Introduction
Due to its undoubtable rapid economic growth, swift urbanization and consequent enlarged energy needs, large parts of
China have been suffering from severe and persistent environmental issues including major air pollution episodes (Song
et al., 2017). Developing and implementing effective air quality control policies is essential in combating such pollution
problems and requires timely as well as dependable information on emission levels (Zhang et al., 2012; van der A et al.,
2017). Understanding and monitoring the local long-term trends of different atmospheric pollutants is paramount in
updating, and predicting, pollution emission scenarios (Kan et al., 2012). Satellite atmospheric observations have
recently become an important information source on the atmospheric state, not only for the academic community but also
for public authorities and international environmental agencies (Streets et al., 2013; Lu and Liao, 2016). Recent reductions
of the two major pollutants emitted mainly by industrial sources, nitrogen and sulfur dioxide, have already successfully
been observed and quantified from space-born instruments over China (Wang et al., 2010, 2015; Liu et al., 2015, 2017).
Sulfur dioxide, SO2, is released into the atmosphere through both natural and anthropogenic processes. In the
former category lie chemical processes, such as the reaction of hydrogen sulfide, which is naturally occurring in crude
petroleum and natural gas as well as arising from the breakdown of organic matter, with atmospheric oxygen; seasonal biomass
burning events, which may be foreseen to some extent if not modelled; and volcanic degassing and unexpected
eruptions (see for example Seinfeld and Pandis, 1998). In the latter category fall the combustion of coal and oil fuel, which
account for more than 75 % of global SO2 emissions (Klimont et al., 2013), a figure found to be similar when
focusing on the Chinese domain (Smith et al., 2001, 2011). Lu et al. (2011) showed that SO2 emissions over
China, calculated from all major anthropogenic sources as well as scheduled biomass burning events by the agricultural
sector in order to clear vegetation and rejuvenate croplands, increased from ∼ 24 Tg in 1996 to
∼ 31 Tg in 2010, including fluctuations due to the onset of environmental protection measures as well
as the international economic crisis. The balance between encouraging China's economic development and dealing with its
environmental side effects often causes irregular changes in the SO2 emitted amounts, further dependent on the
province observed.
Satellite SO2 observations have proven to be a reliable way to monitor emissions from space and are increasingly
used in order to update bottom-up emission inventories (Streets et al., 2013). Numerous works have already amply demonstrated
the ability of satellite sensors to observe regional anthropogenic emission sources, for example by studying the
SO2 load over China using Ozone Monitoring Instrument, OMI, Aura observations. Krotkov et al. (2016) have shown how using long-term
atmospheric data records from the same instrument (OMI/Aura) can provide consistent spatiotemporal coverage, enabling the
analysis of both anthropogenic and natural emissions. For the North China Plain, of direct interest to this work, it was
shown that, despite it exhibiting the world's most severe SO2 pollution, since 2011 a decreasing trend with a 50 %
reduction in emissions has been verified from space. It is of course not only the changing economy and
enforcement of legislation that affect air quality; Witte et al. (2009) calculated a 13 % reduction in sulfur dioxide emissions due
to strict pollutant control for the August–September 2008 Olympic and Paralympic Games held in Beijing observed from
space. Li et al. (2010) further demonstrated that the OMI/Aura observations are capable of verifying the effectiveness of
China's SO2 emission control measures on power plants, while the imbalance in coal consumption between the
different provinces in China was also shown by Jiang et al. (2012). This inter-province diversion was further examined in
van der A et al. (2017), who showed how provinces enforcing desulfurization devices on their power plants have
a decreasing SO2 trend, whereas emerging provinces, which have built new power plants to accommodate the rapid
urbanization of the Chinese population, contribute with high emissions to the country's estimates.
Quite recently a new technique has used OMI/Aura observations as a means to detect large point sources of SO2
emissions from diverse origins, presented by Fioletov et al. (2013, 2016). Satellite observations were used not only to identify but
also to group SO2 emissions into emissions by volcanoes, power plants, smelters, and the oil and gas industry. The
technique has been evolved (Fioletov et al., 2017) into directly assessing traditional statistically obtained emission
levels using OMI as well as OMPS/NPP SO2 columns, with excellent validation results.
Following the aforementioned findings, in this work we aim to present a new spatially resolved SO2 emission
inventory on a monthly timescale for the years 2005 to 2015 based on satellite observations and modern chemical transport
modelling simulations. The technique used here has recently been applied in both Europe (Zyrichidou et al., 2015)
and China (Gu et al., 2014) for NOx emissions based on both GOME/ERS-2 (Global Ozone Monitoring Experiment/second European Remote Sensing satellite) and OMI/Aura observations. We aim
to show how it can be applied also to SO2 emissions and how the new top-down emissions compare against
traditional bottom-up emission inventories.
Data description
The mathematical analysis used in this work in order to extract an updated SO2 emission inventory is fully
described in Sect. 3. The main gist is that three input pieces of information are required: an original, also known as
a priori, emission inventory; the satellite observations of the SO2 load; and SO2 profiles provided by
an air quality chemistry transport model. The quality of these three pieces of information ensures the accuracy of the
updated, a posteriori, SO2 emissions estimates. Since the mathematical formulism also requires quantifiable
error estimates on these three input parameters, using the new OMI/Aura Royal Belgian Institute for
Space Aeronomy (BIRA) SO2 dataset (Theys et al., 2015,
2017) ensures that the satellite observations used here are fully characterized in this manner. In Sects.
to the three input datasets are presented and discussed appropriately.
The SO2 MEIC v1.2 emissions in megagrams per month for March 2010. The relative strength of the four sectors is shown
here: industry (a), power (b), residential (c) and transportation (d). Note the different colour bars used.
![]()
The MEIC emission inventory
The Multi-resolution Emission Inventory for China (MEIC v1.2) model has been developed for the years 2008, 2010 and 2012 by
the School of Environment, Tsinghua University, Beijing, China, and is downloadable from
http://www.meicmodel.org/ (last access: 20 March 2018).
SO2 emissions, in megagrams per month, are calculated on a monthly basis for
four sectors – power, industry, residential and transportation – at a spatial resolution of
0.25∘ × 0.25∘. The domain applicable spans from
15∘ to 55∘ N and from 102∘ to 132∘ E.
For the requirements of the methodology applied here the error in these emissions has
been assumed to rise to 50 % of the actual reported value since the MEIC inventory does not include such an error
estimate, nor were we able to procure such a value from the literature.
An example of the SO2 MEIC v1.2 emissions in megagrams per month for March 2010 is shown in Fig. 1. The
relative strength of the four sectors is shown as well, with industry in the top left panel, the power sector in the top
right, the residential emissions in the bottom left and transportation in the bottom right. Different colour scales in the
panels were used for the different emission strengths. In Zhang et al. (2015) the 2010 MEIC v1.2 emissions were used
as spin-up information in order to perform sensitivity simulations with different SO2 emission reduction
scenarios. It was shown that reducing SO2 emissions from one region has a small effect on SO2
concentrations over the other regions. The national mean SO2 concentration however is most sensitive to
SO2 emissions from northern China, in this work called the greater Beijing area. This strengthens the
importance of providing accurate and updated emission levels over that region in China even though it is considered to be
the best represented within existing inventories since the large population and industry density render the evaluation of
emission levels easier than in remote, less populated, regions.
The OMI/Aura <inline-formula><mml:math id="M75" display="inline"><mml:mrow class="chem"><mml:msub><mml:mi mathvariant="normal">SO</mml:mi><mml:mn mathvariant="normal">2</mml:mn></mml:msub></mml:mrow></mml:math></inline-formula> observations
The Ozone Monitoring Instrument (OMI) is a nadir-viewing instrument on board the NASA Aura satellite flying in
a Sun-synchronous polar orbit with an Equator-crossing time of around 13:30 LT in the ascending node launched in
July 2004. The OMI imaging spectrograph measures backscattered sunlight in the ultraviolet–visible range from 270 to
500 nm with a spectral resolution of about 0.5 nm (Levelt et al., 2006). The OMI spatial swath is around 2600 km wide, achieving near-complete global
coverage in approximately 1 day. The OMI ground pixel size varies from 13 km × 24 km at nadir
to 28 km × 150 km at the edges of the swath. Since June 2007, the radiance data of OMI for some
particular viewing directions have been corrupt, a feature known as the OMI row anomaly
(http://projects.knmi.nl/omi/research/product/rowanomaly-background.php, last access: 20 March 2018). Hence, the suggested OMI observations are
excluded de facto from the analysis.
In this work, we employ the retrieved SO2 vertical column densities (VCDs) using the
BIRA algorithm (Theys et al., 2015) which are calculated using the differential optical absorption
spectroscopy (DOAS) technique (Platt and Stutz, 2008) to the measured spectra in the 312–326 nm wavelength range. This step is followed by data filtering for
the row anomaly issue and a background correction to account for possible biases in the retrieved slant columns. The
obtained quantity is converted into a SO2 VCD using an air mass factor, AMF, which accounts for changes in
measurement sensitivity due to observation geometry, ozone column, clouds and surface reflectivity. The anthropogenic
SO2 profile required in the AMF calculation has been extracted from the Intermediate Model of the Global and
Annual Evolution of Species, IMAGESv2, global tropospheric chemistry transport model (Stavrakou et al., 2013, and
references therein) on a daily basis and for the overpass time of OMI. All details on the BIRA OMI SO2 algorithm
can be found in Theys et al. (2015), updated recently in Theys et al. (2017) in preparation for TROPOspheric Monitoring Instrument (TROPOMI) instrument. The
dataset has already been employed in different studies: in van der A et al. (2017) in order to estimate the effectiveness
of current air quality policies for SO2 and NOx emissions in China; in Koukouli et al. (2016) in
order to quantify the anthropogenic SO2 load over China using different satellite instruments and algorithms; and in
Schmidt et al. (2015) in order to study the 2014–2015 Bárðarbunga–Veiðivötn fissure eruption in
Iceland, among others.
The domain considered extends from 18∘ to 50∘ N and from 102∘ to 132∘ E and covers
eastern China. Daily observations were filtered for high solar zenith angle (SZA) of > 70∘, cloud fraction
> 0.2 and row anomaly flagging as per Theys et al. (2017). The filtered data were then averaged onto
a 0.25∘ × 0.25∘ monthly grid using a 0.75∘ smoothing average box. For further details on
this pre-processing, refer to Koukouli et al. (2016).
Within the OMI BIRA SO2 product, error contributions resulting from each step of the retrieval to the final
vertical column error are provided separately, including their random and systematic parts (Theys et al., 2017). This
allows the estimation of the total error in the column averages, an important feature in this analysis where the
instantaneous OMI observations are gridded and then averaged on a monthly mean basis. The formulation of the error in the
vertical SO2 column is derived by basic error propagation, shown in
Eq. (1):
σNV2=σNsM2+σNsbackM2+Ns-NsbackσMM22,
where σNs, σM and σNsback are the errors in the slant column
(Ns), the air mass factor (M) and the reference correction (Nsback),
respectively. When averaging the observations, the systematic and random components of each given error source need to be
discriminated, and so Eq. (1) evolves into Eq. (2):
σNV2=1M2σNs_syst2+σNs_rand2N+ΔNs2M2σM_syst2+ΔNs2M2σM_rand2N,
where N is the number of ground pixels considered in the average and σNs_syst is the
systematic uncertainty in the slant column density, SCD, which also includes the systematic uncertainty associated with the
background correction. The VCD is denoted by NV, the SCD by Ns, the
SCD minus the SCD_correction by ΔNs, the AMF by M, the VCD precision by
σNV, the SCD precision by σNs_rand, the AMF precision by
σM_rand and the AMF trueness by σM_syst. The error analysis is accompanied by the
total column averaging kernel (AK) calculated as the weighting function divided by the air mass factor, M (Eskes and
Boersma, 2003). The weighting function characterizes the sensitivity of the extracted atmospheric column to changes in the
true profile, and its importance in the analysis of satellite observations, alongside their correct comparison to other
datasets, has long been established (see for example Rodgers, 2000; Ceccherini and Ridofli, 2010; Zhang et al., 2010).
In Sect. the importance of the AKs in co-analysing satellite observations and modelling results in this
work is discussed extensively.
(a) The monthly mean OMI/BIRA SO2 columns in D.U. for March 2010. (b) The associated systematic error (left) and random error (right) in D.U. calculated using Eq. (2).
![]()
An example of the OMI SO2 product used in this work is shown in Fig. 2, for the month of March 2010. The
retrieved SO2 VCD in Dobson units (D.U.) is shown in the upper panel, with the systematic component to the error
in the bottom left and the random component in the bottom right.
In the original work of Martin et al. (2006), which was based on GOME/ERS-2 observations and GEOS-CHEM model data
at a resolution of 2∘ by 2.5∘, the authors conclude that the major limitations in their work were the coarse
horizontal resolution of GOME – which is not the case here for OMI – and the lack of direct validation of the GOME
tropospheric NO2 product – again, not the case here as the OMI BIRA SO2 measurements have already been
verified against other satellite observations (Bauduin et al., 2016; Koukouli et al., 2016) as well as long-term
ground-based measurements in polluted locations (Theys et al., 2015; Wang et al., 2017). However, we would be amiss not
to mention the issue of the possible horizontal transport of SO2 during its lifetime in the lower
troposphere, which would alter the linear relationship inherent in Eq. (3). Hains et al. (2008) calculated the SO2 lifetime
on a global scale to be 19 ± 7 h, whereas Lee et al. (2011) have updated this estimate, at northern US
mid-latitudes where anthropogenic emissions dominate, to 16–40 h with a maximum in winter and a minimum in
summer. Using OMI/Aura observations over the highest-emitting power plant locations in the US, Fioletov et al. (2015),
have provided shorter lifetime estimates of between 4 and 12 h. Even though it is hence not inconceivable that with
moderate wind speeds SO2 may have traversed a grid point on our 0.25∘ × 0.25∘ grid, on
the monthly mean scale that this work is based on it is impossible to evaluate the magnitude to this possible smearing
effect.
The CHIMERE model output
A multi-scale model for air quality forecasting and simulation, CHIMERE (http://www.lmd.polytechnique.fr/chimere/;
last access: 20 March 2018),
provides SO2 profiles over the Chinese domain of
18–50∘ N, 102–132∘ E for the mean overpass hour of OMI/Aura over the domain. The model version is CHIMERE v2013b
(Menut et al., 2013) at a spatial resolution of 0.25∘ × 0.25∘ and on eight vertical levels in
ppb, i.e. seven vertical layers, spanning from the surface up to 500 hPa, for the year 2010. The meteorological input
was provided by ECMWF (http://www.ecmwf.int/; last access: 20 March 2018) operational data. The anthropogenic emission inventory in this CHIMERE
run was a mix of the MEIC v1.2 inventory for mainland China and the Intercontinental Chemical Transport Experiment – Phase B (INTEX-B) emission inventory,
https://cgrer.uiowa.edu/projects/emmison-data (last access: 20 March 2018) for areas outside
China. The biogenic emissions are provided by the MEGAN database (http://lar.wsu.edu/megan/; last access: 20 March 2018). For the background of
the particular CHIMERE set-up refer to Mijling and van der A (2012), whereas more specific details on the CHIMERE v2013b
run used here may be found in Ding et al. (2015).
The March 2010 SO2 columns in D.U. as integrated in height from the original CHIMERE model ppb levels: (a) without rescaling to the effective pressure and without convolution with the OMI AKs; (b) with rescaling and with convolution with the OMI AKs.
![]()
An example of the convolution of the CHIMERE SO2 profile with the OMI averaging kernel to produce the convolved
CHIMERE total SO2 column for the grid of 38.0∘ N, 113.25∘ E. (a) The original CHIMERE SO2
profile of eight levels in ppb is shown in blue; the same profile but in Dobson units per layer is given in red, whereas the profile in
D.U. but of 58 OMI AK levels is given in black. (b) The OMI AK profile. (c) The original CHIMERE profile in
D.U. per layer is shown in black, as in (a), and the convolved CHIMERE profile in D.U. per layer is shown in olive
green. The original CHIMERE total SO2 column is 1.50 D.U., whereas after convolution with the OMI AK it decreases to
0.885 D.U.
![]()
The uncertainty of the CHIMERE SO2 columns is assumed to rise to 25 %. Estimating mathematically modelling
errors is quite challenging due to the large number of modelling processes and input parameters that have no defined
error, such as the boundary and initial conditions, the species emissions, rate constant uncertainties, and even
unresolved aspects of atmospheric physics and chemistry (Deguillaume et al., 2008; Boersma et al., 2016). Typically such
uncertainties are deduced from comparisons to other CTMs (Pirovano et al., 2012) and/or to independent observational
datasets (Lee et al., 2009). Even so, due to the innumerous differences in mathematically expressing atmospheric processes
in the former case and between model simulations and observations in the latter case, calculating a definite value remains
elusive. In Fig. 3, the March 2010 CHIMERE integrated SO2 column is shown as an example for the domain in
question.
Before proceeding to the convolution of the CHIMERE profiles to the OMI AKs and subsequent vertical integration, we investigated
whether the differences in orography heights assumed by the CHIMERE and OMI datasets in the respective algorithms may
introduce artefacts into the final CHIMERE VDCs. Zhou et al. (2009) have shown that, for the case of NO2 profiles retrieved
from OMI measurements over the Po Valley and the Alps, the difference in orography between satellite pixel and chemistry transport modelling (CTM) grid
may lead to either over- or underestimation of the NO2 VCDs by between 10 and 25 %. Theys et al. (2017), in
order to utilize more realistic a priori SO2 profiles, employed CTM model profiles at
1∘ × 1∘ resolution and used the hypsometric equation (Eq. 3) to scale them down to the future
TROPOMI/S5P 7 km × 3.5 km spatial resolution. In this equation, a new effective pressure, Peff
– which differs from the model surface pressure, PERA – is calculated under the assumption that the surface
temperature, TERA, varies linearly with height with a lapse rate of Γ = -6.5 Kkm-1,
gas constant of R = 287 Jkg-1K-1 and gravitational acceleration of
g = 9.8 ms-2. This variation depends on the difference between the orography height of CHIMERE,
hCHIM, and the OMI-reported height per observation, heff. The surface pressure and temperature
have been extracted from the ERA-Interim dataset (https://www.ecmwf.int/en/research/climate-reanalysis/era-interim;
last access: 20 March 2018)
at a daily temporal and 0.75∘ × 0.75∘ spatial resolution (Dee et al., 2011).
In the case of SO2 anthropogenic emissions, this whole issue may be significant in locations where the surface
height changes significantly within our 0.25∘ × 0.25∘ grid, whereupon the OMI pixel may have viewed
an entirely different atmospheric state, by more than ∼ 1 km in the vertical. In this work and for the
entire 10 years of OMI observations, only 3 % of the entire domain of 15 609 grid points show an overestimation of
hCHIM heights above 500 m and fewer than 0.5 % of the grid points show an overestimation of
heff heights.
Peff=PERATERATERA+Γ(hCHIM-heff)-g/RΓ
Even so, and for the sake of completeness, the CHIMERE profiles were re-scaled accordingly to the new pressure levels, calculated
from Peff and the CHIMERE pressure parameters as applied in Eqs. (2) and (6) of Zhou et al. (2009). Grid
points with associated CHIMERE heights of greater than 1500 m, which represent 7.5 % of the domain, almost
exclusively in the westernmost part (west of 110∘ E) where the Tibetan Plateau rises, are excluded from this
re-scaling due to interpolation issues. Those pixels are in any case excluded in the analysis for the new emission
database further on due to their non-existent SO2 contributions. Overall, the non-seasonally dependent
differences found in the CHIMERE columns before and after scaling were of the order of ∼ 10–12 %, on the low
side of the estimates for NOx by Zhou et al. (2009), who were however faced with far greater topological
variabilities in the locations of their study. As a consequence, we consider the convolution of modelling profiles to the
satellite AK a far more important factor in the solidity of the proposed methodology than anything else.
An extremely small fraction of our domain showed significant variation of above 0.5 D.U. in absolute differences, of
fewer than ∼ 0.05 % of the pixels for the entire domain irrespective of month, due to numerical uncertainties
introduced by the re-shaping, re-scaling and altering between the different altitude domains of the CHIMERE and OMI
profiles. Hence, for the main aim of this paper, which is to update the SO2 emission spatial inventory over
eastern China and not to provide absolute SO2 emitted quantities, we deem this difference well within the final
emission inventory error budget discussed below in Sect. .
We then proceed in convolving the re-scaled CHIMERE profiles with the OMI column averaging kernel as discussed in Eskes
and Boersma (2003) and Boersma et al. (2008a). The CHIMERE model profiles were already in
a 0.25∘ × 0.25∘ monthly grid, whereas the OMI observations are daily measurements in a variable
pixel size, from 13 × 24 km2 at nadir to 28 × 150 km2 at the edges of the
swath. Hence, the CHIMERE profile for each grid was convolved with each of the corresponding OMI AKs that fall within the
same 0.25∘ × 0.25∘ grid and then averaged (see Fig. 3, bottom). On average, the convolution of
the CHIMERE re-shaped profiles with the OMI AKs introduced a seasonally dependent decrease in the SO2 modelled
levels, between ∼ 0–5 % (for the summer months) and 10–15 % (for the autumn–winter months) for the entire
domain, as expected.
An example of this entire process is provided in Fig. 4 for the grid box 38.0∘ N, 113.25∘ E, a location
slightly to the west of the greater Beijing area with a moderate orography height of ∼ 1 km. In the left panel
the original CHIMERE SO2 profile of eight levels in ppb is shown in blue; the same profile but in Dobson units per
layer is given in red, whereas the profile in Dobson units but for the OMI AK levels is given in black since the OMI
algorithm performs calculations on a 58-level pressure grid. The y axis ranges up to ∼ 5 km, which is
approximately the vertical range of the CHIMERE model. In the middle panel the OMI AK profile is presented. In the right
panel the original CHIMERE profile in Dobson units is shown again in black so as to compare easily to the convolved
CHIMERE profile, in olive green. In the insert of this panel, the total SO2 load in D.U. for the two profiles is also
given. The re-shaped CHIMERE total SO2 column is 1.50 D.U., whereas after convolution with the OMI AK it
decreases to 0.885 D.U., while the actual load is also re-structured in order to approach the atmosphere sense by the
satellite instrument. It is hence shown that even though the total column has not changed the vertical distribution of
that column does change to reflect the sensitivity of the satellite observations, which peaks higher up in the boundary
layer and lower troposphere.
The seasonal variability of the a posteriori emissions calculated in this work (e–h) in gigagrams per season compared to the
a priori MEIC v1.2 emissions (a–d) in Ggseason-1 as well as their percentage differences (i–l)
in percent. From top to bottom; spring, summer, autumn and winter of reference year 2010.
![]()
Monthly mean time series for the a posteriori emissions in teragrams per month calculated in this work (dark blue points)
between 2005 and 2015. Insert: the reference year 2010 is shown to include the MEIC v1.2 a priori emissions in maroon
diamonds. The light blue shaded area depicts the calculated a priori error (Eq. 7). (a) The entire domain studied
(18–50∘ N, 102–132∘ E), (b) the greater Beijing region (30–40∘ N, 110–120∘ E),
(c) the northeast region (40–50∘ N, 120–130∘ E) and (d) the southwest region (25–35∘ N, 100–110∘ E).
![]()
Mathematical nomenclature
Top-down and a posteriori emissions estimates
The inversion methodology applied here is the one first presented in Martin et al. (2003) and further applied in Martin
et al. (2006), Boersma et al. (2008b), Lamsal et al. (2010), Lin et al. (2010), Gu et al. (2014) and Zyrichidou
et al. (2015), among others. The main premise of the methodology resides in the mass balance equation (Leue et al., 2001)
and requires three input parameters: the a priori emission field, Ea (Sect. ); the
satellite-derived SO2 field, Ωt (Sect. ); and the model SO2 field, Ωa
(Sect. ). Using those, as per Eq. (4), the top-down emission inventory, Et, is
calculated. Using standard propagation error analysis, the error in the top-down emission field may be calculated
through Eq. (5), where the error in the a priori emissions, εa, is required, as well as the error
on the model estimates, εΩa, and the satellite retrieval error, εΩt. These error
levels have been discussed in the equivalent sections.
Et=Ea⋅ΩtΩaεt2=ΩtΩa⋅εa2+EaΩa⋅εΩt2+EaΩtΩa2⋅εΩa2
The calculated top-down emission inventory, Et, may be combined with the a priori emission
inventory, Ea, to provide an a posteriori emission inventory, Ep, following the
maximum-likelihood theory and a log-normal distribution of errors. In Eq. (6) the calculation of the a posteriori emission
inventory is given, and its associated relative error is given in Eq. (7). Hence, in this methodology, the original bottom-up
emission inventory is combined with the top-down satellite observations, weighted by their respective errors and using
modelling outputs as background field, in order to constrain, update and provide new emissions estimates. It also follows
that since the a priori emission field is weighted by the top-down emission field error, and vice versa, the
a posteriori will depend mostly on the a priori should the errors of the top-down be too large, and vice versa. In that
way, it is assured that, at locations where the satellite observations are too sparse or the information content in the
SO2 load too low, the a posteriori emission field will revert back to the a priori.
lnEp=lnEa(lnεt)2+lnEt(lnεα)2(lnεt)2+(lnεa)2(lnεp)-2=(lnεt)-2+(lnεa)-2
We should clarify at this point that the calculations of Eq. (4) to Eq. (6) are performed for domain space; i.e.
for the sake of completeness these equations should have an i,j indicator everywhere designating the lat.–long. location of the
gridded domain space. The i,j were not included because it was deemed the equations would become too complicated
unnecessarily. However, the relative error calculated by Eq. (7), which represents the geometric SD about the expected
value as per Martin et al. (2003), is calculated on the final, total top-down error, εt, and
a priori error, εa, which are calculated as the known summation of error terms,
ε2=εi,j2+εi,j+12+…+εi+1,j2+εi+1,j+12+….
In the very recent paper by Cooper et al. (2017) an iterative version of the mass balance methodology (Martin et al.,
2003) was shown to provide results of similar accuracy to the more computationally demanding adjoint method (used for
e.g. in Stavrakou et al., 2013) in estimating satellite-born NOx emissions, which encourages the usage of the
mass balance technique when one cannot employ modelling results that calculate an adjoint matrix as well.
Roadmap of this analysis
The statistical methodology described above will be applied to the entire 11 years of OMI/Aura observations, from 2005
to 2015. Since the CHIMERE v2013b simulations were performed using the 2010 MEIC v1.2 inventory, the year 2010 will
be used as a reference year in the following analysis. The first step is to present the 2010 updated emissions over the
entire domain and how these compare against the a priori emissions; secondly, monthly mean time series of different
locations within the domain are shown, and the changes of the SO2 emissions over the years are discussed. Finally,
comparisons against pre-existing bottom-up emission inventories are presented.
Results and statistics
Updated emissions over China
In Fig. 5 the seasonal variability of the a posteriori emissions calculated with the methodology above is shown in the
middle column for spring, summer, autumn and winter (top to bottom). The equivalent MEIC v1.2 a priori inventory on the
same seasonal basis is shown in the left column, and the percentage differences of the two in the right column. The main
take-away message from this pictorial representation of the inventory is that the new inventory is producing higher
emissions for the entire domain for all seasons, which are stronger in winter and have positive biases that span from
∼ 10 % to ∼ 35 % accordingly (Table 1). Note from the fifth column of the Table the amount of grid
points that actually provide information out of an original 8414 grid cells for the domain considered in this work,
i.e. the grid cells of the MEIC v1.2 inventory. In the final column of the table, the percentage differences between the
two inventories are calculated in two ways: the first value depicts the difference between the first and third columns,
i.e. on the sum of emissions for the entire domain. The second value, in square brackets, has been calculated as the mean
of the per-grid-point percentage differences within the domain; hence it contains the geographical deviations of the
emission inventories as well. In order to further delve into this geographical variability, we present in Fig. 6 time
series of emissions over four domains of interest: the entire domain studied
(18–50∘ N, 102–132∘ E), the greater Beijing region (30–40∘ N, 110–120∘ E), the southwest region
(25–35∘ N, 100–110∘ E) and the northeast region (40–50∘ N, 120–130∘ E). The
two regions in the corners of the area studied were chosen since high SO2 levels were observed by OMI, resulting
in increased emissions in the a posteriori inventory, which do not appear in the original MEIC v1.2 dataset.
The average SO2 emission levels over China for the four seasons of the year 2010 as presented in Fig. 5.
A priori
A priori error
A posteriori
A posteriori error
# cells
% difference
[Ggseason-1]
[Ggseason-1]
[Ggseason-1]
[Ggseason-1]
Spring
6.36
0.135
7.77
1.57
6975
18.0 (24.0)
Summer
5.96
0.132
6.46
1.01
5765
8.0 (14.0)
Autumn
6.77
0.137
7.68
1.40
7126
13.0 (20.0)
Winter
7.07
0.140
9.12
2.66
7254
29.0 (34.0)
Details of the existing emission databases used for comparative purposes.
Database
Years
Spatial resolution
Temporal
Main reference
Publicly available from:
available
resolution
REASv2.1
2000 to
0.25∘ × 0.25∘
monthly
Kurokawa et al. (2013)
https://www.nies.go.jp/REAS/
2008
(last access: 20 March 2018)
INTEX-B
2006
0.5∘ × 0.5∘
yearly
Zhang et al. (2009)
https://cgrer.uiowa.edu/projects/emmison-data
(last access: 20 March 2018)
EDGAR v4.3.1
2010
0.1∘ × 0.1∘
monthly
Crippa et al. (2016)
http://edgar.jrc.ec.europa.eu/
(last access: 20 March 2018)
Annual SO2 emissions over the domain 15–50∘ N, 102–132∘ E in
teragrams per year. First column, the year; second column, this work; third column, the REASv2.1; fourth column, EDGAR v4.3.1; and fifth column, the INTEX-B database.
Year
This work
REASv2.1
MEIC v1.2
EDGAR v4.3.1
INTEX-B
Tgannum-1 for the 15–50∘ N, 102–132∘ E domain
2000
15.86
2001
15.94
2002
17.53
2003
19.70
2004
21.77
2005
35.27 ± 1.75
24.68
2006
35.33 ± 1.76
24.45
32.08
2007
37.58 ± 1.76
24.40
2008
35.75 ± 1.76
26.96
29.80
2009
31.74 ± 1.75
2010
32.14 ± 1.74
26.26
33.34
2011
33.50 ± 1.75
2012
31.30 ± 1.75
26.48
2013
32.05 ± 1.74
2014
28.32 ± 1.72
2015
23.34 ± 1.71
The seasonal variability of the a posteriori emissions calculated in this work (e–h) in Ggseason-1 compared to the EDGAR v4.3.1 emissions (a–d) in Ggseason-1 as well as their absolute differences (i–l). From top to bottom; spring, summer, autumn and winter of the reference year 2010.
![]()
In Fig. 6 the monthly mean time series for the a posteriori emissions in teragrams per month (dark blue lines) are
presented for the four domains of interest, so as to enable a more in-depth discussion of the new inventory. The light
blue shaded area depicts the extracted a posteriori error in the emissions, and the inset sub-figures depict the reference
year 2010, with the a posteriori levels shown in blue and the MEIC v1.2 emissions in maroon. The pre- and post-2010 drifts
are also calculated since the year 2010 is considered a turning point as far as regulating SO2 emissions
is concerned (Wang et al., 2015; van der A et al., 2017, and references therein). A very similar picture was shown for all
domains: a near-stable decrease in emissions within the statistical error of the analysis for the pre-2010 levels and
a stronger and statistically significant decrease for the post-2010 levels.
For the entire domain (Fig. 6a) a posteriori emissions in all months show an increase for the year
2010 compared to the a priori MEIC inventory, apart from the summer (JJA) ones, with the highest increases for the winter
months. The pre-2010 drift is calculated at the limit of statistical significance, at
-0.51 ± 0.38 Tgmonth-1, whereas the post-2010 drift is stronger and significant at
-1.52 ± 0.36 Tgmonth-1. For the greater Beijing region (Fig. 6b) a small
increase in emissions, nearly constant in all months of 2010, is found with the post-2010 drift to also be negative at the
-0.44 ± 0.11 Tgmonth-1 level. Two special regions of interest, with low emission levels in general,
were revealed by the OMI observations, in the northeast and the southwest of the domain, and are examined in the third
and fourth panels, respectively. The first 3 months of the year 2010 in the a posteriori emission database show quite
higher levels than the MEIC v1.2 compilation, whereas the rest of the months show the same level for the NE
(Fig. c), whereas in the SE (Fig. d) the first months of the year have an increased SO2 emitting signature.
Comparison with existing emission inventories
Apart from the MEIC v1.2 emission inventory discussed in Sect. – which is currently publicly available for the years
2008, 2010 and 2012 – there exist other emission inventories that are frequently used in chemical transport models as
input: the Regional Emission inventory in Asia (REAS) v2.1 (Kurokawa et al., 2013); the 2006 Asia Emissions for INTEX-B
(Zhang et al., 2009); and the Emissions Database for Global Atmospheric Research, EDGAR v4.3.1 (Crippa et al.,
2016). Comparing with similar published works is not as straightforward as one would assume since in this work
a sub-domain of what is termed China in other publications is used. For example when calculating the total annual
SO2 emissions reported by the REASv2.1 database for the year 2000, those are found to be 25.62 Tgannum-1
when allowing the entire domain provided in the database; they are found to be only 15.86 Tgannum-1
when restricting in the domain we are studying. As a result, large differences and erroneous comparisons may be presented if one simply
compares emissions estimates as reported in published works. For similar comparative studies, we refer the interested reader to Table 3 of Lu
et al. (2010) and Table 8 of Kurokawa et al. (2013); however great care is needed when
quoting absolute SO2 emission levels.
In Table 2 the details of the three databases are given. Since we are interested in evaluating the SO2 emission
as spatial patterns and not point source levels, we focused on these three databases, which
are provided at actual spatiotemporal resolutions. As a first inspection, in Table 3, the annual SO2 emissions for the domain
15–50∘ N, 102–132∘ E in teragrams per year are presented. We should point out
that, due to the fact that our methodology is based on the MEIC v1.2 emission inventory, within the domain stated there
are large areas with no emissions, mostly over the sea and the Korean Peninsula. In the following comparisons, only the common
pixels between all inventories are used for the calculations.
Several issues arise; firstly, for the common years between this work and the REAS v2.1, i.e. 2005 to 2008,
the differences span between ∼ 30 and ∼ 60 %, with REAS v2.1 underestimating the emission levels
in the domain studied. For the one common year between REAS v2.1 and MEIC v1.2, namely 2008, this underestimation still
holds but is smaller, of the order of ∼ 10 %. Similarly, for the one common year between REAS v2.1 and INTEX-B,
namely 2006, REAS v2.1 underestimates by ∼ 30 %. All of this points to an underestimation of SO2 levels
in the domain considered by the REAS v2.1 database.
Comparing the 2006 INTEX-B emissions to the ones calculated in this work, we find a difference of the order of
∼ 10 %, whereas comparing to the 2010 EDGAR v4.3.1 emissions the difference is almost insignificant, at
∼ 3.5 %. Since the EGDAR v4.3.1 emissions are provided on a monthly basis, in contrast to the INTEX-B ones, we
can evaluate our spatial patterns as well. After regridding the EDGAR v4.3.1 emissions
at a 0.25∘ × 0.25∘ spatial resolution on a monthly basis, the seasonal variability of the inventory
is compared to the one presented in this work in Fig. 7.