Introduction
The forests of Amazonia cover 6.77 millionkm2
. It is the world's largest continuous area of
tropical forest and reservoir of aboveground organic carbon
. Changes in the carbon dynamics of this ecosystem
thus have global significance . However, the natural
variability of CO2 exchange in Amazonia, as well as its short-
and long-term response to natural and anthropogenic disturbance across
scales, is still poorly understood and a topic of active
research.
There is intense debate about the timing and magnitude of the seasonal
cycle of CO2 fluxes across Amazonia. Studies employing remote
sensing data as a proxy for canopy photosynthetic activity have
suggested a widespread enhancement of gross primary productivity
of the Amazonian rainforest during the dry season
. Yet direct and continuous measurements of net
ecosystem exchange (NEE) between the atmosphere and forest canopy at
a local scale (from 1 ha to 1 km2 scale) based on
eddy-covariance (EC) systems do not support such large-scale
behaviour. Several EC observations in central eastern Amazonia
and north-eastern Amazonia also
indicate that tropical forest areas take up CO2 during the dry
season, but similar EC studies in central Amazonia have suggested an
opposite seasonality . Finally, remote
sensing measurements of the vertically integrated columns of
CO2 (XCO2) retrieved from the GOSAT satellite suggest
stronger CO2 uptake during the wet season in southern
Amazonian forest than during the dry season . These
measurements thus reveal a large heterogeneity in space of the phase
of the seasonal cycle of NEE within Amazonia. However, most dynamic global vegetation model (DGVM) simulations predict stronger uptake
during the wet season throughout Amazonia
, although
limitations related to mortality or land use restrict the ability
of these generic global models to simulate CO2 fluxes and
carbon stocks of Amazonian forest .
Uncertainty associated with potential spatial heterogeneity is also
apparent in the estimates of the interannual variability (IAV) of
CO2 fluxes in Amazonia in particular during years with
extreme climatic conditions. Remote sensing observations during the
severe Amazonian drought of 2005 suggested a widespread enhancement of
photosynthetic activity, or greening, across Amazonia
. The resilience of forests to water stress
suggested by the “drier-yet-greener” papers was originally
attributed to a combination of deep rooting, hydraulic redistribution
and more available solar radiation . However, the
validity of enhanced vegetation index satellite data has been
recently challenged by and by losses in canopy
functioning detected in radar-based measurements .
The observations from optical satellite sensors remain controversial
because other studies did not find such an impact of droughts on
Amazonian forest . Moreover,
observations of microwave backscatter from QuickSCAT have suggested
large-scale persistent negative effects of the drought of 2005 on
forest canopy structure . Biometry measurements,
consisting of periodic measurements of the allocation of photosynthetic
products to wood growth, provide another perspective on the effects of
drought on Amazonian forest trees. In a large-scale, long-term biometric
study, found a reversal of the carbon sink due to
the effect of the drought of 2005 on tree mortality. This is consistent
with a synthesis of yearly estimates of natural fluxes (NEE plus
biomass-burning emissions) from an ensemble of DGVMs compiled at
http://www.globalcarbonatlas.org.
The scientific community has used atmospheric inversions for more than 2
decades in an effort to improve the knowledge of CO2 fluxes at a large
scale. Whereas EC or biometric studies give flux estimates that are valid at
the
local scale , atmospheric inversion offers the possibility
to derive measurement-based estimates for the whole of Amazonia, with spatial
resolutions larger than 500 km, provided that atmospheric
observations can adequately sample the Amazonian flux signal. Inversions use
available measurements of atmospheric CO2 to provide corrections to
prior surface flux estimates using an atmospheric transport model and
statistical inversion methods. The method estimates statistically optimal
fluxes within the boundaries of uncertainties in the measurements, the
transport model and prior flux estimates . The
flux corrections spread beyond the vicinity of the measurement footprint, as
defined by the transport model, through hypotheses on the spatial and
temporal correlation of the uncertainties in the prior fluxes. We define,
hereafter, the tropical South America (TSA) region as the continental land
encompassed between 16.25∘ N–31.25∘ S and
84.38–28.18∘ W, which covers the whole Amazonian forest.
show that the different inverted seasonal cycles and IAVs
of natural CO2 fluxes from several state-of-the-art global
atmospheric inversions are characterized by a large scatter over a very
similar tropical area of South America. This is explained by the variety of
prior estimates used by the different global inversion systems and by the
large-scale corrections that are applied in regions poorly covered by
observation networks, such as TSA, in order to balance the global CO2
budget rather than to match local measurements. For these reasons,
atmospheric inversions have not been included in the review of the carbon
cycle in South America made by . and
applied the principle of atmospheric inversion to exploit
vertical CO2 profile data from airborne measurements in Amazonia.
Their studies, based on measurements near Manaus in central Amazonia
and Santarém in eastern Amazonia ,
constitute important efforts to constrain surface CO2 fluxes at
regional scale, measuring and exploiting some of the few atmospheric data
sets available for South America. Their results suggested CO2 efflux
from the ecosystem during the wet season in eastern Amazonia. By analysing
vertical CO2 profiles collected approximately every 2 weeks over
the period 2010–2011, the recent study of provided
a basin-scale picture that not only confirms this regional signal but also
suggests an opposite pattern in southern and western Amazonia. Their study
reported on the first data-driven estimate of CO2 fluxes for the
whole Amazon basin and it provides insight into the sensitivity of this
important ecosystem to moisture stress. It suggests the importance of
conducting such estimates over longer time periods.
Location of the surface stations used in this study. Blue indicates
surface stations used in MACCv10.1; red shows the surface stations in South
America added to the previous setup of MACCv10.1. Filled circles are
stations with continuous measurements; open circles are sites with
discrete air sampling.
Our goal here is to study the seasonal cycle and IAV of NEE over Amazonia
during 2002–2010. This period offers the opportunity to investigate
significant anomalies in the interannual variability of carbon fluxes,
particularly those associated with the severe droughts of 2005 and 2010 and
the extreme rainfall registered across the Amazon basin in 2009
. The study is based on the global Monitoring of Atmospheric Composition and Climate (MACC) inversion system
initially described by (hereafter CH2010). We used
version 10.1 of the MACC CO2 inversion product released in
August 2011. We also use a similar inversion in which we add four
ground-based atmospheric measurement sites surrounding the north-east of
Amazonia to the assimilated data (Fig. 1). Despite the limitations of the
state-of-the-art global inversion approach in South America, highlighted
above and by , our analysis of these MACC inversions can
help characterise the temporal variations in the NEE over Amazonia for
several reasons. First, it relies on a detailed evaluation of the inversion
results over and within this region, hoping that some reliable inversion
patterns can be isolated. Such a detailed evaluation has not been conducted
in the above-mentioned intercomparisons of the global atmospheric inversions
in TSA. It makes sense to conduct it here on the MACC inversions since the
MACC system uses a variational inversion which solves for the fluxes at
∼ 3∘ and 8-day spatial and temporal resolution. Second, the use
of the stations located in the region can strengthen the robustness of the
inversion results through a significantly increased sampling of the
atmospheric signature of the fluxes in Amazonia. In particular, we are the
first to use continuous measurements from French Guyana. The assessment of
the impact of these stations on the inverted NEE (based on the comparison
between our different MACC inversions with and without these stations) can
help identify the reliable patterns of the inversion.
The rest of this paper is structured as follows. We present each component of
the standard MACCv10.1 inversion setup and the use of the additional sites
around Amazonia in Sect. 2. The results of the inversions, with a focus on
the impact of these additional sites, and their comparison to an independent
flux estimate are presented in Sect. 3. In Sect. 4, we discuss the results
and conclude the study.
The inversion method
This study builds on MACC, the global atmospheric inversion framework (whose
first version is described in detail in CH2010), to correct a prior estimate
of NEE from the model ORCHIDEE Organizing Carbon and Hydrology in
Dynamic Ecosystems, and of ocean fluxes, based on the
assimilation of in situ measurements of atmospheric CO2 mole
fractions into a global atmospheric transport model. The approach relies on
a Bayesian framework to estimate the conditional probability of the “true”
NEE and ocean fluxes given the statistical information from the prior fluxes
and the set of in situ measurements of atmospheric CO2 (hereafter
observations). Assumption of unbiased Gaussian distribution of the
uncertainties in the prior fluxes and of those underlying the simulation of
the observations using the transport model allows us to derive an updated
estimate of NEE and ocean fluxes (hereafter the posterior fluxes) that also
has an unbiased Gaussian distribution. The statistically optimal fluxes
(i.e. the mean of the posterior distribution of the fluxes) are found by
calculating the minimum of the cost function :
J(x)=(x-xb)TB-1(x-xb)+(yo-H(x))TR-1(yo-H(x)),
where x is the control vector and mainly denotes the NEE
(defined as the difference between the gross CO2 uptake
through photosynthesis and output through total ecosystem
respiration) and air–ocean exchanges that are optimized at a chosen
spatial and temporal resolution. xb represents the
prior NEE and ocean fluxes, and yo is the vector of
observations. H is the operator projecting x into
the observation space and is based on an atmospheric transport model
and fossil fuel and biomass-burning CO2 emission
estimates.
B and R are the covariance matrices of the normal
distribution of the uncertainty in xb (the “prior
uncertainty”) and of the sum in the observation space of the other
uncertainties when comparing H(xb) to
yo respectively (the “observation errors”). The latter
includes the measurement, model transport and model representation errors.
A complete solution to the inversion problem requires the estimation of the
uncertainty in the optimized fluxes (the “posterior uncertainty”), which is
a function of the prior and of the observation errors. As explained below in
Sect. 2.1, this estimation was not performed in this study. The following
sections present a brief description of each component of the inversion
configuration used in this study with a focus on parameters that are specific
to this study, while CH2010 provides more details on the parameters which
apply to all the MACC inversion configurations.
Inversion modelling setup
The link between CO2 fluxes and observations in the MACC inversion is
simulated by the global circulation model of the Laboratoire de
Météorologie Dynamique (LMDZ) version 4,, which is
the atmospheric component of the coupled climate model of the Institut Pierre
Simon Laplace (IPSL-CM4). Tracer transport is simulated by LMDZ at a horizontal
resolution of 3.75∘ × 2.75∘
(longitude × latitude) and with a vertical resolution of 19 levels
between the surface and the top of the atmosphere. LMDZ is nudged to winds
modelled by the European Centre for Medium-Range Weather Forecasts (ECMWF).
Prior NEE in MACCv10.1 was estimated at
3.75∘ × 2.75∘ and 3 h resolution from a global
simulation of the ORCHIDEE model at 0.7∘ resolution by
. ORCHIDEE was forced with the atmospheric conditions of
ECMWF reanalysis ERA-Interim . The ORCHIDEE NEE did not
take into account disturbance from land use or wildfires. Prior
ocean–atmosphere CO2 exchanges were obtained from the climatology of
air–ocean CO2 partial pressure difference by .
To complement these fluxes that were controlled by the inversion, the
H operator also included fixed estimates of the fossil fuel
and biomass-burning CO2 emissions. Fossil fuel emissions were
obtained from the EDGAR-3.2 Fast Track 2000 database
, scaled annually with the global
totals of the Carbon Dioxide Information Analysis Center. CO2 emissions from biomass burning were taken from
the Global Fires Emission Database version 2
GFEDv2,. Assuming that the vegetation
recovers rapidly from fire events, the CO2 emissions from
fires that affected the vegetation in a given year were offset by an
equivalent compensatory regrowth CO2 uptake evenly
distributed throughout the year.
The inversion controlled 8-day mean daytime and nighttime NEE and 8-day mean
ocean fluxes at the spatial resolution of the transport model. The analysis
in this study focuses on NEE and thus the impact of the inversion on ocean
fluxes is not detailed here, but Sect. 3.2 still uses an illustration of this
impact to raise insights into the corrections from the inversion over land.
At the grid scale, uncertainties in the prior NEE are estimated to be
proportional to the heterotrophic respiration fluxes from ORCHIDEE. Spatial
correlations of the uncertainties in B decay exponentially as
a function of the distance between corresponding pixel-based estimates of the
fluxes with a length scale of 500 km for NEE (1000 km for
ocean fluxes). Temporal correlations of the uncertainties decay exponentially
as a function of the lag time between the corresponding 8-day mean daytime or
nighttime estimate of the fluxes with a timescale of 1 month but without
correlation between daytime and nighttime uncertainties. The resulting
correlations in B are estimated as the product between the
temporal and the spatial correlations. This setup of the correlations for
B is based on the estimates by and
of differences between the NEE simulated by ORCHIDEE
and EC flux measurements (mostly located in the Northern Hemisphere).
List of surface stations over South America added to the
previous setup in MACCv10.1.
In the inversion framework, the misfits between simulated CO2 mole
fractions and the measurements that are not due to uncertainty in the prior
NEE or ocean fluxes must be accounted for in the covariance matrix
R. Uncertainties in fire and anthropogenic CO2 emissions
are assumed to have negligible impact at the measurement locations used here.
Therefore, they are ignored in the setup of R. Following CH2010,
the measurement errors are assumed to be negligible in comparison to the
uncertainties in the transport model. Model transport and representation
errors are modelled as half the variance of the high-frequency variability of
the deseasonalised and detrended CO2 time series of the measurements
that are assimilated at a given station. The resulting values of these model
errors for the stations in South America will be discussed in Sect. 3.1.
There is a moderate confidence in the adequacy of these error statistics
assigned in the global inversion system for the specific TSA area studied
here, both because B was designed mostly with statistics gathered
in the Northern Hemisphere and because R may not well account for
the uncertainty in the atmospheric convection model, while this could be high
in Amazonia . We also investigate here variations of the
fluxes within TSA at spatial scales that are not much larger than the
e-folding correlation length in B, and these variations in the
inversion results may be affected by our simple hypothesis of isotropic
correlations in the prior uncertainty. This lack of confidence in the input
error statistics weakens our confidence in the posterior error statistics
that can be derived based on the inversion system, even though they may be
realistic at zonal scale for the tropics . In
this context, and given the relatively high computational burden of the
posterior uncertainty computations for grid-point inversion systems (using
Monte Carlo approaches with ensembles of inversions,
, ), we do not derive
these posterior uncertainties for our domain and its sub-domains.
However, we will see at the beginning of Sect. 3 that the inverted fluxes are
more consistent with the CO2 atmospheric observations in TSA than the
prior fluxes and that their difference to the prior fluxes over TSA (i.e.
the flux increments generated by the inversion in order to better fit with
the observations) are significant. This indicates that the inverted fluxes
are strongly driven by the atmospheric data and as such are worth analysing.
This also suggests that the inversions yield a large uncertainty reduction
for TSA.
Assimilated data
MACCv10.1 assimilated measurements of atmospheric CO2, expressed as
dry air mole fractions in µmolmol-1 (abbreviated ppm) from
128 surface sites: 35 continuous measurement stations and 93 sites with
measurements of CO2 from discrete air samples collected approximately
weekly. Twenty-nine sites are located in the tropics, but only two had continuous
measurements over the analysis period and none of them were in TSA. In a
similar inversion conducted specifically for this study, called INVSAm
hereafter, we added new data from four surface sites located in the TSA
region. Figure 1 shows the measurement sites used by MACCv10.1 and the four
stations added in INVSAm. In the following of this section, we focus on the
description of these four stations and on the selection and representation of
their data. Details on the data selection and representation at the sites
used by MACCv10.1 are provided in CH2010.
Arembepe (ABP) (12.77∘ S, 38.17∘ W; 1 m a.s.l.) and
Maxaranguape (MAX) (5.51∘ S, 35.26∘ W; 15 m a.s.l.) are
coastal stations. The ABP site is located at the edge of the beach, where
vegetation consists mostly of grass and beach plants. Data were collected at
approximately 3 m above the ground and consisted of weekly
measurements of atmospheric CO2 with discrete air samples,
specifically under on-shore wind conditions when wind speed >2 ms-1. Air samples were collected preferentially during the
afternoon to avoid the influence of recycled air transported from land to the
ocean by land breeze during the night and early morning and transported back
to land by sea breeze during the morning. The MAX site is located on a cliff
right next to the coast and is surrounded by grass and beach plants. At MAX,
CO2 was measured with a continuous analyzer at approximately
3 m above the ground, and data were reported as 30 min averages.
This site is strongly under marine influence: winds are in general >10 ms-1, and wind direction varies preferentially between
100∘ and 140∘ at its location, so that
the measurements were taken mostly under on-shore wind conditions. Wind and
CO2 measurements at MAX indicate high CO2 variations when the
wind comes from land. These variations may be strongly influenced by the
emissions from the nearby city of Maxaranguape .
However, as in ABP, this does not occur during the afternoon, when the wind
conditions are dominated by sea breeze .
The Guyaflux site (GUY) (5.28∘ N, 52.91∘ W; 40 m a.s.l.) is
located at approximately 11 km from the coast and is surrounded by
undisturbed tropical forest. At GUY, measurements were taken at approximately
55 m above the ground . They were made with a continuous
analyser, and data were reported as hourly averages. The Santarém site
(SAN) (2.85∘ S, 54.95∘ W; 78 m a.s.l.) is located in the
tropical Tapajós National Forest, near km 67 of the Santarém–Cuiabá
highway, at approximately 750 km from the coast. Measurements were made at
eight
vertical levels ranging from ∼1 to ∼ 62 m above the ground with
continuous analyzers, but only data from the highest level were used in
INVSAm. Data were reported as hourly averages.
Figure 2 illustrates the temporal coverage of the observations available in
TSA during the simulated period (2002–2010). There is little overlap among
the site records due to calibration problems, interruption of the
measurements (e.g. at MAX) and the fact that some stations have been
installed only recently (e.g. at GUY). The longest records were from ABP (3
years: 2007–2009) and SAN (4 years: 2002–2005). Data from the four new
sites in TSA have been calibrated on the WMO-X2007 CO2 scale managed
by the ESRL/NOAA.
Top: location of assimilated surface stations in
South America and climatological wind speed/direction for
February (a), July (b) and annual mean
(c), averaged over 1981–2010 between the surface
and a level of 600 hPa (source: NCEP/NCAR reanalysis).
Sensitivity of surface atmospheric CO2 mole fractions
measured on 20 February 2009 at 10:00 UTC, at Guyaflux
(07:00 LT) (d) and Santarém (06:00 LT) (e), to a constant
increment of surface fluxes during the 2 days prior to the
measurement. Sensitivity values are expressed in log scale.
Open circles: sites with discrete air samplings. Filled circles are
measurements taken with continuous analysers.
Prevailing winds in the lower troposphere across TSA convey air masses
entering from the Atlantic Ocean near the Equator, across the continent and
back into the southern Atlantic Ocean generally south of 20∘ S. There
are no critical seasonal variations of the mean winds in the area so that
this typical behaviour applies throughout the year. The climatology of wind
fields from the NCEP/NCAR reanalysis (over the period 1981–2010) for
February, July and annual mean, shown in Fig. 3, illustrates this typical
circulation pattern. This confirms that the variations of CO2 at
coastal stations (ABP, MAX) are mainly influenced by air–ocean exchanges and
fluxes in distant lands. These stations should thus provide more information
on the atmospheric CO2 content upwind of TSA than on the fluxes
within Amazonia. Figure 3 also shows that GUY and SAN receive a signal from the
ecosystems of the north-eastern Amazon basin. Despite GUY being not far from
the coast considering the Amazon-wide scale, this site is still located
inland, in an area covered by undisturbed tropical wet forest. SAN is
located considerably further inland than GUY. Typical influence functions of
fluxes for observations at GUY and SAN (the observation “footprints” in
Fig. 3b and c respectively) illustrate that the sensitivity of instantaneous
mole fractions to the fluxes rapidly decreases with the distance mainly due
to the typically moderate horizontal wind speeds, so that they should bear
a strong signature of local fluxes, i.e. of the NEE in north-eastern Amazonia.
This and the fact that the geographical distance between the sites in the
TSA region ranges from 1000 to 2600 km, i.e. up to 5 times the
correlation length scale in the matrix B, could suggest that the
area well constrained by the sites in the TSA region through inversion is
limited. However, as illustrated in Fig. 3, the station footprints also have
modest values over very extensive areas, which may also result in significant
large-scale constraint from the inversion on the land flux estimates. This
will be analysed below in Sect. 3.2.
We assimilated observations from the South American sites between 12:00 and
15:00 local time, when the boundary layer is well developed and
likely to be well represented by the transport model
. Such a selection of the afternoon data results
in ignoring the measurements under off-shore flow at MAX and thus the
potential for capturing a clear signature of the regional NEE at this site
such as at ABP. However, this potential is rather low since under off-shore
flow conditions the signal at MAX is also connected to the local
anthropogenic emissions, and the inversion cannot reliably exploit such a
signature of the regional NEE when the dynamics of the planetary boundary layer are poorly
represented by the atmospheric transport model. Observations were also
screened for low wind speed (>2 ms-1), thus removing the
effect of local emissions (and sinks) that may not be well captured by the
transport model at resolution 3.75∘ × 2.5∘. Under
such on-shore flow conditions, the model correctly simulates CO2 in
the grid cells corresponding to the horizontal location of the coastal sites,
even though these grid cells bear a significant NEE due to the overlapping of
both land and ocean. This reduces the need for ad hoc changes of the model
grid cells to better represent CO2 at the coastal sites (e.g.
). In a general way, we choose to represent the four
measurements sites using the model horizontal grid cell in which they are
located since, for each site, it yields better statistical fit between the
prior simulations and the selected measurements than when using neighbour
grid cells.
Analysis of an alternative estimate of the NEE for the evaluation of the inversions
Our analysis of the inversion results is compared to the independently
derived NEE estimated by (hereafter J2011). J2011 used model
tree ensembles (MTE), a machine-learning technique, to upscale FLUXNET
eddy-covariance observations, based on remote sensing, climate and land-use
data as drivers, thereby producing gridded estimates of NEE and other surface
fluxes at the global scale at 0.5∘ resolution. As discussed in J2011,
large uncertainties affect their annual mean NEE estimates and associated
seasonal and interannual variations. This is likely particularly true in TSA
region, where few FLUXNET measurements are available. Yet its comparison to
the NEE from the inversion could give useful insights for the analysis of the
latter.
Results
In this section we first analyse the statistical misfits between observations
and simulated mole fractions from prior and posterior fluxes at the sites in
the TSA area, as a measure of the efficiency of the inversion in reducing the
misfits to the measurements. This is a first indicator of the significance of
the corrections applied to the fluxes. We then examine the amplitude and
spatial distribution of the increments from both inversions to give a further
indicator of this significance and to characterise the impact of
assimilating the measurements from the sites in South America. Finally we
focus on the impact of the inversions on the seasonal patterns and IAV of NEE
which are the aim of this study. This analysis is supported by the comparison
to the product of J2011.
Comparison to observed CO2 mole fractions
Comparison of assimilated CO2 observations (blue) and
corresponding simulated mole fractions using prior fluxes (red),
INVSAm (green) and MACCv10.1 (purple). Measurements were collected at
Arembepe (a), Guyaflux (b), Santarém
(c) and Maxaranguape (d). Data shown here
correspond to daily average mole fractions between 12:00 and 15:00 LT, when wind speed > 2 ms-1. Note that
the timescale differs between plots.
The time series of assimilated observations and the corresponding simulated
CO2 mole fractions using the prior fluxes, the inverted fluxes from
MACCv10.1 and that from INVSAm at the four sites in the TSA region are
plotted in Fig. 4. The statistics of the misfits between these measured and
simulated CO2 mole fractions are summarised in Fig. 5. At each site
in the TSA region, the smallest quadratic mean and standard deviation of the
misfits between the simulations and the observations were obtained with
INVSAm, which is a logical consequence of the assimilation of these
observations. However, the misfits are also strongly decreased at all sites
when comparing MACCv10.1 to the prior simulation. While, compared to the
prior simulation, MACCv10.1 strongly decreases the standard deviation of the
misfits at MAX and ABP, it does not significantly reduce it at GUY and SAN.
The decrease of the misfits at all sites in MACCv10.1 is thus explained by
the strong decrease of the bias in these misfits. Indeed, both inversions
critically reduce a large-scale bias over TSA, since the presence of a few
marine stations on the globe is enough to introduce this effect by correcting
the global growth rate of CO2 (CH2010). However, the information from
the local network significantly impacted the seasonality of the simulated
CO2 in the TSA region.
Taylor diagram of the statistics of misfits between
observations and simulated CO2 mole fractions between 12:00
and 15:00 LT at Guyaflux (square), Santarém (circle), Arembepe
(diamond) and Maxaranguape (triangle), when wind speed
> 2 ms-1, using prior fluxes (red), INVSAm (green)
and MACCv10.1 (purple). Radial distance from the origin: ratio of SD of
simulated mole fractions and SD of the observations. Angle measured
from the y axis: coefficient of correlation. Numbers next to the
symbols: bias (in ppm). Grey circles: SD of the misfits (in ppm).
The resulting optimized mole fractions from INVSAm generally shifted from a
minimum to a maximum around June every year at SAN or from a maximum to a
minimum around October (both in 2004 and 2006) at MAX with respect to the
prior simulation and MACCv10.1 (Fig. 4c) and in agreement with the
observations. While yielding a phase of seasonality at GUY comparable to that
of the prior simulation and MACCv10.1 and comparable to that of the data,
INVSAm exhibits a significant rescaling of the seasonal variations in the
period from May to September at this site (Fig. 4b) compared to these two
other simulations, in agreement with the observations. At SAN, during the
austral fall–winter, while the misfits are negative with MACCv10 they become
positive with INVSAm. The positive increments from the assimilation of data
at SAN (no other data are assimilated in TSA in 2002 and 2003) are thus too
high.
Subsequently, when compared to MACCv10.1, INVSAm improves the amplitude of
the seasonal variations of the simulated mole fractions with respect to the
prior simulation at GUY and MAX and does not impact it at SAN. At ABP, the
seasonality is less visible in both the measurements and the inversion
posterior simulations and it is difficult to assess whether INVSAm improves
it compared to MACCv10.1, but both inversions dramatically decrease the large
amplitude of the prior seasonal variations, consistent with the data. The
best correlations with the observations are obtained with INVSAm at all sites
(Fig. 5). The values of these correlations remained generally low, ranging
from 0.23 at GUY to 0.81 at ABP. These correlations are based on comparison
of daily CO2 mole fractions while the inversions control 8-day mean
fluxes, which strongly limits the ability to impact the mole fractions at
higher temporal resolution and can thus explain the low correlation
values. Correlations between time series of observed and simulated monthly
mean mole fractions are higher than those for daily values, ranging from 0.76
at GUY to 0.92 at ABP for INVSAm, with which, again, these correlations are
the highest.
The significance of the reduction of the misfits between the mole fractions
observed and simulated from the inversion is seen from the comparison between
the standard deviations of these misfits and the estimate of the standard
deviation of the observation errors (i.e. of the transport model errors) for
hourly values in the configuration of the R matrix (Table A1 in
the Supplement). According to this comparison, the prior misfits are much
larger than the observation errors at ABP, MAX and GUY but are slightly
smaller than these at SAN. Misfits between MACCv10.1 and the observations are
similar to the prior misfits at SAN and GUY and are much smaller than the
prior misfits (and smaller than the 95 % confidence interval of the
observations) at the coastal ABP and MAX sites. Misfits are further decreased
when assimilating the data from the South American sites: they are about the
standard deviation of the observation errors at all sites but GUY (where they
are twice as large).
These results suggest that the assimilation of data in the TSA region helped
improve the phasing of the seasonal variations, whereas MACCv10.1 did not
impact it. MACCv10.1 mainly improved the amplitude of the seasonal variations
at the coastal sites and decreased the biases. INVSAm improved the amplitude
of the seasonal variations at GUY. More generally, unlike MACCv10.1, INVSAm
led to an improvement of the variability of the simulated CO2 at the
inland sites, which are more sensitive to the NEE in Amazonia.
Characterisation of the monthly to annual mean inversion increments to the prior fluxes
Spatial distribution of 2002–2010 mean flux corrections at
the transport model resolution
(3.75∘ × 2.50∘) to ORCHIDEE from INVSAm (left)
and MACCv10.1 (right) over an area larger than TSA region: mean for
February (a, d), July (b, e)
and the full period 2002–2010 (c, f). Flux increments
over land and ocean are represented with two distinct colour scales
and units: green–yellow for land, in gCm-2h-1; blue–red for ocean,
in mgCm-2h-1. Red symbols are surface stations in South America added to
the previous setup of MACCv10.1, where filled circles indicate
locations of sites with continuous measurements; open circles indicate
locations of sites with discrete air sampling. Black symbols are the surface
stations used in MACCv10.1.
Figure 6 shows the spatial distribution of the mean corrections applied
during the period 2002–2010 by INVSAm and MACCv10.1 over land and ocean,
across an area that covers the TSA area and neighbour regions. Complementary to this,
Fig. S1 shows the spatial distribution of the corrections over land in the
TSA region for the full 2002–2010 period and for the 2002–2005 and
2006–2010 sub-periods. Both give results for the full years and for the
months of February and July. As such, these figures are indicative of the
amplitude and spatial extent of the corrections from the inversions and of
the impact of the assimilation of the measurements in South America. Figure
S1 even dissociates the impact of assimilating data at SAN and MAX and that
of assimilating data at MAX, ABP and GUY by splitting the results between the
time periods when these two different sets of data are available. The
analysis of the annual mean corrections and of mean corrections for February
and July should also give first insights on the significance of the
corrections applied to the seasonal cycle and IAV of the NEE in the TSA
region.
Figure 6 depicts the increments from both inversions, showing large patterns
which are nearly zonal (or along the prevailing winds) and overlap
continuously over land and ocean. Since there is no correlation between the
uncertainty in ocean and land fluxes in the B matrix, and given
the typical length scale of the correlations in this matrix, this can be
directly connected to the signature of atmospheric transport. The contiguous
zonal patterns have alternate negative and positive flux increments. There is
thus an opposition between corrections in the north and in the south of the
TSA region. These corrections are rather negative in the north and positive
in the south (positive in the north and negative in the south) during the
austral summer (winter). As these corrections are stronger during the austral
winter, it results in positive (negative) corrections in the north (south) at
the annual scale. Such dipoles are a typical behaviour of inverse modelling
systems in data-poor regions . However, changes in the
amplitude and latitudinal position of this zonal dipole appear to be the main
impact from the assimilation of data in the TSA region. This dipole structure
may thus yield sensible corrections to the NEE in the TSA area. The dipole
has a high amplitude for MACCv10.1 and even higher for INVSAm. The
increments from INVSAm to the annual fluxes often exceed 150 % of the prior
estimate in terms of absolute values. The highest increments are obtained
during austral winter and when the SAN data are available (during the period
2002–2005, see Fig. S1), which is in line with the fact that this site is
located more inland than the others. Such high control of the data in the TSA
region (even when checking the SAN and MAX or the MAX, ABP and GUY data sets
only) over the zonal patterns of flux corrections also highlights the very
large-extent impact of these data, and of the data in the Southern Hemisphere
in general, despite the relatively small spatial correlation length scales in
the B matrix and the limited area in which the station footprints
are very high. The inversion also generates patterns of corrections of
smaller spatial scale close to the measurement sites in the TSA region when
these sites are used by the inversion. This raises hope that the NEE over the
whole TSA region is strongly constrained by the observations but can also
raise questions regarding the spatial variations of the corrections applied
by the inversion to the NEE within the TSA region, at least when considering
areas at more than 500 km from the measurement sites. However, various
pieces of evidence (Figs. 5 and 6, the analysis of the decrease in misfits to
the observations from the inversion in Sect. 3.1 and the previous analysis
of the high increments to the monthly mean and annual mean NEE over the
entire TSA region) indicate that the corrections from the inversion are
significant.
Diagnostics of the biogenic CO2 fluxes
Seasonality
Monthly mean NEE anomaly integrated over (a) the TSA
region and (b) over pixels dominated by TBE forests in
ORCHIDEE for 2002–2010. The shaded areas denote dry seasons,
defined as months with precipitation < 100 mm, based on
monthly totals from TRMM data over 2002–2010. Estimates from prior
fluxes (red), INVSAm (green), CH2010 (purple) and J2011 (dashed
blue). (c–d) Monthly mean NEE
integrated over the zones 1 (c) and 2 (d) that are
defined in Fig. 8.
Figure 7a illustrates the mean seasonal cycle of NEE from the prior fluxes,
J2011, MACCv10.1 and INVSAm over TSA. The mean for the full period 2002–2010
was removed because uncertainties in the long-term mean can be large for the
inversions as well as for the J2011 product and because this long-term mean
can differ significantly between the different estimates. Removing the mean
allows us to focus on the seasonal variations. Hereafter, positive values of
NEE indicate anomalous CO2 release to the atmosphere; negative values
indicate anomalous uptake by the ecosystems. The shaded area indicates the
dry season, defined as months with precipitation <100 mm according
to data from the Tropical Rainfall Measuring Mission (TRMM 3B43 (v6)
product), averaged over January 2002 to June 2010. The results of Fig. 7a are
calculated considering all the plant functional types (PFTs) represented in
ORCHIDEE over the TSA region. The vegetation map of ORCHIDEE, originally at
a spatial resolution of 0.72∘, was aggregated according to the
transport model grid, and Fig. 8 illustrates the dominant PFTs in terms of
area for each transport model grid cell.
Dominant PFTs for each transport model grid cell
(i.e. 3.75∘ × 2.50∘) according to the
ORCHIDEE vegetation map over the study region. Open circles show
location of sites with discrete air sampling; filled circles show
location of sites with continuous measurements. Zones 1
and 2 indicate areas for which the NEE is presented in Fig. 7c and
d respectively.
Both the prior simulation and the inversions predict a maximum of NEE (i.e.
likely a maximum of CO2 release) in the dry season and a minimum of
NEE (i.e. likely a maximum of CO2 uptake) in the wet season
(Fig. 7a). This behaviour is also seen in J2011. However, J2011 place the
maximum of NEE during the transition between the wet and dry season while the
prior simulation and the inversions place it at the end of the dry season.
Even though the inversions seem to delay or lengthen this maximum, such
a modification is not significant and their seasonal phasing is likely
strongly constrained by the patterns of the prior fluxes. In particular,
according to the comparison between INVSAm and MACCv10.1, the assimilation of
data from the four stations in the TSA region does not seem to impact this
phasing.
The inland data are prone to bear a stronger signature from fluxes in
tropical broadleaf evergreen and raingreen (TBE) forests (Fig. 8), while the
mean seasonal behaviour over the whole TSA region could be mainly related to
other PFTs. Therefore, we isolate the results for the area of TBE forests,
this area being defined by the selection the model grid cells dominated by
this vegetation type. The configuration of the prior uncertainties in the
inversion does not account for PFTs, so that the spread of the flux
corrections in the inversions is not forced a priori to depend on vegetation
type. We still expect that the variations in the measurements, when their
footprint covers different distributions of PFTs, reflect differences in NEE
of the PFTs. Consequently, the spatial patterns of the increments from the
inversion may be consistent with the spatial patterns of NEE induced by the
distribution of the different vegetation types. The mean seasonal cycle of
NEE for the area of TBE forests within the TSA region is given in Fig. 7b.
The restriction of the analysis to the TBE forest does not show any clear
correlation between NEE extremes and the phasing of wet and dry seasons
neither when considering the NEE from the prior nor when considering the NEE
from both inversion estimates. This is different from J2011, who indicate
a maximum of the NEE a few months before the beginning of the dry season and
a minimum of the NEE at the beginning of the wet season. The prior and the
inversions indicate several local extremes of NEE throughout the year that
may reflect the overlapping of significantly different seasonal cycles for
different sub-regions within TBE forests.
The strong spatial heterogeneity of the time variations of the NEE in TBE
forests has been discussed in the introduction. Figure S2 illustrates it this
with results of local NEE mean seasonal cycle estimated from EC measurements
across TSA. This figure also shows the mean seasonal cycle of the
precipitation at these sites to illustrate the spatial heterogeneity of the
drivers of NEE within TSA.
To examine whether the inversion captures this spatial variability of the
fluxes, we analyse the seasonal variations of the NEE estimates for the two
zones depicted in Fig. 8. Zone 1 was located in north-eastern Amazonia, close
to the measurement stations SAN and GUY. Zone 2 was located in central
eastern Amazonia. Both zones are mainly covered by TBE forests, according to
the vegetation classification of ORCHIDEE. According to ,
eastern Amazonia is drier and shows a stronger seasonality than western
Amazonia. However, we do not identify a clear pattern of NEE seasonal
variations that could be driven by the rainfall seasonality in any of the two
sub-regions, except for J2011 in Zone 1 (Fig. 7c), since the other estimates
again exhibited maxima and minima of NEE during both dry and wet seasons.
Actually, in Zone 2 (Fig. 7d) the dry season cannot be clearly
identified. In this zone, the prior flux and the inversions indicated several
maxima and minima of NEE, but J2011 exhibit, again, a clear seasonal cycle
with a maximum in June and a minimum October as in Zone 1. While J2011 showed
nearly the same amplitude and phasing of monthly mean NEE variations in both
zones and over TBE forests (Fig. 7b), prior and inversions estimates of the
seasonal variations differed both in phasing and amplitude among zones 1, 2
and the whole TBE forest area.
Divergent patterns are found in INVSAm with respect to MACCv10.1, which
remains closer to the prior fluxes, even though the departure of MACCv10.1
from the prior NEE is significant in Zone 2 and for the whole TBE area
(Fig. 7b and d). The comparison of these inversion results shows that
significant flux corrections due to the assimilation of data in South America
are applied in Zone 1 (Fig. 7c), i.e. in north-eastern Amazonia, where
stations SAN and GUY are located. The influence of SAN over this zone is
clearer when splitting the analysis period of the mean seasonal cycles
between 2002–2005 and 2006–2010 (not shown). The differences between INVSAm
and MACCv10.1 are more accentuated during the period 2002–2005 when SAN is
active. However, there are still significant changes between these two
estimates during 2006–2010. The changes between MACCv10.1 and INVSAm in Zone
2 (Fig. 7d) are also significant, even though Zone 2 seems hardly observed by
the TSA observation network. As analysed in Sect. 3.2, the control of the
long-range dipole (of its amplitude and latitudinal position) by the
measurements in region TSA explains such an impact of these measurements on
the results in Zone 2, as well as that of measurements outside South America,
which explains the departure of MACCv10.1 from the prior NEE in Zone 2. Zone
2 is actually located close to the frontier between the northern and southern
patterns of the dipole in the TSA region. A latitudinal shift of the frontier
through the assimilation of data in north-eastern Amazonia can thus easily
imply that positive (negative) increments from the inversion are reverted
into negative (positive) increments.
In an attempt at getting clearer seasonal patterns in some of the other
sub-regions of Amazonia, two additional zones have been analysed, located in
south-western and south-eastern Amazonia, where the dry season is potentially
earlier and more extreme (Fig. S2a, d). Both sub-regions encompass areas
where the impact of the droughts of 2005 and 2010 was the highest according
to . The results, however, do not provide any further
information than Fig. 7c and d and are not shown. J2011 still exhibit the same
amplitude of the seasonal cycle and the same location of maximum and minimum
NEE as in zones 1 and 2 despite the extent of the dry season. Prior fluxes
and inversions still showed maxima and minima during the dry season in some
cases, and the inversions introduce only slight modifications to the
amplitude and phasing of the NEE relative to the prior simulation. This is an
expected result due to insufficient data in the southern part of TSA to
constrain fluxes in that region.
Interannual variability
(a) Annual NEE anomaly compared to the mean of
2002–2010; estimates for the whole study region. (b)
Annual NEE anomaly compared to the mean of 2002–2010; estimates for
the area dominated by TBE forests.
Figure 9a depicts the annual NEE anomalies of the prior simulation,
MACCv10.1, INVSAm and an additional inversion called FLAT, compared to their
mean NEE over 2002–2010, aggregated over the whole TSA region (considering
all PFTs). FLAT corresponds to a new inversion using, as a prior estimate, a
“flat prior” whose annual anomalies are null over the TSA region. Using the
standard prior NEE as a basis, the flat prior is built by offsetting the
annual budgets of NEE over the TSA region so that they equal the mean annual
NEE over TSA and over the 2002–2010 period from the standard prior NEE. The
spatial variability and the temporal variability at scales smaller than 1
year are conserved between the standard NEE and the flat prior, since the
offsets are applied homogeneously in space and time within TSA and within 1
year. FLAT assimilates the data from the four surface sites in TSA in
addition to those used by both MACCv10.1 and INVSAm. Of note is that even if
increments on the NEE annual budget of a given year from an inversion are
weak, the changes in the corresponding annual anomaly from the inversion can
be high because the inversion modifies the 2002–2010 average against which
the anomaly is computed. Prior fluxes, MACCv10.1 and INVSAm display only
small positive anomalies during the drought years (2005, 2010) compared to
other years. FLAT displays a negative anomaly (i.e. a strong uptake) in
2010, but it indicates a larger positive anomaly in 2005 than that of other
estimates. However, the strong NEE negative anomaly of 2009 in the
prior fluxes, MACCv10.1 and INVSAm is also in FLAT, which suggests that this
pattern is strongly driven by the atmospheric measurements and raises
confidence in it.
As in Sect. 3.3.1, we isolated the results for the TBE forests area
(Fig. 9b). In this case, prior fluxes and both MACCv10.1 and INVSAm
estimates show diverging annual mean responses of forests to
drought, with a positive anomaly in 2005 and a negative anomaly in
2010. For 2009, when climatic conditions were abnormally humid across
South America, the inversion estimates consistently show a small
positive anomaly, opposite to the response for the whole TSA region. The small anomalies in all inversions suggest a weak
sensitivity of the NEE of TBE forests to interannual variations and
that most of the IAV over the study area is not related to TBE
forests.
Annual NEE anomaly compared to the 2002–2010 mean for Zone 1
(a, c) and Zone 2 (b, d) as defined in
Fig. 8. Estimates from prior fluxes (red), INVSAm (green), MACCv10.1
(purple) and J2011 (grey).
Finally, we analyse the results in the two sub-regions shown in Fig. 8 in an
attempt to identify potential differences in the regional responses. NEE
estimates from the prior, INVSAm and MACCv10.1 show various responses of
forests to drought in these zones. In Zone 1 (Fig. 10a) all these estimates
present a positive anomaly in 2005 and a negative anomaly in 2010, while in
Zone 2 (Fig. 10b) they yielded negative anomalies during both years. J2011
exhibit abnormal anomalies much smaller than these NEE estimates (Fig. 10c
and d). This prevents us from gaining insights into the IAV from the
comparison of J2011 to the other estimates. However, the product of J2011
must be used cautiously, especially when evaluating IAV of NEE. J2011 relied
on a limited number of EC stations across the Amazon basin, with short time
series, to estimate MTE based on spatial gradients among the sites and then
extrapolated to temporal gradients. This is valid assuming that spatial and
temporal NEE patterns have the same sensitivity to climate, which may be
incorrect . The example of the divergences of the results
between MACCv10.1 and INVSAm in 2003 in Zone 2 illustrates, again, some weak
ability to precisely constrain the fluxes in such a small area, which is
quite distant from the measurement sites in TSA. Indeed, the analysis of the
maps of increments from MACCv10.1 and INVSAm for the annual mean NEE in 2003
(not shown) demonstrates that the assimilation of data at SAN during this
year shifts the northern border of the pattern of negative corrections in
MACCv10.1 from north of Zone 2 to south of Zone 2. Since, on average
over 2002–2010, both inversions apply positive increments in this zone (see
Fig. 6), this leads to a clear negative annual anomaly in Zone 2 and for the
year 2003 for INVSAm.
Discussion and concluding remarks
Amazonian forests play a key role in the global carbon balance, but there are
large uncertainties on the evolution of this terrestrial sink. Uncertainties
stem from incomplete knowledge of the processes behind land–atmosphere
CO2 exchange in this region. Improving our understanding of the
seasonal and interannual variations of Amazonian forests is thus a priority.
In an attempt to gain insight into how these temporal variations of
CO2 fluxes vary across Amazonia, we analysed global inversions and
incorporated new measurements of atmospheric CO2 mole fractions in
TSA into one of these inversions. The analysis of the global inversions at
such spatial scales, which are generally ignored in global inversion studies,
is justified by the use of a variational inversion system solving for the
fluxes at ∼ 3∘ and 8-day resolution. We showed that the two
inversions applied large corrections to the estimates of NEE from a
vegetation model that they used as prior information. The inverted NEE was
strongly controlled by the assimilation of CO2 measurements both
outside and within the TSA region, and this control was characterized by
zonal patterns of alternate positive and negative corrections, which we call
“zonal dipole”, in addition to more local patterns in the vicinity of the
sites that were assimilated.
Despite an overall improvement by the inversion of the seasonal
variations of the simulated CO2 mole fractions when compared
to the measurements in TSA, several issues arose when analysing the
seasonal cycles of NEE from the inversion. The seasonality of the mean
NEE over the whole TSA region remained basically unchanged between the inversion estimates (Fig. 7a). The prior and inversion estimates of
this mean seasonal cycle of NEE at the TSA scale are not in line with
J2011 and disagree with the intuitive assumption that the seasonal
cycle should be correlated with rainfall and solar radiation,
especially in the tropical forest area. Furthermore, they do not
exhibit a clear seasonal pattern over TBE forests at basin scale or
within the analysed sub-regions. J2011 display a clear homogeneous
seasonal cycle all the TSA region, which does not give confidence in
its ability to distinguish regional heterogeneity. The proximity of Zone 1 to the stations in north-eastern Amazonia (SAN
and GUY) (Fig. 8) suggests better confidence in the flux
corrections applied by INVSAm to the prior fluxes in that zone than
elsewhere in TSA region.
The reliability in the seasonal patterns of the inverted fluxes is thus not
high, which seems to confirm that the zonal dipoles of increments from the
inversion are artificial patterns that balance the overall correction in
the Southern Hemisphere and are not necessarily consistent with the
actual NEE in the TSA region. This is directly connected to the lack of
CO2 measurements in the TSA region both in space and time. The
limited overlap among the TSA observations is a critical issue since
measurements are often only available at a single site at once and,
consequently, temporary model errors at this site can get far more weight in
the inversion than if it had been balanced by information from other sites.
Furthermore, the lack of confidence in the INVSAm results in Zone 1, which is
relatively close to the GUY and SAN, suggests a low reliability in the
statistics of the uncertainty in the prior NEE (in the inversion
configuration), on which the extrapolation of the information from the
vicinity of these sites to the whole north-east of the TSA region relies.
This further supports the choice of avoiding computing posterior
uncertainties in the inverted NEE as discussed in Sect. 2.1.
Such considerations also weaken the analysis of the IAV based on the
inversion while J2011 do not provide a reliable IAV of the NEE in TSA,
which could have supported such an analysis. However, some patterns of the IAV in
the NEE seem consistent among the different inversion estimates when the
atmospheric measurements have a strong control on it: across the TSA region
the estimates from the prior fluxes, MACCv10.1, INVSAm and FLAT indicate
small positive flux annual anomalies (CO2 release) during the drought
in 2005 and a strong negative (CO2 sink) anomaly in 2009, presumably
related to lower temperatures and more humid conditions in 2009. However, in
2010 there is a divergence of the results between the FLAT estimate and the
others.
In the TBE forests, the highest source anomaly in 2005 seen in the
prior fluxes, MACCv10.1 and INVSAm may be related to reduced
photosynthesis during the drought, as found by ,
and/or tree mortality caused by the squall event of January 2005
. However, in 2010 these results indicate
a small sink anomaly. This anomaly seems inconsistent with the
hypothesis of a higher negative impact of the drought in 2010, which
was more intense in terms of water stress and more geographically
extensive . However, it seems consistent
with the recent results of , who found that the
Amazon basin was carbon neutral during that year.
Even though some seasonal or interannual patterns from the inversion look
realistic, our study mainly reveals some critical issues that hamper the
ability to derive an accurate estimation of the temporal variability of NEE
and of its spatial heterogeneity across Amazonian forests. A denser
monitoring network across the basin with continuous time series, as initiated
by , is needed to well constrain the fluxes in the region.
In addition, the simulation of atmospheric transport may need to be handled
with models that are better adapted to the local meteorological conditions.
Regional transport models with higher spatial and temporal resolution and
improved parameterisations of key atmospheric processes for the region
e.g. deep convection, have been developed
. The combination of a denser observation network and
state-of-the-art regional modelling tools would overcome some of the critical
limitations encountered here for the study of the temporal variability of
biosphere CO2 fluxes in Amazonia. Such regional inversion will
require reliable regional configurations of the input error statistics, which
could rely on extensions of the flux eddy-covariance measurement networks in
Amazonia. Finally, adaptive strategies for the representation of the
observations in the model simulations as a function of the sites and of the
meteorological conditions could help loosen the selection of
the data for the assimilation.