Correct representation of seasonal leaf dynamics is crucial for terrestrial
biosphere models (TBMs), but many such models cannot accurately reproduce
observations of leaf onset and senescence. Here we optimised the
phenology-related parameters of the ORCHIDEE TBM using satellite-derived
Normalized Difference Vegetation Index data (MODIS NDVI v5) that are linearly
related to the model fAPAR. We found the misfit between the observations and
the model decreased after optimisation for all boreal and temperate deciduous
plant functional types, primarily due to an earlier onset of leaf senescence.
The model bias was only partially reduced for tropical deciduous trees and no
improvement was seen for natural C4 grasses. Spatial validation demonstrated
the generality of the posterior parameters for use in global simulations,
with an increase in global median correlation of 0.56 to 0.67. The simulated
global mean annual gross primary productivity (GPP) decreased by
Leaf phenology, the timing of leaf onset, growth and senescence, is a
critical component of the coupled soil–vegetation–atmosphere system as it
directly controls the seasonal exchanges of carbon, C, as well as affecting
the surface energy balance and hydrology through changing albedo, surface
roughness, soil moisture and evapotranspiration. In turn leaf phenology is
largely governed by the climate, as leaf onset and senescence are triggered
by seasonal changes in temperature, moisture and radiation. Leaf phenology is
therefore sensitive to inter-annual climate variability and future climate
change (Cleland et al., 2007; Körner and Basler, 2010; Reyer et al.,
2013), as well as to increasing atmospheric CO
In order to improve predictions of the impact of future climate change on vegetation and its interaction with the global C and water cycles, it is crucial to have prognostic leaf phenology schemes in process-based terrestrial biosphere models (TBMs) that constitute the land component of Earth system models (ESMs) (Kovalskyy and Henebry, 2012b; Levis and Bonan, 2004). Many such models exist in the literature, especially for temperate and boreal forests (e.g. Arora and Boer, 2005; Caldararu et al., 2014; Chuine, 2000; Hänninen and Kramer, 2007; Knorr et al., 2010; Kovalskyy and Henebry, 2012a) and have been included in most TBMs. However, model evaluation studies have shown that there are biases in the growing season length and magnitude of the leaf area index (LAI) predicted by TBMs when compared to ground-based observations of leaf emergence and LAI (Kucharik et al., 2006; Richardson et al., 2012) or satellite-derived measures of vegetation greenness and LAI (Kim and Wang, 2005; Lafont et al., 2012; Maignan et al., 2011; Murray-Tortarolo et al., 2013). This can result in systematic errors in model predictions of the seasonal carbon, water and energy exchanges (Kucharik et al., 2006; Richardson et al., 2012; Walker et al., 2014).
As is always the case prior to model parameter calibration, it is unclear whether the misfit between modelled and observed measures of leaf phenology is the result of inaccurate parameter values, model structural error, or both. In order to answer this question, the parameters first need to be optimised using data assimilation (DA) techniques, and if the models cannot reproduce the data within defined uncertainties we expect to gain insights into possible directions for model improvement. DA is also a useful way to better characterise and possibly reduce uncertainty in model simulations, and to determine the relative influence of parametric, structural and driver uncertainty (e.g. Migliavacca et al., 2012).
Many studies have optimised the parameters of phenology models for a range of species with ground-based observations of the date of leaf onset (Blümel and Chmielewski, 2012; Chuine et al., 1998; Fu et al., 2012; Jeong et al., 2012), the “green fraction” derived from ground-based digital photography (Migliavacca et al., 2011) or with spring onset dates derived from carbon fluxes taken at flux tower sites (Melaas et al., 2013). Melaas et al. (2013) went further and demonstrated the transferability of parameters in time and between sites by including multiple sites in the optimisation. All of these studies have used DA to test different phenology model structures, thereby contributing significantly to the debate about whether a simple classical temperature-driven budburst model is sufficient, or whether more complex chilling and/or photoperiodic cues are needed to best predict leaf onset. Several studies also investigated the impacts of optimising phenology on the resulting C and water budgets (Migliavacca et al., 2012; Picard et al., 2005; Richardson and O'Keefe, 2009).
Peñuelas et al. (2009) noted that medium- to coarse-resolution satellite
data might be more appropriate for optimising the phenology in TBMs, due to
the large difference in scale between the resolution of a typical model grid
cell (1 Can we constrain the phenology-related parameters and processes of a
typical process-based TBM at global scale using satellite “greenness” index
data? Does this produce a generic parameter set that results in improved
simulations of the seasonal cycle of the vegetation, or are further model
structural developments required? What is the impact of the optimisation on mean patterns and trends in
vegetation productivity (as represented by the mean fraction of absorbed
photosynthetically active radiation (fAPAR), amplitude and growing season
length (GSL)) at regional and global scales?
To achieve this we performed a global, multi-PFT, multi-site optimisation of the phenology model parameters for the six non-agricultural deciduous PFTs of the ORCHIDEE TBM. The phenology models in ORCHIDEE are common to many process-based TBMs. Note there is no specific phenology model associated to evergreen PFTs, where leaf turnover is simply a function of climate and leaf age.
Some of the carbon cycle-related parameters of ORCHIDEE (including phenology-related parameters) have previously been optimised using in situ flux measurements (e.g. Kuppel et al., 2014; Santaren et al., 2014; Bacour et al., 2015). Here we focus purely on improving the timing of both spring onset and autumn senescence of ORCHIDEE at global scale, by using a novel approach to assimilate normalised medium-resolution satellite-derived vegetation “greenness” index data (MODIS NDVI collection 5) that are linearly related to the simulated daily fAPAR. The aim of a multi-site (MS) (i.e. model grid cell) assimilation is to find a unique parameter set for each PFT that results in a similar improvement as a single-site (SS) optimisation, as the range of posterior parameter values for individual sites/species can be large (Richardson and O'Keefe, 2009). We hypothesise that the MS approach may average out the site-based variability, and thus provide one consistent PFT-generic parameter vector that can be used for global simulations (e.g. Kuppel et al., 2014).
ORCHIDEE is a global process-oriented TBM (Krinner et al., 2005) and is the
land surface component of the IPSL-CM5 Earth System Model (Dufresne et al.,
2013). In this study we used the “AR5” version that was used for the IPCC
Fifth Assessment Report (Ciais et al., 2013). The model calculates carbon,
water and energy fluxes between the land surface and the atmosphere at a
half-hourly time step. The water and energy module computes the major
biophysical variables (albedo, roughness height, soil humidity) and solves
the energy and hydrological budgets. The carbon module controls the uptake of
carbon into the system and respiration following cycling of C through the
litter and soil pools. Carbon is assimilated via photosynthesis depending on
light availability, CO
Standard PFTs used in ORCHIDEE, their short name and corresponding phenology model (see Appendix A for a full description of the phenology models). Evergreen and agriculture PFTs do not have a specific phenology model in the ORCHIDEE TBM.
The seasonal cycle of the terrestrial vegetation is observed daily, cloud
cover permitting, at a global scale and medium-scale spatial resolution
(250 m) from several polar orbiting spectroradiometers. Studies have shown
that considerable discrepancies exist between so-called “high-level”
satellite products such as LAI or fAPAR, especially when considering their
magnitude (D'Odorico et al., 2014; Garrigues et al., 2008; Pickett-Heaps et
al., 2014). This is because radiative transfer models are used to derive
these products, which introduces uncertainty due to undetermined parameters
or potentially incomplete descriptions of the radiative transfer model
physics. Instead therefore, we considered a vegetation greenness index, the
Normalized Difference Vegetation Index (NDVI), that is directly related to
the near infrared (NIR) and red (RED) surface reflectance,
NDVI observations are derived from the MOD09CMG collection 5 (v5) surface red
(620–670 nm) and near-infrared (841–876 nm) daily global reflectance
products available at 5 km from the MODerate resolution Imaging Spectrometer
(MODIS) on-board the NASA's Terra satellite. The reflectance data were
cloud-screened and corrected for atmospheric and directional effects (related
to the change of reflectance with observation geometry) following (Vermote et
al., 2009), and the corresponding NDVI was calculated for the 2000–2008
period. The time series were interpolated on a daily basis, in order to
account for any missing values due to cloud, using a third degree polynomial
and considering the 10 nearest valid acquisitions, with a maximum allowed
difference of 15 days. The NDVI values were then spatially averaged at
the model forcing spatial resolution (0.72
The ORCHIDEE Data Assimilation System (
The posterior parameter covariance can be approximated from the inverse of
the second derivative (Hessian) of the cost function around its minimum,
which is calculated using the Jacobian of the TBM model with respect to fAPAR
at the minimum of
The posterior parameter covariance can then be propagated into the model
state variables (fAPAR or net C flux) space given the following matrix
product and the hypothesis of local linearity (Tarantola, 1987):
ORCHIDEE parameters optimised. For each PFT the prior values, minimum and maximum values (in squared brackets) and multi-site posterior mean values (in bold) are given. Note the prior uncertainty on the parameters is defined as 40 % of the full parameter range.
Figure 1 shows a general schematic of how the parameters of the phenological
equations used in ORCHIDEE (Botta et al., 2000) control the timing of the
seasonal cycle of the LAI as well as the rate of leaf growth and fall. The
parameters that are optimised for each PFT are given in Table 2 and are
briefly described here. A more detailed description can be found in
Appendix A. The start of the seasonal cycle of temperature-driven PFTs is
constrained by optimising the growing degree day threshold,
GDD
Schematic to show how the optimised parameters control the timing of
the leaf phenology in ORCHIDEE. The dotted arrow shows that the temperature
and moisture threshold for senescence also affects the rate of leaf fall for
grasses by slowing down the turnover rate once this threshold has been
reached (whereas for trees only the
The end of the seasonal cycle is constrained by optimising the critical leaf
age for senescence,
For phenology models that are driven by soil moisture conditions (“MOI”
models – see Appendix A and Table 1) the parameter that controls leaf onset
is the “minimum time since the last moisture minimum”
(Moist
The prior parameter values are taken from the ORCHIDEE standard (non-optimised) version and are detailed in Table 2. The maximum and minimum bounds of the parameters were set based on literature and “expert” knowledge. Prior uncertainty on the parameters was taken to be 40 % of the parameter range following Kuppel et al. (2012).
The six deciduous, non-agricultural PFTs of ORCHIDEE are optimised in this
study. For each of the PFTs that were optimised we selected 30 sites
(where one site is equal to one model grid cell at 0.72
Global distributions of fractional cover for the six PFTs optimised in this study. Red upright triangles mark the location of the optimisation sites, and yellow upside-down triangles mark the location of the validation sites.
In this study ORCHIDEE is used in forced offline mode and is driven by
3-hourly ERA-Interim meteorological fields (Dee et al., 2011), on a regular
0.72
For each PFT optimised, the 15 optimisation sites (see Sect. 2.3.3) were first optimised simultaneously (i.e. all sites were included in the same cost function), over the 2000–2008 period using the multi-site (MS) approach detailed in Kuppel et al. (2012). Following (Santaren et al., 2014) we tested the ability of the algorithm to find the global minimum of the cost function by starting the iterative minimisation algorithm (see Sect. 2.3.1) at different points in the parameter space, choosing 20 random “first guess” sets of parameters and performing a MS optimisation for each. The results of these tests are presented in Sect. 3.1.
A single site (SS) optimisation was then performed for each of the same 15 optimisation sites. The assimilations were exactly the same as for the MS optimisation, except each site was optimised separately. The posterior parameter vector resulting from the “best” random first guess MS optimisation (taken as the greatest % reduction in the cost function) was used as the first guess for the SS optimisation. The first guess with the greatest % reduction in the cost function was equivalent to the first guess that resulted in the lowest value of the cost function, as the % reduction was calculated using the value of the cost function using the default (prior) ORCHIDEE parameters.
The same MS posterior parameter vector for each PFT was then used to perform a simulation at each of the 15 extra spatial validation sites (see Sect. 2.3.3) over the same time period. In addition, prior and posterior simulations at all 30 optimisation and validation sites were extended to cover the 2009–2010 period in order to perform a temporal validation.
Finally two global-scale simulations (with increasing atmospheric CO
The prior and posterior RMSE and correlation coefficient,
The curve-fitting method of Thoning et al. (1989) was used to fit a function to the daily time series of observations and model output as described in Maignan et al. (2008). The function consists of two parts; a second-order polynomial that is used to account for the long-term trend, and a fourth order Fourier function to approximate the annual cycle. The residuals of the fit to this function were filtered with two low pass filters in Fourier space (80 and 667 cut-off days) and then added back to the function to produce a smoothed function that captures the seasonal and inter-annual variability and long-term trend. The detrended curve can be calculated by subtracting the trend from the smoothed function. The start of season (SOS – leaf onset) and end of season (EOS – the start of leaf senescence) were defined as the upward and downward crossing points of the “zero-line” of the de-trended curve per calendar year (see Fig. 1 in Maignan et al., 2008). These values were calculated for all grid cells with only one seasonal cycle per year (this includes grid cells in the SH where the growing season spans 2 calendar years). The GSL was calculated as the number of days per calendar year when the detrended curve is greater than zero. Therefore unlike the SOS and EOS, the GSL was also calculated for grid cells that contain multiple growing seasons within a calendar year.
For the trend analysis, a linear least squares regression was used to calculate the long-term trend in the annual fAPAR amplitude, growing season length (GSL) and the mean fAPAR time series.
The global simulations were evaluated with the same MODIS NDVI data that were used at the site level for the optimisation, following the protocol of Maignan et al. (2011). The following metrics were used for evaluation of both the prior and posterior simulations.
The correlation between the normalised simulated fAPAR and MODIS NDVI weekly time series.
The bias (in days) between the modelled and observed SOS and EOS dates (model – observations) were also examined so as to investigate the impact on the timing of the phenology more directly (a positive bias indicates the model date is later than the date calculated from the observations).
The above metrics were calculated for each grid cell. Following this a global median value was calculated, as well as a median correlation per PFT.
We initially tested the ability of the MS optimisation to find the global
minimum of the cost function (
Metrics to describe the ability of the optimisation algorithm to
find the global minimum of the cost function for the MS optimisation using 20 different random “first guess” parameters. The 2nd column shows the
final value (after 25 iterations of the BFGS optimiser) for the random test
that resulted in the lowest value of
However the picture is different for natural C4 grasses (NC4). Only 2 out of the 20 random first guess tests resulted in a > 10 % reduction in the cost function, and although the spread of final values of the cost function was low and close to the minimum value (Table 3, column 2), the final value was between 2 and 10 times higher than that achieved for other PFTs (Table 3, column 1). This suggests that the optimisation algorithm cannot find a better fit to the data than with the default parameter values. It is possible that the BFGS algorithm is not adequate for exploring the parameter space for NC4 grasses, but given that none of the random tests resulted in a noticeable reduction in the cost function, it is more likely that the model sensitivity to the parameters is lower than for other PFTs. This in turn suggests that the phenology model structure itself is inadequate for this NC4 grasses.
There is an improvement in the model–data fit after both SS and MS optimisations for all temperate and boreal broadleaved and needleleaved deciduous forests (TeBD, BoND, BoBD) and for natural C3 grasses (NC3), largely resulting from an earlier onset of senescence in the model and therefore a substantially shortened growing season length (Figs. 3 and 4). The shift in the start of leaf growth is much smaller, which is not surprising as the prior model more closely matches the observations. Of the four PFTs listed above, only TeBD trees have a slightly later leaf onset as a result of the optimisation. Figure 3 shows the full time series at one site of both the normalised and un-normalised fAPAR and NDVI, together with the simulated LAI, for the BoBD PFT. This site is provided as an example of the typical changes in temporal behaviour seen for the four PFTs listed above. Figure 4 shows the mean seasonal cycle of the normalised fAPAR/NDVI across all sites and years (2000–2008) for each of the four PFTs and demonstrates that the patterns seen in Fig. 3 are similar for all the boreal and temperate deciduous PFTs.
Time series (zoom to 2003–2007) for one example BoND PFT site
(72
The mean seasonal cycle of the normalised modelled fAPAR before and
after optimisation, compared to that of the MODIS NDVI data, for the
temperate and boreal deciduous PFTs (TeBD, BoBD, BoND and NC3).
Black
The optimisations resulted in a significant reduction in the RMSE
(34–61 %) and increase in correlation (posterior
Prior and posterior median (across sites) RMSE and
There is a discernible slowing down in the rate of leaf growth towards the end of the leaf onset after the assimilation, which particularly results in an improved fit to the observations for the TeBD, BoND and NC3 PFTs. However it is noticeable that although parameters that partially control the rate of leaf growth and fall are included in the optimisation, the model generally grows and sheds leaves too fast compared to the observations, except for the BoBD PFT, which results in the model having an unnatural “box-like” temporal profile (Fig. 4 and see Sect. 4.3 for further discussion).
The MS optimisation (red line in Figs. 3 and 4) largely results in a similar
reduction in RMSE and increase in
For both the tropical broadleaved raingreen (TrBR) and natural C4 grass (NC4)
PFTs, the median prior fAPAR simulation performs reasonably well compared to
the observations, with RMSEs of 0.29 and 0.23 and
Example time series (2002–2008 period) for the tropical deciduous
PFTs for which phenology is driven by moisture availability. The two panels
compare the normalised simulated fAPAR to the normalised MODIS NDVI (black
curve) prior to (blue curve) and after the optimisations (orange
curve
There is no change in the median RMSE and
Table 4 also shows the RMSE and
The global median correlation between the model and the MODIS NDVI data have
increased from 0.56 to 0.67 after the optimisation, demonstrating an overall
improvement in the simulated fAPAR time series. As seen at the site level, the
largest increase in correlation between the modelled and observed time series
is for the boreal PFTs. There is also a modest improvement for natural C3
grasses (Table 5). Figure 6 shows the spatial distribution of the correlation
for the both posterior simulation and the difference after optimisation
(posterior – prior). The difference map shows that
Global maps showing the correlation between the simulated fAPAR and MODIS NDVI data in the weekly time series for the posterior simulation (left column) and the difference after optimisation (posterior – prior) (right column). Note POST refers to the posterior simulation after optimisation, and PRIOR to the simulation using the standard parameters of ORCHIDEE.
The prior (blue), MS posterior (red) and SS posterior (orange)
parameter values (circles) and uncertainty (error bars – variance calculated
in Eq. 3) for each parameter and each PFT. For the SS optimisations the
circle and error bars represent the mean and standard error of the mean of
all sites, and the crosses give the posterior values for each site. Refer to
Table 1 for a description of the PFTs and Fig. 1 and Appendix A for a
description of each parameter. The
Median prior and posterior correlation between modelled fAPAR and
satellite NDVI daily time series and inter-annual anomalies of the annual
mean. The metrics are computed for each PFT for grid cells that contained
>
The global median end of season (EOS) bias (model – observations) between the model and MODIS data was reduced dramatically as a result of the optimisation (prior: 33 days; posterior 5 days). Note that a positive bias indicates the model date is later than the date derived from the MODIS data. Again, boreal PFTs and NC3 grasses showed considerable improvement, as expected from the site-level behaviour (Sect. 3.2), but grid cells containing high fractions of temperate and boreal evergreen trees were also positively affected (Table 6). The bias in the start of season (SOS) dates also decreased (prior and posterior global median bias of 22 and 14 days, respectively), with improvements seen for all PFTs except TeBD trees and crops (Table 6).
Median prior and posterior bias between model- and
observation-derived start of season (SOS) and end of season (EOS) dates
(model – observations). The metrics are computed for each PFT for grid cells
that contained >
The prior value, prior range and posterior value from the MS optimisation for each parameter (per PFT) are shown in Table 2. Figure 7 shows the prior and posterior parameter values for both the SS and MS optimisation for each parameter and each PFT. In Fig. 7 the mean and standard error of the mean of the SS posterior parameters are shown (circle with error bars), together which the value obtained at each individual site (crosses). For the prior simulation and MS optimisation the error bar corresponds to the standard deviation of the parameter value (calculated using Eq. 3). The MS uncertainty is lower than spread of SS posterior values, suggesting that it underestimates the true uncertainty of the posterior parameters. This may be the case given the assumptions of linearity of the model and of Gaussian and uncorrelated errors.
The lower posterior values of
The earlier start of senescence is overwhelmingly caused by an increase of
The phenology of C3 grasses is also controlled by soil moisture availability.
The moisture-related leaf onset parameter, Moist
For the tropical raingreen forests (TrBR), there is a decrease in the
posterior SS and MS values of Moist
The difference (posterior – prior) of the simulated annual mean
(over the 1990–2010 period)
Although the phenology of C4 grasses is governed by both temperature and
moisture conditions, the fact that there is no change in the value and
uncertainty of the temperature-related parameters,
Figure 8 shows the change (posterior – prior) in the mean annual GSL, fAPAR
amplitude and mean simulated fAPAR globally over the 1990–2010 period. As
expected from the site-level results for temperate and boreal PFTs (Fig. 4
and Table 4), there was a strong decrease in the mean GSL in the high
latitudes and grasslands across much of the NH (median of
Figure 9 shows the linear trend (yr
Linear trend (yr
We chose not to compare the simulated trends with that of the MODIS NDVI.
This was partly because the 2000–2010 period is likely too short to
calculate a robust trend, as the influence of inter-annual variability will
be stronger; indeed the trends over the longer 1990–2010 period are more
geographically distinct. Secondly it was not the aim of this study to
validate the modelled trends, nor would it be appropriate, because we have
used a version of the model that does not include land use change,
disturbance and other effects that will contribute to changes in vegetation
greenness at global scale. The results presented here serve to highlight that
different parameter values can change the strength, sign and location of the
trends. However, it is worth noting that the sign of the simulated trend does
not always match the MODIS data, especially for drier, warmer semi-arid
regions. For example, the browning trend seen in the Kazakh Steppe in the
MODIS data is stronger in ORCHIDEE and extends further east into Mongolia
(results not shown). Similarly, the model predicts a decline in fAPAR across
much of sub-Sahelian Africa that is not seen in the MODIS NDVI. The
optimisation did not result in a change in the trend direction for this
region, though generally it did reduce its strength. Overall the greening
trend in the NH (>
Whilst the SS optimisations result in a reduction in the model–data misfit across a range of sites, this study has shown that a MS optimisation can achieve a similar improvement with one unique parameter vector. This is an important result, and reinforces the conclusions of Kuppel et al. (2014) that the MS posterior parameters can be used with confidence to perform global simulations. Previous site-level optimisations of phenology models have resulted in a wide range of parameterisations of growing degree day sum, chilling requirements and light availability (Migliavacca et al., 2012; Richardson and O'Keefe, 2009); therefore it is difficult to know which values to use for regional-to-global scale simulations. In most cases the MS optimisation averages out the variability due to specific site characteristics. The validation at the site and global scale using daily MODIS NDVI data demonstrates the generality of the MS posterior parameter vectors given the ORCHIDEE model structure. This gives confidence in using these values in regional-to-global scale simulations of the carbon and water fluxes and for future predictions with this model.
Although the optimisation resulted in a dramatic improvement in the seasonal leaf dynamics for temperate and boreal ecosystems, the impact in the inter-annual variability (IAV) as a result of the optimisation for any variable – mean annual fAPAR, mean spring/autumn fAPAR, GSL, SOS, EOS – was minimal (results not shown). This was disappointing as generally there is a low correlation in the IAV between ORCHIDEE and the MODIS data, and IAV in spring phenology has been shown to be a dominant control on C flux anomalies (Keenan et al., 2012).
As data assimilation schemes are expected to result in a reduction in the prior data–model misfit it is useful to assess if the posterior parameter values are indeed realistic, which is conditioned on the prior range as well as the interaction (correlation) between the parameters. However validating these values remains difficult, as there is no database that corresponds directly to phenology model parameters, which may have a different meaning in different models.
Leaf lifespan (LL) however can more easily be measured and therefore can be
found in the literature and in plant trait databases (Kattge et al., 2011;
Wright et al., 2004). Although leaf lifespan is not the same as the critical
age for senescence parameter (
Most of the posterior values for the moisture-related phenology parameters
for the tropical broadleaved raingreen and natural C4 grasses (TrBR and NC4)
PFTs were also at their upper or lower bounds, even with liberal parameter
bounds due to lack of prior knowledge. For example the posterior value of
Moist
In order to better evaluate the model parameterisations, we suggest that the
phenology observation and modelling community could engage in an elicitation
exercise (O'Hagan, (1998, 2012);
The onset models used for temperate and boreal PFTs in ORCHIDEE are mostly simple spring warming models, though some include a chilling requirement (Table 1). Their comparative ability to reproduce the observations could add to evidence that more complex representations including light availability may not be needed (Fu et al., 2012; Picard et al., 2005; Richardson and O'Keefe, 2009). Other studies however have showed improved performance when a photoperiod term was included for species with a late leaf onset (Hunter and Lechowicz, 1992; Migliavacca et al., 2012; Schaber and Badeck, 2003). Several authors (e.g. Fu et al., 2012; Linkosalo et al., 2008) have pointed out that whilst the simple warming onset models may perform well under current climate conditions, future predictions may require additional complexity; for example the model-defined chilling period may not be sufficient in warmer conditions. More importantly perhaps, any model that requires a fixed start date from which thresholds are calculated may be inconsistent under increased temperatures, as warming will start before the defined start date (Blümel and Chmielewski, 2012). Ideally therefore these models should also be tested under future warming scenarios, although (Wolkovich et al., 2012) showed the magnitude of species' phenological response to temperature increases is lower in warming experiments compared to historical observations.
The coarse-resolution observations used in this study will include the
spatial variability in the timing of phenological events for different
species within a PFT, or even due to specific site characteristics (edaphic
conditions, terrain/slope, local meteorological effects) within the same
species (Fisher and Mustard, 2007). During senescence in particular,
spatial variability in the rate of leaf fall, and to some extent the leaf
colouration, will likely contribute to the decline in “greenness” seen in
NDVI observations. ORCHIDEE does not explicitly represent this variability,
instead vegetation is represented as a mean stand (which is also responsible
for the unnatural “box-like” temporal profile seen for some sites/PFTs) and
thus the posterior value of
The processes that govern leaf phenology in ORCHIDEE cannot reproduce the observations as well in regions where moisture availability, and not temperature, is likely the dominant control. The optimisation has revealed structural deficiencies as the probable cause of inaccurate simulations, rather than incorrect parameter values. In addition to problematic, “edge-hitting” posterior parameter values (see Sect. 4.2) the prior and posterior model incorrectly simulates a strong decline in vegetation productivity in the Sahelian region, which is opposite to that seen in satellite observations (results not shown, but see Traore et al. (2014a). Traore et al. (2014a) suggested that incorrect trends predicted by ORCHIDEE could be related to the phenology models, however this study shows that is not the case. Even without a comparison to observations, ORCHIDEE does not always appear to have the correct response to precipitation, with two periods of simulated growth seen in one rainy season but without any decline in rainfall. This unexpected model behaviour is not corrected by the optimisation and needs investigating.
It is likely that the phenology models in ORCHIDEE are too simplistic for these regions and/or that the computation of soil water availability or the plant water stress function are inadequate. Such issues are likely encountered in other TBMs as they rely on similar models. Knorr et al. (2010) pointed out that evaporative demand, and not just moisture availability, should be considered in phenology models. (Traore et al., 2014b) evaluated the inter-annual variability of the soil moisture of ORCHIDEE across Africa using satellite-derived estimates, and found that the new 11-layer hydrology version performed better than the 2-layer version that was used in this study. The latest version of ORCHIDEE (Naudts et al., 2015) has a more mechanistic representation of plant water stress using water potential in the soil–plant continuum, which may lead to better predictions of leaf dynamics in drought-prone regions.
Although there are fewer studies focusing on the modelling of plant moisture-availability driven phenology, some models do exist for the dry tropics/semi-arid regions. Such models aim to simulate the vegetation response to soil and groundwater availability or atmospheric demand, both empirically (Archibald and Scholes, 2007; Do et al., 2005), or in a more mechanistic way by including non-linear feedbacks via transpiration (Choler et al., 2010). Similar approaches could be included in TBMs in order to better represent leaf growth and turnover in semi-arid grassland and savannah ecosystems in the dry tropics.
Observed increases in GSL and/or increases in vegetation density have been
shown to result in concurrent impacts on the C surface fluxes on seasonal
time scales (Dragoni et al., 2011; Piao et al., 2007; Richardson et al.,
2009), although the magnitude and sign of the effect on net C fluxes is still
a topic of debate (see Barichivich et al., 2013; Keenan et
al., 2014; Piao et al., 2008; Richardson et al., 2010; White and Nemani,
2003). An in-depth analysis of the impact of the modified leaf phenology on
the C, water and energy cycles and the subsequent feedbacks to the atmosphere
was beyond the scope of this study. However, the changes in leaf phenology
described above resulted in a
Aside from making predictions of the carbon, water and energy budgets, TBMs are routinely used in trend attribution studies. A good example in this context would be the exploration of the causes of “greening” or “browning” trends in vegetation productivity (Hickler et al., 2005; Piao et al., 2006), or the impact of such trends on resource use efficiency (Traore et al., 2014b) or the C cycle (Piao et al., 2007). The fact that the optimisation resulted in changes in the strength and location of these trends (Fig. 9) demonstrates that such analyses are partly dependent upon model parameters, which can be a considerable source of uncertainty (Enting et al., 2012).
This study has demonstrated that a time series of normalised
coarse-resolution satellite NDVI data can be used to optimise the parameters
of phenology models commonly used in TBMs, and crucially that a multi-site
optimisation can find a unique parameter vector that enables better
predictions of the seasonal leaf dynamics at global scale. This type of
model calibration framework is thus imperative for Earth system model
development. The results also highlight that optimising the parameters
allows model developers to distinguish between inaccurate model
representations resulting from incorrect parameter values and model
structural deficiencies. In ORCHIDEE the models used for predicting the leaf
phenology in temperate and boreal regions are able to reproduce the seasonal
cycle of the vegetation well after calibration, but ecosystems driven by
water availability require further modification, particularly for natural C4
grasses. The optimisation also led to changes in the strength and location
of “greening” and “browning” trends in the model, suggesting caution
should be exercised when using un-calibrated models for trend attribution
studies. Furthermore, the observed trends were not well captured in some
regions, which is a key aspect to improve upon when considering future
simulations of climate, CO
The MODIS MOD09CMG collection 5 surface reflectance data are freely
available to download from the Land Processes Distributed Active Archive
Center (LP DAAC) data portal (
In temperate and boreal regions the onset of leaves is mainly driven in the
spring by an accumulation of warm temperatures (see the review of
Hänninen and Kramer, 2007). The well-known growing degree day (GDD)
model (Chuine, 2000) sums up the temperatures above a given temperature
threshold,
This simple model may be refined. For example the GDD threshold has been
reported to depend on a “chilling requirement” for some species; i.e. their
physiology requires cold temperatures to trigger the mechanism that will
allow budburst to occur (e.g. Orlandi et al., 2004). This ensures that the
dormancy has been broken after a required cold period, and thus prevents a
too-early awakening. The Number of Chilling Days (NCD) GDD model initiates
leaf onset earlier with an increase in the number of chilling days, defined
as a day with a daily mean air temperature below a PFT-dependent threshold
accumulated after a given starting date (e.g. Botta et al., 2000; Cannell and
Smith, 1986; Murray et al., 1989). The GDD threshold therefore decreases as
NCD increases. This experimental relationship is a negative exponential with
three PFT-specific parameters (Botta et al., 2000):
The start of the growing season for the boreal needleleaved deciduous (BoND)
PFT occurs when the number of growing days (NGD), i.e. days with a mean daily
temperature above the threshold temperature,
For C3 and C4 natural grasses and crops (NC3, NC4, AC3, AC4), the GDD
threshold is given by a second-degree polynomial of the long term mean
annual air surface temperature
For the tropical broadleaved raingreen (deciduous) (TrBR) PFT, the start of the growing season depends only on the moisture availability criterion (hereafter referred to as the “MOI” model) previously described for grasses and crops.
When the onset of leaves is declared, the allocation module first allocates
carbon from the carbohydrate reserves towards leaves and roots as long as the
LAI is lower than a given threshold, which is a function of the maximum LAI,
LAI
The onset parameter values are listed in Table 2.
To account for the fact that the photosynthetic efficiency of leaves depends
on their age,
The second turnover process is a senescence process (the end of the growing
season and the shedding of leaves) that is based on climatic conditions.
This only exists for deciduous PFTs. For tree PFTs whose senescence is
driven by sensitivity to cold temperatures (TeBD, BoND, BoBD and grasses),
senescence begins when the monthly air surface temperature goes below a
threshold temperature, defined as a second order polynomial of the long-term
mean annual air surface temperature
For phenology models that are driven by soil moisture conditions (“MOI”
models – see Appendix A and Table 1) the parameter that controls leaf onset
is the “minimum time since the last moisture minimum”
(Moist
For grasses, the climatic senescence is controlled by both temperature and moisture conditions. The senescence parameter prior values are listed in Table 2. Note that no senescence models in ORCHIDEE currently include a photoperiod term for either onset or senescence.
This work was supported by the CARBONES project within the European Union's 7th Framework Program for Research and Development. N. MacBean warmly appreciates the helpful comments on the manuscript provided by M. De Kauwe and M. Disney for discussions about elicitation. The authors are grateful to the ORCHIDEE project team for their support, and for the computing resources and assistance provided by the IT staff at LSCE. Finally we would like to thank the reviewers and editor for their useful suggestions for improving the manuscript. Edited by: T. Keenan