Introduction
Earth-observing strategies focusing on carbon cycle systematic monitoring
from satellites, flask and in situ networks are leading to an increasing number of near-real-time
observations available to systems such as those developed in the framework of
the European Union Copernicus Atmosphere Monitoring Service (CAMS). CAMS uses
the Numerical Weather Prediction (NWP) Integrated Forecasting system for
Composition (C-IFS) of the European Centre for Medium range Weather Forecasts
(ECMWF) to produce near-real-time global atmospheric composition analysis and
forecasts, including CO2 along with
other environmental and climate relevant tracers
. The purpose of
the real-time CO2 analysis/forecasting system is to provide timely
products that can be used by the scientific community among other users. For
example, those working on new instruments, field experiments, satellite
retrieval products, regional models requiring boundary conditions, or
planning flight campaigns.
The present monitoring of global atmospheric CO2 relies on observations
of atmospheric CO2 from satellites – e.g. Greenhouse Gases Observing
Satellite (GOSAT, www.gosat.nies.go.jp); Orbiting Carbon Observatory 2
(OCO-2, http://oco.jpl.nasa.gov) – and flask and in situ networks – e.g. National
Oceanic and Atmospheric Administration Earth System Research Laboratory
(NOAA/ESRL, www.esrl.noaa.gov/gmd); Integrated Carbon Observation
System (ICOS, http://icos-atc.lsce.ipsl.fr); Environment Canada
(www.ec.gc.ca/mges-ghgm) – which are assimilated by global tracer
transport models to infer changes in atmospheric CO2
e.g. or by flux inversion systems
e.g. to estimate the large-scale surface fluxes of
CO2.
The current C-IFS CO2 analysis is produced by assimilating CO2 data
retrieved from GOSAT by the University of Bremen , as well as
all the meteorological data that is routinely assimilated in the operational
meteorological analysis at ECMWF. have shown that
the atmospheric data assimilation system alone cannot completely remove the
biases in the background atmospheric CO2 associated with the
accumulation of errors in the CO2 fluxes from the model. This happens
because currently the CO2 surface fluxes in the C-IFS data assimilation
system cannot be constrained by observations. The model biases in atmospheric
CO2 also present a problem for the data assimilation system because its
optimization relies on the assumption that both model and observations are
unbiased. It is therefore imperative to remove any large biases before
assimilating observations. In this paper, we present a method to reduce the
atmospheric CO2 model biases by adjusting the CO2 surface fluxes in
a near-real-time CO2 analysis/forecasting system, such as the one used
by C-IFS at ECMWF.
Many different methods already exists to adjust CO2 fluxes by using
observations of atmospheric CO2 within flux inversion systems
. However, these
are not all suitable for the C-IFS real-time monitoring system. Flux
inversion systems adjust the fluxes by either inferring the model parameters
in Carbon Cycle Data Assimilation Systems also known as CCDAS
, or the fluxes
themselves . CCDAS has the advantage of working in
prognostic mode once the model parameters have been optimized. Nevertheless,
it can also be prone to aliasing information to the wrong model parameter
when the processes that contribute to the variability of atmospheric CO2
are not properly represented in the model or missing altogether. Estimating
directly the CO2 fluxes does not rely on the accurate representation of
complex/unknown processes in the CO2 flux model, but the resulting
optimized fluxes do not have predictive skill. Both approaches generally use
long data assimilation windows of several weeks to years in order to be able
to constrain the global mass of CO2 by relying mainly on high quality in
situ flask and continuous observations which are relatively sparse in time
and space. This general requirement for long assimilation windows is
incompatible with the current NWP framework (e.g. a 12 h window is currently
used in the C-IFS). In addition to that, the CO2 observations from flask
and most in situ stations used by these flux inversion systems are not
available in near-real time.
Considering all the aspects mentioned above, a biogenic flux adjustment
scheme (hereafter called BFAS) suitable for the NWP framework is proposed
which aims to combine the best characteristics of both flux inversion
approaches. Namely, the mass constraint from the optimized fluxes is used to
correct the biases of the modelled CO2 fluxes while keeping the
predictive skill of the modelled fluxes at synoptic scales. The main
objective of BFAS is to reduce the large-scale biases of the background
atmospheric CO2. This should improve the representation of the
atmospheric CO2 large-scale gradients, and thereby also lead to a better
forecast of atmospheric CO2 synoptic variability.
The details of the flux adjustment scheme are provided in
Sect. . Section describes the C-IFS experiments
done to test the impact of BFAS on the atmospheric CO2 forecast. From
the experiments, different aspects of the flux adjustment can be monitored
(i.e. the scaling factors and the resulting budget) as shown in
Sect. . The resulting atmospheric CO2 forecast fit
to observations after applying BFAS is presented in Sect. . The
potential use of BFAS for model development and the possibility of including
BFAS in the data assimilation system are discussed in
Sect. . Finally, Sect. gives a summary
of the flux adjustment achievements and possible developments for the future.
Methodology
Any atmospheric CO2 analysis/forecast system requires a flux adjustment of
some sort in order to constrain the budget of sources/sinks a the surface and
avoid the growth of mean errors in the atmospheric background
. The scientific question addressed in this
paper is how to use the best information we have in near-real time to adjust
the fluxes in a way that reduces the bias of the atmospheric CO2 in the
model with the minimum deterioration of the synoptic skill to predict
day-to-day variability.
documented the configuration of the CO2
forecasting system and showed that the large biases in atmospheric CO2
are consistent with errors associated with the budget of CO2 surface
fluxes, in particular the net ecosystem exchange (NEE) modelled by the
CTESSEL carbon model within the C-IFS.
There are three main reasons for modelling NEE fluxes online as opposed to
using offline fluxes such as optimized fluxes from flux inversion systems
directly in the model: (i) the coupling of CO2 biogenic fluxes with the
atmospheric model can lead to improvements in both the understanding of
interactions between ecosystems and the evolution of CO2 in the
atmospheric boundary layer and the
forecast skill of energy and water cycle fluxes in NWP models
; (ii) the use of offline fluxes would entail a
loss of information and the introduction of topographical inconsistencies
when downscaling fluxes from low resolution (e.g. typically a few degrees in
optimized fluxes) to high resolution (e.g. currently 9 km in ECMWF NWP
model); (iii) the non-availability of these offline fluxes in near-real time
implies the interannual variability of the NEE fluxes
cannot be represented.
The challenge remains of how to reduce the large-scale biases associated with
the modelled fluxes in real time. Because these biogenic fluxes are modelled
online, a one-off scaling of the fluxes using a climatology of the annual
global budget or re-scaling locally
the NEE in order to get a better fit with the seasonal cycle
are not suitable
methods, as we do not know the annual budget of the model in real-time.
Optimized fluxes from flux inversion systems constitute the best available
estimate of the CO2 fluxes given the observed variations of CO2 in
the atmosphere at global scales. Thus, they can provide a reference benchmark
for the modelled fluxes. The large-scale biases in the CO2 fluxes can be
diagnosed by computing the budget (i.e. integrated) differences between
modelled fluxes and optimized fluxes over continental and supra-synoptic
spatial and temporal scales (≥ 1000 km, 10 days). Working with budgets
over scales beyond the synoptic scale allows the detection of large-scale
biases without interfering with the synoptic skill of the model.
It is important to note that there are uncertainties and limitations that
should be considered when using optimized fluxes. Optimized fluxes are
computed with flux inversion systems at low resolutions (∼ hundreds of
km) compared to the NWP resolution used for the CO2 forecasts
(∼ tens of km), and they are most reliable at continental and
supra-synoptic scales. Moreover, they have the limitation of not being
available in near-real time, unlike the meteorological observations or
CO2 satellite retrievals . Because of that, a
climatology of the optimized fluxes has to be used as a reference.
Finally, optimized fluxes only provide information on the total CO2 flux
because flux inversion systems are not able to attribute the CO2
variability to the different processes controlling the fluxes, such as
vegetation, anthropogenic sources and fires. Generally, from all these
fluxes, the land CO2 fluxes from vegetation and soils in models are
associated with high uncertainty . For this reason,
the Global Carbon Project provides the CO2 budget from land vegetation
– also known as the land sink – as a residual to close the carbon budget
(see www.globalcarbonproject.org/carbonbudget, Le Quéré et al., 2015). Following the
land sink residual approach, the optimized NEE can be computed as the
residual of optimized fluxes by subtracting the other prescribed fluxes. A
set of 10-day mean budgets of this residual NEE from optimized fluxes is then
computed daily for different regions and vegetation types over a period of
10 years to build the NEE climatology that can be used as a reference. In
order to account for the inter-annual variability of NEE, the reference
climatology is also adjusted with an inter-annual variability factor obtained
from the model.
The flux adjustment scheme essentially estimates the bias of the modelled NEE
budget with respect to the reference NEE budget for each region and
vegetation type as a scaling factor α:
α=fOfM,
where f is the 10-day mean NEE budget computed daily over a specific
vegetation type and region, fO is the reference budget based on
the MACC-13R1 optimized fluxes , and
fM is the budget of the modelled fluxes.
Figure shows how the BFAS scheme interacts with the
model to produce the flux-corrected atmospheric CO2 forecast. First of
all, the uncorrected NEE fluxes from the model are retrieved. Then their
budget is compared with the budget of the NEE climatology from the optimized
fluxes adjusted with the NEE anomaly from the model. The scheme produces maps
with scaling factors of the biogenic fluxes before the forecast run.
Subsequently, these maps are then used to scale the forecast of NEE. There
are three major building blocks required for the computation of these scaling
factors:
the computation of the NEE budget using temporal and spatial
aggregation criteria (e.g. 10 days, vegetation types, different
regions);
a reference NEE data set used to diagnose the model biases
(e.g. optimized fluxes from global flux inversion systems
such as the MACC-13R1 data set from );
the partition of the NEE adjustment into the two modelled ecosystem
fluxes that make up the NEE flux:
i.e. gross primary production (GPP) associated with photosynthesis and
ecosystem respiration (Reco) documented
by .
These different aspects are discussed in further detail below in
Sects. to .
Schematic showing how BFAS fits in the atmospheric CO2
forecasting system. BFAS is called before each forecast to compute the
scaling factors for the model NEE (i.e. GPP + Reco) based on
the past archived forecasts. The maps of the scaling factors are then passed
to the model which applies the adjustment to the output biogenic CO2
fluxes from the land surface model. After combining the adjusted NEE fields
with the other prescribed CO2 fluxes, the resulting bias corrected
fluxes are passed to the transport model to produce the atmospheric CO2
forecast.
Computation of NEE budget
The biases of the NEE fluxes that we aim to correct are partly linked to
model parameter errors that depend on vegetation type and to errors in the
meteorological/vegetation state which are region-dependent (e.g. radiation,
LAI, temperature and precipitation). In addition to that, the global
optimized fluxes used as reference do not currently have a strong constraint
from observations at small spatial and temporal scales due to the sparse
observing network of atmospheric CO2. Therefore, the NEE biases are not
diagnosed at the model grid-point scale, but as biases in the NEE budget over
continental regions for different vegetation types and over a period of
10 days. The 10-day regional budget provides an indicator on the large-scale
biases. Moreover, 10 days is a period that can be used in the current
framework of the C-IFS global atmospheric CO2 forecasting system.
Figure shows how the uncorrected NEE from the
past forecasts can be combined to compute the 10-day mean budget before each
new forecast. The 1-day forecasts initialized from the previous seven days
are used together with the last 3-day forecast available in order to create a
10-day window around the initial date of the new forecast. This 10-day time
window is slightly shifted to the past because otherwise forecasts longer
than 3 days would be required to compute the budget, while errors in the
meteorology affecting the fluxes grow with forecast lead time.
found that forecast errors associated with the
location of extra-tropical weather systems affecting the cloud cover and
temperature gradients – which in turn will affect the NEE errors – are very
small at day 1. These errors continue to be small up to day 3, but they can
grow rapidly with forecast lead time seefor details on the IFS
forecast error evaluation. The different regions have been
selected according to latitudinal band characterized by seasonal cycle
(Northern Hemisphere, tropics, and Southern Hemisphere), continental region,
and vegetation type.
Schematic to illustrate how the 10-day NEE budget from the model is
computed in BFAS for the forecast at day D by retrieving the past forecasts
of accumulated NEE. Note that the retrieved NEE (computed by adding GPP and
Reco) has not been corrected by BFAS. The computation uses a set
of 7 previous 1-day forecasts (initialized at D-8, D-7, D-6, …
until D-2) together with the latest 3-day forecast from the previous day
(i.e. D-1) as shown by the blue boxes.
In the C-IFS the vegetation types follow the BATS classification
, which is widely used in meteorological and
climate models. The vegetation classification is designed to distinguish
between roughness lengths for the computation of the momentum, heat and
moisture transfer coefficients in the modelling of the fluxes from surface to
atmosphere. However, the BATS vegetation types are not always suitable for
the modelling of the CO2 fluxes. For example, the interrupted forest
type which constitutes around 25 % of the high vegetation cover
encompasses many different types of vegetation, including tropical savanna
and a combination of remnants of forest or open woods lands with field
complexes. This could be an important source of error in some regions. For
this reason, BFAS allows the introduction of new vegetation types for
diagnosing the NEE biases. Tropical savanna, which covers large areas in the
tropical region, has been added as a subtype of the interrupted forest
vegetation type by using the Olson Global Ecosystem classification
https://lta.cr.usgs.gov/glcc/globdoc2_0.
Figure shows the distribution of the
dominant vegetation types used in BFAS. Land cover maps from GLCC version 1
(https://lta.cr.usgs.gov/glcc) are used to compute the land cover
of the dominant high and low vegetation types at each grid point. In BFAS,
only one dominant vegetation type is used to classify each grid point, and
this must cover more than 50 % of the grid box. Model grid points with
less than 50 % vegetation cover are not used. The comparison of the
modelled NEE with the optimized NEE fluxes is done by computing 10-day
budgets for each of the 16 vegetation types (see Table )
and 9 different regions (see Fig. ).
Dominant vegetation types based on the BATS classification used in
the C-IFS and extended to include the tropical savanna subtype (in purple, as
defined by the classification) within the interrupted
forest type (in light blue). The vegetation type codes are described in
Table . The nine regions used in the computation of the
NEE budget are delimited by the black lines.
Reference NEE budget
The residual NEE from optimized fluxes provides the reference for the flux
adjustment scheme. Currently, there is no operational centre providing
CO2 optimized fluxes at global scale in near-real time. We have chosen
to use the MACC optimized fluxes which are
delivered around September each year for the previous year. The MACC
optimized CO2 fluxes are regularly improved and their high quality has
been recently shown by .
provides an evaluation of the inverted CO2 fluxes for 2010.
The computation of the residual is done by subtracting the prescribed fluxes
used in the C-IFS CO2 forecast over land from the total optimized flux.
The prescribed CO2 fluxes from biomass burning and anthropogenic
emissions in the CO2 forecast are not the same as the ones used as prior
fluxes in the MACC flux inversion system. Not only they are from different
sources, but they are also used at different resolutions. This means that
there might be fires represented in one and not the other, or with different
emission intensities, as it is the case for anthropogenic hotspots at high
vs. low resolutions. Thus, in order to avoid the transfer of
inconsistencies between the prescribed and prior fluxes into the NEE
residual, the regions with very high anthropogenic emissions (larger than 3×106 g C m-2 s-1) and fires are filtered out.
A climatology of these reference NEE fluxes is created using the last 10 available years
and it is updated every time a new year is available. Thus, allowing for slow
decadal variations in the NEE reference. Figure shows a
comparison of the optimized flux budget in 2010 and its climatology for the
crop vegetation type in North America. The inter-annual variability of the
optimized flux budget is depicted by the standard deviation around the
10-year climatological mean value. The reference NEE climatology is then adjusted to
account for the inter-annual variability of the land sink fluxes as
follows:
Percentage of land grid points at model resolution TL255
(∼ 80 km) for each dominant vegetation type, i.e. more than half of
the grid point is covered by that vegetation type. A land grid point is
defined by a land sea mask value greater than 0.5.
Vegetation
Vegetation type
Percentage of
code
land points
1
Crops, mixed farming
9.9
2
Short grass
7.6
7
Tall grass
6.3
9
Tundra
6.3
10
Irrigated crops
2.2
11
Semidesert
13.5
13
Bogs and marshes
0.8
16
Evergreen shrubs
0.5
17
Deciduous shrubs
2.4
3
Evergreen needle leaf trees
5.7
4
Deciduous needle leaf Trees
2.4
5
Deciduous broadleaf trees
4.0
6
Evergreen broadleaf trees
12.1
18
Mixed forest/woodland
3.3
19
Interrupted forest
9.5
21
Tropical savanna (new type)
4.8
–
Remaining land points without vegetation
8.7
fO=fOclim+γσfOclim,
where f is the 10-day NEE budget for a specific region and vegetation
type, fO is the reference budget, fOclim and
σ(fOclim) are the climatological mean and standard
deviation of the optimized flux budget, respectively, from 2004 to 2013, and
γ is the corresponding standardized anomaly of the NEE budget from the
model with respect to the same period. γ can be positive or negative.
It represents the inter-annual variability factor used to adjust the
reference climatological NEE budget and it is given by
γ=fM-fMclimσfMclim,
where fM is the model NEE budget, fMclim is the
climatological mean budget from the model and σ(fMclim) is
the standard deviation of the model NEE budget denoting the typical amplitude
of its inter-annual variability for the same period as the climatology of the
optimized flux budget (i.e. 2004 to 2013).
Time series of 10-day mean NEE budget (GtC day-1) associated
with the crop vegetation type in North America from the MACC-13R1 optimized
flux data set in 2010 (red line) compared to its climatology (2004–2013)
(yellow line). The yellow shading represents the standard deviation of the
optimized flux budget (for the same period) used to compute the inter-annual
variability adjustment applied to the reference climatology.
Positive/negative values correspond to a source/sink of CO2.
The γ inter-annual variability factor is multiplied by the standard
deviation of the optimized residual NEE budget – representing the typical
amplitude of inter-annual variability – in order to offset the reference
climatological NEE budget. In this way, the inter-annual variability of the
reference NEE follows the inter-annual variability of the model NEE with the
same anomaly sign, while keeping its amplitude constrained by the standard
deviation of the optimized flux budget.
Note that the use of this factor is optional. By setting it to zero,
the model budget can be constrained by the optimized flux climatology.
The rationale for applying this factor in the C-IFS system is based on
the fact that inter-annual variability of the NEE budget
is strongly linked to the inter-annual variability of climate variables
such as precipitation and temperature . Since information
on these climate variables is readily available in the C-IFS system, it is worth
exploring its impact on the CO2 forecast.
A preliminary assessment of the impact of including the inter-annual variability
factor was performed by comparing experiment with and without the factor.
Results confirmed a small but positive impact (see Supplement).
Details on the computation of this factor are given in the next section.
The inter-annual variability factor
The computation of the inter-annual variability factor γ
requires a model climate consistent with the
forecast (i.e. same meteorological analysis, same model version and same
resolution). Producing a consistent model climate is not a trivial
requirement, because both the operational model version and analysis system
can change frequently with new updates and new observations, and high-resolution forecasts spanning a period of 10 years (i.e. 2004 to 2013) are
expensive. A feasible solution has been found where the standardized NEE
anomaly from the model is computed using the operational ensemble prediction
system (ENS) forecasts and hindcasts which are part of the ECMWF monthly
forecasting system . Every
Monday and Thursday the operational ENS is not only run for the actual date,
but also for the same calendar day of the past 20 years. These hindcasts have
the same resolution and model version as the ENS forecasts and they
constitute a valuable data set used for the post-processing
of the NWP forecasts from the medium-range (10 days) up to 1-month lead
times . The ensemble of forecasts is made of
5 members (10 members since 2015) using perturbed initial conditions
and stochastic physics in order to represent forecast
uncertainty .
As the hindcasts are not performed daily, it is not possible to aggregate
consecutive 1-day forecasts into a 10-day period to compute a mean budget as
shown in Fig. . In order to circumvent this, the
mean budget is computed by averaging the 1-day forecast NEE from all the
ensemble members available in the hindcasts. This is done for each year from
2004 to 2013 to preserve consistency with the NEE climatology from the
optimized fluxes. The model climate fMclim given by the 10-year
mean budget and its typical inter-annual variability σfMclim can then be obtained by calculating the mean
value and standard deviation, respectively, over that period. Similarly, the
model budget fM is calculated from the NEE ensemble mean of the
ENS forecast for the current date using the same number of ensemble members
as the ENS hindcasts. The standardized anomaly γ is finally obtained
by subtracting the 10-year mean budget from the current budget and dividing
the anomaly by the standard deviation. Since the hindcasts are available
every Monday and Thursday, γ is only updated twice a week. These
updates are routinely monitored during the forecast (see
Sect. ).
Time series of GPP and Reco flux scaling factors in blue
and red lines, respectively, for the crop vegetation type in 2010 in the
different regions (see map in Fig.
depicting the extent of the crops within each region).
Partition of NEE adjustment
The final stage in the flux adjustment is the attribution of the NEE
correction to the different biogenic fluxes in the model. The residual NEE
from optimized fluxes only provides information on the total flux from the
land ecosystem exchange. While in land vegetation models, NEE is the
combination of two opposing fluxes: gross primary production (GPP) and the
ecosystem respiration (Reco). Given that we have no information
on whether the NEE error is associated with the GPP or the Reco
fluxes, a strategy has to be defined in order to partition the NEE correction
into GPP and Reco. The underlying strategy used here is to have
the smallest flux adjustment possible. Namely, the scaling factors should be
as close to 1 as possible.
The first step is to distinguish between the positive and negative values of
the NEE scaling factor (α). A positive NEE scaling factor implies the
budget of the NEE in the model has the correct sign but the wrong magnitude.
In that case, the scaling of the flux will be smallest if the dominant
component of NEE is scaled. That is to say, the flux correction will be
applied to GPP during the growing season and to Reco during the
senescence period. Whereas if the scaling factor is negative – i.e. the
modelled NEE has the wrong sign – only the flux with smallest magnitude is
corrected (GPP or Reco) to ensure the scaling factor of the
modelled fluxes is always positive.
The scaling factor α is then converted into a scaling factor for the
dominant component of the NEE flux. If the magnitude of GPP is larger than
the magnitude of Reco, then the scaling factor for GPP and
Reco are defined as follows:
αGPP=αNEE-RecoGPPαReco=1.0.
Similarly, if
|Reco|>|GPP| then
αGPP=1.0αReco=αNEE-GPPReco.
This partition of the flux adjustment is a modelling choice based on
minimum flux adjustment criteria. Other solutions might be possible given
additional information on either GPP or Reco budgets.
The αGPP and αReco factors are computed
for each vegetation type and region and then re-mapped as two-dimensional fields using
the dominant vegetation type map in Fig. .
The resulting maps for αGPP and αReco
are subsequently passed to the carbon module in the land surface model in
order to scale GPP and Reco.
Impact of the flux adjustment
The impact of BFAS is shown by comparing the atmospheric CO2 from the
BFAS forecast to the CTRL forecast, and to the benchmark forecasts with
optimized fluxes (OPT and OPT-CLIM) at several observing sites. Four sites
from the NOAA/ESRL atmospheric baseline observatories
www.esrl.noaa.gov/gmd/obop, are used to
evaluate the reduction of the large-scale biases in the well-mixed background
air. In addition, four Total Carbon Column Observing Network stations
GGG2014 TCCON data,see Table and
www.tccon.caltech.edu are also used to assess the
impact on the atmospheric CO2 column-average dry molar fraction.
Finally, three continental sites from the NOAA/ESRL tall tower network
www.esrl.noaa.gov/gmd/ccgg/towers, are used
to investigate the impact of BFAS on the synoptic skill of the forecasts. The
results are grouped into the impacts on bias reduction and synoptic skill in
the following two sections. A comprehensive evaluation of the uncertainty
reduction in the BFAS simulation based on all the
in situ flask and continuous observations, as well
as the NOAA/ESRL aircraft vertical profiles is also
provided in the Supplement.
Time series of the standardized anomaly of the modelled NEE budget
(γ in Eq. ) for crops in 2010 in the different
regions. Positive values indicate larger/smaller CO2 sources/sinks than
normal based on the mean climatological budget; whereas negative values
correspond to smaller/larger CO2 sources/sinks than normal.
Biases in atmospheric CO2
Figure demonstrates that BFAS is very effective at
reducing the atmospheric CO2 biases in the background air at all the
NOAA/ESRL continuous baseline stations. The biases in the CTRL forecast
range from -1.9 to -4.5 ppm; whereas, the BFAS forecast
has biases of -0.5 ppm or less over the whole year. These values are
close to the annual biases of the OPT and OPT-CLIM experiments ranging
between -0.4 and 0.5 ppm. The monthly biases in BFAS can be larger than
its annual biases. For example, there is a bias of up to -1 ppm from March
to September in the Southern Hemisphere (Fig. c, d).
This bias is thought to originate in the tropical regions and
transported to the Southern Hemisphere as shown by a preliminary comparison
with IASI CO2 (not shown here).
The bias starts to grow at the end of the growing season during summer time.
This is also the case for the high latitude station at Barrow, where there is
a negative bias of a few ppm from the last week of July to the end of
September as shown in Fig. a. In summary, BFAS is not
able to completely remove the negative model bias at the end of the growing
season. In the Northern Hemisphere at the end of winter and throughout spring
(from March to May) there is a positive model bias, i.e. the atmospheric
CO2 is overestimated in the model. Although the OPT and OPT-CLIM
simulations also have a slight positive bias in winter, this positive bias is
enhanced in the BFAS simulation.
At the TCCON sites (Fig. ), the atmospheric CO2
column-average dry molar fraction also shows the same large bias reduction in
BFAS with respect to CTRL. The magnitude of the BFAS annual biases in the
atmospheric column is generally less than 1 ppm, slightly higher than the
OPT and OPT-CLIM biases (less than 0.5 ppm), but much lower than the CTRL
biases (from 1.5 to 3.3 ppm). The results at the TCCON sites are consistent
with those from the NOAA/ESRL baseline sites. Namely, in the Northern Hemisphere there is a growing overestimation of the atmospheric CO2 at
the end of winter (around March). While at the end of the growing season in
both the Northern Hemisphere and the Southern Hemisphere (August and March, respectively) there
is a growing negative bias, i.e. an overestimation of the sink. One
hypothesis that could explain why BFAS is not able to achieve as small a bias
as the forecast with optimized fluxes lies in the fact that the optimized NEE
used as a reference in BFAS is computed as a residual after removing the
effect of fires and anthropogenic fluxes. Inconsistencies in the fire and
anthropogenic emissions used by the optimized fluxes and the model will lead
to errors in the optimized residual NEE. These inconsistencies are mainly
associated with the use of different resolutions. Further investigation is
required to address this issue.
Correlation coefficient of different forecast (FC) experiments
(see Table ) with observations at three NOAA/ESRL tall towers
for daily mean dry molar fraction of atmospheric CO2 in March 2010.
The dash symbol means the correlation is not significant.
NOAA/ESRL
Latitude,
Sampling
BFAS
CTRL
OPT
OPT-CLIM
Tower site
Longitude,
level
FC
FC
FC
FC
(ID)
Altitude
(m)
Park Falls,
45.95∘ N,
30
0.843
0.338
0.794
0.797
Wisconsin
90.27∘ W,
122
0.931
0.508
0.893
0.883
(LEF)
472 m
396
0.919
–
0.875
0.881
West Branch,
41.72∘ N,
31
0.748
0.496
0.590
0.590
Iowa
91.35∘ W,
99
0.833
0.436
0.767
0.720
(WBI)
242 m
379
0.851
0.356
0.887
0.876
Argyle,
45.03∘ N,
12
0.857
0.839
0.808
0.893
Maine
68.68∘ W,
30
0.875
0.835
0.816
0.938
(AMT)
50 m
107
0.861
0.668
0.816
0.927
Daily mean atmospheric CO2 dry molar fraction (ppm) from
NOAA/ESRL continuous baseline stations (black circles) at
(a) Barrow, Alaska, USA (71.32∘ N, 156.61∘ W),
(b) Mauna Loa, Hawaii, US (19.54∘ N, 155.58∘ W),
(c) Tutuila, American Samoa, USA (14.25∘ S,
170.56∘ W), (d) South Pole, Antarctica (89.98∘ S,
24.8∘ W) and the different forecast experiments: BFAS (cyan), CTRL
(red), OPT (green) and OPT-CLIM (blue). See Table for a
description of the different experiments. The mean (bias) and standard
deviation (SD) of the model errors are shown at the top of each panel.
Synoptic variability of atmospheric CO2
The C-IFS CO2 forecast has been shown to have high skill in simulating
the synoptic variability of atmospheric CO2
see, except during the spring months,
coinciding with an early start of the CO2 drawdown period in the model.
For this reason, we have examined the impact of BFAS on the synoptic
variability of daily mean atmospheric CO2 at three continental NOAA/ESRL
tower sites in March. Over this period, the day-to-day variability of
atmospheric CO2 at those sites is associated with the advection of
atmospheric CO2 by baroclinic synoptic weather systems as they impinge
on the large-scale continental gradient of atmospheric CO2.
Table clearly demonstrates that with BFAS the synoptic
forecast skill is greatly improved at all sites, with correlation
coefficients between simulated and observed atmospheric CO2 exceeding
0.8. The improvement is particularly striking at Park Falls (Wisconsin, USA)
and West Branch (Iowa, USA) at the centre of North America, where the
correlation coefficients in CTRL are very low (i.e. below 0.5). The OPT and
OPT-CLIM forecasts have generally high correlation coefficients, comparable
to BFAS. Only at the level closest to the surface, the values are slightly
lower than BFAS. This can be explained by the fact that the MACC-13R1
optimized fluxes do not comprise synoptic variability. Thus, when the
synoptic variability of the fluxes contributes to the atmospheric CO2
variability, the correlation coefficients are smaller.
The positive impact of BFAS on the CO2 synoptic variability is
illustrated in Fig. . The large synoptic
variability is characterized by the advection of CO2-rich anomalies
(with up to 10 ppm amplitude) as shown by the CO2 peaks on 10–12 March
at Park Falls, and 8–9, 12–13 and 16–17 March at West Branch. These
CO2 anomalies originate from the advection across the large-scale
continental gradients of atmospheric CO2 which ultimately reflect the
large-scale distribution of CO2 surface fluxes
. In the case study here, the CO2-rich air
is located to the south of the observing stations, as shown by the
distribution of the monthly mean atmospheric CO2 depicting the
large-scale gradients across the continent at the level corresponding to the
height of the tall towers (Fig. a and b). In the CTRL
forecast, there is no monthly mean gradient south of the stations
(Fig. c). This explains why without BFAS the synoptic
variability is very small and largely underestimated throughout March. While
in BFAS the gradient south of the observing stations is very pronounced
(Fig. d), following a similar pattern to OPT and OPT-CLIM.
There are still some differences between the three simulations. OPT-CLIM
results in stronger gradients than OPT and BFAS enhances the gradient even
further, leading to a slight over-estimation of the synoptic variability.
These differences in the patterns of the atmospheric CO2 are directly
linked to the differences in the CO2 surface fluxes
(Fig. ). As expected, the flux adjustment from BFAS results
in a flux pattern similar to OPT-CLIM and OPT, with a stronger source to the
south of the observing stations. Whereas in CTRL there is a large sink area
south of the observing stations, in the region of the Gulf of Mexico,
consistent with the CTESSEL early growing season .
Discussion
All the results from the BFAS experiments indicate that BFAS is highly
beneficial to the C-IFS CO2 forecasting system, both in terms of reducing
the atmospheric CO2 biases and improving the synoptic skill of the
model. As shown in Sect. , the scheme is simple and it is easy
to implement and run. Because BFAS essentially works as a layer on top of the
model, it can adapt to model changes with great flexibility. For all these
reasons, BFAS is now part of the operational global C-IFS analysis and
forecasting system.
Notwithstanding all the advantages of BFAS listed above, there are also
caveats that need to be considered, further tested, and addressed. A
discussion of the current limitations of BFAS is provided in this section,
together with the potential use of BFAS for model development, data
assimilation purposes, and the implications for users.
Current limitations in BFAS
Optimized fluxes have uncertainties of their own and represent the
large-scale variability of the CO2 surface fluxes on supra-synoptic
time-scales. They only estimate the total flux and the NEE residual approach
can transfer biases from other fluxes into the NEE. The use of a climatology
also precludes the correction of the inter-annual variability in the model.
Daily mean atmospheric CO2 column-average dry molar fraction
(ppm) observed at four TCCON stations (see Table ) as shown by
the black circles, and simulated by the different forecast experiments: BFAS
(cyan), CTRL (red), OPT (green) and OPT-CLIM (blue). See Table
for a description of the different experiments. The mean (δ) and
standard deviation (σ) of the model errors are shown at the top of
each panel.
Daily mean atmospheric CO2 dry molar fraction (ppm) in
March 2010 from NOAA/ESRL tall towers (black circles) at (a) Park
Falls (Wisconsin, USA, 45.95∘ N, 90.27∘ W) and
(b) West Branch (Iowa, USA, 41.72∘ N, 91.35∘ W)
and the different forecast experiments: BFAS (cyan), CTRL (red), OPT (green)
and OPT-CLIM (blue) (see Table for a description of the
different experiments).
The aggregation criteria of budget errors can be very challenging because the
error can originate from different aspects of the model. Clearly, errors in
model parameters associated with vegetation type are a good candidate.
However, in the future errors in climate forcing, errors in LAI, missing
processes and other potential sources of error should also be considered.
The partition of the NEE flux adjustment into the modelled biogenic fluxes
(GPP and Reco) is currently ad hoc, leading to the transfer of
errors from GPP to Reco and vice-versa. This problem could be
addressed by using other independent data sets of GPP and Reco
e.g. that contain additional information on how to
partition the NEE adjustment.
BFAS for model development
BFAS can run in both online and offline modes. Thus, it can provide a tool to
diagnose regions that contribute to the errors in the global budget resulting
in large-scale errors of atmospheric CO2. The maps of biogenic flux
scaling factors can be used to compute maps of flux adjustment (e.g. adjusted
NEE – original NEE) which can then be used to diagnose model errors. The
synthesis of the mean adjustments into monthly model biases for different
vegetation types can then guide the effort to develop the carbon model
further. For example, in regions where the bias is consistent between
different months, the corrected NEE could be used to re-tune model parameters
such as the reference ecosystem respiration or the mesophyll conductance,
previously optimized by using a subset of FLUXNET
data. Specific vegetation types can be identified where model improvements
could be achieved by using information from BFAS. For instance, crops have
the same large Reco scaling (>1.5) over all the Northern Hemisphere regions during winter months when the ecosystem respiration is the
dominant component of NEE. This underestimation in the ecosystem respiration
can be addressed by modifying the value of the reference respiration
parameter used for crops. In this case, the same procedure used by
could be applied to optimize the specific model
parameter using the BFAS adjusted fluxes as pseudo-observations together with
the FLUXNET data.
Further information on error sources in fluxes can be obtained by comparing
the corrected fluxes with the eddy covariance observations available in
near-real time from the Integrated Observation System (ICOS) Ecosystem
Thematic Centre (ETC, http://www.europe-fluxdata.eu). For example,
preliminary comparisons have shown that there are large differences in the
model-observation fit between needle leaf evergreen (pine) trees in the
boreal and Mediterranean regions. This is consistent with results from
, and it highlights the need for a new
sub-classification of the evergreen needle leaf forests in regions with
Mediterranean climate.
Monthly mean atmospheric CO2 dry molar fraction (ppm) at the
model level approximately corresponding to the highest sampling height of the
Park Falls and West Branch NOAA/ESRL tall towers (see black triangles) in
March 2010 from (a) OPT-CLIM, (b) OPT, (c) CTRL
and (d) BFAS experiments (see Table for a description
of the different experiments).
Monthly mean total CO2 flux (µmol m-2 s-1)
in March 2010 from (a) OPT-CLIM, (b) OPT, (c) CTRL
and (d) BFAS experiments (see Table for a description
of the different experiments). The black triangles depict the location of the
NOAA/ESRL tall towers plotted in Fig. .
BFAS in the data assimilation framework
Currently, BFAS is only designed to be used as a bias correction computed
before each forecast by using a reference data set based on optimized fluxes.
In the future, BFAS could be adapted to work within a data assimilation (DA)
framework in the C-IFS. To start with, the specification of uncertainties
associated with both the reference data set and the model fluxes and the
covariance of those uncertainties would allow a more optimal estimation of
the flux adjustment. These uncertainties can be obtained from the flux
inversion systems for the optimized fluxes and from the ECMWF ENS forecasts
for the model fluxes.
Including BFAS in the C-IFS DA framework needs further exploration. The C-IFS
uses a short time window (currently 12 h) to assimilate meteorological
observations from very dense observing networks. With the short time window
it is not possible to properly constrain the slowly varying global mass of
the long-lived greenhouse gases due to the sparseness of their observing
system. For instance, the current GOSAT and OCO-2 CO2 observations do
not cover high latitudes in winter. However, if we combined the assimilation
of optimized fluxes (which already contain the global mass constraint) with
observations linked to local fluxes (e.g. solar-induced chlorophyll
fluorescence products from satellites, NEE eddy covariance observations and
in situ atmospheric CO2 observations) it might be possible to obtain an
optimal estimate of more local scaling factors, while still respecting the
global mass constraint. The possibility of optimizing the scaling factors in
the DA system within the weak constraint framework
also needs to be explored in the future.
Aspects to be considered by users
The implementation of BFAS is straightforward. Therefore, it could be easily
used by other models. The only requirements are: (i) a reference budget which
can be obtained from a climatology of optimized fluxes (e.g. the MACC product
can be easily obtained from
www-lscedods.cea.fr/invsat/PYVAR14_MACC/V2/Fluxes/3Hourly and it is
well documented); (ii) past 10-day NEE simulated by the forward model;
(iii) the NEE anomaly of the forward model with respect to its climate based
on a 10-year simulation. The use of the NEE anomaly is optional, as its
impact is relatively small (see Supplement).
The underlying motivation of BFAS is to improve the CO2 analysis and
forecast for users (e.g. those working on flux inversion systems, planning
field experiments, or requiring boundary conditions for regional models). For
this reason, it is paramount to provide information on all the input data
going into BFAS. These are primarily continental-scale climatological budgets
from modelled NEE and optimized fluxes. There is also some input from the
anthropogenic emissions and the biomass burning emissions to extract the NEE
as a residual from the optimized fluxes. The documentation of the specific
components used in the C-IFS BFAS system and their uncertainties can be found
in , , ,
and . The input data streams used in
BFAS can be obtained from http://copernicus-support.ecmwf.int for C-IFS
NEE and GFAS biomass burning fluxes; from the EDGAR database
http://edgar.jrc.ec.europa.eu for the anthropogenic fluxes; and from
www-lscedods.cea.fr/invsat/PYVAR14_MACC/V2/Fluxes/3Hourly for the MACC
optimized fluxes.
Since the BFAS product contains information from the optimized fluxes, users
should be aware that the optimized fluxes assimilated most available
background air-sample monitoring sites (listed in the supplement of
, see
https://acp.copernicus.org/articles/15/11133/2015/acp-15-11133-2015-supplement.pdf).
A specification of the overall uncertainty associated with the BFAS
simulation and the resulting reduction with respect to the control simulation
is given in the Supplement.
Summary
This paper addresses the challenge of designing an online bias correction for
an atmospheric CO2 analysis/forecasting system. The overarching aim is to
deliver an atmospheric CO2 analysis and forecast that can be useful to the
scientific community, e.g. working on data assimilation of atmospheric CO2
observations, the development of the CO2 observing system and providing
boundary conditions for CO2 regional modelling. Tuning model parameters
and/or re-scaling fluxes offline are not sufficient to guarantee a bias
reduction in the system. Thus, an online adaptive system is required because
errors in the meteorology can evolve as a result of regular operational NWP
model upgrades and these affect the NEE budget in the model. This is achieved
in the new biogenic flux adjustment scheme (BFAS) by a simple scaling of the
10-day NEE budgets for different vegetation types and regions using a
climatology of the MACC optimized fluxes as a
reference, adjusted to preserve the model inter-annual variability.
This paper shows that BFAS has a positive impact on the atmospheric CO2
forecast by greatly reducing the atmospheric CO2 biases in background
air and improving the synoptic variability in continental regions affected by
ecosystem fluxes. The improvement in the synoptic skill of the forecast is
associated with underlying changes in the large-scale gradient of the NEE
fluxes where optimized fluxes provide information.
BFAS has been recently implemented in the C-IFS operational CO2 forecast
and analysis system, because of its simplicity, adaptability to model changes
and beneficial impact.
In this paper, the C-IFS model is just providing an example on how this method can be
applied efficiently in an operational forecasting system. Other models could
easily adopt such a system as there are only a few components required for
its implementation (see Sect. ).
As a diagnostic tool, BFAS has also the potential to provide feedback for
model development. The use of BFAS in the data assimilation framework will be
explored in the future.