Introduction
There is large political and scientific interest in developing methods for
improving and verifying estimates of fossil fuel and cement CO2
emissions. Consequently, there is an increasing deployment of urban
CO2 monitoring networks with the objective of quantifying city
emissions through the atmospheric inversion approach . , upon
which this study builds, recently reported first estimates of fossil fuel
CO2 emissions of the Paris urban area during a 2-month period. They used
three ground-based CO2 measurement sites at the north-eastern and
south-western edge of the area and an inversion system based on
a 2 km×2 km horizontal resolution transport model. The
monitoring stations in Gonesse (GON), approximately 15 km north of Paris'
city centre, and in Montgé-en-Goële (MON), 35 km north-east
(NE) of Paris' city centre, were deployed by the CO2-MEGAPARIS project and
operated from August 2010 to July 2011. The monitoring station in
Gif-sur-Yvette (GIF), 20 km south-west (SW) of Paris' city centre, is
part of the Integrated Carbon Observation System France long-term network.
The main principle of the atmospheric inversion proposed by
consists in constraining CO2 emission budgets of the urban area by
assimilating atmospheric CO2 mole fraction gradients between pairs of
sites located upwind and downwind of the city. The use of cross-city
gradients, rather than individual mole fractions, aims at eliminating the
variability of CO2 caused by the transport of remote and natural
fluxes outside the urban area. It assumes that the signal from these fluxes
has a relatively large spatial and temporal scale compared to the distance
and transport duration between the measurement sites. These signals and the
potential signal from natural fluxes within the urban area cannot be
sufficiently well controlled by the monitoring network, in particular because
their large day-to-day variations cannot be filtered as a smooth baseline in
the time series of CO2 concentrations at individual sites an
approach frequently used in regional atmospheric inversions, e.g.
in. On the contrary, such signals can be as high as the signal
caused by the emissions within the urban area . Uncertainties in remote and natural fluxes can thus highly impact
the skill for inverting the urban emissions. In the simulations by
, the ratio between the signal from the natural and remote
fluxes and the signal from the urban emissions is high when analysing
individual measurements. It, however, strongly decreases when analysing
gradients. This weak impact of natural fluxes on inversions is on the one hand
due to the fact that the dense and compact Paris urban area exhibits little
vegetation within its bounds. On the other hand, it is due to the fact that,
upwind of the city, the signal from fluxes outside this urban area is
sufficiently diffused in space so that it is relatively homogeneous over the
Paris urban area and constant during the duration of the transport over this
area.
The selected cross-city gradients also provide a characterisation of the
increase in the CO2 mixing ratios of air parcels that pass over the
city. It is assumed that these gradients represent emissions from the entire
city and are not highly sensitive to the distribution of the emissions. This
assumption is in line with the inversion system of that
controls the city-scale emission budgets and the temporal variation of fossil
fuel CO2 emissions, but not their spatial distribution. However,
this method should not be seen as a sort of mass balance, given that, in
practice, the inversion is not set up to ensure that the upwind and downwind
concentrations correspond to the same air masses that travelled from the
upwind to the downwind site. Furthermore, since the atmospheric boundary
layer evolves significantly in space and time and due to the atmospheric
diffusion during the transport over the Paris area, such cross-city gradients
cannot perfectly represent the CO2 enrichment of air parcels passing
over the urban area. In addition, temporal variations of the emissions during
the transport of air masses over the Paris area prevent relating a given
gradient to the emissions at a given time. The gradients need to be
interpreted using a transport model and knowledge of the spatio-temporal
variations of the emissions at hourly scale. In that sense, the assimilation
of gradients is affected by transport modelling uncertainties and by
uncertainties in the variations of the emissions at high spatial and temporal
resolution, such as is the case for any inverse modelling approaches.
The inversion assimilates cross-city CO2 gradients during the
afternoon to correct prior estimates of 6 h fossil fuel CO2
emission budgets of the Paris metropolitan area (Île-de-France
administrative region). These prior estimates are derived from the AIRPARIF
inventory for the year 2008 . AIRPARIF is a non-profit agency
that is accredited by the French Ministry of the Environment to monitor the
air quality in Île-de-France. Even though they have a limited impact on
the inversion when gradients are assimilated, the system of
also inverts biogenic fluxes and corrects prior estimates of the biogenic
fluxes from C-TESSEL, the land-surface component of the ECMWF (European
Centre for Medium-Range Weather Forecasts) numerical weather forecasting
system . In order to model the CO2 gradients, the
inversion uses an estimate of the fossil fuel CO2 emission and
biogenic flux distribution at 2 km × 2 km and hourly
resolution, coupled to a 2 km × 2 km resolution
configuration of the chemistry transport model CHIMERE .
developed and tested this inversion set-up for 2 months in
autumn 2010. The values of the AIRPARIF 2008 inventory were used to derive
the prior estimates for the corresponding dates in 2010.
reported a significant improvement of the fit between modelled and measured
CO2 gradients by the inversion and reasonable patterns of corrections
applied to prior emission values. The small number of monitoring sites and
the stringent criteria for selecting gradients leads to a large number of
periods, ranging from one to several days, during which the inversion does
not assimilate any atmospheric CO2 data. As a consequence, averages
of the inverted emissions over 1 month were found to be more reliable than
1-day to 1-week mean results.
The aim of this study is to derive a full year of monthly mean emission
estimates for the Paris area, based on the inversion system described by
and on the availability of measurements at MON, GON and GIF
during the period mid-2010 to mid-2011. The 1-year-long inversion allows
a better evaluation of the method by analysing the seasonal variation and the
annual budget of the inverted emissions. In particular, the annual budget can
be compared to the AIRPARIF emission assessment for 2010 . This
assessment is based on an inventory model that has been improved since the
release of the 2008 inventory. The 2010 inventory applies to a time period
which better corresponds to the inversion period than the 2008 inventory used
for the inversion. Therefore, it provides some independent information to
check the corrections applied by the inversion to the prior estimate of the
annual budget derived from the 2008 inventory.
Preliminary tests of the inversion during the 1-year period, however,
revealed that the selection of gradients, as proposed by , do
not conform fully with the underlying assumptions of having gradients
dominantly influenced by urban emissions. Notably, negative CO2
gradients between downwind and upwind sites were frequently measured when
using the gradient selection criteria of . This led us to
revise the selection of CO2 data to form gradients. The revision
consists primarily in a tighter filtering of wind directions to select
gradients in order to avoid situations when air parcels leaving the upwind
site or reaching the downwind site do not overpass a significant part of the
city and the vicinity of the other site. Section presents
a summary description of the inversion configuration and the revised gradient
selection. Section analyses the inversion results for
different configurations. In particular, it assesses the impact of the
stricter gradient selection and the sensitivity of the results to the prior
emission estimates, to the emissions' spatial distribution, and to the
atmospheric transport modelling so as to evaluate how robustly the emissions
are constrained by atmospheric CO2 data. These results are discussed
in Sect. .
Inversion configuration
The inversion method described by is based on the Bayesian
approach. The control vector x gathers the CO2 flux budgets.
xb is the vector of the prior estimates of these budgets,
independent of atmospheric observations. The observed CO2 mole
fraction gradients selected for the inversion are assembled into y0,
which defines the observation space y. The linear observation
operator H:x↦y=Hx+yf projects the control vector x into the observation
space y through the linear operator H (combining the
description of the fluxes' spatial distribution and the atmospheric transport
model) and the addition of CO2 gradients yf caused by
fluxes that are not controlled by the inversion, such as the remote fluxes
characterised by the CO2 boundary conditions of the regional
transport model. The uncertainties in xb and the
observation errors, i.e. errors in the measurements y0 and from the
observation operator H, are assumed to have unbiased Gaussian
distributions and are characterised by the prior uncertainty covariance
matrix B and the observation error covariance matrix R.
xa, the optimal posterior estimate of x, knowing
its prior estimate xb and measurements y0, can be
obtained from e.g.
xa=xb+(B-1+HTR-1H)-1HTR-1(y0-yf-Hxb).
The uncertainty in xa has unbiased Gaussian distribution
and is characterised by the posterior uncertainty covariance matrix
A:
A=(B-1+HTR-1H)-1.
The inversion uses measurements during a given 30-day period to derive fluxes
during the same 30-day period. Independent inversions are made for 12 consecutive
30-day periods starting on 1 August 2010 to cover the entire
observation period from August 2010 to July 2011. The 6 h mean
inverted emissions during each period serve as the basis for the analysis of
emissions in the Paris area at the monthly scale. Even though these 30-day
periods do not correspond exactly to the calendar months, the names of the
calendar months are used to label them.
We briefly recall descriptions of the components of
Eqs. ()–() as laid out by in
the next sections (Sect. –). As
detailed in Sect. , two modifications, however, are
brought to the definition of the observation space y and thus to the
observation operator H. The modifications result in two new
inversion configurations that are denominated initial i.e.
close to and reference configuration hereafter.
Section presents the set-up of the sensitivity tests,
where the prior estimates of the control variable, xb, and
components of the observation operator H are modified with
respect to the reference configuration.
Control vector x and the prior estimate of the flux budgets xb
x contains 6 h mean fossil fuel CO2 emission budgets
for windows 00:00–06:00, 06:00–12:00, 12:00–18:00, 18:00–24:00 (local
time is used hereafter) for each day for the Île-de-France region. Most of
the emissions in this region are concentrated in the urban agglomeration of
Paris (Fig. b). Thus, this choice of x approximately
consists in controlling the emission budget of this urban area. x
also contains 30-day mean biogenic CO2 fluxes for each of the four
6 h windows of the day (00:00–06:00, 06:00–12:00, 12:00–18:00,
18:00–24:00) for nine areas that make up the northern France modelling domain,
including one that encompasses the Paris region (see Fig. ).
The inversion optimises the diurnal cycle of both the fossil fuel CO2
emissions and biogenic fluxes through resolving these fluxes for the
different 6 h windows of the day. However, it controls the day-to-day
variability of the fossil fuel CO2 emissions but not the one of the
biogenic fluxes. The inversion controls scaling factors of the flux budgets
provided by the emission inventories and the ecosystem model simulations
through the linear part of the observation operator (H; see
below). For the sake of simplicity we state hereafter that it controls the
flux budgets themselves.
(a) Map of the northern France modelling domain with the
monitoring sites (red crosses) and the boundary of the Île-de-France
region (green line). Model grid (black lines): 2 km ×2 km
spatial resolution in the centre, 2 km ×10 km and
10 km × 10 km spatial resolution in the surroundings.
(b) Fossil fuel CO2 emission for January 2011 for each grid
cell of the Île-de-France region according to AIRPARIF 2008.
The initial and reference inversion configurations use our best available
knowledge of the flux budgets – the AIRPARIF 2008 inventory (since the
AIRPARIF 2010 monthly mean budgets were not available for this study) and the
C-TESSEL simulation – to define the prior estimate xb. The
sensitivity tests, described in Sect. , investigate the
impact of using different prior estimates for the Paris fossil fuel
CO2 emissions.
Configuration of the observation vector y and measurement vector y0
The specific definition of the observation space y and of the
corresponding measurement vector y0 depends on the measurement
availability, on the range of wind directions used to select gradients, and
on the meteorological forcing of the CO2 transport model. Two
different meteorological products are used to define the wind direction for
the gradient selection (see forward Sect, ).
Afternoon (12:00–16:00) CO2 data availability during the
CO2-MEGAPARIS project (August 2010–July 2011) for the different
monitoring sites used in this study. Grey vertical lines: available hourly
observed data. Red and blue lines: observations that are actually assimilated
when using the reference (stringent) gradient selection criteria. Selection
when using ECMWF (red lines) and Meso-NH/TEB (blue lines) wind fields for the
wind estimation.
The three monitoring sites are located roughly along a NE–SW direction at
the edges of the urban area in mixed urban–rural environments
(Fig. ) at heights of 9 m (MON), 4 m (GON)
and 7 m (GIF) above ground. The NE–SW direction corresponds to the
dominant wind directions in Île-de-France. Technical details about the
measurements are given in , and
. Here, we briefly summarise the main aspects. The
CO2-MEGAPARIS sites GON and MON were equipped with a ring-down cavity
analyser from Picarro (model G1302), while an automated gas chromatograph
analyser Agilent HP6890; see was used at GIF. All
measurements are quality-controlled and calibrated against the World
Meteorological Organization mole fraction scale WMO-X2007 .
The instrumental reproducibility of the Picarro 5 min averages is
better than 0.17 ppm, while measurement accuracy is estimated at
0.38 ppm . The precision of the chromatograph
analyser in GIF is estimated at 0.05 ppm for 5 min averages
. In our study, we bin measured CO2 data into
1 h means. The accuracy for these hourly means is better than
0.4 ppm at the three sites which is negligible compared to the
modelling uncertainties (see forward Sect. ).
Figures and – in the
Appendix illustrate the temporal coverage of the measurements available
during the CO2-MEGAPARIS period (August 2010–July 2011) at each
measurement site. They also show which data are used to finally form
gradients. Some significant data gaps can be noticed, e.g. during June 2010
and 2011 at GON, September 2010 at MON, and January, November and December 2010
at GIF. The regular 1-day gaps correspond to instrument calibrations.
CO2 at the measurement sites is significantly influenced by both the
Paris urban emissions and the remote fluxes (i.e. by fluxes outside the
modelling domain, whose influence is simulated by the transport of the
CO2 conditions imposed at the model boundaries, and by biogenic and
fossil fuel fluxes within the modelling domain but outside the Paris urban
area). It is assumed, that, due to atmospheric diffusion, the signature of
the remote fluxes upwind of the city on the concentrations in our domain has
horizontal and vertical spatial scales and a temporal scale of variability
that are large enough so that it does not evolve during the transit of an air
parcel above the city. In other words, it is assumed that the remote fluxes
do not cause CO2 gradients between downwind and upwind stations when
the wind blows from the upwind to the downwind sites. This critical
assumption is supported by the fact that the simulated CO2 gradients,
caused by remote fluxes, are negligible. However, this does not necessarily
imply that the measured gradients are not influenced by the actual fluxes
. This assumption is also supported by the much better fit
between observed and modelled CO2, when observations are defined by
cross-city gradients instead of CO2 mixing ratios at individual sites
. By assimilating CO2 gradients rather than individual
CO2 mole fractions, we thus expect to prevent the inversion from
being sensitive to the uncertainties in the estimate of the remote fluxes.
Local sources in the vicinity of the measurement sites are difficult to
represent in the model. In order to limit their impact,
selected gradients only if the wind speed is above a given threshold of
2 ms-1. Similar to most inversion studies that used rural
measurement sites e.g.,
assimilated data only during the afternoon, since the model seemed to poorly
represent vertical transport during other periods of the day. Specifically,
used differences in simultaneous hourly-averaged CO2
measurements between the peri-urban stations during the afternoon
(12:00–16:00) to define the measurement vector y0. When (at a given
hour) the wind at GIF, given by the meteorological simulation (see below
Sect. ), is from the SW, i.e. from 160 to 260∘, and is
above 2 ms-1, GIF is the upwind site; and
assimilate hourly CO2 mole fraction differences between MON and GIF
and between GON and GIF. When the simulated wind at MON is from NE, i.e. from
0 to 135∘ and exceeds 2 ms-1, MON is the upwind site and
assimilate the CO2 differences between GON and MON
and between GON and GIF.
Using this configuration, assimilated CO2 gradients
between GON and MON – stations that are only separated by a short distance.
The enhancement of CO2 between these two sites therefore reflects the
emissions from a small portion of the north-eastern suburbs of Paris than
emissions from the entire urban area. Model–data misfits for such gradients
relate far more to the uncertainties in the high-resolution mapping of the
emissions than to uncertainties in the budget of the city emissions. In
addition, these gradients are strongly affected by emissions from the
Charles-de-Gaulle airport which is located between the two sites and an
important local source of CO2 that is not representative of the main
CO2 sources in the Paris urban area. Thus, gradients between GON and
MON are not adapted to constrain city-scale emissions. Furthermore, in order
to retain a significant fraction of measurements in the inversion,
used a loose range of wind directions to define upwind and
downwind conditions. This loose range could allow the assimilation of
gradients when air masses leaving the upwind site or reaching the downwind
site hardly cross a significant portion of the Paris urban area, or, more
generally, when air masses are not really transported from the upwind to the
downwind site. This loose selection of gradients for constraining fluxes was
not identified as a major source of systematic error. Through this
configuration, primarily aimed at decreasing the impact of
remote fluxes on CO2 mole fractions while keeping a large amount of
data for the inversion. Both choices, the assimilation of GON–MON and the
loose wind filtering to select gradients, however, lead to estimates of
spatially integrated emissions of the city constrained by measurements that
are influenced only by emissions from a small fraction of the city. This
would not be an issue if the spatial distribution of emissions provided by
AIRPARIF was perfectly accurate. On the other hand, any significant error in
the emissions' spatial distribution may induce a large error on the city-wide
emission inversion. Indeed, if the assumed spatial distribution of the
emission bears significant errors (which is likely the case), the inversion
corrections, driven by model–data misfits due to errors in emissions from
a small part of the city, will become inconsistent with the errors at the
city scale, raising large, so-called aggregation errors .
As mentioned in the introduction, preliminary tests of inversions using the
configuration of for the period August 2010–July 2011
demonstrated the need for an improved configuration where the selection of
CO2 conforms better with our assumptions on gradients. In this study,
two critical changes are applied. They consist in assimilating GIF–GON and
GIF–MON and in discarding GON–MON gradients when the wind is from NE.
Furthermore, a stricter (narrower) range of wind directions in selecting
CO2 gradients is used. Discarding GON–MON gradients suppresses the
large amount of negative gradients in the measurement vector y0. The
impact of discarding GON–MON gradients on the inversion results is not
analysed deeper in the following. It relates to specific details of the Paris
network configuration. Here, we focus on the impact of using a narrower
range of wind directions for the gradient selection. The stricter selection
of wind directions consists in assimilating a gradient between two sites only
if the modelled wind at the upwind site is within ±15∘ of the
transect between the downwind and upwind site. The specific choice of
±15∘ is somewhat arbitrary. On the one hand it ensures the
selection of a significant number of gradients. On the other hand it ensures
that air masses leaving the upwind site or reaching the downwind site are
transported over a large part of the urban area and in a direction that is
close to the transect between downwind and upwind sites. Thus, the gradients
GIF–GON, GIF–MON, MON–GIF and GON–GIF are assimilated only if the wind is
from 20–50∘, 35–65∘, 215–245∘ and 200–230∘, respectively.
We use the term SW gradients for the gradients GON–GIF and MON–GIF
and NE gradients for the gradients GIF–MON and GIF–GON.
We apply other significant changes to the gradient selection criteria of
. First, we increase here the minimum wind speed threshold at
the upwind site from 2 to 3 ms-1. This change is driven by the
fact that, as noticed by , large model–data misfits persist
after inversion for wind speeds close to 2 ms-1. This suggests
that a threshold of 2 ms-1 was not sufficient in avoiding
a large contamination of the measurements by poorly modelled local sources.
Furthermore, in this study, a single valid 1 h mean gradient during
a given afternoon is not selected for the inversion. This avoids constraining
the emissions of a given day based on a single observation that potentially
bears a large transport model error. At last, an analysis of the impact of
individual observations on the corrections applied by the inversion to the
prior monthly flux estimates (i.e. impact of the product between the
gain matrix K=(B-1+HTR-1H)-1HTR-1
and the model–data misfit for each individual gradient; see , for more
details) was conducted for the initial and
reference inversion experiments. It revealed that, for the initial inversion,
during November, two gradients had far more impact on the correction to the
emissions budget of this month than the other gradients. The gradients
removed had both an impact of approximately -0.3 MtCO2 on this
budget (i.e. approximately -0.6 MtCO2 in total). In both cases,
these high impacts were connected to high prior model–data misfits during
weak vertical mixing episodes. Again, similar to many inversion experiments
e.g,, in order to avoid giving too much weight to
individual measurements the two corresponding gradients were removed from the
initial inversion experiment. However, such gradients are not selected by the
tighter wind direction filtering of the reference inversion.
Three configurations of the observation space y are used in this
study: yini, yref and
ylag. The first one, yini, corresponds to
the initial inversion configuration. It includes all the new options
discussed above, except the narrowing of the wind direction ranges for the
gradient selection. The selection of GIF–GON, GIF–MON, MON–GIF and GON–GIF
gradients in yini is based on the wind direction ranges at
GIF and MON as proposed by . The second one,
yref, corresponds to the reference inversion configuration.
It includes all the new options and selects gradients based on the new wind
direction ranges at GIF, GON and MON defined in this section. The comparison
between the initial and reference inversions is used to assess the impact of
using tight wind direction ranges on retrieved emissions and to evaluate if
the selected gradients now conform better with our assumptions (see
Sects. –). ylag is
only used for a single experiment whose results are briefly discussed in
Sect. . This observation vector consists in
spatio-temporal gradients, i.e. mole fraction differences between a downwind
site at a given time and an upwind site 2 h before. Given a mean wind
speed of 7 ms-1 in the lower planetary boundary layer of the
Paris area during the afternoon over the 1-year period and a distance of
about 40 km between the upwind and downwind site, the typical time for
air being transported from the upwind to the downwind site is approximately
2 h. The wind selection in this experiment is similar to that of the
reference experiment. It uses simulated wind fields at the time of the upwind
mole fraction measurement involved in the gradient. At a given site the
assimilation window is also reduced so that a given gradient does not involve
any measurement outside the 12:00–16:00 window, despite the use of a time
lag in the gradient. The use of spatio-temporal gradients instead of spatial
gradients appears more in line with the concept of a mass balance approach
which constrains emissions based on mole fraction variations in air parcels
that are transported over the Paris area. Due to atmospheric diffusion and
variations in the planetary boundary layer, the spatio-temporal gradients
still need to be interpreted using a high-resolution transport model.
However, with such a configuration of the observation vector, the number of
data that can be assimilated is further decreased as the assimilation window
at both the upwind and downwind sites is reduced.
Observation operator H
This section describes the observation operator H:x↦y=Hx+yf. The linear operator
H can be decomposed into three operators
(H=HsampHtransHmap)
consisting in the fluxes' spatio-temporal distribution
(Hmap), the atmospheric transport simulated using CHIMERE
(Htrans), and the sampling of simulated 4-D CO2
field like the observations (Hsamp). yf
gathers influences on the gradients which are not controlled by the
inversion, such as the signature of the model boundary conditions. In the following, we
present the implementation of these operators and vectors used in
and the initial and reference inversions of this study,
respectively, as well as alternative options used for sensitivity tests.
Atmospheric transport modelling and sampling
The atmospheric transport Htrans and the signature of
sources and sinks on CO2 concentrations that are not controlled by
the inversion (yf) are modelled using a northern France
configuration of CHIMERE. It has a 2 km × 2 km
spatial resolution for the Paris region, and
a 2 km × 10 km and
10 km × 10 km spatial resolution for the
surroundings (see Fig. a). It has 20 vertical hybrid
pressure-sigma (terrain-following) layers that range between surface and the
mid-troposphere, up to 500 hPa. In the initial and reference
inversion of this study, as in , CHIMERE is driven by
operational analyses of ECMWF's Integrated Forecasting System, available at
approximately 15 km × 15 km spatial resolution and
3 h temporal resolution. In this case we will denote
Htrans=HECMtrans and
yf=yini-ECMf,yref-ECMf or ylag-ECMf
depending on the type of gradient selection used.
conducted meteorological simulations on our modelling domain
using a 2 km × 2 km resolution configuration of the
non-hydrostatic mesoscale model Meso-NH. Meso-NH, jointly developed by
Météo-France and Laboratoire d'Aérologie , is coupled to
3-hourly analysed meteorological fields from AROME-France (Application of
Research to Operations at Mesoscale) and to the
land–surface–atmosphere interaction model SURFEX . SURFEX
includes the urban and vegetation scheme TEB . Therefore, in
contrast to the ECMWF meteorological forcing, Meso-NH/TEB includes some urban
parameterisation, which may have a large impact on the transport over the
city. showed, by comparison to lidar systems operated on
a short-term basis in the Paris area, that Meso-NH/TEB captures the diurnal
cycle of the boundary layer height relatively well, as well as the
differences in this height between peri-urban and urban locations.
A test of sensitivity is conducted to assess the impact of the uncertainties
in the meteorological product on the inversions (in particular the
uncertainties in the wind and in the boundary layer height). The
meteorological product is used to drive the atmospheric transport model and
to select the cross-city gradients. This test consists in using hourly mean
outputs of Meso-NH/TEB to drive CHIMERE and to select gradients. Meso-NH/TEB
simulations, originally conducted over a slightly different grid
seetheir Fig. 1a., are interpolated onto the CHIMERE grid.
When using Meso-NH/TEB, Htrans and yf are
denoted by HMNHtrans and
yref-MNHf, respectively.
In order to build the linear part of the observation operator H
and yf, the operator Hsamp is applied.
Hsamp extracts the selected gradients between the
monitoring sites from the simulated 4-D CO2 mole fraction fields as
described in Sect. . The underlying selection of the
horizontal and vertical positioning of the monitoring sites in the CHIMERE
grid is the same as in . Because the gradient selection
depends on modelled wind speed and direction, the observation space y
and thus yf and Hsamp depend on the
meteorological simulations (ECMWF or Meso-NH/TEB). We denote
Hsamp by Hini-ECMsamp,
Href-ECMsamp,
Hlag-ECMsamp or
Href-MNHsamp depending on the inversion cases.
Emissions outside Île-de-France and model boundary conditions
yf encompasses the signature of fossil fuel CO2
emissions outside the Paris region but within the modelling domain, that of
the modelling domain's CO2 boundary conditions and that of the 30-day
simulations initial conditions. The signature of the emissions outside the
Paris region but within the modelling domain are simulated by CHIMERE using
fossil fuel CO2 emissions from the EDGAR database .
Daily CO2 mole fraction fields provided by the global inversion of
are used as CO2 boundary conditions at the lateral
and top edges of the modelling domain and as initial conditions for the
CO2 mole fraction fields at the beginning of each 30-day period. The
global inversion of is based on the simulation of the
CO2 transport by the LMDZ model and on the
assimilation of ground-based measurements from a global network.
Mapping of the Paris fossil fuel CO2 emissions and biogenic fluxes
Hmap is built on hourly biogenic flux and emission maps at
the horizontal resolution of the CHIMERE transport model. In both the initial
and reference inversions, as in , the description of the Paris
fossil fuel CO2 emissions at 1 h and
2 km × 2 km resolution in Hmap is
based on the hourly AIRPARIF 2008 inventory. The temporal profiles and
spatial distributions of this inventory are analysed in . We
just recall that emissions are available at 1 h and
1 km × 1 km resolution for three typical days
(weekday, Saturday, Sunday) of five typical months (January, April, July,
August, October) of the year 2008. In order to build hourly estimates for the
1-year period August 2010–July 2011, we follow AIRPARIF's recommendation and
use January emissions for all five months from November to March, April data
for all three months from April to June, and October data for both September
and October and, for a given day in 2010 or 2011, we use the values from the
same day in 2008.
For sensitivity tests (see Sect. ), the emission
component of Hmap is alternatively built based on
a national emission inventory for 2005 compiled by the Institut für
Energiewirtschaft und Rationelle Energieanwendung (IER) of the University of
Stuttgart, Germany. disaggregated reported emission totals for
France for 2005 into a 1 × 1 arcmin grid with the use of various
proxies for the distribution of emitting activities such as population
census, traffic intensity and land cover. We used monthly, weekly and hourly
temporal profiles for different emission sectors from the IER inventory for
Europe as described by to disaggregate annual emissions to
hourly emissions. The IER and AIRPARIF emission inventories are two largely
independent datasets.
In all experiments, the component in Hmap that corresponds
to the biogenic control variables is based on net ecosystem exchange
simulated by C-TESSEL at 3 hourly and 15 km × 15 km
resolution. The simulated net ecosystem exchange is interpolated hourly onto
the CHIMERE grid (at 2 to 10 km resolution). The C-TESSEL model does
not have a specific implementation for urban ecosystems and due to its
moderate horizontal resolution, it is not expected to provide a precise
representation of biogenic fluxes within the urban area and within its
vicinity. However, as reminded in the introduction , the signal from C-TESSEL
in the CO2 gradients between the peri-urban sites simulated by
is low. Therefore, the natural fluxes are not expected to
critically affect the inversion of fossil fuel CO2 emissions in our
study (see forward Sect. ). We denote
Hmap by HAPmap if the hourly
fossil fuel CO2 flux maps are built using AIRPARIF 2008; and by
HIERmap if the hourly fossil fuel CO2
flux maps are built using IER.
Building the H matrix
In order to apply Eqs. () and (), H
is built based on the different operators described above. Each column of
H corresponds to the response of the selected CO2
gradients to a control variable. Each column of this matrix is computed by
applying the H operator (i.e. the series of operators described
above) to a control vector containing only zeros except for the corresponding
control variable which is set to 1.
Let nx denote the number of control variables (156 elements) for a given
month of inversion, nf the dimension of the 3-D flux field in the
input of the CHIMERE model (i.e. the number of model horizontal grid cells
times the number of hours during 1 month of inversion, i.e,
118 × 118 × 720), nc the dimension of the 4-D
field of CO2 in output of the CHIMERE model (i.e. number of model grid
cells times the number of hours during 1 month of inversion, i.e.
118 × 118 × 20 × 720), and, at last, ny the
number of gradients selected for a 1-month inversion. The dimension of
H is nx×ny, while the dimension of
Hmap is nx×nf, of
Htrans is nf×nc and of
Hsamp is nc×ny. nx application of
HsampHtransHmap are
needed to build the H matrix. Once the H is built,
since both nx and ny are relatively small, we can easily afford the
computations in Eqs. () and () which involve
the inversion of matrices of size nx×nx or ny×ny and
multiplication of such matrices with H.
Prior error covariance matrix B
We set up the prior error covariance matrix B as in
. Assuming that there is no correlation between the
uncertainties in the fossil fuel CO2 emissions and the uncertainties
in the biogenic fluxes, B is modelled as a diagonal block matrix
with two blocks: one corresponds to the uncertainties in the Paris fossil
fuel CO2 emissions, and the other one to the net ecosystem exchange
in the modelling domain. For each block, we make separate assumptions on the
variance of the uncertainty in the individual control variable on the one
hand and on the temporal and spatial correlations between these uncertainties
on the other hand.
Regarding the Paris fossil fuel emissions, we assume a 50 % relative
uncertainty (in terms of standard deviation) in the prior estimates of
individual 6 h emission budgets. We assume that we can decompose
these prior uncertainties for a given month into uncertainties in the mean
diurnal cycle of the emissions and into uncertainties in the day-to-day
variations of the emissions. Therefore, we compute the temporal
autocorrelations of the prior uncertainties in the 6 h emission
budgets, ct(t1,t2) (where t1 and t2 are two 6 h windows of
2 days of the month of inversion), as the product of the correlations of the
uncertainties in the mean diurnal cycle between the four 6 h windows
of the day, cw(w1,w2) (where w1 and w2 are the two 6 h
windows of the day corresponding to t1 and t2 respectively), and
correlations of uncertainty in the day-to-day variations between different
days, cd(d1,d2) (where d1 and d2 are the 2 days corresponding to
t1 and t2). We assume that the correlations of the uncertainty in the
mean diurnal cycle between the 6 h windows of the day are positive:
cw(w1,w2)=0.4 for two consecutive windows (for example,
w1=00:00–06:00 and w2=06:00–12:00) and cw(w1,w2)=0.2 for two
non-consecutive ones (for example, w1=00:00–06:00 and
w2=12:00–18:00). The correlations of uncertainty in the day-to-day
variations between different days are modelled using an exponentially
decaying function with a characteristic time of 7 days:
cd(d1,d2)=e|d2-d1|7.
The standard deviation of the prior uncertainty in the 30-day budgets of net
ecosystem exchange for a given area and 6 h window of the day is
assumed to be about 75 % of the prior estimate of this budget from
C-TESSEL. In practice, from our computations, it appears that the resulting
value of this uncertainty decreases when the surface of the corresponding
area increases. Spatial and temporal correlations between the uncertainties
for the various 6 h windows of the day and areas are assumed to be
negligible due to the large size of the corresponding areas and due to the
differences in the processes dominating the ecosystem exchanges between
daytime and night-time.
Observation error covariance matrix R
The observation errors encompass instrumentation errors and errors in the
observation operator H. The latter combines transport model
errors, representation errors, aggregation errors, errors from the boundary
conditions and errors from the emissions in the modelling domain but outside
the Paris area. One of the main sources of transport errors is linked to
errors in the wind and planetary boundary layer height in the meteorological
forcing of the transport model. Representation errors are associated with the
variations of CO2 within the 2 km × 2 km
horizontal resolution grid cell of the model which encompass the peri-urban
sites. They should be relatively small since there is no major CO2
source in these grid cells. Aggregation errors are mainly associated with
uncertainties in the spatial and temporal distribution of the emissions
within the Paris area and 6 h windows. Aggregation errors are
critical to account for in our inverse modelling system given that for
a given 6 h window, we control one scaling factor for the emissions
over the whole Paris area.
used the diagnostics of to estimate the
variances of the observation error. It was assumed that these errors have the
same statistics for any hourly gradient and that there is no correlation
between errors for different hourly gradients. Their corresponding estimate
of the standard deviation of the observation error was 3 ppm. Here, since
a similar inverse modelling framework is used, and even though the revised
gradient selection should decrease the aggregation errors (see
Sect. ), we assume that our misfits between the measured
and modelled gradients bear the same observation error as in this study.
R is thus modelled as a diagonal matrix with a (3 ppm)2
variance for all elements in the diagonal.
Principles of the sensitivity tests
Several tests are conducted to check the reference inversion results'
sensitivity to changes in different components: (a) the prior estimate of the
Paris fossil fuel CO2 emissions, (b) the spatio-temporal distribution
of the fossil fuel CO2 emissions within the Paris region and within
a given 6 h window, (c) the meteorological forcing driving both the
atmospheric transport model and the selection of the observations. These
changes are representative of typical uncertainties in these components. Most
of these uncertainties are, in principle, accounted for in the configuration
of the prior uncertainty covariance matrix B and observation error
covariance matrix R, respectively. Their impact on the robustness
of the inversion results should be given by the posterior uncertainty
covariance matrix A. However, they may not be correctly reflected
by the statistical representation that is based on Gaussian and unbiased
distributions and by the rather simple models used to set up the covariance
matrices. Therefore, the sensitivity tests provide a useful alternative
evaluation of the robustness of the inversion results.
Regarding the prior estimate of the Paris fossil fuel CO2 emission
budgets, as an alternative to the AIRPARIF 2008 budgets, we use what is from
hereafter called flat priors, i.e. prior fossil fuel CO2 emission
estimates that are not informed about month-to-month variations. Three sets
of flat priors are built by rescaling the AIRPARIF 2008 budgets using
monthly, daily or 6 h scaling factors. In the first case, the flat
priors have constant monthly values, but retain the relative temporal
variations of the 6 h budgets within a month. In the second case the
flat priors have constant daily values, but retain the relative temporal
variations of the 6 h budgets within a day. In the third case, the
flat priors have constant 6 h mean values. This change of prior
estimate can potentially have a large impact on the results since the system
assimilates data only during the afternoon. Consequently, such a change
imposes a direct constraint on two 6 h windows of the day only (the
06:00–12:00 and the 12:00–18:00 windows) while the constraint on the other
two windows (the 00:00–06:00 and the 18:00–24:00 windows) relies indirectly
on the description of the temporal correlations in the prior uncertainty
covariance matrix B.
Summary of the different inversion configuration and
Île-de-France (IdF) annual fossil fuel CO2 emissions from
different inventories and inversion results. Priors that are flat at the
monthly, daily and 6-hourly scale are denoted M, D and H, respectively (see
Sect. for the details). Posterior estimates are derived
from inversions using the operator and prior estimate indicated in the
corresponding line of the table.
Inversion
Acronym
H
yf
xb
IdF annual
fossil fuel CO2
emissions (MtCO2)
Hsamp
Htrans
Hmap
Prior
Post
Initial
ini
Hini-ECMsamp
HECMtrans
HAPmap
yini-ECMf
AP08
51.9
37.4
Reference
ref
Href-ECMsamp
HECMtrans
HAPmap
yref-ECMf
AP08
51.9
40.9
Sensitivity
FLAT_4.3H
Href-ECMsamp
HECMtrans
HAPmap
yref-ECMf
H
51.9
47.1
Tests
FLAT_4.3D
D
51.9
41.1
FLAT_4.3M
M
51.9
41.4
FLAT_3.0H
H
36.0
37.1
FLAT_3.0D
D
36.0
33.0
FLAT_3.0M
M
36.0
33.0
FLAT_5.0H
H
60.0
52.2
FLAT_5.0D
D
60.0
45.3
FLAT_5.0M
M
60.0
45.3
INV_mapIER
Href-ECMsamp
HECMtrans
HIERmap
yref-ECMf
AP08
51.9
39.0
INV_IER
Href-ECMsamp
HECMtrans
HIERmap
yref-ECMf
IER
60.1
45.5
INV_MNH
Href-MNHsamp
HMNHtrans
HAPmap
yref-MNHf
AP08
51.9
*
Time lag
lag
Href-ECMsamp
HECMtrans
HIERmap
yref-ECMf
AP08
51.9
46.8
Emissions for the year 2010 as given by
41.8
* Meso-NH/TEB data are only available up to June 2011.
For each set, different flat priors are tested by taking different values for
the monthly budgets. These values cover a case of relatively high emissions
(5 MtCO2month-1), a case of relatively low emissions
(3 MtCO2month-1), as well as an intermediate case
corresponding to the annual budget from the AIRPARIF 2008 inventory
(4.3 MtCO2month-1). A prior estimate based on the budgets
from the IER inventory is also used for the sensitivity tests. As explained
in Sect. , HIERmap is used as
an alternative Hmap to HAPmap,
while Href-MNHtrans,
yref-MNHf and
Href-MNHsamp are used as an alternative
Htrans, yf and Hsamp to
HECMtrans, yref-ECMf,
Href-ECMsamp. Table summarises
the acronyms and settings of the different sensitivity tests.
Annual and seasonal bias, standard deviation (SD), root mean square
error (RMSE) and coefficient of determination (r2) of prior model–data
misfit and posterior model–data misfit for the initial inversion (experiment
ini) and the reference inversion (experiment ref),
respectively. JJA denotes July, August and September; SON
September, October and November; DJF December January and February; MAM March, April and
May. All values, except for r2, are given in ppm.
Bias
SD
RMSE
r2
ini
ref
ini
ref
ini
ref
ini
ref
prior
post
prior
post
prior
post
prior
post
prior
post
prior
post
prior
post
prior
prior
Annual
2.50
0.33
3.04
0.36
3.60
2.21
3.77
2.20
4.38
2.23
4.84
2.22
0.53
0.80
0.53
0.81
JJA
2.20
0.23
2.70
0.39
2.31
1.59
2.54
1.62
3.18
1.60
3.70
1.66
0.13
0.45
0.03
0.34
SON
2.41
0.28
3.73
0.38
3.49
1.98
2.95
2.05
4.23
2.00
4.74
2.07
0.35
0.75
0.27
0.61
DJF
2.35
0.48
2.79
0.44
4.21
2.51
4.55
2.29
4.82
2.55
5.33
2.32
0.61
0.84
0.55
0.85
MAM
3.01
0.26
3.29
0.20
3.65
2.38
4.01
2.63
4.72
2.39
5.17
2.62
0.22
0.56
0.07
0.50
Results
analysed the skill of the inversion by comparing the fit
between measured and modelled CO2 gradients, which is a first
indicator of the reliability of the inverted emissions. In
Table , statistical comparisons between the selected measured
gradients and results from the initial and reference inversion, respectively,
are provided. It demonstrates that both inversions strongly increase the
consistency between model and measurements compared to the prior simulations
(Table ). Of note is that the statistics of both the prior
and posterior model–data misfits are smaller for the initial inversion than
for the reference inversion. This is explained by the fact that the initial
inversion selects gradients for which the signal (mainly the impact of the
city emissions) is smaller than for the gradients that both inversions
select. However, we avoid a more detailed analysis of the model–data misfit
in the following.
Using the loose wind ranges of the initial inversion to select gradients, the
root mean square of the biogenic signal in these hourly gradients, averaged
over the 1-year CO2-MEGAPARIS period, is 1.1 ppm, which is, as
indicated in the introduction, much lower than the signal from the Paris
emission. When using the tighter wind ranges of the reference inversion, the
root mean square of the biogenic signal in these gradients is even smaller
(0.8 ppm). Therefore, the changes in the inversion configuration
proposed in our study decrease the impact of the uncertainty in the natural
fluxes. This weak impact was already demonstrated by . We thus
do not analyse further the results that correspond to the control of the
natural fluxes in the following.
Rather, the presentation of the results focuses on the estimates of the
monthly fossil fuel CO2 emission budgets from mid-2010 to mid-2011,
expressed in MtCO2month-1 (strictly speaking MtCO2 per
30 days; see Sect. ). The uncertainties (in terms of
standard deviation) in prior and posterior estimates of the monthly emissions
are based on the modelling of the B matrix, described in
Sect. , and on the derivation of the A matrix
(Eq. ). However, the robustness of the results is evaluated
using independent knowledge of the emissions rather than using these
theoretical indicators. reports an estimate of
41.8 MtCO2yr-1 for the annual emission budget of
Île-de-France in 2010. This number is used as an indicator for the
evaluation of the 1-year budget of the estimates from the inversion.
According to , the residential and the service sector account
for 43 % of the Paris fossil fuel CO2 emissions in 2010. These
emissions are almost entirely linked to heating. Heating in the industry
sector also contributes significantly to Paris' emissions. The heating is
mostly dedicated to ambient air in the buildings and to sanitary water.
Therefore, we expect a large increase in the emissions from summer to winter
and a high correlation between these emissions and the temperature during the
cold season. An independent analysis of both daily gas use and hourly
electric consumption within Île-de-France indicates a heating energy use
that is highly correlated to the daily mean temperature when this temperature
is below 19 ∘C, and essentially independent of the daily mean
temperature when this temperature is above 19 ∘C (unpublished
analysis led by a co-author of this study, François-Marie Bréon). For
the evaluation of the results, our emission estimates are thus compared with
monthly averages of an independent measure, which we call heating degrees
hereafter. It is defined as the positive difference between the daily mean
temperature and 19 ∘C (set to 0 for days when the temperature is
higher than 19 ∘C). The ratio between January and July emission
estimates from the AIRPARIF 2008 inventory seem surprisingly low given these
considerations. Furthermore, the prior estimates based on this inventory make
use of a single emission value from November to March, which does not account
for the large temperature variations during this period. Therefore, an
amplified seasonal cycle of the emissions that better correlates with heating
degrees is expected through the atmospheric inversion.
Monthly fossil fuel CO2 emissions prior and posterior
estimates from inversions in MtCO2. Left-hand panels: prior estimates
(dashed grey line) ± the standard deviation of uncertainties (shaded
area) and posterior estimates (green line) ± the standard deviation
(green bars). (a) Results using the initial inversion configuration.
(b) Results using the reference inversion configuration. Monthly
mean heating degrees (dashed black line) for the centre of Paris as obtained
from ECMWF's operational analysis. Numbers at the top show CO2 mole
fraction gradients assimilated for SW or NE winds, respectively. Prior and
posterior annual emission estimates are displayed in the bottom left-hand
corner of each panel. Right-hand panels: results using the
(c) initial and (d) reference configuration of the
inversions but assimilations of only subsets of selected gradients (see
Sect. .) Colour-coded numbers at the top are those of the
assimilated gradients. Prior estimates (dashed grey line) ±
the standard deviation of uncertainties (shaded area). Symbols are shifted
slightly to prevent overlap.
Initial inversion using loose wind direction criteria for the gradient selection
Figure a shows prior and posterior estimates of the Paris
emissions from the initial inversion (experiment ini,
Table ). Monthly mean heating degrees for the centre of
Paris, derived from the temperature given by the ECMWF's operational
analysis, are also shown in this figure. The posterior estimates are lower
than the prior ones for all months. The inversion decreases the annual
emissions from 51.9 to 37.4 ± 2.1 MtCO2yr-1. This
number is smaller but closer to the AIRPARIF inventory 2010 used for
evaluation (41.8 MtCO2yr-1; see Table ).
Posterior fluxes are the lowest in August (1.6 MtCO2month-1),
increase steadily until they peak in February
(4.6 MtCO2month-1), drastically drop in March
(2.8 MtCO2month-1) and vary between 2 and
3 MtCO2month-1 from April to July. Compared to the prior estimate, the inversion
yields larger emissions in winter and increases the amplitude of the seasonal
variations. For the period analysed, monthly mean heating degrees were
highest in December (Fig. a). Hence, fossil fuel CO2
emissions are expected to be the highest in December rather than in February.
There is no clear correlation between monthly heating degrees and emission
estimates during the November–March period.
The number of assimilated gradients varies considerably from one month to
another, which influences the month-to-month variations of the inverted
emissions. For instance, 163 observations are assimilated in March, compared
with only 34 in November. Figure a also shows that, for most
months, the numbers of selected gradients are not apportioned equally amongst
the NE and SW wind directions. For instance, there are no NE gradients to
constrain August emissions, while less than half of the gradients in March
are SW gradients. The different upwind conditions for NE and SW gradients
could play a role in the month-to-month variability of the inverted
emissions, in case the gradient approach does not entirely remove the
influence of remote fluxes.
We investigate the impact of assimilating these two different gradient types
on monthly fossil fuel CO2 flux estimates by conducting inversions
based on NE gradients, SW gradients, or even GIF–MON, GIF–GON, MON–GIF or
GON–GIF only (Fig. b). The difference in inverted December
emissions when assimilating only SW gradients compared to assimilating only
NE gradients is large, even though a large number of both types of gradients
is available during this month. Compared to the prior estimate, the inversion
of SW gradients increases the December emissions. The opposite, however, is
true for the inversion of NE gradients. This behaviour seems to be driven by
both the assimilation of GIF–MON and GIF–GON gradients. An analysis of the
average temperature in Paris (not shown) shows lower temperatures for NE wind
conditions than for SW wind conditions. The heating emissions in Paris should
thus be higher for NE wind conditions. Therefore, the temperature variations
cannot explain the differences in December emissions between assimilating SW
gradients and assimilating NE gradients.
Time series of mean afternoon (12:00–16:00) CO2 mole
fraction differences (ΔCO2) between downwind and upwind sites
for December 2010 (upper panels) and May 2011 (lower panels). Measured
CO2 mole fraction differences (red); prior (blue) and posterior
(green) CO2 mole fraction differences. (a, c) Selection of
gradients relies on wind direction ranges of the initial inversion
configuration. (b, d) Wind direction ranges are limited to a narrow
SW–NE wind corridor across the city following the reference inversion
configuration (see Sect. ).
The differences between the results when using NE gradients or SW gradients
are not as large for other months as in December. However, they can still be
significant, e.g. April (Fig. b). These differences cannot
be explained by a lack of data for a given type of gradient, except in
August, when there are no NE gradients. For January and February, differences
of 1 to 1.5 MtCO2 (i.e. about 35 % relative differences) are
obtained between inversions assimilating only MON–GIF compared to
assimilating only GON–GIF, although both are SW gradients and gather more
than 40 observations during each month. This large mismatch between the
different inversions when using different data subsets undermines the
reliability of the inversion results, in particular in December. The seasonal
profile of the retrieved emission when assimilating only SW gradients is far
better correlated with heating degrees than when the inversion uses both SW
and NE gradients, as emissions reach their maximum in December. Results seem
nearly as sensible when using MON–GIF or GON–GIF as when using all SW
gradients.
Figure a and c illustrate that, even if we discard GON–MON
and increase the threshold of the wind speed, there are episodes when
measured gradients show negative values. They, however, should be positive
owing to the city's emissions. Most of the negative gradients are found when
the wind is from NE, such as in December (Fig. a),
suggesting that such gradients may not represent the emissions of the entire
city.
Reference inversion
The estimates of the monthly Paris fossil fuel CO2 emissions by the
reference inversion (experiment ref, Table) are
given in Fig. c. All negative observed gradients outlined
above were obtained at the limit of the wind direction range proposed by
. As illustrated for December and May by comparing
Fig. a to b and Fig. c to d, respectively,
the reference inversion, which uses a stricter range of wind directions
(Sect. ), removes the negative gradients. Despite the loss
of 65 % of the data compared to the initial inversion, the reference
inversion still predicts a large uncertainty reduction for monthly fossil
fuel CO2 emission estimates, from 9 % in November to 50 % in
October.
The Paris monthly fossil fuel CO2 emission estimates from the
reference inversion correlate well with monthly heating degrees (r2=0.67
for the whole period, r2=0.45 for November–February), which was not the
case of the initial inversion (r2=0.54 for the whole period, r2=0.07 for November–February). In general, the reference inversion decreases
the fossil fuel CO2 budget from the prior estimate, except in
December, which becomes the peak of emissions (Fig. c). The
emissions decrease from February to March, which does not correspond to
a relative change in heating degrees, is significantly smaller in the
reference than in the initial inversion. The seasonal variations of the
reference inversion are strongly improved compared to initial inversions. The
annual budget from the reference inversion (40.9 MtCO2) is close
to that from AIRPARIF 2010 (41.8 MtCO2; see
Table ).
The stricter gradient selection further leads to a much better agreement
between emission estimates when using different subsets of gradients (compare
Fig. d with b). Although significant differences in December
and April emissions are still apparent between using NE gradients or using SW
gradients, and in January and February emissions between using GON–GIF or
MON–GIF, these differences are smaller than in the initial inversion. Now,
even when assimilating only NE gradients, the four months with the largest
inverted emissions correspond to the four coldest ones with the highest
heating degrees of the year (November to February), though the assimilation
of NE gradients still leads to smaller emissions in December than in
November, January and February. One may argue that the improvements of the
reference over the initial inversion reflect the assimilation of a smaller
dataset, and therefore are due to smaller corrections. However, results from
the reference and initial inversion are closer to each other if only SW
gradients are assimilated than if only NE gradients are assimilated. The
highest correlation with heating degrees are obtained when only SW gradients
are assimilated.
The theoretical posterior uncertainties in the monthly budgets of the
emissions are generally much lower than the prior uncertainties (with more
than 30 % uncertainty reduction for most of the months) in the reference
inversion. However, even though our analysis above gives more confidence in
the results from the reference inversion than in the results from the initial
inversion, the inverse modelling system diagnoses smaller posterior
uncertainties in the latter than in the former.
Results for the Lagrangian experiment (see Sect.
and Table ). As for Fig. c, but a 2 h time
lag between downwind and upwind measurements is introduced.
Results from the inversion using a 2 h lag between the upwind and the
corresponding downwind measurement are shown in Fig. . The
corrections applied to the prior estimate from AIRPARIF 2008 by this
inversion are qualitatively consistent with those of the reference inversion.
The amplitudes of the corrections, however, are much smaller as the large
decrease in the number of assimilated gradients in this inversion compared to
the reference configuration (only 4 % of available measurements are used
by reducing the time window of eligible upwind or downwind measurements; see
Sect. ) clearly limits the weight of the observational
constraint.
Sensitivity tests
Sensitivity of the Paris emission budgets to prior estimates xb
Monthly fossil fuel CO2 emission estimates using flat priors for
xb (all experiments FLAT_ in
Table ) are reported in Fig. . Although
differences in prior monthly budgets between the FLAT_3.0, FLAT_4.3 and
FLAT_5.0 experiments amount to 2 MtCO2month-1, posterior
differences between their monthly budgets are generally much lower. In
addition, the posterior monthly emissions when using flat priors are
comparable to those from the reference inversion – with a very similar
month-to-month variation. The differences between posterior monthly emissions
from the FLAT_mM and FLAT_mD (m=3.0, 4.3 or 5.0) inversions and
the reference inversion are generally smaller than
1 MtCO2month-1, except during September, November, May and
July when very few (4 to 24) gradients are assimilated
(Fig. a and b). Larger differences are obtained between the
reference inversion and FLAT_mH (m=3.0, 4.3 or 5.0) inversions, which
use prior estimates that are flat at 6 h scale
(Fig. c). The FLAT_mH experiments yield larger
posterior monthly budgets than the FLAT_mD and FLAT_mM experiments.
Sensitivity of monthly fossil fuel CO2 emissions on
xb. Monthly fossil fuel CO2 estimates ± the
standard deviation of their uncertainties are shown for inversions that use
3 MtCO2month-1 (black), 4.3 MtCO2month-1
(red), and 5 MtCO2month-1 (blue) monthly prior emissions.
(a) Priors are flat at monthly scale (FLAT_mM, m=3.0, 4.3 or
5.0 MtCO2month-1). (b) Priors are flat at daily
scale (FLAT_mD). (c) Priors are flat at 6 h scale
(FLAT_mH); see Sect. for details. Fluxes obtained by
the reference inversion (green). Numbers at the top denote the number of
assimilated CO2 mole fraction gradients. Symbols are shifted slightly
to prevent overlap.
Posterior annual budgets from FLAT_mM and FLAT_mD inversions range
between 33 and 45.3 MtCO2yr-1, encompassing the budgets from
the reference inversion and AIRPARIF 2010 (Table ). In
particular, the inverted annual budget from FLAT_4.3M and FLAT_4.3D,
whose prior estimate have the same annual budget as the prior estimate from
the reference inversion, is equal to 41.1 MtCO2. This is very
close to the annual budgets from the reference inversion and AIRPARIF 2010.
However, the annual emission budgets from the FLAT_mH inversions range
from 33 to 52.2 MtCO2, which is biased compared to both the
reference inversion and AIRPARIF 2010.
Sensitivity of the Paris emission budgets to mapping and variations at hourly scale
Figure compares the estimates from the reference inversion,
which uses Hmap=HAPmap, to the
estimates from the sensitivity test with
Hmap=HIERmap (INV_mapIER and
INV_IER; see Table ). Thus, this experiment also includes
results when using the IER inventory to build both the 6 h budgets in
xb and
Hmap=HIERmap (INV_IER). It
provides estimates when the inversion relies entirely on the IER inventory to
define these parameters and ignores the existence of the AIRPARIF inventory.
This situation is similar to that in cities, where no local inventory is
available. We have less confidence in the posterior estimates from such an
inversion, since the IER inventory does not rely on the same amount of local
data as the AIRPARIF inventory.
INV_mapIER regularly predicts lower monthly budgets than the reference
inversion, except in June, July, September and November. The corresponding
differences are relatively small and do not exceed
0.5 MtCO2month-1, except in January and February. Similar to
the reference inversion, INV_mapIER predicts the highest emissions in
December. However, its estimates for January and February fluxes are
particularly low, e.g. January estimates (3.9 MtCO2month-1)
roughly equal that for May (3.9 MtCO2month-1) or October
(3.8 MtCO2month-1), and are smaller than that for September
(4.1 MtCO2month-1). This results in an annual budget of
39 MtCO2yr-1 that is still closer to the one from AIRPARIF
2010 than to the one from AIRPARIF 2008 (Table ).
The monthly prior emissions from AIRPARIF 2008 and IER differ substantially.
In particular, from November to May, the IER inventory estimates up to
3 MtCO2month-1, (approximately 40 %) higher fossil fuel
CO2 emissions for the Paris region than AIRPARIF 2008. At the annual
scale, estimates differ by 8.2 MtCO2yr-1
(Table ). The differences between the two inventories are due
to both the differences in the emission model and the driving activity data
used. The two inventories correspond to two different years (2005 vs. 2008).
However, this hardly explains the amplitude of the difference between the two
inventories by itself. The decrease in the total emission in France between
2005 and 2008 was approximately 5 % see, e.g.. Here,
the difference in total emissions in Île-de-France between the IER and
AIRPARIF 2008 inventory, however, is about 14 %. Results of the inversion
using IER for both the prior emission budgets and the emissions' spatial
distribution (INV_IER) reflect these large prior discrepancies. Indeed,
monthly and annual budgets of Paris' fossil fuel CO2 emissions
estimated by INV_IER are larger than that from the reference inversion and
from INV_mapIER (Fig. ). The differences in posterior
February emissions from IER_INV and the reference inversion exceed
2 MtCO2month-1. The discrepancies are even larger, when
comparing the monthly emission estimates from INV_mapIER and IER_INV,
since the change of Hmap from
HAPmap to HIERmap
has a tendency to decrease the posterior emission estimates.
Sensitivity of monthly fossil fuel CO2 emissions on
Hmap. Red: monthly fossil fuel CO2 emission
estimates ± the standard deviation of their uncertainties obtained from
the reference inversion (green), INV_mapIER (red), and INV_IER (blue),
respectively. Monthly fossil fuel CO2 emissions prior estimates by
AIRPARIF (black) and IER's monthly estimates (light blue) ± the standard
deviation of uncertainties (shaded grey area). Note the different scale of
the ordinate compared to Figs. , and
.
The IER inventory indicates higher emissions in March than in November.
Posterior estimates from INV_IER still indicate that the highest emissions
are in November–February. Due to a residual influence from the IER prior
estimate, INV_IER predicts the highest emissions in February. The December
emission estimate is close to the February emission estimate, and is the
second highest 1-month mean estimate from INV_IER. Finally, the annual
posterior emission from INV_IER is closer to that from AIRPARIF 2010 than
to that from the 2008 inventory, despite the far higher prior annual estimate
from the IER inventory (Table ).
Sensitivity to Hsamp, Htrans and yf
INV_MNH is compared to the reference inversion to analyse the impact of
using Meso-NH/TEB instead of ECMWF as meteorological simulation for both the
gradient selection in the observation vector (Hsamp=Href-MNHsamp vs. Hsamp =
Href-ECMsamp) and the forcing of CHIMERE
(Htrans=HMNHtrans vs.
Htrans=HECMtrans and
yf=yref-MNHf vs. yf=yref-ECMf). Meso-NH/TEB data are only available up
to June which explains that the analyses here are restrained to the period
August 2010 to June 2011.
The time series of the gradients that are selected for the assimilation using
Href-MNHsamp and
Href-ECMsamp, respectively, are shown in
Fig. (and Figs. –). The
significant differences in selected gradients apparent in
Fig. (and Figs. –) are
driven by small differences in simulated wind fields between ECMWF and
Meso-NH/TEB. Small differences in wind direction and speed are often
sufficient to cross the thresholds defining the gradient selection
(Figs. –). This differences result in
a significantly different set of assimilated gradients and in a different
apportionment according to prevailing NE or SW wind directions
(Fig. ).
Despite this, Fig. reports similarities in inverted monthly
emissions from INV_MNH and the reference inversion. Differences in monthly
posterior emission estimates are less than 0.5 MtCO2month-1
when assimilating all selected gradients (Fig. ). The four
highest emitting months are still November to February for INV_MNH.
However, larger differences to the reference inversion estimates are found
for December and May, resulting in the loss of the peak in December and in an
unexpected peak in May in INV_MNH (Fig. a). This
disagreement is related to the assimilation of NE gradients. As shown in
Fig. b, emission estimates from INV_MNH and the reference
inversion are very similar when only SW gradients are assimilated. By
contrast, large differences are obtained in December and May when only NE
gradients are assimilated (Fig. c). The larger fraction of
selected NE gradients compared to selected SW gradients when using
Meso-NH/TEB instead of ECMWF could explain the loss of the emission peak in
December. There is no peak in December when using either Meso-NH/TEB, or
ECMWF and NE gradients only. Nevertheless, when assimilating SW gradients,
the consistency between INV_MNH and the reference inversion is surprising,
given the significantly different SW gradient selection.
Sensitivity of monthly fossil fuel CO2 budgets on
meteorological data. Displayed are the estimates ± the standard deviation
of their uncertainties obtained from the reference inversion (green) and
INV_MNH (black), respectively. Numbers at the top denote the colour-coded
number of assimilated gradients. (a) Assimilation of both SW and NE
gradients. (b) Assimilation of SW gradients.
(c) Assimilation of NE gradients.
Discussion and conclusions
Summary and general analysis of the results
We have analysed estimates of monthly mean fossil fuel CO2 emissions
from the Paris urban area from August 2010 to July 2011 using continuous
CO2 measurements from three stations and a city-scale atmospheric
inverse modelling framework derived from . The inversion
modelling is based on a mesoscale configuration of CHIMERE, on the AIRPARIF
high-resolution CO2 emission inventory for 2008, on the C-TESSEL
simulation for the biogenic fluxes in northern France, and on the principle
of constraining the 6 h city-scale budget of the emissions using
cross-city CO2 gradients. As demonstrated by the analysis of the
inversion results, this study has critically improved configuration of
by (i) discarding GON–MON gradients since they are not
related to the emission of the entire city, and (ii) by using stricter
criteria on the wind direction and wind speed for the selection of gradients.
The analysis suggests an improvement of the city's seasonal to annual
emission budget from the reference inversion compared to the prior estimate
that is based on the AIRPARIF 2008 inventory. The inversion derives an annual
emission budget (for August 2010–July 2011) of
40.9 MtCO2yr-1, which is closer to the independent estimate
from the AIRPARIF 2010 inventory (41.8 MtCO2yr-1) than to the
prior estimate (51.9 MtCO2yr-1). Although the reported
estimate from the AIRPARIF 2010 inventory does not exactly correspond to the
mid-2010 to mid-2011 period, changes between the 2008 and 2010 inventories
reflect improvements in the inventory model and actual changes of the Paris
emissions. Therefore, the fact that the corrections applied by the inversion
to the prior estimate from AIRPARIF 2008 are consistent with the differences
between the AIRPARIF 2008 and 2010 inventories gives confidence in the
inversion.
The seasonal variations of the monthly inverted emissions also appear more
realistic than that of the prior emission estimates. The seasonal amplitude
of the emissions revealed by the reference inversion is higher than that of
the prior estimate of the emission, derived from the five typical months of
the AIPARIF 2008 inventory. This increase in amplitude makes sense, given
that a large fraction (43 % according to the AIRPARIF 2008 inventory) of
the Paris emissions are due to domestic and commercial heating. It is
supported by the fact that the seasonal variations in the AIRPARIF 2010
inventory are higher than that derived from the AIRPARIF 2008 inventory. The
inverted seasonal cycle of the emissions correlates well (r2=0.45) with
the heating degrees in autumn–winter (November–February). The four months
with the highest inverted emissions correspond to the four coldest months
(November to February) – with a peak in both the emissions and the heating
degrees in December. By contrast, the prior estimate of the emissions derived
from AIRPARIF 2008 does not differentiate monthly budgets from November to
March.
The sensitivity tests indicate that the uncertainties assigned to the prior
estimates of the 6 h mean emissions, to the spatio-temporal
distribution of the emissions within the Paris area and 6 h windows,
and to the meteorological simulations (for the cross-city gradient selection
and for the forcing of CHIMERE) have a moderate impact on the monthly mean
emission estimates once the inversion is driven by the most stringent
selection of the measurements. This weak sensitivity of the inverted
emissions to the uncertainties assigned to the inverse modelling components
is important for the credibility of the inversion approach in view of
applying this approach to an independent method to verify inventories. Here,
the inverted emission budgets are sensitive to each of the above-mentioned
components. However, even though we assimilate a relatively small number of
data, this sensitivity is generally much smaller than the differences between
inverted and prior estimates at monthly to annual scale. Furthermore, the
plausible seasonal variations of the emissions revealed by the reference
inversion is robust to most sensitivity tests.
The inversions generally return smaller emissions than the prior estimates.
This is even the case when using a prior estimate that is flat at the monthly
scale only, and that has an annual emission budget of
36 MtCO2yr-1, i.e. a budget that is smaller than that from
AIRPARIF 2010. The inversion decreases the annual emission budget when using
the diurnal cycle of the emissions from AIRPARIF 2008 as prior estimate of
the 6 h mean emissions. In contrast, the annual emission budget is
increased when a flat diurnal cycle and a prior estimate of the annual
emission budget that is smaller than that from AIRPARIF 2010 (i.e. of
36 MtCO2yr-1) is used. This can reveal an error in the mean
diurnal cycle of the emissions from AIRPARIF 2008, which the inversion could
not correct for, since data are only assimilated during the afternoon.
Moreover, we define the uncertainties in the prior emission estimates in
terms of relative rather than absolute uncertainty. Consequently, using the
diurnal cycle of the emissions from AIRPARIF 2008 in the prior estimate of
the 6 h mean emissions and higher (smaller) prior emissions at
monthly to annual scale leads to higher (smaller) prior uncertainties, and
thus to a stronger (weaker) constraint from the atmospheric measurements,
resulting in a stronger (weaker) decrease in the emissions. One could argue
that this artificially helps getting a robust convergence of the sensitivity
tests using different prior estimates and it likely plays a role at the
annual scale. This could be problematic, since having a fixed value for the
relative uncertainty in the prior estimates is not suitable when these
estimates become very small. However, for some months, the convergence
between inversions utilising different flat priors is obtained by both
positive and negative corrections. This is the case in January and
February 2011 for FLAT_mM experiments (Fig. a). The
convergence can also be obtained with positive corrections that are larger
when prior uncertainties are smaller, e.g. in December 2010 for FLAT_mM
experiments (Fig. a). Figure c gives
several examples where the monthly budget of the prior estimate that are flat
at the 6 h scale determines if the corresponding corrections are
positive or negative. This figure also illustrates the fact that the
amplitude of the correction to the monthly estimates is not highly correlated
with the corresponding prior uncertainty. Furthermore, the fact that higher
prior emission estimates are assigned higher prior uncertainties cannot
explain the level of convergence of the sensitivity tests. In particular, it
can not explain the robustness of the retrieved seasonal cycle of emissions
when using flat priors. It neither explains the fact that the annual budget
from INV_IER is closer to AIRPARIF 2010 than to AIRPARIF 2008. INV_MNH
selected a significantly different set of gradients. However, it still
constrains the inverted emissions towards the same levels of emissions as in
the reference inversion (typically differences in monthly emissions are
< 5 %).
The improvement of the reference inversion compared to the initial inversion
demonstrates the need for a narrow definition of the wind direction ranges,
and more generally the need for a very careful selection of CO2 data.
This reveals the asset of following, as much as possible, the concept of
assimilating cross-city CO2 gradients to control the emissions at the
whole city scale, and to filter out the poorly modelled influence of fluxes
outside the Paris urban area. The assimilation of gradients cannot perfectly
cancel this influence because firstly one cannot set up the inversion system
to ensure that the selected gradients correspond to the concentration
variations of air masses that travel from the upwind to the downwind sites
(at least due to uncertainties in the atmospheric transport) and secondly
because the signal from fluxes outside Paris varies during such a transport
(due to atmospheric diffusion). However, results from and
from this study demonstrate that the assimilation of gradients succeeds in
decreasing this signal. These studies also show that such a decrease
strengthens the inversion results by limiting the problem of the
uncertainties in the remote fluxes for regional inversions which is
particularly critical in the Paris area as shown by and the
problem of the uncertainties in natural fluxes for urban CO2 emission
inversions. The positive insights from the evaluation of our results also
strengthens the confidence in this relatively simple concept of estimating
monthly budgets of the city emissions, even if it relies on the assimilation
of a relatively small amount of data.
Problems to be solved
The different inversion tests still raise concerns for the inversion of the
cities' monthly emission budgets. We expected that cross-city gradients would
be weakly sensitive to the uncertainties in the distribution of the emissions
within the Paris region and the 6 h windows, which explains why we
control, for a given 6 h window, a single scaling factor for the
emissions of the entire urban area. The inversion results, however, are
significantly affected by changes in the emission distribution. This does not
necessarily question the control of a single scaling factor for the emissions
of the whole urban area since reasonable results are obtained using the
emission distribution from AIRPARIF but it reveals the need to rely on
robust, high-resolution emission maps such as those produced by local
agencies like AIRPARIF. However, many cities do not have such local
inventories.
have shown that the selection of afternoon data provides
little constraint on night-time emissions. This is problematic since the
diurnal cycle is highly uncertain in inventories. The differences between the
results from FLAT_mD and FLAT_mH indicate that the poor
representation of the diurnal cycle in FLAT_mH has a large impact on the
inverted monthly emissions. As highlighted above, the inversions, based on
the diurnal cycle from AIRPARIF 2008, generally tend to decrease the prior
estimates which can also be viewed as an impact from errors in this diurnal
cycle. New approaches and techniques are needed to provide a direct
constraint on the night-time emissions or to better extrapolate the
information from daytime to night-time data to solve for this problem. The
poor representation of the day-to-day variations in the flat prior of
FLAT_mD does not seem to impact the results from this inversion, which
are close to that of FLAT_mM. Even though there is a large number of
days, sometimes even weeks, during which no gradients are assimilated, the
inversion does not strongly rely on the prior day-to-day variations within
the months to correct the monthly mean emission budgets. However, there is
a critical lack of data, which is primarily due to the small number of sites
available for this study, and thus to the relatively small wind sectors by
which we select cross-city gradients. This lack of data hinders the results
of all inversions for specific months such as September, November, April, May
and July, when less than thirty 1 h mean gradients are assimilated. The
month-to-month variability is thus often driven by the variability of the
data availability. Results at the monthly scale are thus not systematically
consistent with the different sensitivity tests. Monthly estimates can be
weakened by missing or over-weighting high variations in the emissions over
short time periods (e.g. due to a cold event). One can hope that this
limitation could be overcome by an expansion of the observation network with
stations all around the Paris urban area, which could ensure a continuous
monitoring of the cross-city CO2 gradients.
In December, the number of assimilated data are relatively high for both the
reference inversion and INV_MNH. However, while the inversions increase
the emissions compared to AIRPARIF 2008 during December when using ECMWF data
and all gradients (SW and NE gradients), this is not the case when
assimilating subsets of the cross-city gradients only, or when using
Meso-NH/TB. Consequently, there is no peak of inverted emission estimates in
December. Neither is this a robust feature of the reference inversion. The
absence of an emission peak in December is associated with the assimilation
of NE gradients (i.e. due to the assimilation of NE gradients only, or, to
the use of the Meso-NH/TEB meteorology which selects a larger fraction of NE
gradients than its ECMWF counterpart).
More generally, the assimilation of NE gradients seems to raise concerns
while more satisfying results are obtained when using SW gradients. This
applies also to the initial inversion, for which the NE direction corresponds
to wider wind direction ranges. Thus, the problem cannot be related to a very
specific source NE of Paris. When the wind blows from NE, the signature of
emissions from remote, highly urbanised and industrialised areas
(north-eastern France, Benelux and western Germany) should impact the
CO2 fields in the Paris area. On the opposite, the regions between
the Atlantic Ocean and Paris are mostly rural. While the computation of
gradients is an efficient way of limiting the signatures of the fluxes
outside the Paris area on assimilated data, and while it effectively reduces
these signatures to a small component in the simulated gradients, it does not
ensure a total removal of such signatures in the measurements which may bear
a more spatial heterogeneity than the modelling framework. The large-scale
signature of the remote natural fluxes from SW may be more easily modelled or
filtered out by the computation of gradients in the Paris area than the
signature of emissions from NE. This could explain why the assimilation of NE
gradients is more problematic than that of SW gradients. This could reveal
another limitation in assimilating cross-city gradients. The high, temporal
variability of the ratio of assimilated NE gradients to SW gradients may be
problematic for the monitoring of the month-to-month variability of the city
emissions. Similarly, the small biogenic signal in the simulated gradients
may be due to the use of an ecosystem model with moderate horizontal
resolution. Measured gradients might be impacted by urban ecosystems that
cannot be represented with this model. Due to the high density and
compactness of the Paris urban area, we can assume that such urban ecosystems
should have a low impact on the inversion of Paris emissions. This should be
further investigated based on urban ecosystem modelling and monitoring
.
The last major issue is the limited confidence in the posterior uncertainties
computed by the inversion system. We purposely avoided analysing them in
detail in Sect. . They provide qualitative insights on the
behaviour of the inversion, i.e. posterior uncertainties remain close to the
prior ones for night-time emissions, which are poorly constrained by only
using afternoon CO2 data . The posterior uncertainties
also vary as a function of the number of assimilated data. The different
estimates from the sensitivity tests generally lie in the 68 % confidence
(1σ) interval of the reference inversion. However, the posterior
uncertainties generally look very low for specific months, despite the lack
of confidence in the specific monthly estimates as discussed above, and
despite the very limited number of assimilated data. During February and
March, the posterior uncertainties from the reference inversion are lower
than 0.69 MtCO2yr-1. The large emission decrease of
1.32 MtCO2yr-1 from February to March is surprising. The
relative difference between the posterior and prior uncertainties when moving
from the initial inversion to the more reliable reference inversion
demonstrates how misleading the interpretation of theoretical uncertainties
can be when several mathematical assumptions in the inversion are not met in
practice. However, even though the configuration is far from perfect, the
misfits between posterior estimates and observations are still smaller than
between prior estimates and observations. This gives a stronger confidence in
posterior emission estimates than in the posterior uncertainties of these
emissions. Sensitivity tests with the analysis of the posterior estimates
were only conducted to give a better picture of the strength of the
measurement constraint.
Perspectives
Despite these concerns, the results from this study are promising and several
methodological improvements were found. The inversion test of assimilating
spatio-temporal gradients accounts for the time air parcels need to pass from
the upwind site over the urban area to the downwind site. Such gradients
should bear a smaller signal from fluxes outside the urban area than spatial
gradients, which should help in isolating this signal from the city
emissions. The lack of data, however, prevented this inversion from
significantly departing from its prior emission estimate. Such a strategy
would be more appropriate if a larger amount of data was available, but it is
impractical for our limited network: it exacerbates the loss of data from
already strict gradient selection criteria and degrades the overall emission
retrieval compared to the reference inversion. For the same reason, it would
be inappropriate with our limited network to narrow the wind direction ranges
to select gradients to less than 15∘ of the transect between the
downwind and upwind sites, even though, in principle, it would strengthen the
decrease of the signature from fluxes outside the urban area.
The expansion of the network, in particular a full encirclement of the city
with at least 8 sites (given that the wind ranges for the selection of
gradients between one upwind site and one downwind site cover 30∘ in
this study) should strengthen the results and could allow for application of
such new techniques that result in a stricter gradient selection. However,
relying on such a measurement expansion may not be sufficient. Exploiting
more information from the available dataset without violating or undermining
our assumptions on the selection of cross-city gradients is a requirement to
strengthen the observational constraint of the inversion. The Paris
observation network has been set back since September 2014 in the framework
of the Carbocount City and Le CO2 Parisien projects. Both projects aim to
deploy more measurement sites than the CO2-MEGAPARIS project. However,
relying on such a measurement network expansion may not be sufficient. New
methods should be developed to exploit urban measurements which
would allow to solve for the spatial distribution of the emissions, which
does not seem possible with the current monitoring
network of peri-urban sites. This in turn could help in assimilating data
that do not necessarily bear the signature of the emissions from a large part
of the city. Finally, developing methods to exploit morning, evening and
night-time data would be necessary to constrain night-time fluxes. This is
not necessary to improve the knowledge of the emissions based on atmospheric
inversion, but this is necessary to develop accurate tools for the
operational monitoring and verification of the emissions based on this
approach.
Even though it applies to the specific case of monitoring the CO2
emissions from Paris, this study demonstrates the potential of an approach
which can be adapted to a wide range of cities. The urban surroundings,
spread, size, topography and meteorology of some cities increase the
difficulty in catching cross-city gradients, and different strategies may be
more adapted for such cases. The atmospheric inversion of the city emissions
is still an emerging activity, but the present results already raise some
confidence in this concept, especially since many other resources (combining
atmospheric CO2 inversions with air-quality monitoring, the
development of new measurement types) could help to overcome the remaining
challenges.