A new global anthropogenic SO 2 emission inventory for the last decade: a mosaic of satellite-derived and bottom-up emissions

. TS1 TS2 Sulfur dioxide (SO 2 ) measurements from the Ozone Monitoring Instrument (OMI) satellite sensor have been used to detect emissions from large point sources. Emis-sions from over 400 sources have been quantiﬁed individually based on OMI observations, accounting for about a half 5 of total reported anthropogenic SO 2 emissions. Here we report a newly developed emission inventory, OMI-HTAP, by combining these OMI-based emission estimates and the conventional bottom-up inventory, HTAP, for smaller sources that OMI is not able to detect. OMI-HTAP includes emis-10


2
F. Liu et al.: A new global anthropogenic SO 2 emission stantially greater than natural ones on a global basis (Smith et al., 2001 TS3 ) owing to the high concentrations of sulfur contained in fossil fuels. In response to the rapid growth in fuel consumption driven by economic development in developing countries, particularly China, India, and international shipping, global SO 2 emissions increased from 2000 to (Smith et al., 2011. Meanwhile, stricter environmental legislation has promoted the introduction of new emission control with the fuel quality directive and desulfurization endof-pipe abatement, in particular earlier (since the 1980s for 10 power plants) in the US and Europe (Crippa et al., 2016) and more recently in China . Additionally, shipping emissions over the Sulphur Emission Control Areas (SECA) reduced since 2005 following the International Convention for the Prevention of Pollution from Ships (MAR-15 POL) Protocol, which further strengthened measures in 2012 and 2013 (Alföldy et al., 2013). This has led to a decline in global SO 2 emissions since about 2006 (Klimont et al., 2013). SO 2 emissions usually are estimated using a bottom-up 20 mass balance method. Bottom-up emissions are equal to the amount of sulfur in the fuel (or ore) minus that removed or retained in bottom ash or in products (Smith et al., 2011). The magnitude of emissions is subject to uncertainties, particularly when information on sulfur contents of fuels and 25 ores or sulfur removal is not available. The spatial distribution of emissions is even more uncertain, as emissions within a region are in most cases allocated by spatial proxies rather than actual locations of emission sources owing to a dearth of data. In addition, developing SO 2 emission inventories for 30 a specific year may become outdated if applied to other years when technologies and fuel use change rapidly. SO 2 observations from space-based platforms provide valuable global information on the spatiotemporal patterns of SO 2 emissions (Krotkov et al., 2016) that may complement existing bottom-up emission inventories and help to identify hotspots. Satellite-measured SO 2 has been used to monitor and characterize regional emission trends (van der A et al., 2017), volcanic emissions (Theys et al., 2013;Carn et al., 2016), and anthropogenic emissions from large point 40 sources like smelters (Carn et al., 2007), power plants , and oil sands (McLinden et al., 2012). Additionally, satellite retrievals of SO 2 vertical column densities have been used to quantify the strength of SO 2 emissions (Fioletov et al., 2015. 45 Chemical transport models (CTMs) have been employed to exploit SO 2 observations as a constraint towards improving SO 2 inventories using inverse modeling techniques (Lee et al., 2011;Wang et al., 2016). However, the derived emissions are usually determined at the coarse spatial resolu-50 tion of CTMs (e.g., 2 • latitude by 2.5 • longitude in Lee et al., 2011) and are subject to large uncertainties at finer spatial scales. Alternative CTM-independent approaches have been proposed to resolve SO 2 signals around individual large sources with simple model functions such as Gaussian dis-55 tributions (Fioletov et al., 2011). More recently, SO 2 emission rates and lifetimes were fitted simultaneously from the satellite-observed downwind plume evolution and meteorological wind fields for volcanoes (Beirle et al., 2014) and anthropogenic sources (e.g., Fioletov et al., 2015. 60 The satellite-based approaches used to estimate emissions are generally limited to larger sources, typically > 30 Gg yr −1 , for the highest spatial resolution observations currently available from the Ozone Monitoring Instrument (OMI). Here, we develop a 65 methodology to provide a comprehensive emission inventory that combines information about large SO 2 source from satellite-derived emissions with the conventional bottom-up emission estimates for smaller sources. An overview of the satellite-derived and the bottom-up inventories used in this 70 study is provided in Sects. 2.1 and 2.2, respectively. The methodology and features developed for our merged inventory are detailed in Sect. 2.3. Section 3 describes the model and in situ measurements used for evaluating our merged inventory, respectively. Section 4 details the validation results. 75 The validation focuses on 2010 for which the bottom-up inventory used by this study is most valid and a large number of in situ measurements are available. The validation for other years is performed to evaluate the emission trend of large sources that can be detected by OMI. Section 5 compares our 80 inventory with other existing bottom-up inventories. Section 6 presents a summary of the performance of the new inventory and the future work plans for maintaining and improving the inventory.

Satellite-derived emission inventory
The global OMI measurements allow for quantification of SO 2 emissions from anthropogenic sources. OMI is a UV-VIS nadir-viewing satellite spectrometer (Levelt et al., 2006(Levelt et al., , 2017 on board the NASA Aura spacecraft launched in 2004. 90 We use the OMI-based emission catalogue of nearly 500 sources from  to develop a new global SO 2 emission database in this study. The OMI-based emission catalogue is based on version 1.3 level 2 (orbital level) OMI planetary boundary layer (PBL) SO 2 products retrieved 95 with the principle component algorithm (PCA) algorithm (Li et al., 2013) and the updated air mass factors (AMFs) for each site (McLinden et al., 2014). The OMI SO 2 observations are rotated according to wind directions such that all observations were aligned in one direction (from upwind to 100 downwind; Valin et al., 2011;Fioletov et al., 2015). The location of the source is derived by comparing the difference between the average downwind and average upwind SO 2 column (McLinden et al., 2016). The rotated observations are assumed to be a single point source convolved with a 105 Gaussian function (Beirle et al., 2014) and fitted by a three-dimensional parameterization function of horizontal coordinates and wind speeds (Fioletov et al., 2015) in order to estimate emissions. Only observations contained within a rectangular area (hereafter called the fitting domain) are used for the fit. The fitting domain spreads ± L km across the wind direction, L km in the upwind direction and 3 × L km in the downwind direction. The value of L is chosen to be 30 km for small sources (under 100 Gg yr −1 ), 50 km for medium sources (between 100 and 1000 Gg yr −1 ), and 90 km for large sources (more than 1000 Gg yr −1 ; . Note that we prescribe values of the lifetime and the parameter describing the spread of the emission plume to obtain more robust fitting results. Additional information on the algorithm and uncertainties in the emissions are available from . The source types are further authen-15 ticated through a combination of satellite imagery and external databases based on site coordinates. The annual SO 2 emission, site coordinate, source type (power plant, smelter or source related to the oil and gas industry) for each anthropogenic source in the catalogue for the period from 2005 to 20 2014 are used here.

Bottom-up emission inventory HTAP
We use the up-to-date global anthropogenic emission inventory developed by the Task Force Hemispheric Transport Air Pollution (HTAP v2.2, available at: http://edgar.jrc.ec. 25 europa.eu/htap_v2 TS4 ) for sources that satellites are unable to detect. The HTAP v2.2 emission database is a state-ofart inventory compiling the latest available official and regional emission data and has been widely used in global and regional modeling experiments (e.g., Bian et al., 2017;30 Paulot et al., 2016;Ojha et al., 2016). It provides annual and monthly gridded air pollutant emissions with global coverage at a spatial resolution of 0.1 • × 0.1 • for the 2008 and 2010 (Janssens-Maenhout et al., 2015).
The gridded HTAP v2.2 SO 2 emission maps are provided 35 for six categories (energy, industry, residential, ground transport, aviation, and shipping). Some of the emissions from the energy and industry sector are identified as point sources and allocated to their exact locations; others are treated as areal sources and distributed to grid cells based on spatial prox- for North America (Pouliot et al., 2015), Monitoring Atmospheric Composition and Climate -Interim Implementation (MACC-II) for Europe (Kuenen et al., 2014), the 2012 ver-sion of MIX for Asia , and the Emission Database for Global Atmospheric Research version 4.3 for 55 the rest of the world (EDGAR v4.3; Crippa et al., 2016). Although the data provided in each inventory aims to actually represent 2008/2010 at the spatial resolution of 0.1 • × 0.1 • , the dataset was not consistently compiled with activity statistics of 2008/2010 (as is the case in EDGAR v4.3).

60
The National Emissions Inventory (NEI) of the US EPA is compiled bottom-up every 3 years and updated for the years in between with total consumption-based trends. The 2010 data for the US are based on the 2008 NEI with yearspecific updates made for power plants equipped with con-65 tinuous emissions monitoring systems and on-road mobile sources (Pouliot et al., 2014); for other sources a trend has been applied based on the trend in the sector-specific country totals. The 2010 data for Canada are based on the 2008 National Emission Inventory of Environment and Climate 70 Change Canada with updated emissions for point sources (Pouliot et al., 2014 -Maenhout et al., 2015).
In addition, re-sampling is applied to obtain gridded maps with a uniform spatial resolution of 0.1 • × 0.1 • based on the 80 MACC-II inventory at 1/8 • × 1/16 • resolution and the MIX inventory at 0.25 • × 0.25 • resolution. As pointed out by the HTAP report (Janssens-Maenhout et al., 2015), such inconsistency between the different inventories may yield uncertainties in strengths and locations of emissions. For example, 85 emissions from large point sources with changing emission patterns cannot be accurately derived from a linear extrapolation in time, because such extrapolation is not able to reflect sudden changes, such as shutting down of certain sources.

90
The OMI-based and the HTAP emission inventories are merged to construct a harmonized inventory that we refer to as OMI-HTAP. OMI-HTAP is particularly developed for the years 2008 and 2010 when HTAP is available. For other years, emissions from large sources that can be detected by 95 satellites are updated in OMI-HTAP. For other sources including those from the aviation and shipping sectors, the 2008 HTAP v2.2 inventory is used for construction of the OMI-HTAP inventory for years prior to 2008 as well as 2009; similarly, the 2010 HTAP v2.2 is used for years after 2010. 100 The emissions from these sources can be further updated using more recent bottom-up inventories with multi-year estimates. Consistent with the HTAP inventory, the OMI-HTAP inventory provides monthly gridded SO 2 emissions with global coverage at a spatial resolution of 0.1 • × 0.1 • for 105 different sectors. In order to estimate monthly emissions from OMI, its annual emissions are scaled by the HTAP monthly variations averaged over the fitting domain for the corresponding sec- 15 tor. That is, the OMI-based emissions are regarded as a single source within a particular fitting domain; areas not included within any fitting domain use HTAP emission grid maps. Figure 2 displays the 2010 OMI-HTAP SO 2 inventory (top) and compares it with the 2010 HTAP inventory (bot-20 tom and Fig. S1). The two inventories are consistent in total amount with a slightly larger (1 %) estimate from the OMI-HTAP inventory. However, they differ in the spatial distribution of emissions. Reasonable agreement is found in total emissions over China and most Western and Central Eu- and 56 % larger, respectively. Smaller OMI-HTAP estimates are concentrated over US and India, with OMI-HTAP estimates 31 % smaller.
Uncertainties in the OMI-based estimates may contribute to the differences. These uncertainties can be grouped into 35 three categories: in the retrieval of the OMI SO 2 vertical column density (VCD); those that come from the fit of the OMIdetected SO 2 downwind plume; and those related to the wind information. The overall uncertainty in annual emissions is estimated to be around 50 % , with 40 the primary contributors of the air mass factor calculation when determining VCD (27 %) and the wind height (20 %). Combining these components, we estimate that, on the other hand, uncertainties inherent in the total magnitude of bottom-up emissions may also contribute to the differences such as when bottom-up emissions are not routinely updated. The uncertainties of emissions from the industry sector are estimated to range from 15 % to 70 % over countries depending on how well the statistical infrastructure is maintained by individual countries (Janssens-Maenhout et al., 2015 and references in there). In addition, the uncertainties of spatial distribution may cause the differences. In fact, emissions from some emitting sectors in bottom-up inventories are not tracked with individual point sources but spread out over 10 larger areas instead. The country-specific emissions in HTAP are allocated where possible to the locations of point sources (e.g., public electricity plants), but a large fraction (e.g., some smelters of which the location are not available) remains distributed over the countries with spatial proxies (e.g., urban 15 population) of which the representativeness is only qualitatively known.
Bottom-up US SO 2 estimates are considered to be accurate, as over half of the emissions are directly measured by continuous emission monitoring systems. However, the 20 emissions from the source types without continuous monitoring devices, including some power plants (ranging from 10 % to 20 % for the period of 2005-2014; US EPA, 2014) as well as other industrial and residential sources were not tracked as point sources in HTAP, but distributed over a larger area 25 making use of spatial proxies. Moreover, updates on the fuel quality and technologies in these sources since 2008 were not accounted for. The discrepancy over the US is most likely related to such sources.
HTAP estimates 9 % and 12 % declines of SO 2 emissions 30 for energy and industry sectors in the US, respectively, from 2008 to 2010; this is less than the reported 27 % and 20 % decline by EPA (EPA Air Pollutant Emissions Trends Data; available at https://www.epa.gov/air-emissions-inventories/ air-pollutant-emissions-trends-data) TS5 ; HTAP estimates 35 are larger than the OMI-HTAP estimate for 2010. For 2008 with better information on the fuel quality and technologies in HTAP, the discrepancy between the two inventories over the US is much smaller (17 %). This is further supported by the excellent agreement for the largest individual US sources, 40 for which emissions are based on direct stack measurements using continuous emission monitoring systems (Figs. 1 and 3, Fioletov et al., 2015, 2017 ).
In other regions, uncertainties in bottom-up inventories could be larger owing to the lack of local emission mea-45 surements including continuous emission monitoring. For instance, local emission measurements in India are sparse and discrepancies between estimates from different bottom-up inventories can be as large as 50 % . The sulfur content of Indian fossil fuels adopted by HTAP was 50 based on assumptions in the MIX inventory. This inventory includes detailed information on China; however, there is much less information available for India owing to limited reporting in the literature (e.g., Reddy and Venkataraman, 2002). In addition, the fuel use is usually based on officially 55 Figure 3. Annual mean surface SO 2 concentration in 2010 based on the GEOS-5 model driven by the OMI-HTAP inventory, 2010 (a), and the differences between the modeled SO 2 using the OMI-HTAP and the HTAP inventory, 2010 (b). SO 2 concentrations using the HTAP inventory are subtracted from those in the OMI-HTAP inventory to derive the differences.
reported statistics, which may not be accurately documented. Some fuel consumption in South Asia is not included in official statistics, such as the burning of kerosene for wick lamps or fuel oil for diesel generators (Lam et al., 2012), which may be even more uncertain.

60
Long-standing experience (e.g., Hoesly et al., 2018a) in the development of emission inventories suggests that bottom-up inventories may miss some significant sources. The larger values over the Middle East, Mexico, and Russia in OMI-HTAP are due to the inclusion of emissions from 65 the OMI-identified sources missing from HTAP (McLinden et al., 2016). This helps to make OMI-HTAP a more complete inventory for these regions.
The locations of emissions in HTAP sometimes deviate from those in OMI-HTAP. This is probably caused by dif-70 ferent geographical allocation methods in two inventories, in particular the use of spatial proxies instead of real point source locations. In the OMI-based estimates, the location of each individual source is obtained from the OMI observations and then manually verified with satellite images in 75 Google Earth; this can lead to high accuracy. In the HTAP inventory, spatial proxies like total, rural, and urban population densities, road network and combinations were adopted to downscale emissions that lack geographical information; this may produce uncertainties when emission locations are 80 decoupled from spatial proxies (Liu et al., 2016. Sec-6 F. Liu et al.: A new global anthropogenic SO 2 emission tion 6 provides further discussion regarding the spatial mismatch of emission sources in HTAP and OMI-HTAP.

GEOS-5 model
We use the NASA Global Modeling and Assimilation Office 5 (GMAO) Goddard Earth Observing System version 5 data assimilation system (GEOS-5 DAS) (Rienecker et al., 2008) to simulate global surface SO 2 in this study. The aerosol module in GEOS-5 is based on the Goddard Chemistry Aerosol Radiation and Transport (GOCART) model (Chin et al., 2002). The model simulation is driven by GMAO atmospheric analyses from the Modern-Era Retrospective Analysis for Research and Applications, version 2 (MERRA- 2;Gelaro et al., 2017) in what is referred to as a replay mode where the aerosol fields do not feed back to the system. In other words, we run the GEOS-5 aerosol module in forecastmode with initial conditions from a previous run of the system, and the resulting aerosol fields do not impact the radiation within the model as they do in a full model run. The replay mode is run at a resolution of 0.5 • × 0.5 • and 72 ver-20 tical layers between the surface and about 80 km. We ran the system using either the HTAP or OMI-HTAP inventory within the aerosol module. We allow a 1-month spin up of aerosol fields for each experiment. For both the HTAP and OMI-HTAP emissions, we allocate the non-25 energy emissions (from industrial, residential, and transportation sectors) to the lowest GEOS-5 layer and the energy emissions from power plants to levels between 100 and 500 m above the surface (Buchard et al., 2014). All the simulations include aircraft and ship emissions from the HTAP 30 v2.2 inventory, biomass burning emissions from the Quick Fire Emission Dataset (QFED) inventory (van der Werf et al., 2010), production from dimethyl sulfide (DMS) oxidation (Kettle et al., 1999). Volcanic SO 2 emissions are derived from Total Ozone Mapping Spectrometer (TOMS), OMI, and 35 Ozone Mapping and Profiler Suite (OMPS) SO 2 retrievals (Carn et al., 2015) and the Aerocom inventories (Diehl et al., 2012).
While the main focus here is on 2010, we also conducted GEOS-5 simulations for 2006 and 2014 in order to evaluate  Figure 3 illustrates the annual mean surface SO 2 simulation using both inventories for 2010. Not surprisingly, the differences (Fig. 3b) show spatial patterns similar to the emission changes (Fig. 2b). The concentrations in the lowest model layer (from ground up to around 50 m) are evaluated 50 using surface SO 2 observations in the following analysis. (Dentener et al., 2006) CE2 . The aerosol module incorporated in GEOS-4 is based on the NASA GOCART model as described in Chin et al. (2002) and contains components for 55 dust, sea salt, black and organic carbon, and sulfate aerosols.

TS7
We evaluate the modeling surface concentrations of SO 2 over the US, Europe and East Asia for the 2006, 2010, and 2014 using in situ measurements from air quality net-60 works. We use stations from the US EPA Air Quality System (AQS; available at https://www.epa.gov/aqsTS8) for the US, the European air quality database (AirBase; available at https://www.eea.europa.eu TS9 ) for Europe, and the Acid Deposition Monitoring Network in East Asia (EANET, avail-65 able at http://www.eanet.asiaTS10) for East and Southeast Asia. For our analysis, we only include stations that had quality-controlled data for at least 75 % days for an individual year. We further exclude stations located in mountainous regions with an elevation of over 1000 m, as we expect 70 model limitations in describing pollutant concentrations over complex terrain (Liu et al., 2018). Additionally, we exclude stations located in regions with volcanoes as the dominant SO 2 source, e.g., Hawaii; the aim of this evaluation is to assess the performance of HTAP and OMI-HTAP, and volcanic 75 emissions have not been considered in either inventory. This leaves 248, 818, and 32 stations across US, Europe, and East Asia, respectively.
Sites in US-AQS and EU-AirBase are typically closer to urban areas. These sites may not be representative of the 80 model grid-cell mean when impacted by local pollution. To increase representativeness of grid box values, the available in situ measurements are averaged over the model's 0.5 • × 0.5 • grid cells before comparison with the model output. east Asia for simulations with HTAP and OMI-HTAP respectively. Here, we focus on the differences between modeled SO 2 using the OMI-HTAP and HTAP inventories. However, we note that the scatter between modeled and observed values may be attributed to the representativeness error re-100 lated to the incompatibility between in situ measurements and grid-cell averaged values simulated by the model. In addition, the slightly longer SO 2 lifetime simulated by the model as compared with in situ measurements and uncertainties in emissions may further contribute to the discrepancy (Buchard et al., 2014). Additional details on the evalua-5 tion of GEOS-5 SO 2 simulations can be found in Buchard et al. (2014). The implementation of OMI-HTAP improves the GEOS-5 performance with respect to observed surface SO 2 concentrations. We calculate normalized mean bias (NMB) to quan-10 tify the differences between modeled and observed SO 2 con-centrations; NMB is defined as The reduction in NMB for 2010 is highlighted for the US in Fig. 4a, b with values of 0.41 using HTAP and −0.03 us-15 ing OMI-HTAP. The reduction is particularly significant for grid cells with emission changes when comparing the two inventories, with NMB values of 0.70 and 0.06 for simulations with HTAP and OMI-HTAP, respectively. Improvements in Europe and Asia are much more subtle; most ob-20 Figure 5. Annual averaged SO 2 surface concentrations from AQS measurements in 2010 (a) and their differences between the modeled SO 2 using the HTAP (b) and the OMI-HTAP inventory, 2010 (c). AQS measurements are subtracted from the modeled SO 2 to derive the differences. The outline of circles corresponding to the grid cells with differences in emissions between the HTAP and the OMI-HTAP inventories is highlighted in black.
servations are made in grid cells with no differences between OMI and OMI-HTAP. For example, no significant changes are detected for Asia as most EANET sites are located far away from areas with modified emissions in OMI-HTAP. Figure 5 further illustrates the spatial distribution of  air-pollutant-emissions-trends-data TS11 ) and much larger than the decline of 9 % in HTAP (see details in Sect. 2.3). Improved agreement between observations and simulations is also shown for Europe. In particular, for grid cells with emission changes, the correlation coefficient increases 25 from 0.29 (HTAP) to 0.44 (OMI-HTAP). A plausible explanation for the improvement is the more reasonable spatial distribution of all large emission point sources in OMI-HTAP as detailed in Sect. 2.3.

30
In this section, we highlight the improvements obtained with OMI-HTAP for tracking emission changes driven by trends in the OMI data. Global anthropogenic SO 2 emissions substantially decline in OMI-HTAP. The US, Europe, and China are the primary contributors to the emissions reductions, 35 showing declines of 47 %, 27 %, and 23 % in the OMI-HTAP SO 2 emissions during 2005-2014 respectively. These declines are attributed in part to the installation of flue-gas scrubbers for coal-fired power plants. In addition, emissions from the world's largest smelters decreased due to phase 40 out of operations in some plants (e.g., Ilo, Peru; Flin Flon, Canada) or installation of scrubbers (e.g., La Oroya, Peru) (see more details in Sect. 5. 2 of Fioletov et al., 2016). In contrast, India experienced a rapid rise in emissions with a growth of 39 % in OMI-HTAP emissions during 2005-2014, 45 potentially surpassing China as the world's largest emitter of anthropogenic SO 2 .
The capability of OMI-HTAP (in particular OMI) to capture the emission trends is examined in Fig. 6. We compare the GEOS-5 simulations using both HTAP (grey dots) and 50 OMI-HTAP (blue dots) with in situ surface measurements for 2006 (Fig. 6a) and 2014 (Fig. 6b). The agreement between the observed and modeled SO 2 is better with simulations using OMI-HTAP, with larger correlations and smaller biases. This is particularly true for 2014 with a large gap 55   2014 (b). The grey and blue dots denote values using the HTAP and the OMI-HTAP inventories for corresponding years, respectively. The inset plots compare emissions from two inventories by region. Note that Europe only includes European countries with AirBase sites and Asia only includes East and Southeast Asia in the plot. The values of correlation coefficient (R) and normalized mean bias (NMB) are color coded by black and blue for grey and blue dots, respectively. Note that the plots use logarithmic scales, but R and NMB are calculated based on original data.

Intercomparison of bottom-up inventories
In this section, we compare OMI-HTAP with bottom-up emission inventories that are widely used within the climate and air-quality modeling community. The discussion is focused on inventories that are incorporated into HTAP (here-5 after called incorporated inventories), including the global EDGAR v4.3 inventory (Crippa et al., 2016), the European MACC-II inventory (Kuenen et al., 2014), and the Asian MIX inventory . Two additional regional inventories, the European Monitoring and Evaluation Pro-10 gramme (EMEP, Mareckova et al., 2013) at 0.5 • × 0.5 • resolution and Regional Emission inventory in Asia version 2 (REAS 2, Kurokawa et al., 2013)  We first focus on emission locations. For each OMIdetected source, if the bottom-up estimate is less than 20 % of the OMI-based estimate (out of the uncertainty range of 25 satellite-derived emission estimates) in the fitting domain (see the definition in Sect. 2.1), the source is considered to be missing from the bottom-up inventory. Otherwise, the location of the grid cell with the maximum emission within the fitting domain is identified to compare with that in the 30 OMI-based emission catalogue  used by OMI-HTAP. A source found within the fitting domain is classified as matched when the locations in the OMI-based emission catalogue and the bottom-up inventory are the same; otherwise, the source is classified as relocated 35 and the distance between the OMI-detected and the bottomup inventory source is calculated. The comparison is performed for four regions separately, i.e., North America, Europe, Asia, and the rest of the world (other). Note that emissions from countries that are only partly covered by the ei-40 ther the European or Asian inventories (e.g., Russia, Turkmenistan, Uzbekistan, and Kazakhstan) are categorized as other in this study to stay consistent with HTAP. Figure 7 summarizes the differences of emission locations between the OMI-based emission catalogue (and thus OMI-45 HTAP) and bottom-up inventories. HTAP shows the best agreement with OMI in North American, the region where it is expected to have good information about large SO 2 emission sources in bottom-up. The average distance between sources in HTAP and the OMI-based emission catalogue is 50 merely 4 km for North America. This is significantly less than the mean distances differences of 20, 22, and 15 km for Europe, Asia, and other regions, respectively.
It is interesting to note that sources are not always consistently located in HTAP and its incorporated inventories. 55 The average mismatch of locations between the OMI-based emission catalogue and HTAP is significantly larger than that between the OMI-based emission catalogue and the incorporated inventories for both Europe (20 km for HTAP vs. 12 km for MACC-II) and Asia (22 km for HTAP vs. 17 km for 60 MIX). The enhanced distances for HTAP are associated with a loss of spatial accuracy by the upscaling of incorporated inventories to a coarser grid (e.g., MACC-II for Europe has a higher resolution than HTAP) and by the re-sampling of grids that are not a multiple of 0.1 • . Re-sampling is ap-65 plied to merge grid maps at different spatial resolution (i.e.,  -Maenhout et al., 2015). This potentially misallocates emissions and thus increases the number of relocated sources (grey in Fig. 7).

5
Additionally, the incorporated inventories show better consistency in terms of location than other inventories developed for the same regions (i.e., EMEP for Europe and REAS for Asia) as compared with the OMI-based emission catalogue. For MACC-II, the improved consistency arises 10 from its fine spatial resolution of 1/8 • × 1/16 • , higher than that of 0.5 • × 0.5 • for EMEP. For MIX, the better consistency is attributed to the improved spatial patterns associated with the incorporation of local high-resolution emission datasets, such as the China Coal-fired Power Plant Emissions 15 Database (CPED,  and an Indian emission inventory for power plants developed by Argonne National Laboratory (Lu et al., 2011).
We further examine individual sources with annual bottom-up SO 2 emissions exceeding 70 Gg yr −1 that are ex-20 pected to produce a statistically significant signal in OMI data (Fioletov et al., 2011) but are not found in the OMIbased emission catalogue of nearly 500 sources . These large sources that are indicated by different bottom-up inventories mentioned previously in this section 25 are shown in Fig. 8b-e as solid and open circles for power plants and other types of sources, respectively. There are 74 such sources in total with 15 from HTAP, 31 from EDGAR, 3 from MACC-II, 14 from MIX, and 11 from REAS.
Bottom-up sources are likely not be seen by OMI if they 30 are located in regions with large systematic bias and retrieval noise for OMI PBL SO 2 data. These conditions occur, for instance, at high latitudes and over the South Atlantic and South America (from southern Peru southward) that are affected by the South Atlantic Anomaly that increases detector 35 noise in OMI observations (Fig. 8c). Additionally, bottom-up sources located in close proximity to other significant sources like volcanoes (Indonesia in Fig. 8e) could be absent from the OMI-based emission catalogue, as OMI may have difficulty in separating emission signals from individual sources.

40
In general, information on emissions from large sources individually may not be consistent among bottom-up inventories; sources identified as significant in one inventory may be missing from another, depending on the quality of the point source database used as input. Bottom-up emissions 45 from large point sources are derived from distributing country total emissions for the corresponding sector to individual facilities, when emissions at the facility level are not available. Emissions from large sources are potentially represented with too strong of an intensity concentrated over a 50 limited number of specific locations in the country. In this way, fewer point sources identified by bottom-up inventories in total lead to more sources with strong emission intensity, which may explain why more sources (31) in EDGAR are missing from the satellite-derived emission catalogue com-55 pared with those (15) in HTAP. Figure 9 compares emissions from global/regional inventories considered in this section to those from unitbased inventories for the power plants shown in Fig. 8be (solid circles). The considered unit-based power plant 60 databases include Emissions & Generation Resource Integrated Database (eGRID) for the US (US EPA, 2014), CPED  for China, and the European Pollutant Release and Transfer Register (E-PRTR; available from https:// www.eea.europa.eu/data-and-maps/data/lcp-4 TS12 ) for Eu-65 rope. It is interesting to see that power plant emissions estimated by global/regional inventories are on average biased high by a factor of 6 as compared with those from unitbased databases. This supports our hypothesis that emissions from some of these sources are distributed over too 70 few point sources in global/regional inventories, as emissions from unit-based databases are expected to be more accurate  (Fioletov et al., 2011). SO 2 sources identified that were found to be missing from bottom-up inventories are in blue. Locations of large sources indicated by bottom-up inventories but not detected by OMI (unmatched) over (b) North America, (c) South America, (d) Europe, and (e) Asia. The background is the global mean SO 2 distribution (in DU) map for 2005-2014. The area affected by the South Atlantic Anomaly is shown as a white oval.
due to the use of continuous emissions monitoring systems and unit-level fuel consumptions/emission factors (Liu et al., 2016).

Conclusions and future work
In this work we developed a merged emission inventory, 5 OMI-HTAP, by combining OMI satellite-based emission estimates for about 500 larger point sources  and a state-of-art bottom-up inventory HTAP v2.2 for smaller sources. Consistent with the HTAP inventory, the OMI-HTAP inventory provides monthly gridded SO 2 10 emissions with global coverage at a spatial resolution of 0.1 • × 0.1 • . OMI-HTAP is available for the period from 2005 to 2014, but is most accurate for 2008 and 2010, the years for which HTAP v2.2 was developed. We plan to include more recent years in the near future and use other bottom-up in-15 ventories in which multi-year estimates are provided.
The accuracy of OMI-HTAP has been evaluated by comparing modeled surface SO 2 concentrations with the measurements from ground-based air-quality monitoring networks focusing on the year 2010. GEOS-5 simulations us-20 ing OMI-HTAP showed considerably better agreement with in situ measurements compared with those using the bottomup inventory. The reduction in model bias is highlighted for the US, with the normalized mean bias decreasing from 0.41 (HTAP) to −0.03 (OMI-HTAP) for 2010. The improvements 25 obtained with OMI for tracking emission changes over the years 2006-2014 is similarly confirmed by evaluation with ground-based data. Figure 9. Comparison of SO 2 emission estimates from unit-based and regional emission inventory for power plants that are not detected by OMI.
The OMI-HTAP emission database developed in this work has several advantages as compared with conventional bottom-up inventories. To our knowledge, it is the first inventory with inclusion of nearly 40 OMI-detected sources that are not included in previous widely used bottom-up in-5 ventories. It enables more accurate emission estimates for regions with such missing sources, e.g., the Middle East and Mexico. OMI-HTAP SO 2 emissions estimates for the Persian Gulf, Mexico, and Russia are 59 %, 65 %, and 56 % larger than HTAP estimates in 2010, respectively. Unlike satellite 10 observations, bottom-up inventories typically cannot provide high-quality local information on point sources for all countries. For instance, the European Union (EU) has reported total SO 2 emissions for each country for a few decades, but the directive for reporting emissions from point sources with 15 corresponding public database started in 2007 and the quality of data varies over EU countries. In developing countries, such data infrastructure has not been built up yet.
OMI-HTAP provides dynamic emissions for over 400 OMI-based large sources since 2005, allowing for updates 20 to the emissions over time. Such updates based on satellite measurements are more consistent than those compiled in bottom-up inventories with annual activity statistics. The US, Europe, and China show declines of 47 %, 27 %, and 23 % in the OMI-HTAP during 2005-2014, respectively.

25
The exact location of each large point source in OMI-HTAP is obtained from satellite observations and crosschecked by Google Earth manually. The location information contributes to correction of mislocated emissions arising from the downscaling approach adopted by bottom-up 30 inventories or inaccurate locations provided by point source databases which sometimes use the administrative or even postal address but not the coordinate of the stack as the location of the facility.
Although satellite data provide good information on the 35 locations and trends for larger sources, they are currently not sufficient for providing complete information on SO 2 emissions and therefore much be merged with bottom-up inventories. We plan to combine satellite-based emission estimates with other bottom-up inventories in which multi-year esti-40 mates are provided, e.g., EDGAR v4.3.1 (Crippa et al., 2016) of the Joint Research Centre and the Community Emissions Data System (CEDS; Hoesly et al., 2018b) of Pacific Northwest National Laboratory, to better present emissions for small sources that cannot be detected by satellites or to use 45 the historic trends for extrapolating backwards in time.
We anticipate that our approach can be used with higher spatial and temporal resolution satellite observations that will be available in the near future. This will complement and improve merged inventories by providing more ac-50 curate satellite-based emissions estimates, potentially with diurnal and seasonal variability. Improved global satellite observations are anticipated from new sensors in low Earth orbit (LEO). The recently launched TROPOspheric Monitoring Instrument (TROPOMI) on the LEO ESA 55 Sentinel-5 Precursor satellite (Veefkind et al., 2012) featuring approximately 7 × 7 km 2 . The recently launched LEO NASA/NOAA JPSS-1/NOAA-20 OMPS instrument also has greater resolution (up to 10 × 10 km 2 ) than its predecessor on the NASA/NOAA Suomi National Polar-orbiting Partner-60 ship (SNPP) spacecraft (50 × 50 km 2 ). Zhang et al. (2017) showed that higher spatial resolution observations increase the detection limit of SO 2 sources. This is particularly important in the future, as emissions may continue to decrease due to emission control measures.

65
Upcoming geostationary Earth orbiting (GEO) satellite instruments will enable emissions estimates for different times of the day at relatively high spatial resolution. Planned GEO atmospheric composition instruments include the Korean Geostationary Environmental Monitoring Spectrometer 70 (GEMS; Kim et al., 2012), NASA Tropospheric Emissions: Monitoring of Pollution (TEMPO; Chance et al., 2012), and ESA Sentinel-4 . These will have high spatial resolution similar to TROPOMI but on an hourly basis.

75
Finally, the merging inventory methodology proposed in this study is potentially applicable for other air pollutants. It has good potential for application to NO x , as NO x emissions from power plants and cities can be quantified by similar CTM-independent approaches as well (Beirle et al., 2011;80 Liu et al., 2016). However, merging satellite-derived urban NO x estimates with bottom-up inventories is more challenging than point source emissions. Urban emissions are distributed over a larger number of sectors, including large contributions from areal sources such as road transport. An alter-85 native method needs to be explored to reconcile bottom-up and top-down satellite-derived urban emissions.
Data availability. The OMI-HTAP inventory is publicly available for the years 2005-2014 through the Aura Validation Data Center (AVDC) at https://avdc.gsfc.nasa.gov/pub/data/project/OMI_ HTAP_emis/ TS13 . The GEOS-5 model outputs are available upon request from the corresponding author.

5
The Supplement related to this article is available online at: https://doi.org/10.5194/acp-18-1-2018-supplement Author contributions. . TS14 Competing interests. The authors declare that they have no conflict of interest. TS15 10 Acknowledgements. This work was funded by NASA through the Aura and GMAO core programs. We thank the NASA Earth Science Division (ESD) Aura Science Team program for funding of OMI SO 2 product development and analysis (grant no. 80NSSC17K0240). We acknowledge the free use of the HTAP v2.2 and the EDGAR v4.3 emission inventories from European Commission, Joint Research Centre (JRC) and the Netherlands Environmental Assessment Agency (PBL). We acknowledge TNO for providing the MACC-II emission inventory, Tsinghua University for providing the MIX emission inventory, and National 20 Institute for Environmental Studies of Japan for providing the REAS 2 emission inventory. We thank the AQS, AirBase, and EANET networks for making their data available online. We thank the KNMI and OMI SIPS team for providing and processing the OMI data and the GMAO's MERRA-2 team for the datasets used 25 to drive the model simulations. We thank Mian Chin for helpful comments and Zifeng Lu for information on power plants in India.

CE2
What is the significance of this reference? It needs to be moved or removed.
Remarks from the typesetter TS1 Copernicus Publications collects the DOIs of data sets, videos, samples, model code, and other supplementary/underlying material or resources as well as additional outputs. These assets should be added to the reference list (author(s), title, DOI, and year) and properly cited in the article. If no DOI can be registered, assets can be linked through persistent URLs. This is not seen as best practice and the persistence of the URL must be secured.

TS2
The composition of Figs. 2-6 has been adjusted to our standards.

TS3
Do you mean also 2011?

TS4
Please add last access date.

TS5
Please add last access date.

TS7
Please check. The sentence is not in the your files.

TS8
Please add last access date.

TS9
Please add last access date.

TS10
Please add last access date.

TS11
Please add last access date.

TS12
Please add last access date.

TS13
Please provide a reference including creators, title, and date of last access.

TS14
Please note that the section "Author contributions" is mandatory for ACP.

TS15
Declaration of all potential conflicts of interest is required by us as this is an integral aspect of a transparent record of scientific work. If there are possible conflicts of interest, please state what competing interests are relevant to your work.

TS16
Please add page range or article number.

TS17
Please add page range or DOI.

TS18
Please add page range or DOI.

TS19
Please note, reference updated.