Introduction
Carbon and water fluxes at one particular site can strongly vary from
year to year (e.g. Goulden et al., 1996; Yamamoto et al., 1999; Baldocchi
et al., 2001). This interannual variability in net ecosystem exchange (NEE)
and actual evapotranspiration (AET) is observed across different
geographical regions and ecosystem types, and understanding interannual
variability in carbon and water fluxes (IAVcw) is crucial for
projections of future ecosystem changes and feedbacks on climate. However,
little is known about the processes determining this year-to-year variation.
Numerous studies have tried to relate IAVcw to climatic variables and
local ecosystem responses to droughts, fires, and deforestation (e.g.
Goulden et al., 1996; Yamamoto et al., 1999; Aubinet et al., 2002; Hui et
al., 2003; Williams et al., 2008; Sierra et al., 2009; Weber et al., 2009;
Yuan et al., 2009), but no clear picture has yet emerged.
Process-based biogeochemical and vegetation models capture the response of
terrestrial ecosystems to mean climatic drivers reasonably well at diurnal
and seasonal timescales, but not at yearly and longer timescales (Keenan et
al., 2012). At the global scale, some vegetation models reproduce interannual
variability in terrestrial net primary production and atmospheric CO2
growth rates well (Peylin et al., 2005; Ahlström et al., 2012; Sitch et
al., 2015), but large uncertainty exists at smaller spatial scales. Only few
studies have quantified the extent to which these models can reproduce
observed IAVcw at the regional and site scale (Peylin et al.,
2005; Keenan et al., 2012). Despite the uncertainties, such models are widely
used to project future changes in vegetation and ecosystem functioning. Some
of these model simulations suggest the potential for severe vegetation
changes across major global biomes in the future: for example, Amazon forest
die-back/greening, as well as substantial shifts in potential natural
vegetation distributions for boreal and Mediterranean forests (e.g. Lenton et
al., 2008; Rammig et al., 2010; Hickler et al., 2012), and alternative
vegetation states under elevated atmospheric CO2 (e.g. Higgins and
Scheiter, 2012). Such vegetation changes would also feed back to regional and
global climate (e.g. Cox et al., 2000; Naeem, 2002; Sitch et al., 2003; van
den Hurk et al., 2003; Arora and Boer, 2005; Bonan, 2008; Pitman et al.,
2009; Wramneby et al., 2010), and can affect the long-term terrestrial carbon
balance profoundly. Therefore it is crucial that these models accurately
reproduce IAVcw across all spatial scales.
To provide insight into the climate change impacts on the terrestrial carbon
balance in the long term, both short- and long-term vegetation responses to
a constantly changing environment should be better understood and
represented. This implies better model representations of indirect
short-term processes such as the mechanisms governing vegetation phenology
(Cleland et al., 2007; Kramer and Hänninen, 2009; Wolkovich et al.,
2012), dynamic carbon and nutrient allocation (Litton et al., 2007; Epron
et al., 2012; Franklin et al., 2012), photosynthetic temperature acclimation
(Gea-Izquierdo et al., 2010), as well as better
representations of indirect long-term processes such as soil, nutrient and
carbon dynamics. Before addressing these complex process representations
within models, however, it can be useful to test whether IAVcw can be
explained by rather simple relationships with direct environmental drivers,
such as drought, temperature, and radiation, which can affect, e.g.
photosynthesis and soil respiration very directly and instantaneously.
Factorial experiments with a dynamic vegetation model can then be used to
generate hypotheses concerning simple and/or complex interactions of
processes driving IAVcw. These vegetation models can be expected to
capture at least some of the complexity of real ecosystems, and the
factorial experiments can be used, for example, to keep certain
environmental drivers constant (i.e. switching
of their effect, e.g. Hickler et al., 2005) or to implement different
hypotheses concerning the most important processes within an ecosystem. The
latter can also be achieved by data–model intercomparisons with several
models that differ in their process representation
(e.g. Medlyn et al., 2015). In this study, the
factorial model experiments refer to model setups with different process
representations. With this purpose in mind, we used a long time series of
eddy covariance measurements at a well-researched forest site (Loobos, a
Scots pine forest on sandy soils in the Netherlands) and a dynamic global vegetation model, DGVM (Lund–Potsdam–Jena General Ecosystem Simulator, LPJ-GUESS;
Smith et al., 2001), parameterised for the site. The observed interannual
variability in NEE at Loobos is comparable to that found at sites with
similar vegetation composition and climate (Carrara et al.,
2003), but this interannual variability cannot be explained directly from
climate variables (Jacobs et al., 2009;
Kruijt et al., 2009). Previous analyses suggest that temperature is an
important driver of ecosystem respiration at this site, and the remaining
variation could be related to local extremes, such as drought, storm damage,
and snowfall in winter (Moors et
al., 2015). Luyssaert et al. (2007)
thoroughly analysed observational Loobos data and proposed that
photosynthesis variability is the main driver of interannual variability in
NEE, suggesting that short-term ecophysiological responses play an important
role.
In this study, we first tested whether LPJ-GUESS can reproduce the observed
IAVcw and seasonal carbon and water exchange at the Loobos site from
direct environmental factors only. LPJ-GUESS combines detailed vegetation
demographics and dynamics with mechanistic representations of short-term
plant physiological processes. This combination makes the model a good
platform to study IAVcw because we can simultaneously study the
effects of environmental and ecosystem drivers on modelled IAVcw.
Secondly, we tested whether using alternative model formulations and
parameters can explain model error for this single site. We performed these
secondary tests because in the first test we observed systematic biases
during winter periods and drought events. Therefore, we analysed the
photosynthesis response to temperature during winter periods, and we
analysed the response to drought events by comparing alternative plant water
uptake parameterisations.
Results
Default modelling setup
The general site characteristics of Loobos are well represented by the
default modelling setup (S1, Table 2): modelled LAI for Scots pine is 1.5,
declining to 1.4 between 1997 and 2009. This LAI is just below the observed
site average of 1.62 between 1997 and 2009 (minimum 1.44 in 2007, maximum
1.78 in 2009). Modelled LAI for C3 grasses is higher than observed (2.4
and 1.0 respectively), but few measurements of understory grass LAI were
available for validation and none for mosses. Modelled aboveground biomass
estimates are close to available observations.
Modelled and observed site characteristics of Loobos. All
modelled values for biomass are calculated for the period 1997–2009, and
multiplied by a factor 0.82 to exclude root biomass (taken from Jackson et
al. (1996) as a topical value for conifer forests).
Aboveground
LAI
biomass (kg C m-2)
Pinus sylvestris
C3 grass
Observed:
4.98a
1.62b
1.0c
Modelled:
Default/S1
5.95 ± 0.10
1.5
2.4
pstemp
7.18 ± 0.14
1.7
1.9
S2
4.55 ± 0.11
1.1
3.6
S3
4.72 ± 0.11
1.2
2.8
S4
7.64 ± 0.19
1.8
2.6
a 9.23 kg m-2 standing biomass in 1997, annual growth increment of
0.124 kg m-2 (data source: http://www.climatexchange.nl/sites/loobos/).
To convert to carbon mass a factor of 0.5 was used (e.g. see
Sandström et al., 2007; Thomas and Martin, 2012), resulting in an
estimated average aboveground biomass between 1997 and 2009 of
4.98 kg C m-2.b Measured average tree LAI from 1997 to 2009 (unpublished data), minimum
1.44 (2007), maximum 1.78 (2009), standard deviation is 0.10. Dolman et al.
(2002) report maximum LAI of 1.9 for 1997.
c Measurements between 1999 and 2002 (n = 52), standard deviation 0.4
m2 m-2 (unpublished data).
Figure 2 shows the interannual and monthly variability in GPP and AET. Table 3 summarises the goodness-of-fit values for GPP and AET. The model shows
good correlations on daily and monthly timescales (Fig. 2c and d). Monthly
correlations are significant (0.92 for GPP, and 0.87 for AET), indicating
that the model is accurately capturing the seasonal pattern of both fluxes.
This is also visible in Fig. 3a and b. In contrast, we find poor
correlations on the annual timescale: annual totals for GPP and AET are of
the same order of magnitude as observed values, but the observed IAVcw
is not captured well by the model for water or for carbon (Fig. 2a and b).
The modelled data distribution is similar to observations (Table 3, bold
values), but correlation coefficients are low and not significant (0.22 and
0.20 for GPP and AET, respectively).
The monthly scatter plots (Fig. 2c and d) display systematic model biases
during certain periods. Fluxes are underestimated in winter, overestimated
in spring/early summer, and slightly underestimated in autumn (Fig. 2c and d).
In summer (mainly in August and July), large deviations from the 1:1 line
can be seen, which we directly relate to periods with high atmospheric
temperatures and low precipitation. Figure 3 shows these deviations per
month in more detail.
Alternative temperature response function
Observed temperature response
According to the EC data, the vegetation at Loobos is able to keep
assimilating carbon even at temperatures below 0 ∘C (Fig. 4). In
the fitted response curve of half-hourly EC fluxes, maximum GPP for the
lowest temperature class (-5 to 0 ∘C, Fig. 4a) is
1.8 µmol m-2 s-1, which corresponds to 1.87 g C m-2 day-1.
Figure 4b shows temperature-binned daily GPP on sunny days, and the response to
temperatures below -10 ∘C. The lower temperature limit in our
observational data, i.e. where average GPP approaches 0, is found when
temperatures are below -8 ∘C. Note that the number of data
points, however, in temperature class -8 to -10 ∘C is
relatively low (n=2). To further check data for this particular
temperature class, we included half-hourly EC data for two such days
(Figs. S4 and S5). On these days, NEE becomes negative and
strongly responds to radiation, especially around noon. The average
assimilation capacity for all the example dates in Figs. S4 and S5
correspond well with the upper quartile of daily observed GPP as shown in
Fig. 4b. As can be expected, average observed GPP per day is slightly lower
than the maximum capacity for a certain temperature class. The leaf level
measurements (Fig. S6) also show active assimilation when
atmospheric temperatures were below 0, with P. sylvestris needles strongly responding to
radiation. A linear regression through these data points gives a minimum of
-10.1 ∘C.
Observed vs. modelled variability in GPP (a, c) and
AET (b, d) for the default model scenario (S1) on annual (a, b) and monthly timescales (c, d). Dotted line is the 1:1 line. The equation shows linear regression through the origin, with
correlation coefficients. Fluxes are hatched per season for sub-panels (c) and (d): black circles are for winter (December, January, February); black squares are for spring
(March, April, May); black triangles are for summer (June, July, August); +=
autumn
(September, October, November).
All three data sources indicate that carbon assimilation stops when
temperatures fall below -10 ∘C (Fig. 4b), and when a prolonged
period of extremely cold temperatures is observed. The latter was the case
in early January 1997, even on days with high radiation and temperatures
between -6 and -8 ∘C (Fig. 4b, first and second quartile).
Goodness-of-fit values for model scenarios S1–S4 and changed
temperature response function, “pstemp”. Correlation coefficient (r) and RMSE for daily, monthly and annual data. Bold values
represent data distributions that are identical using the Wilcoxon ranking
test.
GPP
AET
annual
monthly
daily
annual
monthly
daily
Run
r
RMSE
r
RMSE
r
RMSE
r
RMSE
r
RMSE
r
RMSE
Default/S1
0.22
125.9
0.92∗
35.7
0.79∗
2.20
0.20
77.7
0.87∗
19.7
0.62∗
1.27
pstemp
0.16
109.3
0.90∗
36.3
0.78∗
2.15
0.21
73.4
0.87∗
19.6
0.62
1.25
S2
0.32
128.6
0.92∗
32.6
0.81∗
1.93
0.19
90.8
0.87∗
17.2
0.65∗
1.03
S3
0.27
198.9
0.92∗
31.4
0.81∗
1.78
0.13
141.9
0.86∗
17.3
0.65∗
0.94
S4
0.24
231.3
0.94∗
51.9
0.85∗
2.45
0.31
168.3
0.88∗
36.2
0.68∗
1.67
∗ Significance tests for Pearson correlation: P
value < 0.05.
Observed (black dotted line) and modelled values for default
(S1, green line) and changed temperature response (pstemp, purple line) runs. (a) Monthly values for GPP (g C m-2 month-1). (b) Monthly
values for AET (mm month-1).
Observed temperature responses at Loobos. (a) Courtesy
of P. Abreu: fitted GPP at a solar light intensity of 1000 W m-2
(GPP1000, µmol m-2 s-1) based on half-hourly EC measurements
(1997–2011) following Jacobs et al. (2007);
(b) daily GPP (g C m-2 day-1) observed at Loobos calculated
from site EC measurements, for days with average daily temperatures
< 0 ∘C and total net radiation received > 2 MJ day-1 (n=175).
Effect of change in temperature scalar, tscalar, on
modelled estimates of (a) GPP (g C m-2 day-1) and (b)
AET (mm day-1). pstempmin for Pinus sylvestris is set to -10 ∘C, other
values remain unchanged. White: observed values; dark grey: modelled
default (S1); light grey: changed tscalar function (pstemp). Results for days
with net radiation > 2 MJ day-1.
Variability during winter on monthly timescale for (a, b) GPP (g C m-2 month-1) and (c, d) AET (mm month-1),
between default settings (S1, a and c) and changed
tscalar (pstemp, b and d) during winter. All days in December,
January, and February are included (i.e. no selection for radiation). All
slopes significantly differed from 1.0 (P<0.05). RMSE values: (a) 22.7, (b) 20.4, (c) 14.7, and (d) 19.7.
Comparison of fluxes for (a) GPP
(g C m-2 month-1) and (b) AET (mm month-1) using different water
uptake functions. Dotted line: observed values. Solid lines: modelled values
for scenarios S1–S4.
Modelled temperature response
Based on the outcome of the literature review and observational data analysis,
this model setup used a lower threshold for P. sylvestris photosynthesis
(pstempmin) of -10 ∘C. The effect of changing the temperature
response in LPJ-GUESS on the seasonal trend of GPP and AET is shown in Figs. 3, 5, and 6. Changing the lower boundary for photosynthesis for P.
sylvestris to -10 ∘C (Fig. 1) results in higher winter estimates for GPP
(Figs. 3a and 5a) and, to a lesser extent, for AET (Figs. 3b and 5b). The
latter can be expected since interception and soil evaporation do not
change and there is only a slight increase in plant transpiration. When
selecting days with high radiation only (Fig. 5), simulations with changed
temperature response follow the distribution of daily observed GPP more
closely. For the entire simulation, the overall error (RMSE, Table 3)
reduces for both AET and GPP, with the exception of GPP at monthly timescales. Correlations (r, Table 3) do not increase for GPP, and are similar
for AET over the entire simulation period. However, the Wilcoxon ranking
test shows that for GPP the modelled data distribution is now matching the
observed data distribution at monthly timescales more closely (P <0.05). In addition, when only winter month data are included (Fig. 6), the slope of the regression substantially improves for GPP from 0.32 to
0.58, while keeping a similar correlation coefficient (0.80 vs. 0.78). This
indicates a better match between modelled and observed results. By changing
the temperature response, the simulation of IAVcw does not improve for the
carbon fluxes, and only marginally for the water fluxes (Table 3).
Alternative plant water uptake parameterisations
Figure 7 shows modelled carbon and water fluxes on a monthly timescale for
the three different water uptake parameterisations (S1–S3) and the control
model setup without soil moisture stress (S4). All three uptake
parameterisations appear to be equally strong in simulating the seasonal
trend with correlations of 0.92–0.94 for GPP and 0.86–0.88 for AET (r,
Table 3). During summer, the linear uptake response curve (S3)
underestimates both AET and GPP more often than the species-specific (S2)
and default uptake (S1) parameterisations. Eliminating water stress (model
setup S4) results in overestimation of fluxes during summer, increased
error, and lower RMSE. Moreover, with this setup both AET and GPP are
overestimated in spring and summer for all years (Fig. 7a), indicating that
water limitation does play an important role in Loobos.
Given the model's very simple two-layer soil hydrology (Sect. 2.2.4) and the
fact that our measured soil moisture data were averaged to correspond with
the model's layer depths (l1 and l2), seasonal soil moisture
patterns are captured reasonably well between the different model setups
when compared to observations (Fig. S3). Modelled soil moisture
in the upper soil layer changes more rapidly than observations suggest, and
modelled moisture recharge in winter increases to higher values than
observed for some years. Soil moisture measurements, however, were not
always available during winter and completely absent from autumn 2000 until
summer 2002. Because plants take up water more conservatively in setup
S3, modelled soil moisture is higher during the growing season for all years
compared to the other two setups, and the bucket never completely empties as
is often the case for the other two setups. Available sap flow data for P. sylvestris (1997,
1998, and 2009) show good correlations with modelled transpiration (Fig. 8,
r=0.68–0.74). For setups S2 and S3, the range of modelled plant
transpiration is lower than the observed plant transpiration (0–1.5 and 0–3 mm day-1 respectively). For setup S1, the range of
modelled plant transpiration matches that of the observations for 1997 and
1998 (0–3 mm day-1). This relates directly to the shape of the
response curve for each setup (Fig. S1), where S2 and S3 reduce
the water supply, S, more strongly than S1 in response to declining soil
water. Correlations for individual years are lowest for 1997, especially for
setups S2 and S3, where modelled transpiration is reduced too strongly in
response to declining modelled soil water between day 100 and 300
(Fig. S3).
On the annual timescale, species-specific uptake (S2) leads to the best
explanation of interannual variability in GPP in terms of correlation
coefficient (Table 3), while for AET there is a small decrease compared to
the default setup. Using the model setup in which soil water is not a
limiting factor (S4), the model also cannot accurately capture interannual
variability in GPP and AET.
Modelled transpiration (mm day-1) for Pinus sylvestris, compared to
observed sap flow (mm day-1). Pearson correlation coefficients
significantly different from 0 (P<0.01) for all separate years as
well as all data points together (ralldata). Sap flow measurements for
1997 and 1998 acquired using tissue heat balance systems, and for 2009 using
Granier thermal dissipation probes. S1 is default uptake, S2 is species-specific uptake, S3 is linear uptake.
Daily modelled (mod, black lines) and observed (obs, red and
blue) soil moisture (as volumetric water content, 1/100 %) for summer of
2003 and 2005. The two depths refer to the two soil layers in LPJ-GUESS:
l1 (0–50 cm) and l2 (50–150 cm). For 2003, the heatwave period is
indicated between the black lines.
Daily observed and modelled fluxes for (a) GPP
(g C day-1) and (b) AET (mm day-1) for July and August in two
different climate years. In summer 2003, a heatwave and corresponding drought
occurred in Europe (e.g. see Teuling et al., 2010). Based
on long-term averages of the Dutch Royal Meteorological Institute (KNMI),
higher temperatures, more sunshine hours, and much less precipitation occurred during this summer, and an official heatwave took place in the
Netherlands during August (KNMI, 2003). The KNMI defines a
heatwave as a period of at least 5 consecutive days in which the maximum
temperature exceeds 25 ∘C, provided that on at least 3 days in
this period the maximum temperature exceeds 30 ∘C. Based on these
criteria, the heatwave duration was from 31 July to 13 August and is marked in
the graph by two dotted black vertical lines. The summer of 2005 had average
temperatures and sunshine but was much wetter, and August was a month with
particularly high precipitation compared to long-term averages
(KNMI, 2005).
Comparing water uptake parameterisations during a dry and wet summer
The summers of 2003 and 2005 were very different, with the 2003 heatwave
over Europe affecting both managed and natural vegetation systems but each
ecosystem showing different responses to the extreme heat (e.g. see
Granier et al., 2007; van der Werf et al., 2007; Teuling et al., 2010). The
2003 heatwave affected the Netherlands (KNMI, 2003) especially in
August, which in combination with a prolonged period of low precipitation
resulted in a drought. We compare the results of the extremely sunny, warm,
and dry August 2003 to those of August 2005, which was a regular but very
wet month. Observed soil moisture at Loobos declined considerably during the
2003 heatwave, and modelled soil water runs out earlier than observations
suggest (Fig. 9, for 2003), with the exception of setup S3 and, to a lesser
extent, for the lower soil layer of setup S2. For 2005, modelled soil
moisture is often too low when using the default setup (S1), and water
content of the upper layer changes more rapidly than observations suggest.
When comparing daily carbon and water fluxes to observations (Fig. 10)
during the wet period (2005), all uptake parameterisations perform well
compared to observed data, with no striking differences between uptake
parameterisations in simulating GPP and AET. During the 2003 heatwave and
drought however, the parameterisations show different responses. During the
first half of the heatwave period (indicated by the two vertical dotted
black lines in Fig. 10), there is a gradual decline in observed daily GPP
and AET at the site. Given the considerable drop in observed soil water
during the heatwave (Fig. 9), reductions in observed GPP and AET look
considerably more gradual (Fig. 10). This suggests the vegetation's possible access of water from deeper layers, or groundwater. The no-water stress
control run (S4) clearly demonstrates there is some water stress at Loobos
(both observed GPP and AET are lower than the model predicts), but all
parameterisations fail to simulate the correct response. The default and
species-specific response curves (S1 and S2) allow PFTs and species to take
up relatively more water at low soil water contents compared to the linear
uptake parameterisation, thereby not restricting photosynthesis as long as
water remains available for uptake. We can observe this effect during the
heatwave period, where the linear uptake function (S3) least underestimates
GPP and AET, because there is more water available for uptake due to
conservative water use, and the effects on the modelled supply are less
strong at lower soil water contents (Figs. S1 and S2). The real observed
response of the Loobos vegetation, however, is not reproduced using either
uptake parameterisation. The sensitivity of GPP and AET to declining soil
moisture during the growing season is visible in Fig. S2 by plotting the
residuals (modelled values minus observed values, so that an underestimation is depicted
with a negative sign) against modelled available soil moisture (Θ).
In general, the linear uptake parameterisation seems to underestimate both
GPP and AET more at higher soil moisture values, so regarding the
observations, this response curve imposes water stress on plants at this
site too strongly.
A comparison of the three different plant water uptake response curves does
not lead to identification of any setup that is clearly superior for
simulating IAVcw compared to the others (Table 3). Species-specific uptake (S2)
results in the smallest errors (RMSE, Table 3) on monthly and daily timescales, but on annual timescale the default uptake (S1) has the smallest
error.
Discussion
Default modelling setup
The model reproduced the daily and monthly carbon and water fluxes equally
well as shown in previous studies with LPJ-GUESS (Sitch et al., 2003;
Gerten et al., 2004; Morales et al., 2005; Zaehle et al., 2005; Hickler et
al., 2006). Fatichi and Ivanov (2014), using a different
process-based vegetation model, similarly found very high correlations on
daily timescales and low correlations on annual timescales for GPP and
evapotranspiration. However, good correlations on shorter timescales can be
expected, given the strong diurnal and seasonal cycles to climatic drivers
(mainly radiation and temperature). While the model produces reasonable flux
estimates at daily and monthly timescales, the small deviations on these
timescales lead to poor estimates of IAVcw and longer timescales,
which Keenan et al. (2012) demonstrated for a wide
range of terrestrial biosphere models.
At some sites where needleleaf evergreen vegetation is the dominant
vegetation type, year-to-year variation in fluxes can be explained by
climatic and environmental drivers (e.g. disturbances) only. For example,
Sierra et al. (2009) applied a process-based stand
vegetation model which showed that some forests are mostly affected by short-term dynamics such as disturbances, and others are more influenced by
climatic controls. Duursma et al. (2009) performed
a model–data comparison using a calibrated empirical photosynthesis model,
and found good fits for GPP on daily to seasonal timescales for several
European FLUXNET sites and, similar to this study, comparably poor fits on
the annual timescale. They attributed part of this mismatch to uncertainty
in the EC data, variations in LAI, and reductions in GPP as a result of soil
drought. Purely observational studies at temperate coniferous forests in
Brasschaat (Carrara et al., 2003, 2004) and Vielsalm (Aubinet et al., 2002) showed that climatic and
ecological drivers (such as changes in LAI, phenology shifts) explain the
majority of interannual variability in observed carbon and water fluxes. Our
results, as well as studies by Jacobs et al. (2009),
Kruijt et al. (2009) and
Luyssaert et al. (2007), suggest that,
in addition to direct climatic and environmental factors, ecological drivers
also operate at the Loobos site.
Uncertainties in the observational data set
For this study, the mismatch between simulated and observed fluxes both at
the monthly and at the annual timescale can only be partly attributed to
uncertainties in the flux data. The magnitude of the error for this data set
is estimated by Elbers et al. (2011) as 8 % of annual
NEE, which is a quarter of the standard deviation of annual NEE, and is
small compared to other flux sites (Elbers et al., 2011, data from
1997 to 2010). Because GPP is estimated from NEE and night-time respiration,
the errors in annual NEE, especially the notorious errors in night-time NEE
due to low turbulence, propagate into GPP estimates. During winter, when
relatively more data are gap-filled, this uncertainty in the data can
contribute to a higher deviation between the modelled and observed results
in this study.
Alternative temperature response function
Observed temperature response at Loobos and similar sites
We presented strong evidence that Pinus sylvestris continues to assimilate during winter in
temperate climates, and even acts as a carbon sink during frost periods
rather than as a source, as most DGVMs currently suggest
(Morales et al., 2005). Falge et al. (2002) even
suggest, based on their analysis of FLUXNET data, that temperate and boreal
conifers should be seen as two separate classes. The observations at Loobos
support this suggestion, as Pinus sylvestris clearly continues to assimilate in winter during
all years, even when daily average temperatures drop below 0 ∘C.
These pine trees grow in a temperate climate, and therefore experience
relatively milder winters compared to the same species at boreal sites.
Plants are known to acclimatise to their growing conditions, so differences
in the seasonal carbon gain within species reflect to a large extent the
light and temperature environment in which they exist (Teskey
et al., 1994). Plants native to a colder climate exhibit higher
photosynthetic rates under colder temperatures, but, at higher latitudes,
Pinus sylvestris is also known to display winter photo-inhibition as a result of lower
winter temperatures (Berry and Bjorkman, 1980).
This winter inhibition of photosynthetic capacity is thought to be a
protective mechanism against damaging combinations of low atmospheric
temperatures and exposure to high irradiances that can be enhanced by snow
cover. If, however, winters are warm enough, photosynthesis in evergreen
forest stands can continue if enough soil water is available to meet the
transpirational demand (Sevanto et al., 2006, and references
therein). How long it takes for the photosynthetic capacity to diminish
during extended cold periods – and possibly recover when temperatures rise
again (e.g. see Suni et al., 2003a, b; Kramer et al., 2008; Kramer and
Hänninen, 2009) – is not known for this site and will be investigated
in a winter measurement campaign of leaf photosynthesis over the next few
years.
Modelled temperature response
The modelled changed temperature response function had a smaller effect on
simulated AET than on simulated GPP (Fig. 5). Simulated AET is calculated as
the sum of plant transpiration, soil evaporation, and canopy evaporation.
Underestimation of canopy evaporation (interception loss) in relation to
precipitation intensity in winter can play a role here. In general, measured
AET fluxes during winter are high for this type of forest. At Loobos,
measured AET peak values during winter are mainly the result of high
interception evaporation (Elbers et al., 2010). Modelled LAI
was slightly lower than observed (Table 2), which results in a lower
precipitation storage capacity for the vegetation than in reality.
Additionally, as the model does not explicitly handle shower intensity, and
prolonged periods of low precipitation intensity occur often at the site
during winter, the model underestimates interception evaporation. This
underestimation of canopy interception likely contributes to
underestimations of AET on the longer timescales as well.
Even when Pinus sylvestris is allowed to continue assimilating at lower
temperatures, the difference between modelled and observed fluxes improves
but is not completely resolved. The shape of the temperature response curve
for Pinus sylvestris (Fig. 1) is modelled as a steep increase from the minimum temperature
(pstempmin) to the optimum temperature (pstemplow), which, to our knowledge,
is not supported by literature but purely empirical. For this study, we
identified a lack of data and literature to verify the exact shape of this
response curve and instead calculated the minimum temperature threshold from
the available data. Smith and Dukes (2013) reviewed the
latest available methods to incorporate photosynthesis temperature
acclimation into global-scale models, and suggest that instead of just
looking at temperature optima, shifts in the slope/intercept of the initial
instantaneous temperature response could be of equal or greater importance,
especially at suboptimal temperatures, and that a combination of data
collection and modelling studies, such as ours, is needed to improve our
understanding and realistically simulate long-term responses of vegetation
to temperature shifts.
The small impact of changing the temperature response function on simulating
IAVcw is of course related to the fact that wintertime fluxes make up
only a small part of the total annual flux (average observed annual GPP for
this data set is 1284 g C m-2), usually less than 10 %. In contrast,
the largest observed interannual difference in GPP for this period is almost
twice as large at 200 g C m-2. Therefore, small improvements in the
winter estimates will not translate directly into good estimates and high
correlation coefficients on the annual timescale.
Alternative plant water uptake parameterisations
The use of three different soil water uptake parameterisations revealed that
the model can satisfactorily simulate GPP and AET during wet summers such as
that of 2005. The model performed well for the years when plant
transpiration for Scots pine could be compared with sap flow observations
(Fig. 8). However, none of the uptake parameterisations capture the observed
response in terms of GPP and AET to a drought such as occurred in the summer
of 2003 (Fig. 10). In addition, none of the three parameterisations
consistently improved all results or improved simulated IAVcw at
Loobos.
Previous studies have demonstrated that LPJ-GUESS is sensitive to
limitations in soil moisture, firstly because the parameters controlling
stomatal conductance are very sensitive to plant water stress
(Zaehle et al., 2005) and secondly because the model does
not account for plant ability to access water from deeper soil layers and
aquifers in water-limiting situations (Hickler et al., 2006; Wramneby et
al., 2008). The debate on how to improve modelling efforts in a mechanistic
way, however, is still ongoing. For example, Hickler et al. (2006) included plant hydraulic architecture in the global
model version of LPJ, thereby changing the calculation of plant water supply
to a more mechanistic scheme. This improved global simulations of AET, but
the updated model requires additional PFT/species-specific parameters that
are often not available and the model still underestimates summer AET at one
Mediterranean test site. Verbeeck et al. (2011) tried
increasing soil depth and used locally varying root profiles to improve
simulations of dry-season GPP for the tropics. Such an approach, however,
does not lead to the desired mechanistic model improvements because it
eliminates simulated water stress completely. Furthermore, high-quality
data on effective rooting depth, soil volume, and deep soil water are rarely
available, and deriving model parameters representing deep tap roots,
sometimes growing through rock fissures or compacted soil layers, is
difficult. These challenges are probably the reason why access to deep water
is, to our knowledge, not captured in any DGVM. Nevertheless, we think that
further efforts should be devoted to improving the current state of the art
in this respect, because access to deep water is probably crucial in many
ecosystems around the world.
The 2003 summer drought simulations at Loobos confirm the strong model
sensitivity to drought: under dry soil moisture conditions the vegetation
shows a much more gradual response in flux reduction compared to the model
runs (Fig. 10). Observed soil moisture values are low and gradually decline
during the heatwave (Fig. 9), suggesting that the vegetation can access water
from deeper layers, or groundwater. Pinus sylvestris is known for its ability to create long
tap roots, especially when growing on sandy soils, so that water uptake is
also possible from sparsely rooted deep soil layers when water becomes
limiting (Jarvis, 2011).
The shape of the water uptake response curves in the model clearly has an
effect on the water uptake (Fig. S1). The exact shape of this
curve, however, is both species and site specific, and remains poorly
defined for global model studies that use broad PFT classifications. For P. sylvestris,
Lagergren and Lindroth (2002) summarised uptake curves from
several studies, and the reported shapes are very similar to the ones used
in this study, most closely resembling wr_rootdist and wr_speciespecific. The reality probably lies in
between the original linear formulation and wr_rootdist, because plants do not reduce
transpiration immediately when soil water content declines: transpiration
remains unaffected until the soil water potential reaches values at which
the xylem can be damaged by cavitation. Next, depending on the strategy of
the tree, transpiration is either reduced due to cavitation or to stomata
closing to prevent cavitation (McDowell et al.,
2008). During droughts, plants may reallocate carbon to roots instead of
leaves or needles, thereby reducing their assimilation potential through
reduced leaf area. Such seasonal changes in carbon allocation and phenology
under drought are currently not explicitly handled in LPJ-GUESS because
allocation occurs annually in the model (on the annual timescale, however,
the ratio of leaves to fine roots is adjusted for water availability). Model
inaccuracies in reproducing this type of vegetation phenology and hence the
simulation of seasonal cycle of CO2 and water can lead to poorly
simulated fluxes compared to observed ones. Future modelling efforts should
focus on root dynamics, and include the effects of groundwater uptake and shifts
in carbon allocation under water stress.
Conclusions
Variability in ecosystem carbon and water exchange is a key aspect of
ecosystem functioning, but, in many cases, the drivers are poorly
understood. Here, we showed that a DGVM, when adapted to the local
conditions, can reproduce daily to seasonal variability in carbon and water
exchange with high correlation coefficients. Similar to other studies,
however, the model cannot reproduce interannual variability. We tried to
identify the driving mechanisms of IAVcw by looking at systematic
biases in the model output. By comparing the model to a long-term data set,
we found that carbon assimilation during winter months at daily average
temperatures below 0 ∘C is important for winter fluxes and not
captured in the current parameterisation of the model, which might also
apply to other, similar, models. Lowering the minimum temperature threshold
for photosynthesis improved the simulation of winter GPP substantially, but
did not greatly improve simulations of IAVcw. In addition, we
demonstrated that the modelled response to drought is too strong for this
site, and that none of the water uptake formulations were consistently
superior in reproducing the observed response of GPP and AET. AET and GPP
during the 2003 heatwave were substantially underestimated by the model,
even when assuming that plants have maximum water supply until the wilting
point is reached. This result and the soil water curves suggest that at this
site, access to deep water is crucial for the vegetation response to extreme
drought. However, our understanding of IAVcw at the Loobos site still
remains incomplete, as we were not able to disentangle the main drivers of
IAVcw here. As future steps we suggest that, firstly, the
representations of water uptake and root growth of plants need further
attention in terms of model testing and parameterisation. This includes the
implementation of a groundwater table and rooting access to it, and
accounting for precipitation duration and intensity to make interception
evaporation in winter more realistic. Secondly, estimating the amount of
water stored deeper in the soil than the soil depth of common DGVMs may be
crucial for simulating the drought response of vegetation even in areas such
as the Loobos site, where this was not expected. Thirdly, we want to further
explore the hypothesis that IAVcw is driven by short-term resource
allocation of the vegetation. If past and current productivity (GPP) drive
future productivity, for example via LAI changes, and these are influenced
by environmental drivers and stressors such as temperature and droughts,
modelling allocation and growth on a daily or monthly time step could be
crucial. Because the process interactions underlying variability in
ecosystem functioning are so complex that analyses with single factors, such
as temperature or precipitation, often do not shed light on the mechanisms,
we think that improvement of the process-based modelling and comparing
these results with observations is an important complementary approach.
Accurate reproduction of site-level fluxes with such models on seasonal
to annual timescales is essential for our understanding of
vegetation–climate interactions and for reducing uncertainties in future
projections.