
increase of FTA from the S3 experiment, and error bars indicate
1σ standard deviation. The increasing seasonal amplitude of
FTA is decomposed into the influence of time-varying atmospheric
CO2 (blue), climate (light green) and land use/cover change (gold).
![]()
Attribution of global and regional <inline-formula><mml:math display="inline"><mml:mrow><mml:msub><mml:mi>F</mml:mi><mml:mtext>TA</mml:mtext></mml:msub></mml:mrow></mml:math></inline-formula> seasonal
amplitude
Models agree on increase of global FTA seasonal amplitude during
1961–2012, but they disagree even in sign in the contribution of the
different factors (Fig. 5). By computing the ratios between amplitude
trends from rising CO2, climate change and land use/cover change with
the total trend for each model, we find that the effect of varying CO2,
climate and land use/cover contribute 83 ± 56, -3 ± 74
and 20 ± 30 % to the simulated global FTA amplitude increase.
All models simulate increasing amplitude for total FTA in the boreal
(50–90∘ N) and northern temperate (23.5–50∘ N) regions, and most models also
indicate amplitude increase in the northern (0–23.5∘ N) and southern tropics
(0–23.5∘ S) (Fig. 6). There is less agreement on the sign of amplitude
change among the models in the southern extratropics (23.5–90∘ S). Individual
models' global and regional trends of FTA amplitude attributable to the
three factors (CO2, climate and land use/cover) are listed in Table S1.
For most models, latitudinal contribution to global FTA amplitude
(computed with FkAi) shows that the pronounced
midlatitude to high-latitude maxima in the NH dominate the simulated amplitude increase
over 1961–2012 (Fig. 8, red dashed line for S3 results). All models also
indicate a negative contribution from at least part of the northern
temperate region.
The four models (CLM4.5BGC, VEGAS, LPX-Bern and ORCHIDEE) that simulate a
more realistic mean global FTA seasonal cycle (Fig. 1) are also
relatively close in global FTA seasonal amplitude, clustering around an
increase of 14 ± 3 % during 1961–2012. Furthermore, they all suggest
that land use/cover change contributes positively to global FTA seasonal
amplitude increase. On the other hand, four of the remaining five models
(OCN, LPJ, JULES, VISIT) show a much larger rate of increase (26 ± 3 %),
but given that these four models underestimate the mean amplitude by
about 50 %, the absolute increase in global FTA seasonal amplitude is
actually similar (about 5 Pg C yr-1) between the two groups of models.
ISAM is an exception; it both underestimates the mean global FTA
seasonal amplitude and has the lowest rate of amplitude increase.
The rising CO<inline-formula><mml:math display="inline"><mml:msub><mml:mi/><mml:mn mathvariant="normal">2</mml:mn></mml:msub></mml:math></inline-formula> factor
Seven of the nine models suggest that the CO2 fertilization effect is most
responsible for the increase in the amplitude of global FTA, while
VEGAS attributes it to be approximately equal among the three factors (Fig. 5).
The CO2 fertilization effect alone seems to cause most of the amplitude
increase in a majority of models, with notable contribution from climate
change and land use/cover change in CLM4.5BGC and VEGAS (Fig. 7). The
effect of rising CO2 appears to be slightly negative for JULES,
possibly reflecting an offsetting of the strong seasonal soil respiration
response found in this model. For each model, rising CO2 in the boreal,
northern temperate and the southern extratropics leads to a similar trend
(Fig. 6). The magnitude of this trend may indicate each model's differing
strength for CO2 fertilization. This is possibly due to similar phases
of FTA seasonal cycle within the three regions that are mainly driven
by climatological temperature contrast. The positive amplitude trend in the
carbon flux of the northern and southern tropics from CO2 fertilization
is similar, and they likely would cancel out each other because their
seasonal cycles are largely out of phase. Latitudinal contribution analyses
reveal that the trend in the northern midlatitudes to high-latitudes is the main contributor
to global FTA amplitude increase when considering CO2
fertilization effect alone (Fig. 8, blue line).
Attribution of the seasonal amplitude trend of regional (boreal
(50–90∘ N), northern temperate (23.5–50∘ N), northern tropics (0–23.5∘ N),
southern tropics (0–23.5∘ S) and southern extratropics (23.5–90∘ S)) net land
carbon flux for the period 1961–2012 to three key factors CO2, climate
and land use/cover. The red dots represent models' global amplitude increase
of FTA from the S3 experiment. The increasing seasonal amplitude of
FTA is decomposed into the influence of time-varying atmospheric
CO2 (blue), climate (light green) and land use/cover change (gold).
![]()
The climate change factor
The effect of climate change on FTA amplitude is mixed: five models
(OCN, LPJ, LPX-Bern, ORCHIDEE and ISAM) suggest climate change acts to
decrease the FTA amplitude, and four models (JULES, VISIT, CLM4.5BGC
and VEGAS) suggest it is an increasing effect (Fig. 5). The high-latitude
greening effect is evident in six out of nine models (Fig. 6),
contributing, on average, 29 % of boreal amplitude increase. The latitudinal
contribution analyses (Fig. 8) also suggest that warming-induced high-latitude “greening” effect is present in all models, but this positive
contribution only exhibits a wide range of influence in about half of the
models (CLM4.5BGC, JULES, VEGAS and VISIT). The latitudinal patterns also
reveal that, once climate change is considered, the contribution from the
northern temperate region around 40∘ N shifts to negative in all models. In
the northern temperate (23.5–50∘ N) region, climate change alone would
decrease the FTA amplitude – this is consistent among the four models
with realistic mean global and northern temperate (Fig. 2) FTA
seasonal cycle simulation, but is not the case for JULES and LPJ (Fig. 6).
Such decrease is possibly related to midlatitude drought
(Buermann et al., 2007), which is consistent with findings by
Schneising et al. (2014), who observed a negative
relationship between temperature and seasonal amplitude of xCO2 from
both satellite measurements and CarbonTracker during 2003–2011 for the
Northern temperate zone. The negative contribution from the temperate zone
counteracts the positive boreal contribution, suggesting that the net impact from
climate change on FTA amplitude may not be as significant as previously
suggested. With changing climate introduced, some models exhibit similar
characteristics of decadal variability in global FTA amplitude
(Fig. 7). OCN and ORCHIDEE appear to be especially sensitive to the climate
variations after the 1990s, resulting in a decrease in FTA amplitude.
It is also apparent from the time series figure that the strong increasing
trend of FTA amplitude from climate change in JULES is mostly due to
the sharp rise from early 1990s to early 2000s, suggesting some possible
model artifacts (Fig. 7). The effect of climate change is more mixed in
both tropics and the southern extratropics.
Trends for seasonal amplitude of global total net carbon fluxes
from S1 (CO2), S2 (CO2+climate) and S3
(CO2+climate+land use) for each individual TRENDY model. All
amplitude time series are relative to their own 1961–1970 mean amplitude.
![]()
The land use/cover change factor
Six of the nine models show that land use/cover change leads to increasing
global FTA amplitude (Fig. 5). Land use/cover change appears to
amplify FTA seasonal cycle in boreal and northern temperate regions for
most models. For some models (VEGAS, CLM4.5BGC and OCN), this effect is
especially pronounced in the northern temperate region where most of the
global crop production takes place (Fig. 6). Note that the effect of land
use/cover change includes two parts: one is the change of land use practice
without changing the land cover type; the other is the change of land cover,
including crop abandonment etc. VEGAS simulates the time-varying management
intensity and the crop harvest index, which is an example of significant
contribution from land use change (Zeng et al., 2014). For
many other models, crops are treated as generic managed grasslands (i.e.,
CLM4.5BGC, LPJ), and land cover change is possibly the more important
factor. During 1961–2012, large cropland areas were abandoned in the eastern
United States and central Europe, and forest regrowth often followed. New cropland
expanded in the tropics and South America, midwestern United States, eastern and central
North Asia and the Middle East. How such changes affect the global FTA
amplitude is determined by the productivity and seasonal phase of the old
and new vegetation covers. For CLM4.5BGC, JULES, LPJ and ORCHIDEE, enhanced
vegetation activity from growing forest in these regions contributes
positively to global FTA amplitude increase (Fig. 9). In contrast,
for LPX-Bern, VISIT and VEGAS in the eastern United States, a loss of cropland leads
to a decrease in the amplitude. Additional cropland in the midwestern United States and
eastern and central North Asia contributes negatively to the FTA amplitude
trend for JULES, LPJ and ORCHIDEE. These regions, however, are major zones
contributing to the amplification of global FTA for LPX-Bern, OCN, VEGAS
and VISIT. One mechanism mentioned previously is agricultural
intensification in VEGAS: in fact, CO2 flux measurements over corn
fields in the US Midwest show much larger seasonal amplitude than over
nearby natural vegetation (Miles et al.,
2012). Similarly, although croplands are treated as generic grassland, they
still receive time-varying and spatially explicit fertilizer input in OCN
(Zaehle et al., 2011). Another plausible mechanism is
irrigation, which can alleviate adverse climate impact from droughts, and
crops may have a stronger seasonal cycle than the natural vegetation they
replace in these regions. The overall effect of land use/cover change for
each model, therefore, is often the aggregated result over many regions that
can only be revealed by spatially explicit patterns. When examining the
latitudinal contribution only (Fig. 8), CLM4.5BGC, LPX-Bern, OCN and VEGAS
are quite similar, even though the spatial patterns reveal that CLM4.5BGC is very
different from the other three models (Fig. 9). For JULES, LPJ and
ORCHIDEE a significant part of land use/cover change contribution comes from
the tropical zone (Fig. 8). While most models indicate that land use/cover
change in the southern tropics (Amazon is probably the most notable region)
decreases global FTA amplitude during 1961–2012, LPJ suggests that it
would cause a large increase in the amplitude instead, possibly related to
its different behavior in simulating the mean seasonal cycle of carbon flux for
that region (Fig. 2d).
Latitudinal contribution of trends for seasonal amplitude of
global land–atmosphere carbon flux from TRENDY models in the three
sensitivity experiments. Fluxes are summed over each 2.5∘ latitude
band (Pg C yr-1 per 2.5∘ latitude) before computing the
FkAi (refer to the
Methodology section for definition). For each 2.5∘ latitude band,
the trend is calculated for the period 1961–2012.
![]()
Contribution from land use/cover change on trends in the seasonal
amplitude of global land–atmosphere carbon flux. For each spatial grid, the
trend is computed as trends of the FkAi (refer to Methodology section for definition) in the S3
experiment (changing CO2, climate and land use/cover) subtracted by
trends in S2 (changing CO2 and climate).
![]()
Discussion and conclusion
Our results show a robust increase of global and regional (especially over
the boreal and northern temperate regions) FTA amplitude simulated by
all TRENDY models. During 1961–2012, TRENDY models' ensemble mean global
FTA relative amplitude increases (19 ± 8 %). Similarly, the
CO2 amplitude also increases (15 ± 3 %) at Mauna Loa for
1961–2012. This amplitude increase mostly reflects the earlier and deeper
drawdown of CO2 in the NH growing season. The models in general,
especially the multimodel median, simulate latitudinal patterns of FTA
mean amplitude that are similar to the atmospheric inversions results.
Their latitudinal patterns capture the temperature-driven seasonality from
the NH midlatitude to high-latitude region and the two monsoon-driven subtropical
maxima, although the magnitude or extent vary. Despite the general
agreements between the models' ensemble amplitude increases and the limited
observation-based estimates, considerable model spread is noticeable. Five
of the nine models considerably underestimate the global mean FTA
seasonal cycle compared to atmospheric inversions, and peak carbon uptake
takes place 1 or 2 months too early in seven of the nine models. The
seasonal amplitude of model ensemble global mean FTA is 40 % smaller
than the amplitude of the atmosphere inversions. In contrast to the
divergence in simulated seasonal carbon cycle, atmospheric inversions in
Northern temperate and boreal regions are well constrained: 11 different
inversions agree on July FTA minimum in the Northern Hemisphere
(25–90∘ N), with no more than 20 % difference in amplitude
(Peylin et al., 2013).
The simulated amplitude increase is found to be mostly due to a larger
FTA minimum associated with a stronger ecosystem growth. Over the
historical period, global mean carbon sink also increases over time,
suggesting a possible relationship between seasonal amplitude and the mean
sink (Ito et al., 2016; Randerson et al.,
1997; Zhao and Zeng, 2014). The increasing trend of CO2 amplitude,
dominated by increasing trend of FTA amplitude, has been interpreted as
evidence for steadily increasing net land carbon sink
(Keeling et al., 1995; Prentice et al., 2000).
However, the increasing amplitude could also arise from (climatically
induced) increased phase separation of photosynthesis and respiration, e.g.,
due to warming-induced earlier greening (Myneni et al., 1997). For the nine models, we
found a moderate relationship between enhanced mean land carbon sink and the
seasonal amplitude increase similar to reported results by in
Zhao and Zeng (2014), with an R-squared value of 0.61 (Fig. 10).
There might be some possibility in constraining change in land carbon
sink with changes in observed CO2 seasonal amplitude; however, extra
caution should be given when interpreting this global-scale cross-model
correlation, as there could be important regional differences that cancel out
in aggregated global values. A factorial analysis of the long-term carbon
uptake could help to determine which factor contributes to what extent to
this correlation. Further research is also needed to explore the mechanisms
behind such a relationship at continental scale, where more data from
well-calibrated CO2 monitoring sites and data on air–sea fluxes and
atmospheric vertical transport could better constrain carbon balance
(Prentice et al., 2001). Changes in residual land
carbon sink estimates are also shown (Fig. 10), with the caveat that it is
not directly comparable with simulated net carbon sink increase if there is
a trend in simulated carbon flux changes associated with land cover
conversion (deforestation, crop abandonment, etc.). Additionally, the decadal
changes in residual and net land carbon sink are far from linear; instead, a
sudden increase in mean land uptake occurred in 1988
(Beaulieu et al., 2012;
Rafique et al., 2016; Sarmiento et al., 2010). With the aid of atmospheric
transport, CO2 amplitude trends at remote sites have benchmarking
potential to constrain the models, especially with more observations and
improved understanding of vegetation dynamics at regional level in the near
future.
Relationship between the increase in net biosphere production
(NBP, equal to -FTA) and increase in NBP seasonal amplitude (as in
Fig. 4's red dots), for the 1961–2012 period for nine TRENDY models. Error
bars indicate the standard errors of the trend estimates. Increase in
residual land sink is estimated by taking the difference between two
residual land sinks, over 2004–2013 and 1960–1969 (an interval of 44 years),
as reported in Le Quéré et al. (2015). This
difference is then scaled by 52/44 (to make it comparable with models' NBP
change for this figure), which is displayed by a black vertical line and
shading (error added in quadrature, assuming Gaussian error for the two decadal
residual land sinks, then also scaled). The cross-model correlation
(R2= 0.61, p < 0.05) suggests that a model with a larger net
carbon sink increase is likely to simulate a higher increase in NBP seasonal
amplitude.
![]()
Models with a strong mean carbon sink (for example JULES and OCN) can have
relatively weak seasonal amplitude, and the LPX-Bern model shows no carbon
sink despite having a strong FTA seasonality. Based on data from Table
8 of the Global Carbon Budget report
(Le Quéré et al., 2014), the net land carbon sink for 2000–2009 is
estimated to be 1.5 ± 0.7 Pg C yr-1 (assuming Gaussian errors). Four
models (JULES, OCN, VEGAS and VISIT) examined in this study are within the
uncertainty range of this budget-based analysis. In spite of their similar
mean land carbon sink, the shape of their FTA seasonal cycle differs.
While VEGAS also shows a similar seasonal carbon cycle compared to
inversions, the other three models exhibit an unrealistically long carbon
uptake period with half the amplitude as the inversions. July and August are
the only 2 months with net carbon release for JULES, whereas OCN and VISIT
both have a long major carbon uptake period from May to September. Given
that the mean global and regional FTA seasonal cycles are relatively
well constrained in the northern extratropics, they can serve as
a benchmark for terrestrial models (Heimann et al.,
1998; Prentice et al., 2001). Insights gained from analyzing modeled
seasonal amplitude of carbon flux may help to understand the considerable
model spread found in the mean global carbon sink for the historical period
(Le Quéré et al., 2015), which is possibly due to
varied model sensitivity to different mechanisms
(Arora et al., 2013). Examining details of
different representations of important processes in models could also help
to better assess the different future projections on both the magnitude and
direction of global carbon flux
(Friedlingstein et al., 2006, 2013).
Unlike many previous studies that focused on comparing the season cycle at
individual CO2 monitoring stations (Peng et
al., 2015; Randerson et al., 1997), we studied the global and large
latitudinal bands. Such quantities often demonstrate well-constrained
seasonality that is relatively robust against uncertainty from atmospheric
transport, fossil fuel emission and biomass burning etc. We found greater
uncertainty for the tropics and southern extratropics regions where
atmospheric CO2 observations are relatively sparse. Tropical ecosystems
are also heavily affected by biomass burning; however, some models used in
this study do not include fire dynamics. For models that simulate fire
ignition/suppression, they are also varied by structure and complexity of
fire-related processes, and many of them are prognostic
(Poulter et al., 2015). It is not clear how fire would
affect the FTA seasonal cycle at global scale, and recent sensitivity
study shows only minor differences among fire and “no fire” scenarios in
CO2 seasonal cycle at several observation stations
(Poulter et al., 2015). These uncertainties, however, are
unlikely to affect our main conclusions because of the limited contribution of
tropics to global FTA amplitude increase. Another possibly important
factor is the impact from increased nitrogen deposition, which may have been
included in the “CO2 fertilization” effect for some models with full
nitrogen cycle (Table 1); however, this can only be explored in future
studies, as the TRENDY experiment design does not separate out the nitrogen
contribution.
Our factorial analyses highlight fundamentally differential control from
rising CO2, climate change and land use/cover change among the models,
with seven out of nine models indicating major contribution (83 ± 56 %)
to global FTA amplitude increase from the CO2
fertilization effect. The strength of CO2 fertilization varies among
models, but for each model, its magnitude in the boreal, northern temperate
and southern extratropics regions is similar. Models are split regarding
the role of climate change, as compared with the models' ensemble mean
(-3 ± 74 %). Regional analyses show that climate change amplifies
the boreal FTA seasonal cycle but weakens the seasonal cycle for other
regions according to most models. By examining latitudinal trends from
FkAi, we found all models indicate a negative
climate contribution over the midlatitudes, where droughts might have
reduced ecosystem productivity. This negative effect offsets the high-latitude greening, which in some models results in a net negative climate
change impact on global FTA amplitude. Such a mechanism casts doubt on
whether climate change is the main driver of the global FTA amplitude
increase. Land use/cover change, according to majority of the models,
appears to amplify the global FTA seasonal cycle
(20 ± 30 %); however, the mechanisms seem to differ among models. Conversion to/from
cropland could either increase or decrease the seasonal amplitude, depending
on how models simulate the seasonal cycle of cropland compared to the natural
vegetation it replaces/precedes. For the same pattern of increasing
amplitude, the underlying causes could include irrigation mitigating
negative climate effect, agricultural management practices and other
mechanisms.
Overall, this study is largely helpful to enhance our understanding of the role
of CO2, climate change and land use/cover change in regulating the
seasonal amplitude of carbon fluxes. In particular, models' disagreement in
spatial pattern of carbon flux amplitude helps to identify optimal locations
for additional CO2 observations in the north. However, this work can be
further improved through utilizing the CO2 seasonal cycle and its
amplitude at different locations as indicators to diagnose model behaviors.
To achieve this, it is necessary to apply atmosphere transport on the
simulated net carbon flux, along with ocean and fossil fuel fluxes, which
would allow direct comparison with observed CO2 amplitude change. In
doing so, it is possible that models may overestimate CO2 amplitude
increase at most CO2 observation stations if the simulated CO2
fertilization effect is too strong.