Introduction
Carbon dioxide (CO2) is commonly recognized as the major greenhouse gas
providing the driving force of recent and future climate change (IPCC,
2013). Its atmospheric concentration has considerably increased (by 40 %)
since the industrial revolution (Petit et al., 1999). This increase, the
rate of which has accelerated in the past decade, is attributed mostly to
anthropogenic sources, such as fossil-fuel (FF) burning (Canadell et al.,
2007). Curbing further growth of CO2 concentration has become a goal of
international agreements such as the Kyoto protocol (UNFCCC, 1998) and the
Paris Agreement on Climate Change (UNFCCC, 2015). Thus accurate knowledge of
anthropogenic CO2 emissions is of paramount importance both for climate
prediction and mitigation policy purposes.
Over the past few decades a lot of effort has been put into the
compilation of global (e.g., Olivier et al., 2005; GCP, 2010; Ciais et al.,
2010a; Janssens-Maenhout et al., 2015) as well as regional (Gurney et al.,
2009; Huang et al., 2011; Kurokawa et al., 2013; Zhao et al., 2012; Wang et
al., 2013) inventories of CO2 emissions from FF burning and other
smaller anthropogenic sources (such as biofuel burning and cement
production). Those emission inventories are based on available statistical
information regarding economic activities and corresponding technologies.
However, it is known that such information can be subject to errors and
biases leading to considerable uncertainties in emission estimates,
especially in the case of rapidly growing developing economies (e.g.,
Akimoto et al., 2006; Guan et al., 2012; Korsbakken et al., 2016). For
example, the uncertainty of available estimates of the total FF CO2
emissions in China is assessed to be about 15–20 % (Gregg et al., 2008).
Much larger uncertainties may be associated with the subnational spatial
distributions and temporal evolution of FF CO2 emissions within a year
(Ciais et al., 2010b). The uncertainties in anthropogenic CO2 emission
inventory data are mostly due to inaccuracies of available data regarding
fuel consumption and fuel chemical composition. Note that the estimation of
uncertainty in emission inventory data is itself a challenging task: in
particular, as different inventories are usually based (at least partly) on
common sources of information, their intercomparison does not necessarily
result in revealing all the uncertainties.
A promising alternative approach to constrain NO2 emissions and to
assess the uncertainty in available emission estimates is inverse modeling
(Enting, 2002); the key idea of this approach is to derive emission
estimates from atmospheric measurement data by optimizing emissions coupled
to a transport model. Such estimates are frequently referred to as
“top-down”, in contrast to “bottom-up” ones based on emission inventories
alone. Numerous studies have successfully used in situ CO2 measurements
in the framework of this approach to constrain surface CO2 fluxes
associated mostly with biospheric and oceanic sources and sinks of CO2
in different regions of the world (e.g., Gurney et al., 2002; Baker et al.,
2006; Schulze et al., 2009; Chevallier et al., 2010; Broquet et al., 2013).
More recently, it was demonstrated that uncertainties in CO2 flux
estimates can be potentially reduced by using satellite CO2
measurements (e.g., Chevallier et al., 2007; Houweling et al., 2004;
Hungershoefer et al., 2010; Kadygrov et al., 2009; Nassar et al., 2011;
Maksyutov et al., 2013; Reuter et al., 2014a). However, less progress has been
made in isolating FF CO2 emissions from other sources and sinks. Major
limitations are due to the fact that the atmospheric variability of
CO2 is strongly affected by biogenic sources and sinks, such as plant
respiration and photosynthesis, and that the signatures of regional FF
CO2 emissions in CO2 observations are typically weak relative to
regional background CO2 concentration, except near hot spots. Promising
approaches suggest separation of FF CO2 emissions from biospheric
fluxes by using available measurements of radiocarbon content (14C) of
CO2 (e.g., Turnbull et al., 2009; Miller et al., 2012; Lehman et al.,
2013; Basu et al., 2016), ground-based CO2 measurements in vicinity of
strong anthropogenic emission sources like megacities (Bréon et al.,
2015), or satellite CO2 retrievals with sampling near hot-spots
(Bovensmann et al., 2010; Silva et al., 2013; Reuter et al., 2014b).
However, neither of these approaches has already been sufficiently
generalized to provide reliable estimates of the budget of anthropogenic
CO2 emissions in an arbitrary industrialized region of the world.
It has also been suggested that anthropogenic CO2 emissions can be
constrained to a certain extent by measurements of “proxy” species, whose
sources are mostly collocated in time and space with CO2 sources
(Rivier et al., 2006; Suntharalingam et al., 2004). The measurements of proxy
species can be either combined with CO2 measurements (Palmer et al.,
2006; Rivier et al., 2006; Suntharalingam et al., 2004; Brioude et al., 2012) or
used alone but with information on a relationship between emissions of
CO2 and of the proxy species from bottom-up emission inventories. In
the second approach, Berezin et al. (2013) estimated multiannual relative
changes of FF CO2 emissions from China by using satellite measurements
of nitrogen dioxide (NO2) and emission inventory data on the ratio of
FF emissions of CO2 and nitrogen oxides (NOx= NO + NO2). A
similar approach was employed by Konovalov et al. (2014) to obtain estimates
of CO2 emissions from biomass burning in Siberia by using satellite
measurements of carbon monoxide (CO) and of aerosol optical depth.
The goal of this study is to examine the feasibility of inferring estimates
of annual budgets of CO2 emissions from FF burning in a given
industrialized region with a typical size of the order of 1000 km by using
satellite measurements of NO2 and CO. In doing so, we develop a special
method by building upon the ideas that were exploited in Berezin et al. (2013) and Konovalov et al. (2014). The method includes several major steps,
namely (1) inferring top-down estimates of total anthropogenic emissions
of NOx and CO from satellite measurements of the corresponding proxy
species by using simulations performed with a mesoscale chemistry transport
model (CTM), (2) applying NOx-to-CO2 (or CO-to-CO2) emission
conversion factors given by bottom-up emission inventories to relate FF
CO2 emissions to the NOx and CO anthropogenic emissions from the
previous step, and (3) cross-validation and optimal combination of estimates of
the FF CO2 emission budgets derived from measurements of different
proxy species. As a result, we obtain a “hybrid” FF CO2 emission
estimate integrating information coming from measurements and bottom-up
inventories. The use of NO2 and CO as proxy species in the context of
our approach is justified because their satellite measurements are known to
contain a strong signal associated with human activities in industrial
regions and have abundantly been used earlier to constrain emissions of,
respectively, NOx (e.g., Martin et al., 2003; Konovalov et al., 2006;
Napelenok et al., 2008; Miyazaki et al., 2012; Gu et al., 2014) and CO
(e.g., Arellano et al., 2004; Pétron et al., 2004; Kopacz et al., 2010; Hooghiemstra
et al., 2012; Krol et al., 2013; Jiang et al., 2015) from various sources,
including anthropogenic ones. Note that although NOx and CO emissions
from FF burning are more sensitive to technological factors than CO2
emissions, different aspects of the combustion technology are expected to
affect NOx and CO emissions in different ways: e.g., while NOx
emissions are strongly dependent on the temperature of combustion (more
NOx is released at higher temperatures), CO emissions can be regarded
as a measure of the incompleteness of combustion processes. So, the
combination of hybrid FF CO2 emission estimates derived from both
NO2 and CO measurements can enable a compensation of a part of the
uncertainties associated with inaccurate knowledge of technology and
conditions of combustion affecting separately NO2 and CO measurement
based FF CO2 emission estimates.
Particular efforts in this study were made to provide adequate confidence
intervals for the hybrid FF CO2 emission estimates. To this end, we had
to ensure that potential errors in our top-down estimates of NOx and CO
emissions are statistically independent from those of the conversion
factors. We also had to ensure that the evaluation of confidence intervals
does not involve any subjective quantitative assumptions regarding the level
of uncertainties in measured and simulated data. Such requirements would be
difficult to satisfy if the top-down estimates of the emission annual
budgets were partly constrained (in the Bayesian sense) with a priori
knowledge on these budgets from a bottom-up emission inventory (as it is
usual in inverse modeling studies). Furthermore, the use of a priori
constraints would make the cross-validation of the estimates of FF CO2
emission budgets based on NO2 and CO measurements infeasible, as both
priors and cross-validation estimates could then be biased in a similar way
due to possible systematic uncertainties in activity data employed in the
emission inventory. Accordingly, a distinctive feature of our method is that
it does not involve any formal a priori constraints to the top-down
estimates of the emission budgets or any quantitative settings specifying
the level of uncertainties in measured and simulated data. This feature is
expected to reinforce the potential of the method to elucidate possible
uncertainties and/or inconsistencies in CO2, NOx, and CO emission
data provided by different bottom-up emission inventories.
In this study, our method is applied for estimation of annual FF CO2
emissions from a group of 12 western European countries, including a
selection of 11 member states of the European Union (EU) that provide the
predominant part (> 70 %) of EU total FF carbon emissions
(Ciais et al., 2010b) and Switzerland. Taking into account availability of
bottom-up emission inventory data necessary for our analysis, the annual
emission estimates were obtained specifically for the year 2008. We believe
that estimation of FF CO2 emissions from the European region could be
considered as a good testing case for our method, taking into account that
uncertainties in corresponding emission inventory data for the EU countries
with well-developed statistics are relatively low (compared to potential
uncertainties in FF CO2 emission data for countries with less developed
statistical infrastructure), although not quite negligible. For example, by
comparing data of several international emission inventories, Ciais et al. (2010b)
estimated the full uncertainty of the bottom-up estimates of the
anthropogenic CO2 emissions in the EU countries to be about 19 % but
less (∼ 7 %) if possible inconsistencies between types of
CO2 sources are taken into account in different emission inventories are
resolved. Note that the uncertainty in bottom-up FF CO2 emission
estimates is expected to be lower than in corresponding NOx and CO
emission estimates because of an important role played for emissions of
those proxy species by the technology and end-of-pipe measures; at the same
time, the ratio of emissions of CO2 and of the proxy species can be
less uncertain than the emission themselves if the emission data are subject
to a strong common bias caused by uncertainties in fuel consumption
statistics.
In the following, Sect. 2 describes data and modeling tools used in our
study. Description of our inverse modeling method and its validation with
synthetic “observations” are presented in Sect. 3. The results of its
application to the real-world situation are presented and discussed in Sect. 4. Finally, major findings are summarized in Sect. 5.
Data and model description
Retrievals from satellite measurements
We used the tropospheric NO2 column retrievals from measurements of the
Earth's backscattered radiation in visible and ultraviolet spectral regions
by the OMI satellite instrument (Levelt et al., 2006) on board the NASA EOS
Aura spacecraft. The Aura satellite (Schoeberl et al., 2006) is in a
sun-synchronous ascending polar orbit with an Equator crossing time of 13:30
local solar time (LST) and an orbital period of 99 min. The OMI instrument
has a swath width of ∼ 2600 km divided into 60 pixels with a size of
13–26 km.
The retrievals used in this study are provided by the Royal Netherlands
Meteorological Institute (KNMI) as the DOMINO version 2 data product
(Boersma et al., 2011) through the TEMIS portal (http://www.temis.nl/). The
product contains Level 2 data, that is, NO2 columns and relevant
geophysical information for each ground pixel observed by the instrument. In
this study, only cloud and surface albedo screened data (retrieved for the
scenes with the cloud fraction less than 30 % and with the surface albedo
less than 0.3) were used. The main steps of the OMI NO2 retrieval
algorithm (see Boersma et al., 2011, for details) include (1) spectral
fitting and the slant column density (SCD) estimation by using the
differential optical absorption spectroscopy (DOAS) method, (2) separation
of the tropospheric and stratospheric parts of the slant columns, and (3) calculation of the tropospheric vertical column by applying the air mass
factor (AMF) to the tropospheric slant column. Each step involves different
uncertainties that may contribute, to various extents, to the uncertainty of
the tropospheric NO2 columns; in particular, the SCD uncertainty is
likely to predominate over other uncertainties in remote areas and is
about 0.7 × 1015 molec. cm-2, while the retrievals for
urban areas are mostly affected by the AMF uncertainty, that is of about 25 % for cloud-free conditions (Boersma et al., 2007, 2011). Several studies
(e.g., Zhao and Wang, 2009; Miyazaki et al., 2012; Vinken et al., 2014) found
that in spite of the considerable uncertainties, the tropospheric NO2
columns retrieved from the OMI measurements provide useful constraints to
anthropogenic NOx emissions in different regions of the world,
including Europe.
We also used the Level 2 retrievals of total CO column amounts from the
measurements performed by the Infrared Atmospheric Sounding Interferometer (IASI) on board the Metop-A satellite (Clerbaux et al., 2009). Metop-A has
the sun-synchronous polar orbit with Equator crossing at 21:30 LST for the
ascending node. The IASI instrument provides global coverage twice a day
(around 09:30 and 21:30 LST), with a swath of about 2 × 1100 km and a
nominal pixel diameter footprint on the ground of 12 km.
The CO column amounts are retrieved from the cloud screened measurements of
the spectrum at the 1–0 rotation vibration band centered at 4.7 µm
(2128 cm-1) by using the Fast Optimal Retrievals on Layers for IASI
(FORLI) algorithm (Hurtmans et al., 2012). The FORLI algorithm provides CO
partial column amounts (for at most 19 layers) fitting the spectral
observations with a priori constraints; the partial columns are combined to
yield the total column amounts. The uncertainty of the IASI CO retrievals
strongly depends on the geographical location and conditions of the
observations (Clerbaux et al., 2009; George et al., 2009; Turquety et al.,
2009); it is estimated to be about 10 % under typical conditions
(Clerbaux et al., 2009). It should be noted, however, that the capability of
the IASI measurements to inform about CO sources depends not only on the
accuracy of the CO retrievals but also on the sensitivity of the spectral
observations to the CO concentration in the boundary layer. A convenient way
to characterize this sensitivity (which is related to the vertical
resolution of the retrieval and depends, in particular, on the difference
between the temperatures of the surface and of the atmospheric boundary
layer) is to consider the trace of the averaging kernel matrix (Clerbaux et
al., 2009); this parameter is called the degree of freedom of the signal (DOFS). Distinguishing between the upper and lower troposphere requires this
parameter to be about 2 (George et al., 2009). Taking these considerations
into account, we used only those retrievals that were characterized by
relatively large DOFS values: similar to Konovalov et al. (2014), the DOFS
threshold was set to be 1.7. The available CO retrievals for individual
pixels were projected to the 0.5∘ × 0.5∘ grid of a CTM (see Sect. 2.2) and averaged over each
day.
We would like to note that instead of (or together with) the IASI
measurements, we had an option of using alternative data from other infrared
sounders, such as MOPITT and AIRS. Our decision to choose the IASI
measurements was made by taking into account their relatively high
sensitivity in the boundary layer (George et al., 2009), as well as previous
studies in which the IASI data were successfully employed for constraining
CO emissions from different sources (Fortems-Cheiney et al., 2009; Krol et
al., 2013; Konovalov et al., 2014). We considered the relatively high
sensitivity of the IASI measurements in the lower troposphere as an
important advantage, especially in the context of the given study involving a
mesoscale CTM. Indeed, the upper troposphere CO content simulated with such
a CTM is likely to be strongly affected by boundary conditions which are
specified by using global CTM simulations and therefore are not dependent on
CO emissions used in the regional CTM (see, e.g., a discussion in
Konovalov et al., 2011). Exploring the potential of the alternative CO data
products goes beyond the scope of the given study.
CTM simulations and initial processing of the model output data
In this study, the relationships between NOx and CO emissions and,
respectively, NO2 and CO column amounts are simulated by the CHIMERE
CTM. CHIMERE is a three-dimensional Eulerian model designed to simulate air
pollution on urban, regional, and continental scales; it allows to take into
account the most important atmospheric processes (such as anthropogenic,
biogenic, and fire emissions, gas-phase and heterogeneous chemistry,
advection, turbulent diffusion, deep convection, dry and wet deposition)
affecting the atmospheric fate of a number of reactive gaseous and aerosol
species (see Menut et al., 2013, and references therein). The model was
earlier successfully used in combination with satellite NO2 and CO
retrievals in several inverse modeling studies of NOx and CO emissions
(e.g., Konovalov et al., 2006, 2008, 2010, 2014; Berezin et al., 2013; Mijling and van der A, 2012; Ding et al., 2015).
In this study, the CHIMERE model was run with one of the standard domains
(called the CONT5 domain) covering a western part of Europe (-13.75–25.25∘ E, 34.75–58.25∘ N) with the
horizontal resolution of 0.5∘ × 0.5∘. The simulations were performed with 12
non-equidistant layers in the vertical (up to the 200 hPa pressure level);
the layers were specified in the hybrid sigma–pressure coordinates such that the
distance between the layers increased with the altitude from ∼ 50 m near the surface to ∼ 2 km in the upper part of the
modeled atmosphere. Gas-phase chemical processes were simulated with the
simplified MELCHIOR2 chemical mechanism (Schmidt et al., 2001), and several
heterogeneous reactions on the surfaces of aerosol particles were taken into
account as described in Menut et al. (2013). Initial and boundary conditions
for several key gaseous species responsible for the oxidation capacity of
the lower atmosphere (e.g., CO, NO, NO2, O3,H2O2,
HCHO) and aerosols were specified using monthly climatological data from
LMDz-INCA global model (Folberth et al., 2006). A full list of these species
is provided in the CHIMERE documentation available on the web site
www.lmd.polytechnique.fr/chimere. An influx of other species, most of which
are very reactive and short-lived (such as OH and HO2), into a
model domain is not specified in CHIMERE. Meteorological data were obtained
from the WRF-ARW (v.3.6) model (Skamarock et al., 2008), which was run with
a horizontal resolution of 50 km × 50 km and with 30 levels extending
in the vertical up to the 50 hPa pressure level for a region covering the
CHIMERE domain and was driven with the NCEP Reanalysis-2 data (National
Centers for Environmental Prediction, 2000). The anthropogenic, biogenic, and
fire emissions of major gaseous and aerosol species were taken into account
in our simulations as described in the next section (Sect. 2.3). The model
was run with different scenarios (specified below in Sect. 3.2) for the
period from 22 December 2007 to 29 December 2008. The spin-up period
included the first 10 days of any run, which therefore were withheld from
the following analysis.
To enable consistency of our simulations with the satellite data employed in
this study, the CHIMERE outputs were processed by taking into account
measurement properties. All the simulated NO2 and CO vertical profiles
corresponding (in time and space) to any pixel that contained, respectively,
the OMI and IASI measurements satisfying to the criteria specified in Sect. 2.1 were projected into the measurement vertical grids and transformed into
tropospheric NO2 columns and total CO columns, CmNO2 and
CmCO, by applying the respective averaging kernels,
ANO2 and ACO.
Specifically, the simulated NO2 profiles were transformed as follows
(Eskes and Boersma, 2003):
CmNO2=ANO2TCm(o)NO2,
where Cm(o)NO2 are the original model outputs (partial
columns) interpolated to the pressure grid of the averaging kernels up to
the tropopause pressure level (specified in the measurement database). Note
that in relatively rare cases (constituting less than 20 % of the total
number of valid observations available for the study region and period)
where the tropopause pressure was smaller than the pressure at the top of
the model grid (200 hPa), the lack of the simulated data at altitudes
exceeding the height of the upper model layer could result in some
underestimation of the modeled tropospheric columns, but such a minor
inconsistency between the modeled and simulated NO2 columns is not
expected to result in underestimation of NOx emissions in our analysis,
owing to application of a debiasing technique described in Sect. 3.2 and
validated in Sect. 3.5.
A slightly different procedure was used to process the modeled CO partial
columns:
CmCO=(ACO)TCm(o)CO-CaCO+ITCaCO,
where CaCO is the a priori CO vertical
profile used in the retrieval procedure, and I is the
identity vector. The missing components of Cm(o)CO for
altitudes above the upper layer of the CHIMERE CTM were taken to be equal to
the respective values from CaCO. Note that the
transformation providing the total CO columns in accordance to Eq. (2) is a
special case of the more general transformation procedure providing partial
CO columns (see Fortems-Cheiney et al., 2009).
The model outputs transformed with different averaging kernels but
corresponding to the same model grid cell and hour as the observations were
averaged. The modeled profiles which had not been matched with the
corresponding observational data were not used in our analysis. With the
satellite data used in this study, each grid cell is provided with observed
or modeled data for at most two different hours of each day. In addition to
the selection criterion based on the DOFS values (see Sect. 2.1), in order
to minimize the impact of model errors that are not associated with
uncertainties in emission data on inverse modeling results, only those days
and grid cells were taken into account when and where the modeled
contribution of anthropogenic NOx or CO emissions in the study region
(specified in the next section) to CmNO2 and
CmCO was larger than one percent of the corresponding
“background” values of the columns (here “background” is defined according
to a simulation made without anthropogenic emissions, i.e., with the biogenic
and open biomass burning emissions specified in the next section, and with
the transport model boundary conditions described above).
Emission inventory data
We used annual anthropogenic emission data for the year 2008 from several
sources: the European Monitoring and Evaluation Programme (EMEP) regional
emission inventory (EMEP/CEIP, 2014; Mareckova
et al., 2014), the Emission Database for Global Atmospheric Research,
version 4.2 (EDGAR v4.2) (EC-JRC/PBL, 2011), and the Carbon
Dioxide Information Analysis Center (CDIAC) (Boden et al., 2011). The EMEP
inventory data were used in our simulations described in Sect. 2.2, and the
EDGAR v.4.2 data were used to relate the emission estimates for the proxy
species with CO2 emissions (see Sect. 3.3). The CDIAC data were
involved in the analysis of uncertainties in our emission estimates (see
Sect. 3.4) along with the EMEP and EDGAR v4.2 data.
The EMEP/CEIP inventory is based on emission data reported under the
Convention on Long-range Transboundary Air Pollution by individual countries
in Europe and in the Middle East, which are expected to use a unified
approach (EMEP/EEA, 2013) applicable on the national level. In this study,
we used the EMEP anthropogenic annual emission data distributed among 11
Selected Nomenclature for Air Pollutants (SNAP) sectors and provided for
several pollutants, such as NOx, CO, non-methane hydrocarbons (NMHC),
SOx, and particulate matter, on a grid with the resolution of
0.5∘ × 0.5∘. Note that the EMEP inventory does
not provide data for CO2 emissions and that the emissions for the
11th sector (comprising biogenic sources and fires associated with
human activities) were replaced in our simulations with data of dedicated
inventories (as described in this section below).
The EDGAR v4.2 inventory is created by using the energy activity data
provided by the International energy agency (IEA, 2010) and by
following the methodology and fuel-specific emission factors based on the
2006 IPCC guidelines (IPCC, 2006). The IEA data were compiled following
harmonized definitions of fuels and activities and applying the same
methodologies across most countries (and some groups of countries outside of
the study region). We used the EDGAR v4.2 data for the national totals of
anthropogenic NOx, CO, and CO2 emissions distributed between
several emission sectors (not necessarily coinciding with the SNAP sectors).
Note that we used the EDGAR v4.2 FF CO2 emission data excluding
CO2 emissions from biofuel burning (that is, the data used were
calculated after “excluding short-cycle organic carbon”), while the
corresponding CO and NOx data included emissions from both fossil-fuel
and biofuel burning.
The FF CO2 emission data provided by CDIAC are based on the energy
statistics that were compiled primarily from the annual energy questionnaire
distributed by the United Nations Statistics Division and supplemented by
official national statistical publications (UN, 2012). The quantity of fuel
was converted into the quantity of CO2 emissions by using the
methodology based on Marland and Rotty (1984). The CDIAC database used in
this study reports only national totals of FF CO2 emissions without
sectorial breakdowns and was used in this study for evaluation of
uncertainties in our results.
Note that CO2 emissions from cement production have been reported in
CDIAC (as well as in EDGAR v4.2) separately from FF CO2 emissions and
were not considered in our study. Excluding this emission source from our
estimates seems to be reasonable, taking into account that cement production,
unlike FF burning, is not associated with considerable emissions of either
NOx or CO, and so satellite measurements of the corresponding proxy
species cannot provide strong constraints on CO2 emissions from cement
production.
The anthropogenic emissions were aggregated into two categories. Splitting
the total emissions among the two categories was deemed to reduce the
generation of aggregation errors (Kaminski et al., 2001) in our top-down
estimate of the total NOx and CO emissions. To this end, we tried to
ensure that, on the one hand, the emissions corresponding to the different categories had
distinct spatial distributions (such as the emissions from power
plants and from transport) and, on the other hand, that the amounts of
annual emissions from each category were of the same order of magnitude.
Specifically, the first category (EHI) included the emissions associated
mostly with energy and heat production and heavy industries. The second
category (TCO) comprised transport, chemical industry, and all other
anthropogenic sources. In the EMEP inventory, the EHI category was defined
by aggregating the sources corresponding to the first, second, and third
sectors of SNAP (combustion in energy and transformation industries,
nonindustrial combustion plants, and combustion in manufacturing industry,
respectively). The sectors 1A1a-c (public electricity and heat production,
other energy industries), 1A2 (manufacturing industries and construction),
and 1A4 (fuel combustion in residential and other sectors) were allocated
into the same category in the case of the EDGAR inventory. The TCO category
aggregated all other anthropogenic sources considered in the EMEP or EDGAR
v4.2 inventories. We expected that, apart from limiting the aggregation
error, consideration of these two categories would allow us to get more
specific information on emission sources. Note that splitting emission
sources between the two categories specified above is, at large, rather
arbitrary: in this study, we did not attempt analyzing the impact of the
source categories definitions on the uncertainty of our emission estimates.
Spatial distributions of NOx (a) and CO (b) total annual
emissions (g cm-2 yr-1) and the fractions (%) of the EHI
(c, e) and TCO (d, f) emission source categories (see the definitions in
Sect. 2.3) according to the EMEP inventory for 2008. The emission data are
shown only for the study region comprising land territories of 12 European
countries.
Figure 1 shows the CONT5 domain (employed in this study) of the CHIMERE CTM
along with the spatial distributions of total annual anthropogenic NOx
and CO emissions from the selection of 12 western European countries
considered in our analysis according to the EMEP inventory for 2008; it also
shows the fractions of the two source categories introduced above. Note that
emissions outside of the selected countries (including ship emissions) are
not indicated (the corresponding territories are left blank), such emissions
constitute minor parts of the total NOx and CO emissions in the whole
model domain shown in Fig. 1 (41 and 30 %, respectively, according to
the EMEP inventory for 2008). The territory of the United Kingdom is not
fully represented in the model domain; however, the emissions from the
missing northern part of this country are rather negligible (∼ 0.5 % of the total emissions in UK).
It is noteworthy that not only the total emissions (see Fig. 1a, b) but also
the fractions of the different emission source categories (see Fig. 1c–f)
exhibit considerable spatial variations. The spatial variability of the
source category fractions indicates that, given sufficiently accurate
observations, an appropriate inverse modeling procedure together with the
dense spatial sampling of the atmosphere by satellites may have a potential
to distinguish between emissions coming from the different sources. It can
also be noted that the fractions of the same source categories of the
NOx and CO emissions considerably differ (cf. Fig. 1c–f). In particular, while the NOx emissions mostly come from the TCO
sources, the CO emissions are distributed between the TCO and EHI sources
much more evenly. This observation indicates that the measurements of these
two proxy species might provide complementary (to a certain extent)
information on human activities associated with CO2 emissions, even if
atmospheric fates of the CO and NOx emissions were identical.
The annual anthropogenic emission data were distributed at shorter timescales by applying monthly, daily, and hourly factors from the standard
emission interface of the CHIMERE CTM (Menut et al., 2013); the factors were
provided for specific pollutants, the SNAP sectors, and countries by IER,
University of Stuttgart (GENEMIS, 1994). The seasonal variations specified
in this way for the two categories of anthropogenic emissions are shown in
Fig. 2. In addition, emissions were vertically distributed within 1 km by
using the profiles (specific for each SNAP sector) provided in the emission
interface of CHIMERE. Note that the vertical profiles did not explicitly
account for aircraft emissions, which are also included in the EMEP
inventory, but are likely to provide a very small contribution (less than 2 %) to anthropogenic NOx and CO emission in Europe (Tarassón et al.,
2004).
The seasonal variations of the spatially averaged (over the study
region) NOx (a) and CO (b) emissions for the EHI and TCO
categories of sources. The variations were calculated as explained in Sect. 2.3. The values shown (unitless) are the monthly emissions normalized to the
total annual emissions divided by 12.
Along with the anthropogenic emissions, our model included biogenic
emissions (in particular, NOx emissions from soils and emissions of
isoprene and some other hydrocarbons from vegetation) and emissions of
gaseous species (NOx, CO, and non-methane hydrocarbons) from open
biomass burning (fires). Biogenic emissions were calculated for each grid
cell, day, and hour within the CHIMERE model by using the European inventory
of soil NO emissions (Stohl et al., 1996) and the emission factors and
parameterizations from the MEGAN (Model of Emissions of Gases and Aerosols
from Nature) model (Guenther et al., 2006). The fire emissions were
specified using the daily data provided by the Global Fire Assimilation
System, version 1.0 (GFAS v1.0) fire emission inventory (Kaiser et al.,
2012). The fire emissions were distributed in the vertical uniformly up to
the altitude of 1 km (similar to Konovalov et al., 2011). Note that
according to the data of the GFAS v1.0 and EMEP emission inventories, the
total emissions of both NOx and CO from fires in the countries
considered (mainly, in Portugal) in 2008 were rather small (∼ 0.5 and ∼ 5 % relative to the corresponding FF
emission estimates given by the EMEP inventory).
Time series of the spatially averaged NO2 (a, c) and CO
(b, d) columns retrieved from satellite measurements (see green curves) and
simulated using the CHIMERE CTM both with and without anthropogenic
emissions in the study region (see red and blue curves, respectively). The
simulated data shown have been debiased: the differences (see brown curves)
between either the annual (a, b) or monthly (c, d) averages of the simulation
and measurement data were subtracted from the original simulation data.
Preliminary comparative analysis of the measurement and simulation data
In this section, we compare the measurement and simulated data and assess to
what extent the variability of the NO2 and CO columns over the study
region is affected by direct anthropogenic emissions in the same region.
Figure 3 shows time series of the daily values of NO2 and CO columns
averaged over the study region (see Fig. 1). The model was run both with and
without anthropogenic emissions in the study region, and the model results
are presented in Fig. 3 after compensating for a systematic difference with
the measurements. The systematic difference (the bias) was evaluated as the
average difference between the model data (obtained by running CHIMERE with
full emissions) and the corresponding measurements. The averaging was
carried out either directly for the whole annual period considered (see Fig. 3a, b) or for each month independently (see Fig 3c, d). Note that the
modeled NO2 and CO columns shown in Fig. 3 were sampled consistently
(both in time and space) with the respective available satellite data and
processed using averaging kernels (see Sect. 2.2 and Eqs. 1 and 2); a very
small difference between the CO columns calculated with and without
anthropogenic emissions in the study region partly reflects the relatively
low sensitivity of the CO retrievals in the boundary layer (compared to the
upper troposphere).
It can be seen that both the NO2 and CO measurements exhibit strong
day-to-day variability. A part of the observed variability is captured by
the model, but the amplitude of the variations is typically smaller in the
simulations than in the measurements. Exact reasons for the stronger
day-to-day variations in the measurements are not known: one possible reason
is that the variations in the measurements may reflect random errors in the
retrieval procedures (see Sect. 2.1), while another possible reason is that
a part of the variations in the measurements may be due to factors which are
not taken into account in our model (such as daily variability in the
boundary conditions). Apart from the day-to-day variations, both the
NO2 and CO columns manifest slower variations. Such variations have a
seasonal component in both the measured and simulated NO2 columns, with
larger values observed in winter than in summer. A regular seasonal
variability is visible also in the simulated CO data; however, similar
variability in the corresponding measurement data appears to be offset by
slower (probably interannual) variability, which is not reflected in the
boundary conditions of CHIMERE. The differences between the measurements and
simulations vary from month to month, thus indicating the importance of
evaluating the biases on shorter than annual timescales; this observation
is taken into account in our inversion procedure described in Sect. 3.2. It
should be noted that the seasonal changes in the monthly biases may partly
be due to errors in the seasonal cycles of the emissions specified in
CHIMERE and in the global models that were used to obtain the a priori
NO2 and CO profiles for the respective retrieval procedures (see
Boersma et al., 2011, and George et al., 2009, for details); such changes may
also be indicative of some errors in the assumed seasonal variations of
other parameters of the retrieval procedures, such as surface
reflectance or atmospheric scattering by clouds and aerosol in the case of
the NO2 retrievals and surface temperature, local emissivity, vertical
distributions of atmospheric temperature, and humidity in the case of the CO
retrievals. Figure 3 also shows that while the anthropogenic emissions in
the study region provide the predominant contribution to the NO2
columns over the same region, the respective signal in the CO columns is
very small.
Spatial distributions of the annually averaged NO2 (a, c, e) and CO (b, d, f)
columns obtained from satellite observations (a, b) and
model runs performed with (c, d) and without (e, f) anthropogenic emissions
in the study region. Red lines (e, f) depict four subregions used in the
uncertainty analysis described in Sect. 3.4; the subregions contain
approximately the same amounts of daily data. Note that the simulation data
have been debiased (in the same way as the data shown in Fig. 3a, b). Note
also that the data which are not taken into account in our inverse modeling
analysis are not shown.
Figure 4 presents the spatial distributions of the annually averaged
NO2 and CO columns derived from OMI and IASI measurements and simulated
with the CHIMERE CTM. Note that only the data taken into account in our
analysis are shown. NO2 columns from both the measurements and
simulations show very strong spatial variability correlating with the
spatial distribution of NOx emissions (cf. Figs. 4a, c and 1a);
this observation is coherent with findings of earlier studies (e.g.,
Konovalov et al., 2006; Napelenok et al., 2008; Mijling et al., 2012)
demonstrating that satellite retrievals of NO2 columns combined with
CTM outputs can provide useful information on the spatial distribution of
NOx emissions on a regional and even local (e.g., cities) scale.
However, the simulations do not reproduce the spatial variability of
NO2 columns perfectly. In particular, the NO2 column amounts over
the hot spots located in the heavily industrialized Po Valley in Northern
Italy, as well as over an industrialized region in the northwestern Germany
and Madrid, are considerably smaller in the simulations than in the
measurements; in contrast, the simulated NO2 column amounts tend
to be larger than the satellite retrievals over Great Britain. These
differences may be due to uncertainties in the spatial distribution of
NOx emissions as well as to measurement and simulation errors.
Consistent with the results shown in Fig. 3, the signal from anthropogenic
emissions appears to be rather weak and “smeared” in the spatial
distribution of the CO columns. There are also big differences between the
retrievals and simulations in some locations. Both the retrieved and
simulated CO column amounts tend to be elevated over areas where the
anthropogenic emissions are particularly large (such as those in Belgium,
Germany, England, or the Po valley in Italy). However, Fig. 4f, showing the CO
columns simulated without anthropogenic CO emissions in the study region and
transformed using averaging kernels (see Eq. 2), bears evidence that an
“anthropogenic signal” in the spatial variations of the measured CO columns
may come mostly from the a priori CO columns employed in the retrieval
procedure. Therefore, the preliminary analysis presented in this section
indicates that the NO2 measurements can potentially provide much
stronger constraints for anthropogenic emissions on a regional scale
compared to the CO measurements.
Method
Preliminary remarks
Our method is first described below for a rather general case (with
arbitrary numbers of proxy species and emission source categories and for an
arbitrary region); some settings specific for this study are either
explained later or have been discussed in Sect. 2. The main steps of the
method were briefly outlined in Introduction. The key step of the method –
namely, the estimation of annual emissions of a proxy species from different
categories of sources (emission sectors) in a region of interest – is
described in Sect. 3.2. This step involves optimization of the emissions for
a given sector by fitting simulations performed with a CTM to satellite observations of a corresponding species. An
important element of the first step is the estimation and elimination of a
possible systematic discrepancy between the simulations and observations
which is not related to uncertainties in a priori emission data. Further
steps leading to the estimation of the budgets of FF CO2 emissions are
described in Sect. 3.3. An important part of the method is dedicated to the
estimation of the confidence intervals for all our emission estimates (see
Sect. 3.4).
Optimization of emissions of proxy species
We estimate annual totals of anthropogenic emissions, Ecs, from
different categories of sources, c (c∈[1, Nc],
where Nc is the total number of categories), for a given proxy
species, s, in a study region. To do that, we combine observations,
Cos, of the species atmospheric column amounts with
respective modeled data, Cms, by assuming (similar
to,
e.g.,
Berezin et al., 2013) that Cms depends on the emissions of
a corresponding species linearly:
Cms≅Cmbs+∑cScsacs(Ecs-Ẽcs),
where Ẽcs are the available (a priori) bottom-up annual
anthropogenic emission estimates for a species s and a source
category c, acs is the vector specifying
allocation of the annual anthropogenic emissions to each cell of model's
grid and each day of the simulations, Scs is the
Jacobean matrix containing sensitivities of the model outputs to the
emissions, and Cmbs are the species amounts calculated in a
“base” model run using the bottom-up emission inventory data. Note that in
this study, Eq. (3) was used specifically to express the modeled
relationships between NO2 measurements and NOx emissions, as well
as between CO measurements and CO emissions.
The annual emission estimates for individual source categories, Ecs,
constitute the control vector of our inverse problem,
Es. The optimum estimate of
Es can be obtained by minimizing the sum of
the squared differences between the observations and simulations as follows:
E^s=argminCos-Cms+ΔsTCos-Cms+Δs,
where E^s is the optimal estimate of the control vector, and
Δs denotes the systematic discrepancies
between the simulations and observations of a given proxy species
s. Note that different components of the vectors
Cos, Cms, and Cmbs are
assumed to represent available values of the respective columns amounts of
the species s in different grid cells and/or different moments of
time in the region and period considered.
The estimation given by Eq. (4) formally implies that the errors are
homoscedastic, normally distributed, and uncorrelated in space and time;
deviations of real data from these ideal assumptions can result in errors in
E^s, but we attempt to take such errors into account in respective
confidence intervals for E^s (see Sect. 3.4). The systematic
discrepancies Δs, which are assumed to be
independent of emission uncertainties and are estimated as explained below,
can, in principle, be due to systematic errors both in the simulations and
observations. For definiteness, Δs is assumed
in this study to be due to biases in the simulations; the vector
Δs is referred to below as simply “the bias”.
Formally, it can be defined as follows:
Δs=〈Cms-Cos〉,
where the brackets denote the averaging over the assumed statistical
ensemble of probable values of
Cms and
Cos in a situation when the anthropogenic
emissions in the study regions are known exactly.
Note that Eq. (4) does not include any formal a priori constraints on the
magnitude of the optimal emission estimates (unlike many other inverse
modeling studies) or any other regularization terms, and, accordingly, our
procedure does not involve any explicit quantitative settings for the a
priori error covariance matrices. In this way, we avoid possible
uncertainties in optimal emission estimates that could be associated with
such settings. Not using a priori constraints on the magnitude of the
optimal emission estimates also enhances the value of the CO2 emission
estimates derived from completely independent measurements of different
proxy species for cross-validation purposes, because otherwise the top-down
estimates of emissions of the proxy species (and, accordingly, hybrid
estimates of CO2 emissions) could be more strongly dependent on the
data of bottom-up inventories providing a priori estimates. Avoiding formal
a priori constraints (or any other regularization) does not necessarily result
in ill-conditioning of an inverse problem, as long as the dimension of the
control vector does not exceed that of the measurement vector (Enting,
2002), and it is definitely so in our case. Although satisfying this
criterion alone cannot guarantee that the problem is well conditioned, the
numerical experiments presented below in Sect. 3.5 indicate that errors in
our emission estimates due to probable errors in input data remain limited
and thus the problem considered in this study is not ill conditioned. The
dimension of the control vector (one or two) is much smaller, in our case,
than that of the measurement vector (including tens of thousands of
observations) because we do not attempt improving the allocation of the
emissions in space and time: the vectors acs are assumed to
be known (in practice, acs are provided implicitly by an
emission interface in a CTM). Similar assumptions are not unusual in inverse
modeling studies involving CTMs (e.g., Pétron et
al., 2004; Müller and Stavrakou, 2005; Huneeus et al., 2012), when the
emissions are corrected for big regions rather than for each model grid cell
individually: indeed, optimization of emissions of chemically reactive
species (like NOx) is, in a general case, a time-consuming task, even
when an adjoint code is available. A drawback of fixing the spatial
distribution of the emissions in inversion is a probable aggregation error
(Kaminski et al., 2001). Similarly, errors in our total annual emission
estimates can also result from fixing the temporal distribution of the
emissions. For example, if the assumed seasonal cycle of the emissions
overestimates them in summer and underestimates in winter, then, taking into
account that more satellite observations are available in summer than in
winter (because of seasonal differences in the atmospheric conditions), our
annual estimates can be biased negatively. We attempted to take into account
possible errors in our estimates due to errors in spatial and temporal
allocation of the emissions in the uncertainty analysis (see Sect. 3.4).
We assess the bias for a given data point i as the average
difference between the simulated and observed columns of a species
s for the month m in which the data point i lies:
Δis≅∑jθjsm-1∑jθjsmCmjs-Cojs,
∈θjs=1,j∈Ωmθjs=0,j∉Ωm,
where Ωm denotes the subset of the available data for a given
month m, and i∈Ωm is the index of a
component (a point in time and space) of the vector Δs. It should be noted that values of
Cms in Eq. (6), like those in Eq. (4),
depend on the control vector, Es. When
combined with Eqs. (3) and (6), Eq. (4) specifies a linear optimization
problem that can be easily resolved numerically. Effectively, information
about optimal values of the emission vector is obtained from spatial and
temporal variations of the observations and simulations within each month.
Equation (6) provides a simple approximation for Eq. (5) by implying that the
systematic differences between different pairs of simulations and
observations corresponding to a given month are about the same; that is, we
assume that the bias is uniform in space and time during a given month. In
reality, however, systematic errors of satellite retrievals and model
results can be different for different grid cells and days. Therefore, this
approximation (which reflects the lack of any a priori information about
the bias) may introduce some extra errors in our emission estimates which
would not appear if the structure of the bias were known exactly. Although we
cannot avoid such errors, we try, at least, to take them into account in the
confidence intervals for our estimates. Note that as long as there is only
one realization of Cms and
Cos for the region and period considered,
an unambiguous separation between their random uncertainties and systematic
errors is hardly feasible anyway.
Summing up the optimal emission estimates for the different source
categories provides the estimate of total emissions, E^sums, of
the species s in the study region. Alternatively, the estimate of
the total emissions can be obtained by applying the estimation procedure
described above to the special case where all emission sources are
aggregated together and Nc= 1. The corresponding
optimal emission estimates are denoted below as E^tots.
Considering the difference between E^sumsand
E^tots provides a useful test for self-consistency of the
inversion procedure: the difference should not exceed the combined
confidence intervals (that are expected to include an aggregation error
among other uncertainties) for E^sums and E^tots.
The estimation method described above requires the knowledge of the product
of the Jacobean matrix, Scs, and of the vector
acs (see Eq. 3), while the knowledge of the Jacobean matrix
itself is not needed. In this study, this product was evaluated as the
difference between the results of a model base run performed with the
standard emission settings as described in Sect. 2.3 and the results of the
special runs (EHI or TCO) performed after decreasing the annual EMEP
emission values for the respective (EHI or TCO) source categories by 10 %.
The product of Scs and acs in the
case where all emission sources were aggregated together (that is, with
Nc= 1) was evaluated as the sum of the products of
Scs and acs for the two
individual (EHI and TCO) emission categories.
Note that we analyzed only the measurements over land in the study region,
and so the measurements outside of the study region (e.g., over ocean) were
not used. Such a limitation affected the amount of data used in the
analysis, but we do not see any reason to expect that it could result in any
biases in our emission estimates, which would not be covered by their
uncertainty intervals (evaluated as explained in Sect. 3.4). Likewise, we do
not expect that any biases in our emission estimates can be caused by
NOx and CO emissions outside of the study region. Indeed, on the scales
considered, it seems reasonable to regard temporal and spatial variations of
NO2 and CO originating from any sources (including ship emissions)
outside of the study region as model errors on top of the modeled variations
of NO2 and CO originating from inside of the study region. Accordingly,
we do not distinguish such variations from other errors and treat their
systematic and random parts as explained in this section (see Eq. 6) and in
Sect. 3.4, respectively.
Estimation of FF CO2 emissions
Following Berezin et al. (2013), we introduce
the conversion factors, Fcs, describing the relationships between
the annual emissions for a given proxy species s and the CO2 emissions:
Fcs=ẼcCO2Ẽcs,
where ẼcCO2 and Ẽcs are the annual
estimates of anthropogenic CO2 emissions and of anthropogenic emissions
for a species s for a given emission source category (sector)
c. Here (as above), the tildes indicate that the emission estimates
are obtained from a bottom-up emission inventory (as opposed to the optimal
emission estimates, E^cs, inferred from the measurements
according to Eq. (4) by using the modeled relationships between the column
amounts of a given proxy species and corresponding emissions).
Application of the conversion factors to the corresponding optimal emission
estimates allows us to obtain the hybrid CO2 emission estimates,
E^scCO2, that are partly constrained by the measurements but also
depend on data of the emission inventory:
E^scCO2=FcsE^cs.
Similarly, we can estimate the total CO2 emissions:
E^s, sumCO2=∑cFcsE^cs.
The alternative total CO2 emission estimate, E^c, totCO2,
can be inferred directly from an estimate of the total emissions for a proxy
species:
E^s, totCO2=FtotsE^tots,
where Ftots is the conversion factor evaluated similar to Eq. (7)
but by using total annual emission estimates based on emission inventory
data, and E^tots are the corresponding estimates inferred from
satellite measurements. Note that the conversion factors that were used to
obtain our hybrid FF CO2 emission estimates reported below in Sect. 4.2
were calculated with the EDGAR v4.2 emission inventory data.
The hybrid CO2 emission estimates derived from measurements of
different species can be used for the cross-validation purposes
(specifically, the different estimates are expected to agree within the
range of their confidence intervals if all uncertainties including
aggregation errors are adequately accounted for in the inversion procedure).
They can also be combined by taking into account the uncertainty of the
individual estimates. Specifically, given Ns individual
emission estimates, E^scCO2, the combined (maximum likelihood)
estimate of the CO2 emissions, Ecomb, cCO2, and its uncertainty
range can be expressed as follows:
E^comb, cCO2=∑s=1NsσscCO2-2-1∑s=1NsE^scCO2σscCO2-2;σcomb, cCO2=∑s=1NsσscCO2-2-1/2,
where σscCO2 are the uncertainties (the standard deviations)
of E^scCO2.
A combined estimate for the total CO2 emissions,
E^comb, totCO2, can be obtained in a similar way by using values
of E^s, totCO2. An alternative combined estimate for the total
emissions, E^comb, sumCO2, can be obtained by summing up values
of E^comb, cCO2 for different source categories c. The
standard deviations σscCO2can be evaluated as described in the
next section (Sect. 3.4). Importantly, according to Eq. (11), the probable
uncertainty of the combined estimate E^comb, cCO2 is smaller
than the uncertainty of any of the individual estimates. It should be noted,
however, that Eq. (11) provides the maximum likelihood estimate only if the
“input” emission estimates derived from measurements of individual proxy
species are statistically independent from each other; otherwise it would be
necessary to take into account their error covariances. Applicability of Eq. (11) to the situation addressed in this study is discussed in Sect. 4.2.
Uncertainties in the emission estimates
Evaluation of credible confidence intervals for our optimal emission
estimates by using a typical error propagation technique requires proper
knowledge of the statistical characteristics of model and measurement
errors. However, in case of simulations and satellite measurements of minor
atmospheric species, such knowledge is usually lacking due to complexity and
multiplicity of factors that may lead to retrieval and model errors. Taking
such considerations into account, instead of using the error propagation
technique, we follow the so-called subsampling approach (Politis et al.,
1999). Subsampling suggests estimating the confidence interval of a sample
statistic (e.g., the variance) by considering variability of that statistic
among subsamples drawn from the original sample without replacement.
To adopt the subsampling approach in this study, the original set (sample)
of input data for a given proxy species s is divided into
nd subsets (subsamples). From each subset, a “partial”
independent emission estimate, E^c,is (i∈[1, nd]), is inferred. The partial estimates can be used to
evaluate the standard error, σcs, of E^cs (that
is, the standard error of the sample estimate) as follows:
σcs≅1nd(nd-1)∑i=1ndE^c,is-E^c(•)s2,
where (•) denotes the mean over all the partial estimates.
Importantly, the estimation given by Eq. (12) requires the partial estimates
to be statistically independent. If this condition is satisfied, the partial
estimates, E^c,is, that are involved in Eq. (12) can be
regarded as independent observations of the same characteristic:
deviations between E^c,is and E^c(•)s
can only be due to errors in the simulated and measured data. In this
sense, Eq. (12) essentially evaluates the standard deviation of the mean of
individual observations (individual top-down emission estimates in our
case) affected by random errors. Note that a simple and robust estimation
technique involving Eq. (12) is basically the same as one of the oldest and
popular techniques within the subsampling approach, known as replicated
sampling (Deming, 1960; Lee and Forthofer, 2006). The standard errors in our
estimates, E^sums and E^tots, for the total
emissions of proxy species can be evaluated in the same way (that is, by
substituting E^sum,is and E^sum(•)s or
E^tot,is and E^tot(•)s into Eq. (12)
instead of E^c,is and E^c(•)s).
The statistical independence of the partial estimates could not be ensured
in our case if different subsets were selected in a quite arbitrary way. The
reason is that the model and observation errors tend to covariate both in
space and time (as confirmed by our analysis discussed below in Sect. 3.5).
Thus, on the one hand, the data included in different subsets should be
sufficiently separated in time and/or space to avoid co-variation of
errors of different partial estimates. On the other hand, the number of the
subsets should not be too small to ensure that the standard error estimate
is sufficiently reliable (note that statistical inference defined by Eq. (12) is based on nd- 1 degrees of freedom). It was also
necessary to take into account that the error structure in temporal and
spatial domains can be different.
In view of these considerations, we opted to divide the original dataset
into four subsets in the temporal domain and four subsets in the spatial domain.
Each of the subsets in the temporal domain included data for only one season
but for the full spatial domain. The gridded data subsets for winter,
spring, summer, and autumn months included 3.9 × 104,
4.1 × 104, 5.4 × 104, and 4.1 × 104
values in the case of NO2 measurements and 2.6 × 103,
1.4 × 104, 2.5 × 104, and 1.2 × 104
values in the case of CO measurements. The spatial subsets were defined as
shown in Fig. 4e, f and each included about 4.3 × 104 and
1.3 × 104 values for the whole year in the cases of NO2
and CO measurements, respectively. The standard error was estimated in
accordance to Eq. (12) independently for both “temporal” and “spatial”
subsets (that is, nd was equal 4 in the both cases), and the
maximum of the two estimates of σcs was selected as the final
estimate of the standard error. Note that such a division allowed us to
retain most of the actual error covariances within a given subsample, as the
areas and time periods covered by each subset were significantly larger than
expected error covariance scales (see Sect. 3.5 for further details). In contrast, selection of the maximum of the two different σcs
estimates may result in overestimation of the confidence intervals that can
be robustly evaluated by applying t values (from the Student's distribution
with three degrees of freedom in our case) to the standard error estimate.
We expect that apart from random errors in the input data, the error
estimate obtained as described above also includes (at least to some extent)
the aggregation error (Kaminski et al., 2001). In this study, that kind of
error may be due to aggregation of similar sources in all the countries
considered into a single component of the control vector. As contributions
of various sources to the CO and (especially) NO2 columns in the
different countries are different, the aggregation error is likely to be
manifested as deviations between the different partial estimates. For
example, if, in a hypothetical situation, an emission estimate inferred from
the full dataset were mostly affected by strong emission sources from only
one country, a partial estimate obtained after leaving the measurements over
that country out would likely be much less affected by the same sources, at
least in the case of emission estimates of such a short-lived species as
NOx.
The confidence intervals estimated using Eq. (12) are also likely to account
for most of estimation errors associated with uncertainties in the diurnal
and weekly variations of anthropogenic emissions, as well as with
uncertainties due to shortcomings in the model representation of chemical
processes (including effects of subgrid-scale chemical interactions).
Indeed, it seems reasonable to expect that different errors of the emission
temporal cycles for different emission sectors, countries, and seasons can be
manifested as quasi-random deviations between the simulations and
measurements in different grid cells and days. Uncertainties in the diurnal
variations of emissions are likely to be manifested additionally in the
differences between the hybrid CO2 emissions estimates inferred
separately from the CO and NO2 measurements, as those measurements are
taken in different times of a day (see Sect. 2.1). Uncertainties in
simulations of chemical processes and subgrid-scale chemical interactions
are likely to have a different impact on the modeled NO2 or CO columns
in different types of environments (e.g., rural or urban) and in different
seasons; therefore, the respective model errors are likely to differ in
different grid cells (and days of an year) and to have a different impact on
the NOx or CO emission estimates for different subregions and seasons.
Accordingly, it indeed seems reasonable to assume that such errors are
mostly taken into account in the emission estimate uncertainties evaluated
with Eq. (12). In addition, as the NO2 and CO behaviors are governed by
essentially different chemical processes, uncertainties due to a “chemical”
part of model errors are likely to contribute to differences between the
CO2 emission estimates based on the NO2 and CO measurements.
Note that it is nonetheless not quite infeasible that some model errors
associated with the representation of chemical interactions can result in
similar (positive or negative) biases across the NOx or CO emission
estimates inferred from the different data subsets. For example, systematic
underestimations of the NOx emissions may be due to persistent positive
biases in the ozone formation rate and in boundary conditions for
tropospheric ozone concentration (as ozone concentration accounts for
partitioning of NOx between NO and NO2) as well as due to other
numerous factors (such as underestimation of the hydrocarbon emissions
or of the ozone photolysis rate) that may result in underestimation of
concentration of hydroxyl radical providing a major sink for NOx and
determining its atmospheric lifetime (Seinfeld and Pandis, 2006). Depending
on atmospheric conditions, effects of different model errors on the emission
estimates may or may not compensate each other. Even though different model
errors are likely to combine and affect the emission estimates in different
ways in the different subregions and seasons, we cannot completely ensure
that the confidence intervals for our CO and (especially) NOx emission
estimates actually account for all possible model errors. More accurate
evaluation of effects of possible errors in the model representation of
chemical processes on NOx and CO emission estimates that can be derived
from satellite measurements by using our inverse modeling method requires
further research (involving, e.g., multi-model inversions) that goes beyond
the scope of this study.
Uncertainties in the seasonal cycles of anthropogenic emissions are likely
to be manifested (in the absence of any other model and measurement errors)
as the differences between the annual emission estimates obtained with the
four data subsets including data for the different seasons. Therefore, we
expect that such uncertainties are also addressed in the confidence
intervals evaluated as explained above. However, compared to the diurnal and
weekly variations, uncertainties in the seasonal variations of emissions
more probably result in common systematic biases of NOx and CO emission
estimates. To get an idea about the magnitude of such biases, we compared
the emission estimates for the two cases involving simulations with
different seasonal cycles. The first case (referred to below as the “cycle”
case) corresponds to the standard seasonal cycles assumed in our model (see
Fig. 2). The second (“flat”) case corresponds to simulations performed with
constant emissions in any month of a year (yet with the same diurnal and
weekly emission temporal profiles as in the cycle case). Note that the
differences between the emissions estimates obtained for these two cases are
likely to strongly exceed the respective uncertainty (because the flat
case is evidently unrealistic).
It should be noted that the qualitative considerations discussed above are
by no means intended to strictly prove that the estimations based on
Eq. (12) actually account for all possible errors. Nonetheless, taking the
above arguments into account and given the fact that both the origins and
the statistical characteristics of errors in the measurement, simulation, and
inventory data involved in our analysis are very poorly known, we believe
that the simple and robust subsampling technique described above provides
sufficiently reliable and robust uncertainty estimates and has no serious
alternative in the situation considered. Some further arguments supporting
reliability of this technique are discussed in Sect. 3.5.
To obtain the confidence intervals for our CO2 emission estimates, we
need to combine the uncertainty of our estimates of emissions of proxy
species with the uncertainty of the corresponding conversion factors.
Ideally, the uncertainty of the conversion factors for source categories
that group different sectors (like EHI and TCO) could be obtained, e.g., by
varying parameters of a bottom-up inventory (Wang et al., 2013) and provided
along with emission data. However, in our knowledge, such information has
unfortunately not yet been made available within any inventory except those
by Wang et al. (2013) for China. As an alternative approach, we suggest that
the uncertainty of the conversion factors can be roughly estimated by
comparing their values based on data of different emission inventories.
Ideally, it would be best to consider an ensemble of several independent
inventories providing the data on emissions of all the species (NOx, CO,
and CO2) involved in our analysis. However, in this study, in view of
the limited practical availability of the necessary data, we realized only a
highly simplified version of such an approach. Specifically, along with the
conversion factors based on the EDGAR v4.2 emission inventory (those values
were used to obtain our “main” CO2 emission estimates as described in
Sect. 3.3 and are denoted in this section simply as Fcs), we
considered “alternative” conversion factor values based on the data of other
inventories, such as EMEP and CDIAC. The alternative conversion factor
values are denoted below as Fcs'. Specifically, we used the EMEP
inventory data for NOx and CO emissions and the CDIAC data for FF
CO2 emissions (see Sect. 2.3). Taking into account considerable
differences in the data sources and methodologies used across the three
inventories (see Sect. 2.3 and the corresponding references for details), we
assume that the main and alternative conversion factor estimates are
sufficiently independent. As the CDIAC emission data had not been originally
distributed among individual emission sectors, the fractions of the two
categories of the CO2 sources were taken to be the same as in the
EDGAR v.4.2 inventory. However, only the original CDIAC and EMEP data were
used to estimate the conversion factors applied to the total
emissions (Ftots').
Using again the subsampling technique, we roughly estimated the standard
error for the conversion factors, σscF, as follows:
σscF=1(Nk-1)Nk∑k=1NkFc,ks-Fc,ks'-Fc(•)s+Fc(•)s'2+Fcs-Fcs'2,
where Fc,ks and Fc,ks are the conversion factors evaluated
individually for each of the 12 countries considered, c is the
country index, Nk is the total number of the countries
considered (Nk= 12 in this study), and (•)
denotes the means over the countries. The country scale is used in Eq. (13),
because the CDIAC data had not been provided on a spatial grid, and thus we
could not consider the same spatial subsamples as those with the data for
NO2 and CO columns. The estimations given by Eqs. (12) and (13) are
based on the same idea, except that unlike Eq. (12), Eq. (13) does not
involve the assumption that the error of a “sample” estimate is completely
random in origin; rather, it takes into account that the error may contain
both random and systematic components. The latter is evaluated in Eq. (13)
as the difference between the estimates Fcs and Fcs'
representing the full study region. Actually, that difference may include a
part of the random error, so Eq. (13) is likely to overestimate σcsF. Further overestimation may be due to the fact that the
differences in Eq. (13) comprise cumulative errors in the both conversion
factor estimates: if the errors were distributed equally between the main
and alternative estimates, a proper value of σcsF would be at
least the factor of 21/2 smaller than the one given by Eq. (13). In
contrast, using the same (EDGAR v.4.2) data to evaluate both Fc,ks
and Fc,ks' may compensate such an enhancement or even entail a
tendency for underestimation in σscF (except for the case
where the conversion factors and their uncertainties are estimated directly
for total emissions, that is, without sectorial breakdowns). Nonetheless, on
the whole, taking the above qualitative considerations into account, we
expect that values of σscF calculated as described above are
more likely to be overestimated than underestimated, thus being conservative
in our approach to provide optimal CO2 emission estimates.
Values of the conversion factors, Ftots and Ftots',
calculated using different inventories for each country considered are shown
in Fig. 5. The differences between the different estimates of the conversion
factors are, in general, considerable and vary across different countries in
the study region. Specifically, the differences for the NOx-to-CO2
emission and CO-to-CO2 conversion factors range from 1.4 to 24.9 %
and from 3.8 to 52.6 % (relative the values based on the EDGAR v.4.2
data), respectively. The differences are smallest for Austria and Germany.
NOx-to-CO2 (a, b) and CO-to-CO2 (c, d) emission
conversion factors obtained using NOx, CO, and CO2 emission
estimates from the EDGAR v4.2 emission inventory (a, c) and from the CDIAC
and EMEP emission inventories (b, d) for the emission totals.
The standard error, σscCO2, representing the uncertainty in
our hybrid estimates of anthropogenic CO2 emissions was estimated by
assuming that uncertainties in the estimates of a proxy species emissions
and in the estimates of the conversion factors are independent:
σscCO2=E^scCO2σcsE^cs2+σscFFcs2.
The standard error, σs, totCO2, for a corresponding total
CO2 emission estimate, E^s, totCO2 (see Eq. 10), was
evaluated in the same way. Taking into account that the uncertainties in the
top-down estimates of emissions of proxy species for different source
categories are likely not independent, the standard error, σs, sumCO2, of E^s, sumCO2 (see Eq. 9) was given by a
similar but slightly more complicated equation:
σs, sumCO2=∑cE^csσscF2+σs, sumCO2|F2,
where σs, sumCO2|F represents the standard error of
E^s, sumCO2 under the condition that the conversion factors are
known exactly (that is, the errors included in σs, sumCO2|F are associated with only uncertainties of our top-down emission
estimates for the proxy species); σs, sumCO2|F was
evaluated by using the same subsampling technique as described above for the
case of estimation of uncertainties in E^cs. The standard
errors given by Eqs. (14) or (15) allowed us to combine the estimates based
on the measurement of NO2 and CO columns by using Eq. (11).
Observation system simulation experiments (OSSEs; tests with synthetic data)
In this section, we examine the capabilities of our method for estimation of
emissions of the proxy species by means of OSSEs. Specifically, we apply our method to synthetic
“observational” data featuring known uncertainties that are evaluated by
considering the misfits between real observation and corresponding simulated
data. Specifically, to generate the synthetic data, we assumed that the
covariances of cumulated errors in the real measurement and simulation data
can be described by the three-dimensional covariance function,
covsρx,ρy,ρt, that can be
approximated as follows:
covsρx,ρy,ρt≅covxsρxcovysρycovtsρt,
where ρx and ρy denote the distances
between a pair of observations in west-to-east and south-to-north
directions, respectively; ρt is the period (the lag)
between different observations; and covxs, covys, and
covts are the respective one-dimensional covariance functions. We
further approximated the covariance functions by using misfits between the
observations and simulations as follows:
cov∗sρ∗∼∑i∑jHijs[ρ∗]Cois-Cmis+ΔisCojs-Cmjs+Δjs,
where the subscript ∗ denotes either x, y, or
t; Hijs[ρ∗] is the selection operator
which is nonzero (unity) only for those pairs of data points that
correspond to a given value of ρ∗;Cos and Cms are the vectors of the observational and simulated
data;
and Δs is the bias. The distances and the lag were
expressed in the numbers of grid cell and days, respectively. The vector
Cos involved in Eq. (17) represents the
actual observational data described in Sect. 2.1. The simulated data,
Cms, were obtained from the
model base run results presented in Sect. 2.4, and the bias was evaluated
on the monthly basis as the zero-order estimate obtained by applying Eq. (6) to the same data (that is, without using top-down emission estimates).
The covariance functions, covxs, covys, and
covts, evaluated according to Eq. (17) were found to have the
following characteristic scales (corresponding to a 2-fold decrease of the
covariance functions): 3 (5) and 2 (3) grid cells and 1 (1) day in the case
of NO2 (CO) columns, respectively, although these scales do not
necessarily reflect the presence of rather long “tails” in the covariance
functions.
Our OSSEs are not expected to disregard any of possible errors in
observational and model data that determine variability of the misfits
between the observations and simulations within 1 month, although it
should be noted that Eq. (16) provides a rather simplified temporal and
special structure of such errors. In particular, our error model does not
allow us to take into account probable error “clusters” that can be
associated with the aggregation error in our optimal emission estimates, as
well as less probable model errors (see Sect. 3.4) that affect the modeled
relationship between the NOx emissions and the NO2 columns but do
not contribute to the variability of the differences between the
observations and simulations. Nonetheless, inversion of the synthetic data
generated even with the simplified error model is useful, as it allows us to
assess the adequacy of our uncertainty estimates obtained with the
subsampling technique in the presence of probable covariances of errors in
the input data, as well as to examine the self-consistency of our procedure
(that is, to see whether or not any systematic deviations of our optimal
emission estimates from the “true” emission values are covered by the
corresponding confidence intervals).
Estimations given by Eqs. (16) and (17) were used to set up a Monte Carlo
experiment in which the vector
Cms (obtained either from the
model base run or from a model run with the emissions perturbed as explained
below) represented the true content of a given species, while the
synthetic observations were generated by adding random errors
(and, in some cases, biases) to
Cms. Samples of the
errors with the covariance structure given by Eq. (16) were generated from a
Gaussian distribution by using a standard method (Press et al., 1992)
involving the Cholesky decomposition of the correlation matrices that were
specified, in our case, using the covariance functions given by Eqs. (16)
and (17). The Cholesky decomposition of a correlation matrix gives a
lower-triangular matrix, L; applying this matrix to a vector of
uncorrelated samples of Gaussian noise, u, gives a vector,
Lu, with the components satisfying the
original correlation matrix. Using the synthetic data, we obtained
uncertain emission estimates which were compared with true emission data
specified in the model. Each Monte Carlo experiment included 100 iterations
performed with the same covariance matrix and with the same bias,
Δs, but with different samples of
random errors. The bias added to
Cms in each experiment for any
given day was specified by linearly interpolating (in time) the monthly
biases shown in Fig. 3c, d, with the magnitude of the monthly bias values
scaled (in different experiments) with a factor (δ) ranging from 0
to 1.
The uncertainties (expressed as the standard error) in the emission
estimates were evaluated both in the “direct” way (as the root mean square
difference between the uncertain and true emission estimates) and by
averaging squares of σcs calculated by using the subsampling
technique described above. The magnitude of errors in the synthetic data was
changed in different experiments by applying a scaling factor (σ)
ranging from 0 to 1 to the covariance matrix given by Eq. (16). An
additional factor (ξ) was introduced to scale the non-diagonal
components of the covariance matrix: ξ equals zero in an “ideal” case
where errors in each grid cells and days are statistically independent from
errors in any other grid cells and days. Along with the experiments where
the “true” emissions were set to be exactly the same as the bottom-up
emissions used in the base run of our model (see Figs. 6, 7), we performed
the experiments where the base case emissions for both NOx and CO were
either uniformly increased by 20 % (see Fig. 8) or increased by 20 % only for the EHI categories but reduced by 20 % for the TCO
category (see Fig. 9). Note that not only anthropogenic NOx and CO
emissions were perturbed in the corresponding model runs but also
respective anthropogenic emissions of all other model species, including
those of NMHCs.
Results of the OSSEs for estimation of NOx (a, b) and CO
(c, d)
emissions: the dependencies of the normalized standard error of the NOx (CO)
emission estimates for the EHI (a, c) and TCO (b, d) source categories
on the level of noise (that is, on the value of the diagonal elements,
σ, of the error covariance matrixes) in the input “synthetic” data.
The noise level is normalized to its magnitude estimated with the real
measurement and simulation data. Apart from the random noise, the synthetic
data included the bias that was specified (for any given day) by linearly
interpolating (in time) the monthly biases shown in Fig. 3c, d. Different
colors show the results obtained with the different levels of error
co-variances (ξ is the scaling factor applied to non-diagonal elements
of the covariance matrix). The standard errors estimated in the “direct” way
(as the RMSE representing the differences between the emission estimates
inferred from the synthetic data and the “true” NOx emission estimates)
and by using the subsampling technique are shown by solid and dashed lines,
respectively.
The same as in Fig. 6 but for the dependencies of the
normalized standard error on the scaling factor, δ, characterizing
the bias applied to the synthetic data (see Sect. 3.5 for further details).
The total NOx (a) and CO (b) emission estimates
(E^tots) obtained in the OSSEs where the
“true” emissions were specified by scaling the bottom-up emissions (employed
in the base case model run) with the factor of 1.2. The emission estimates
(normalized to the respective bottom-up emission estimates,
Ẽtots, based on the EMEP inventory data)
represent the average over the ensemble of 100 Monte Carlo experiments, each
with a different sample of noise in the synthetic data, and are shown as a
function the noise level (σ). Both non-diagonal elements of the
error co-variance matrix and the systematic uncertainties were taken into
account in the OSSEs (specifically, both ξ and δ were set to be
equal to unity; see further details in Sect. 3.5). Note that the value of
1.2 on the axis of ordinates corresponds to a perfect emission estimate in
the case considered. Note also that the confidence intervals shown were
estimated by using the subsampling technique (see Eq. 12) that is expected
to predict a nonzero uncertainty (associated with the bias estimation
procedure) even when the synthetic input data are not affected by random
uncertainties (that is, when σ= 0; see also blue lines in Fig. 7).
The NOx (a) and CO (b) emission estimates obtained similarly to
the estimates shown in Fig. 8 but separately for the two source categories
(EHI and TOC) in the OSSEs, where the “true” emissions in the EHI and TCO
were specified by scaling the corresponding bottom-up emissions (employed in
the base case model run) with the factors of 1.2 and 0.8. Note that the
estimates for the EHI and TCO categories are depicted by using the abscissa
axes at the bottom and at the top of the plots, respectively.
Top-down estimates (E^cs and
E^tots) of the anthropogenic
NOx (a) and CO (b) emissions in the study region in comparison with the
corresponding estimates from the EMEP and EDGAR v4.2 inventories. Our
estimates are shown for the two cases (“cycle” and “flat”) with different
seasonal cycles of anthropogenic emissions.
The results of the OSSEs indicate, in particular (see Figs. 6, 7), that if
errors in the input data for different grid cells and days were
statistically independent (ξ= 0), the uncertainties (evaluated in
the “direct” way with both σ and δ equal unity) of our
top-down estimates of both NOx and CO emissions would be very small,
specifically 0.9 and 0.6 % for the NOx emission estimates in
the EHI and TCO sectors and somewhat larger (13 and 5 %) for the CO
emission estimates for the same sectors. The fact that the uncertainties in
our emission estimates remain rather small in spite of the large
uncertainties in the input data (see Sect. 2.4) clearly indicates that the
inverse problem considered is not ill conditioned. Expectedly, taking the
error covariances into account increases the emission estimate uncertainties
considerably. The uncertainties in the estimates of NOx (CO) emissions
from the EHI and TCO sectors are found to be 4 (28) and 5 (17) %,
respectively. Larger uncertainty levels in the CO emission estimates
compared to those in the NOx emission estimates are an expected result,
reflecting the fact that the constraints to CO emissions provided by the CO
observations are much weaker than the corresponding constraints provided by
the NO2 observations to the NOx emissions. Indeed, an “emission
signal” in the CO data considered (see Figs. 3 and 4) is, on average, much
weaker than that in the NO2 data; moreover, taking into account that
the atmospheric lifetime of CO is much longer compared to that of NOx,
an emission signal from a given grid cell is effectively spread between a
much larger number of grid cells (and days) in CO than in NO2
observations, resulting in large non-diagonal elements of the Jacobian
matrix and potentially leading to a stronger sensitivity of the CO emission
estimates to errors in the input data. Interpretation of changes in the
uncertainty estimates with respect to the ideal case is difficult: it can
only be speculated that the increase in the uncertainties is larger in the
NOx than in CO emission estimates, probably because introduction of
the error covariance is effectively equivalent to aggregation of available
observations into a few “super-observations”, leading to suppression of the
effect of large non-diagonal elements in the Jacobian matrix describing the
relationship between the CO emissions and observations.
Importantly, it is found that introduction of a bias into the synthetic data
does not have a strong impact on the accuracy of the retrieved emission
estimates (see Fig. 7) affected by random errors in the input data. This
result confirms that our inverse modeling scheme is indeed capable of
efficiently filtering out the bias, even if it is not constant during 1
month (as assumed in Eq. 6).
The results of our OSSEs also indicate that the subsampling technique
employed in this study provides reasonable uncertainty estimates, although it
tends to overestimate the actual uncertainties in the experiments
representing the most realistic case (where all the scaling factors equal
unity). We consider a probable overestimation of uncertainties in our
emission estimates as a rather positive feature of our procedure, making
conclusions of this study more reliable.
The results shown in Figs. 8 and 9 demonstrate that the optimal emission
estimates obtained with our inversion procedure are likely not significantly
biased even if the true emissions are considerably different from the
bottom-up emission inventory data. These results also confirm that our
inversion procedure enables efficient separation of the uncertainties in the
model data due to emission errors from other systematic uncertainties in the
model and observation data. Importantly, the fact that the emissions
perturbations are retrieved almost perfectly indicates that the effects of
chemical interactions (nonlinearities) and changes in NMHC emissions on the
relationships between NOx and CO emissions and the NO2 and CO
columns, respectively, are likely rather small in the situation considered,
although it should be noted that such effects can be stronger if the
differences between the bottom-up and true emissions were much larger than
in our experiments (±20 %).
Results
NOx and CO emission estimates
The estimates of anthropogenic NOx and CO emissions from the EHI
and TCO categories (E^cs) as well as from all sources
aggregated together (E^tots) based on actual observations are
presented in Fig. 10. The corresponding numbers are listed in Table 1, which
also shows alternative estimates (E^sums) of the total
emissions. The results are reported for the two estimation cases (cycle
and flat; see Sect. 3.4) that involve different seasonal variations of
anthropogenic emissions in the model (specifically, the seasonal cycles
specified in the standard version of the CHIMERE CTM were used for the
cycle case estimations, while constant monthly anthropogenic emissions
with diurnal and weekly variations were employed for the flat case). Note
that the flat case is obviously unrealistic and is considered here only
for testing purposes; accordingly, if not stated otherwise, below we discuss
estimates obtained for the main (cycle) case. The uncertainties are
reported in terms of the 68.3 % (1σ) confidence intervals. It should
be noted that the confidence intervals were evaluated under the assumption
(see Sect. 3.4) that the NOx and CO emission estimates are not
significantly affected by any systematic errors that cannot be manifested in
the differences between the emission estimates for different subregions and
seasons. If this assumption holds, the uncertainty intervals evaluated with
our subsampling technique may actually correspond to a higher confidence
level, as discussed in Sect. 3.4 and 3.5.
The optimal estimates of the anthropogenic NOx and CO
emissions (Tg NO2 and Tg CO, respectively) from the study region. The
numbers in brackets represent the one-sided 68.3 % confidence intervals
(in % relative to the respective optimal estimate).
Estim.
EHI
TCO
Totals
Species
case
E^1s
EMEP
EDGAR
E^2s
EMEP
EDGAR
E^sums
E^tots
EMEP
EDGAR
NOx
cycle
2.59 (18)
2.55
4.25
4.17 (15)
4.75
3.53
6.76 (11)
6.86 (10)
7.30
7.78
flat
2.82 (23)
4.02 (17)
6.84 (7)
6.97 (7)
CO
cycle
6.99 (54)
7.41
5.55
9.59 (33)
10.84
10.02
16.57 (30)
17.03 (26)
18.25
15.57
flat
5.53 (75)
10.4 (50)
15.93 (25)
16.82 (28)
All of our optimal (top-down) estimates of both the total NOx and CO
emissions are slightly (less than 10 %) smaller than the bottom-up
estimates based on the EMEP inventory data; the differences between the
top-down and bottom-up estimates are not statistically significant. The
relative uncertainties in our estimates of the total emissions range from
10 % (in case of the E^tots estimate for the NOx
emissions) to 30 % (in case of the E^sums estimate for the
CO emissions). A low uncertainty in our estimate of the total NOx
emissions is not surprising, as random uncertainties of a very large number
of individual retrievals used in our inverse modeling analysis tend to
compensate each other, while systematic errors were taken into account in
the framework of our inversion procedure explicitly. Nonetheless, this low
uncertainty estimate should be considered with a certain degree of caution
as it may not fully account for some unknown errors depending on emissions
themselves (e.g., due to uncertainties in a model chemical scheme; see also
Sect. 3.4). Taking into account our preliminary analysis (see Sect. 2.4)
indicating that the contribution of the anthropogenic CO emissions in the
study region into the corresponding CO columns is relatively small and the
results of the OSSEs (see Sect. 3.5), it is also not surprising that the
uncertainties in our CO emission estimates are much larger than those in the
NOx emission estimates.
The differences between our alternative estimates of the total emissions,
E^sums and E^tots, are also small compared to the
uncertainties of those estimates, while the uncertainties in
E^sums are larger than the uncertainties in
E^tots. The difference between the uncertainties in
E^sums and E^tots would be difficult to predict a
priori, particularly because the cost function (see Eq. 4) employed in this
study includes the bias whose estimation may increase uncertainties in the
emission estimates to a various extent, depending on the number of variables
to be optimized. Our emission estimates for the individual source categories
are much more uncertain than the estimates of the total emissions: the
uncertainties range from 15 % in the case of the NOx emission
estimate for the TCO sector up to 54 % in the case of the CO emission
estimate for the EHI sector. The absolute differences of our estimates of
both CO and NOx emissions with the EMEP data are smaller than the
respective uncertainty range. It may be noteworthy that our estimates for
both the CO and NOx emissions from the TCO sector are ∼ 12 % lower than the corresponding EMEP estimates. This observation
indicates that there may be a common bias in the EMEP data for both NOx
and CO emissions in this sector; however, available information does not
allow us to make a firmer conclusion in this regard.
Unlike the EMEP data, the EDGAR v4.2 data strongly disagree with our
estimate for the NOx emissions from the EHI sector. The differences of
our estimates with the EDGAR v4.2 data are also larger than with the EMEP
data in particular in the cases of the E^tots estimates of
both NOx and CO emissions and in the case of CO emission estimate for
the EHI sector, although smaller in the cases of the CO emission estimates
in the TCO sectors. It is noteworthy that the differences between all our
NOx emission estimates with the corresponding EDGAR data are
statistically significant. In general, our analysis indicates that the
NOx and CO emission data provided by the EMEP inventory are more
consistent with the NO2 and CO satellite measurements than those given
by the EDGAR v4.2 inventory. This is an expected result because the
methodology used in the EMEP inventory is specific to national statistical
data available from European countries, while the EDGAR v4.2 inventory uses
another approach which is deemed to be robust at the global scale.
The differences between the estimates obtained with different types of
seasonal variations of anthropogenic emissions are small compared to the
uncertainty in the estimates for the cycle case, although not entirely
negligible. Evidently, these differences cannot explain the significant
disagreement of our NOx emission estimates with the EDGAR v4.2 data.
Nonetheless, our test results indicate that the effect of possible
inaccuracies in the seasonal variations of emissions may not be negligible
and should not be disregarded a priori when examining the significance of
the differences between the top-down estimates of annual emissions and
respective bottom-up inventory data. Note that the uncertainties in the
NOx and CO emission estimates for the individual source categories tend
to be larger for the flat case than for the cycle case, while, in contrast, the uncertainties in the total NOx emission estimates are
larger for the cycle case. Such non-symmetrical differences indicate that
none of the cases considered represent the seasonal cycles in NOx and
CO emissions quite perfectly.
The uncertainty levels in our estimates of both NOx and CO emissions
using actual data are considerably larger than those obtained above in our
OSSEs (see Sect. 3.5) in which synthetic data were generated using a
simplified error model. (Note that to be compared with the confidence
intervals discussed in this section the standard errors presented in
Sect. 3.5 should be multiplied with the t score of about 1.2.) This
result indicates that, as expected, the uncertainties in our emission
estimates are caused not only by random uncertainties in the input
(measurement and simulation) data but also by other factors – such as the aggregation error and spatial variability of the bias – which
could not be adequately taken into account in our tests. Besides, the actual
temporal and spatial structure of both the model and measurement errors is
likely much more complex and irregular than that assumed in Eq. (16).
Anyway, unless the data subsamples defined in Sect. 3.5 are strongly
affected by temporal and spatial covariances of errors in the input data (as
evidenced by our OSSEs, that is unlikely the case in this study), the
confidence intervals provided by the subsampling technique are expected to
be sufficiently reliable even in such a complex real situation as that
considered in this study.
Note that the uncertainties of our top-down estimates of NOx emissions
in the region considered turned out to be comparable with the differences
between similar estimates provided by different emissions inventories, or
even smaller than them. Therefore, our top-down NOx emission
estimates can be considered as an independent alternative to bottom-up
estimates based on emission inventory data alone. Both our NOx and CO
emission estimates could formally be combined (in the Bayesian way) with the
bottom-up (a priori) estimates; the uncertainties in the combined (a
posteriori) estimates would probably be lower than the uncertainties in
either the top-down or a priori estimates taken alone.
In general, our results confirm the findings of previous studies (see the
corresponding references in Introduction) showing that NO2 and CO
retrievals from satellite measurements can provide useful information on
NOx and CO emissions over high emission regions. In this regard, it can
be noted that while previous inverse modeling studies utilized satellite CO
measurements to estimate CO emissions from regions with predominantly
anthropogenic sources involved global CTMs (e.g., Pétron, 2004;
Fortems-Cheiney et al., 2009; Kopacz et al., 2010; Jiang et al., 2015), we
obtained reasonable top-down CO emission estimates by using a regional
model. We regard this fact as a promising development, because the use of
regional models (usually featuring a higher spatial resolution than global
CTMs and employing high-resolution regional emission inventories that are
likely more accurate and detailed compared to global ones) in inverse
modeling procedures can, potentially, provide more detailed and accurate
constraints to CO emissions from various sources. A major difficulty that
needs to be overcome in applications of a regional CTM for estimating
anthropogenic CO emissions by inverse modeling is associated with probable
biases in boundary conditions, especially for CO, which has a long chemical
lifetime compared to the transit time across the European domain; here we
tackled this difficulty by means of a special procedure aimed at eliminating
systematic differences between the measured and simulated data. The results
of the OSSEs presented in Sect. 3.5 (see Figs. 8, 9) indicate that our
estimation procedure successfully relies on the spatial gradients of CO (and
NO2) columns within the European domain to constrain the CO (and
NOx) emissions rather than on the average abundance (which is strongly
driven by the boundary conditions) in the measurements.
The NOx-to-CO2 (g CO2 [g NO2]-1)
and CO-to-CO2 (g CO2 [g CO]-1) emission conversion factors
based on the EDGAR v4.2 emission inventory along with their relative
uncertainties given in brackets as one-sided 68.3 % confidence interval
(in %).
Sectors
NOx-to-CO2
CO-to-CO2
EHI
494.97 (58)
378.63 (38)
TCO
262.06 (30)
92.42 (22)
TOT
389.22 (4)
194.50 (22)
Hybrid estimates of the fossil-fuel CO2 emissions
(E^scCO2,
E^s, totCO2
E^comb, cCO2, and
E^comb, totCO2) from the study region in
comparison with the data of the EDGAR v4.2 and CDIAC emission inventories.
The estimates were obtained either from (a) only NOx and (b) only CO or
(c) from both NOx and CO measurements. The “partial” and
“full” 68.3 % confidence intervals are also shown: the partial intervals
(depicted by narrow brackets and not shown for the combined CO2
emission estimates) are determined only by uncertainties in the top-down
estimates of NOx or CO emissions, while the full intervals also take into
account probable uncertainties in the conversion factors.
Fossil-fuel CO2 emission estimates
Our hybrid FF CO2 emission estimates presented in this section were
obtained by applying the conversion factor values listed in Table 2 to our
top-down estimates of NOx and CO emissions discussed above. The FF
CO2 emission estimates derived from NO2 and CO measurements
(E^scCO2 and E^s, totCO2) as well as the
combined FF CO2 emission estimates (E^comb, cCO2 and
E^comb, totCO2) are shown in Fig. 11 in comparison with the
corresponding data of emission inventories. The same estimates are listed in
Table 3, which, in addition, presents another version of the hybrid
estimates of total FF CO2 emissions, E^s, sumCO2 and
E^comb, sumCO2 (see Sect. 3.3). Note again (see also Sect. 2.3)
that CO2 emissions from the cement production are not included in our
estimates. Two types of the confidence intervals are provided along with the
CO2 emission estimates based on measurements of one proxy species. The
full confidence intervals include the uncertainty in the top-down
estimates of the proxy species as well as the uncertainty in the conversion
factors. The partial confidence intervals were estimated by taking into
account only the uncertainty in the top-down estimates of the NOx and
CO emissions.
The estimates of the fossil-fuel CO2 emissions (Pg CO2) from the study region in comparison with corresponding data (when
available) of the EDGAR v4.2 and CDIAC emission inventories. The numbers in
brackets represent the one-sided 68.3 % confidence intervals (in %
relative to the respective optimal estimate). Along with the “full”
confidence intervals, the “partial” confidence intervals are shown after a
slash (except for the combined estimates) and do not include uncertainties
in the conversion factors.
Inversion
Estim.
EHI
TCO
Totals
settings
case
E^s, 1CO2
EDGAR
E^s, 2CO2
EDGAR
E^s, sumCO2
E^s, totCO2
CDIAC
EDGAR
NOx-based
cycle
1.28 (72/18)
2.10
1.09 (39/15)
0.93
2.37 (43/12)
2.67 (11/10)
2.86
3.03
flat
1.40 (74/23)
1.05 (40/17)
2.45 (44/9)
2.71 (8/7)
CO-based
cycle
2.64 (71/55)
0.89 (42/33)
3.53 (49/35)
3.31 (37/26)
flat
2.09 (88/75)
0.96 (57/50)
3.06 (54/42)
3.27 (38/28)
NOx- and
cycle
1.55 (54)
0.98 (29)
2.67 (33)
2.71 (11)
CO-based
flat
1.56 (57)
1.02 (33)
2.63 (34)
2.73 (8)
The relative differences of NO2- or CO-measurement-based FF CO2
emission estimates with the EDGAR v4.2 CO2 data correspond to the
differences of our top-down NOx or CO emission estimates with the
EDGAR v4.2 data for the respective species. This is not surprising, as the
conversion factors relating CO2 emissions with the respective proxy
species were based on the EDGAR v4.2 inventory. The full relative
uncertainties in our CO2 emission estimates are larger than the
uncertainties in our estimates of emissions of proxy species due to
uncertainties in the conversion factors. Among the uncertainties in the
conversion factors, σscF, they are largest for the
NOx-to-CO2 and CO-to-CO2 emission conversion factors for the
EHI source category (58 and 38 %, respectively).
These uncertainties strongly contribute to the confidence intervals of the
respective CO2 emission estimates. In contrast, the uncertainties are
relatively small in the NOx-to-CO2 and CO-to-CO2 emission
conversion factors for the total emissions (4 and 22 %, respectively);
those uncertainties contribute considerably to the full confidence
intervals only for the total CO2 emission estimates based on the CO
measurements, while the uncertainty in the respective
NO2-measurement-based estimate is mostly due to the uncertainty in the
top-down NOx emission estimates. Note that as discussed in Sect. 3.4
and 3.5 our method is likely to overestimate uncertainties in both the
top-down estimates and in the conversion factors.
Taking into account the full confidence intervals (which are, in some cases,
very wide), all our estimates are in agreement with the EDGAR v4.2 data,
except for the estimates of the total CO2 emissions
(E^s, totCO2) based on NO2 measurements and on both
NO2 and CO measurements. Our hybrid NO2-measurement-based and
combined estimates of the total CO2 emissions (2.67 and 2.71 Pg CO2 with the relative uncertainties of about 10 %) are 12 and
11 % lower than the EDGAR v4.2 estimate (3.03 Pg CO2), respectively.
These differences are statistically significant but at the edge of
significance with the given confidence level. Note that while discussing
statistical significance of the differences between the hybrid and bottom-up
emission estimates, we do not take into account the uncertainty in the
bottom-up inventory data, which has not been reported. The differences
between the same hybrid estimates and the corresponding estimate (2.86 Pg CO2) provided by the CDIAC inventory (7 and 5 %) are slightly
smaller than the differences with the EDGAR v4.2 data and are not
statistically significant. Therefore, our analysis suggests that the
CO2 emissions in the region considered are likely estimated more
accurately by CDIAC than by EDGAR v4.2; however, the difference between the
data of the two inventories in the case considered is small (∼ 6 %).
Note that if the conversion factor uncertainties were not taken into
account, which is not recommended, the difference between our
NO2-measurement-based CO2 emission estimate for the EHI sector and
the respective EDGAR v4.2 estimate would be statistically significant.
However, it is not significant with respect to the full confidence interval.
Considering the emission estimates for the EHI sector along with the total
CO2 emission estimates illustrates a possible way of using our method
for evaluation of bottom-up FF CO2 emission inventory data. That is,
assuming that the confidence intervals for our estimates are sufficiently
reliable, we can argue that a difference between hybrid and bottom-up
estimates that exceeds uncertainties associated with measurement and model
errors may, in a general case, be due to the two following reasons: (1) there are
inconsistencies between bottom-up estimates of emissions of
CO2 and of a corresponding proxy species or/and (2) a bottom-up
CO2 emission estimate is inaccurate. Taking uncertainties in the
conversion factors into account allows examination of the first reason:
evidently, it cannot be ruled out in the case of the emission estimates for
the EHI sector. However, the first reason alone is not sufficient to fully
explain the differences between the hybrid and bottom-up estimates of the
total CO2 emissions.
Comparing NO2- and CO-measurement-based CO2 emission estimates
(which, ideally, should be the same) enables their cross-validation. All
kinds of NO2-measurement-based CO2 emission estimates are found to
be consistent with the respective CO-measurement-based estimates in the
sense that their confidence intervals are intersecting. In principle, this
is an important result confirming that uncertainties in our emission
estimates are not underestimated, since NO2 and CO measurements are
independent from each other. However, it should be noted that the
uncertainties in CO-measurement-based estimates are so large that such
estimates can hardly be useful as a unique source of information on CO2
emissions. Similar large uncertainties are associated with
NO2-measurement-based CO2 emission estimates for the EHI and TCO
sectors, as well as with the total CO2 emission estimates obtained by
summing the NO2-measurement-based estimates for the individual sectors
together. While the uncertainties in the CO-measurement-based estimates are
mostly caused by uncertainties in the top-down estimates of CO emissions,
the uncertainties in the NO2-measurement-based estimates are
mainly associated with uncertainties in the conversion factors.
Importantly, the combined estimates (based on both NO2 and CO
measurements) of the FF CO2 emissions from individual sectors feature
considerably lower relative uncertainties (evaluated with Eq. 11) than the
uncertainties in the estimates based on measurements of only one proxy
species (for example, relative uncertainties of 39 and 42 % for the
NO2- and CO-measurement-based estimates, respectively, are reduced to a
relative uncertainty of 29 % in the combined estimate for the TCO
sector). This fact illustrates the potential usefulness of combining hybrid
estimates based on independent measurements of different proxy species such
as NO2 and CO. The uncertainty of our combined estimate
E^comb, totCO2 of the total CO2 emissions is very
insignificantly smaller than the uncertainty of the corresponding
NO2-measurement-based estimate.
As mentioned in Sect. 3.3, the uncertainty intervals for our combined
estimates evaluated with Eq. (11) can be reliable only if the hybrid
emission estimates derived from measurements of individual species are
statistically independent. We believe that the CO2 emission estimates
derived from NO2 and CO measurements are indeed sufficiently
independent particularly because NO2 (as a part of the NOx
chemical family) and CO experience very different atmospheric processing.
Indeed, while the key role in spatial and temporal variations of CO is
played by the transport processes (and boundary conditions in simulations),
atmospheric evolution of NO2 is very strongly affected by local
photochemistry. Furthermore, the results of our estimations for the cycle
and flat cases (see Table 1) indicate that probable errors in the seasonal
cycles of the NOx and CO emissions are also unlikely to result in
considerable and common biases in the NOx and CO emission estimates.
Thus it seems reasonable to believe that possible model errors for these
species are, for the most part, different in origin and weakly correlated.
Any significant covariance of errors in CO and NO2 measurement data is
also hardly possible, as those measurements are performed with different
satellite instruments and by using different methods (see Sect. 2.1). The
covariance of errors in the conversion factors Fcs for the different
species is likely small, too (given the complexity of data involved in
bottom-up estimates of different proxy species and the fact that NOx
and CO emissions depend on different technological factors and end-of-pipe
measures), although we could not evaluate it confidently with available
information. Therefore, the uncertainties in our combined emission estimates
are based on an (so far) inevitable assumption that errors in the conversion
factors for the different proxy species are statistically independent.
As in the case with the top-down estimates of NOx and CO emissions, our
hybrid estimates of FF CO2 emissions are rather insensitive to the
changes in simulations associated with using different seasonal cycles (cf.
the estimates for the cycle and flat cases). That is, we can conclude
that the impact of uncertainties in the assumed seasonal cycles of
anthropogenic emissions on our hybrid estimates is small. In particular,
such uncertainties can hardy explain the rather considerable difference
between our “combined” estimate of the total CO2 emissions and the
corresponding estimate based on the EDGAR v4.2 inventory.
Summary and conclusions
We examined feasibility of estimation of FF CO2 emissions
by using NO2 and CO column retrievals from satellite measurements. FF
CO2 emissions are an important component of the global carbon balance
and are believed to be a major contributor to global warming. Although such
emissions are usually known with better certainty than CO2 fluxes
associated with the biosphere, there still exist considerable divergences
between data of different bottom-up FF CO2 emission inventories;
typically, such data cannot be evaluated by using atmospheric CO2
measurements and rarely come with a reported uncertainty structure.
We followed the concept of proxy species that suggests constraining FF
CO2 emissions by using atmospheric measurements of minor species
co-emitted with CO2. We developed a general inverse modeling method
aimed at estimation of the budgets of FF CO2 emissions from different
sectors of economy in a given region by using satellite measurements of
proxy species. The method involves (1) obtaining top-down estimates of
anthropogenic emissions for a proxy species from the satellite measurements
and simulations performed with a mesoscale CTM,
(2) using bottom-up emission inventories to relate CO2 emissions
with emissions of the proxy species, and (3) combining CO2 emission
estimates derived from measurements of different proxy species. Important
parts of our method are robust techniques to estimate systematic differences
between the measured and simulated data, as well as uncertainties in
top-down estimates of the proxy species.
Considering NO2 and CO as the proxy species, the method was applied to
a western European region including 12 countries by using the NO2 and
CO column amounts retrieved from, respectively, the OMI and IASI satellite
measurements along with the simulated data from the CHIMERE CTM. The study
region was selected by taking into account that uncertainties in available
bottom-up emission inventory data for the EU countries with well-developed
statistics are likely rather low, compared to potential uncertainties in FF
CO2 emission data for countries with less developed statistical
infrastructure, although such uncertainties are likely not quite negligible
even in the study region. The relationship between FF CO2 emission and
the NOx and CO emissions was represented by the NOx-to-CO2
and CO-to-CO2 emission conversion factors evaluated with the EDGAR v4.2
emission inventory. The estimates were obtained for the total FF CO2
emissions from the region considered as well as individually for FF CO2
emissions aggregated into two different source categories (sectors), such
that the first category (EHI included the emissions associated mostly
with energy and heat production and heavy industries, and the second
category (TCO) comprised transport, chemical industry, and all other
anthropogenic sources. Our FF CO2 emission estimates were compared with
the corresponding data of the EDGAR v.4.2 global emission inventory; in
addition, our total FF CO2 emission estimates for the study region were
compared with the data of the CDIAC FF CO2 emission inventory. The
top-down estimates of NOx and CO emissions were compared with the
respective data from the European EMEP and global EDGAR v.4.2 emission
inventories.
As expected (taking into account findings of several previous studies), the
NO2 column retrievals from OMI measurements provide rather strong
constraints to NOx emissions. Our most reliable top-down estimate of
the total NOx emissions is found to be only insignificantly (by about
6 %) lower than the respective bottom-up estimate based on the EMEP
emission inventory; our estimates for the emissions from the EHI and
TCO
having much larger uncertainties (of about 18 and 15 %,
respectively) are also found to agree with the corresponding estimates based
on the EMEP emission inventory within the uncertainty range. Larger and
statistically significant differences are found between our NOx
emission estimates and the respective data of the EDGAR v4.2 global emission
inventory. In particular, our results suggest that the total NOx
emissions from the study region may be overestimated in the EDGAR v4.2
inventory by ∼ 13 %, while the EDGAR emissions for the EHI
sector are likely overestimated by more than 60 % (relative to our
estimates).
In contrast to the NOx emission estimates, our top-down estimates of
the CO emissions are fully consistent with both the EMEP and EDGAR v4.2
emission data; however, this consistency is partly due to much larger
uncertainties in our CO emission estimates (compared to uncertainties in the
NOx emission estimates). The relatively large uncertainties in the
top-down CO emission estimates (∼ 55 and ∼ 35 % in the estimates for the EHI and TCO sectors, respectively, and
∼ 25 % in the total CO emission estimate) are not
surprising in view of the much lower sensitivity of the satellite CO
measurements to anthropogenic CO emissions in the study region compared to
the sensitivity of the NO2 measurements to the anthropogenic NOx
emissions. Nonetheless, in spite of the large uncertainties (which may be
overestimated by our procedure), the differences between our top-down
estimates of CO emissions and respective EMEP data are rather small (less
than 7 %). Similar to our NOx emission estimates, the top-down CO
emission estimates differ more considerably from the EDGAR v4.2 data.
The top-down estimates of the NOx and CO emissions were used to obtain
different hybrid estimates (combining different information coming from
measurements and bottom-up inventories) of CO2 emissions. The
NO2-measurement-based hybrid estimate of total CO2 emissions is
about 12 % smaller than the respective estimates based on the EDGAR v4.2;
the difference exceeds the estimated uncertainty range (∼ 11 %) of our estimate, although only marginally. In contrast, the difference
between the same hybrid estimate and the corresponding estimate provided by
the CDIAC inventory (∼ 7 %) is not statistically
significant. A large negative difference (more than 60 %) is found
between our NO2-measurement-based CO2 emission estimate for the
EHI source category and the corresponding EDGAR v4.2 estimate. This
difference is, however, not statistically significant and can be mostly
attributed to uncertainties in the NOx-to-CO2 emission conversion
factor for the given source category. Our CO-measurement-based hybrid
estimates of the total FF CO2 emissions are larger than the respective
bottom-up estimates based on both the EDGAR v4.2 and CDIAC data but the
differences are not too big (less than 25 %) and can be well explained by
uncertainties in our estimates. Similar to the case with the
NO2-measurement-based hybrid estimate, the largest difference between
our CO-measurement-based FF CO2 emission estimates and the EDGAR v4.2
data is found for the EHI source category, with our best estimate being
about 26 % larger.
Taking into account the range of uncertainties, all our
NO2-measurement-based CO2 emission estimates were found to be
consistent with the respective CO-measurement-based estimates. This is an
important result confirming the reliability of our approach. The combined
emission estimates (based on both NO2 and CO measurements) for
individual source categories feature considerably smaller uncertainties than
the corresponding partial estimates. Our combined estimate of total FF
CO2 emissions is weighed toward the NO2-measurement-based estimate
and is found to be ∼ 11 and ∼ 5 % lower
than the respective estimates based on the EDGAR v4.2 and CDIAC data. The
difference of our estimate with the EDGAR v4.2 data slightly exceeds the
confidence interval of our combined estimate, while the difference with the
CDIAC data is again not statistically significant.
In general, our analysis demonstrated that NO2 and CO column retrievals
from satellite measurements provide reasonable constraints to FF CO2
emissions at the scale of western Europe. Although relative uncertainties in
our top-down CO emission estimates were evaluated to be considerably larger
than in the similar NOx emission estimates based on NO2
measurements, the CO column retrievals were found to be a useful source of
independent information on FF CO2 emissions, particularly in the cases
where probable uncertainties in the conversion factors for NOx
emissions are larger than uncertainties in the conversion factors for CO
emissions. Differences between hybrid CO2 emission estimates derived
from the satellite data and respective estimates based on bottom-up emission
inventory data can, in principle, be due to various kinds of uncertainties
in the hybrid estimates (including uncertainties in the top-down estimates
of NOx and CO emissions and uncertainties in the conversion factors).
We argue that such uncertainties can be roughly evaluated using the robust
techniques described in this paper. Nonetheless, further research
(involving, e.g., multi-model inversions and ensembles of independent
emission inventories) is needed to ensure that the confidence intervals for
the emission estimates actually take into account all possible estimation
errors, including those associated with uncertainties in the modeled
representation of chemical processes, in the boundary conditions for
reactive species, and in the NOx-to-CO2 and CO-to-CO2
emission conversion factors. Possible future developments of our approach
can also include using NO2 and CO retrievals from measurements
performed by other satellite instruments (such as GOME-2, MOPITT and AIRS)
together with the retrievals from the OMI and IASI measurements (as in this
study), and implementing hybrid CO2 emission estimates into a global
transport model simulating CO2 distribution in the atmosphere in order
to validate them against ground-based and satellite CO2 measurements.
Finally, it should be noted that as FF CO2 emission inventory data for
the western European countries are likely much less uncertain than similar
data for developing regions of the world, applications of our method to
developing regions can be especially fruitful. In this regard, our method
can become an integral part of a policy-relevant global carbon observing
system (Ciais et al., 2014, 2015).