Dynamics

Abstract. We present a top-down ship NO x emission inventory for the Baltic Sea, the North Sea, the Bay of Biscay and the Mediterranean Sea based on satellite-observed tropospheric NO 2 columns of the Ozone Monitoring Instrument (OMI) for 2005–2006. We improved the representation of ship emissions in the GEOS-Chem chemistry transport model, and compared simulated NO 2 columns to consistent satellite observations. Relative differences between simulated and observed NO 2 columns have been used to constrain ship emissions in four European seas (the Baltic Sea, the North Sea, the Bay of Biscay and the Mediterranean Sea) using a mass-balance approach, and accounting for non-linear sensitivities to changing emissions in both model and satellite retrieval. These constraints are applied to 39 % of total top-down European ship NO x emissions, which amount to 0.96 Tg N for 2005, and 1.0 Tg N for 2006 (11–15% lower than the bottom-up EMEP ship emission inventory). Our results indicate that EMEP emissions in the Mediterranean Sea are too high (by 60%) and misplaced by up to 150 km, which can have important consequences for local air quality simulations. In the North Sea ship track, our top-down emissions amount to 0.05 Tg N for 2005 (35% lower than EMEP). Increased top-down emissions were found for the Baltic Sea and the Bay of Biscay ship tracks, with totals in these tracks of 0.05 Tg N (131% higher than EMEP) and 0.08 Tg N for 2005 (128% higher than EMEP), respectively. Our study explicitly accounts for the (non-linear) sensitivity of satellite retrievals to changes in the a priori NO 2 profiles, as satellite observations are never fully independent of model information (i.e. assumptions on vertical NO 2 profiles). Our study provides for the first time a space-based, top-down ship NO x emission inventory, and can serve as a framework for future studies to constrain ship emissions using satellite NO 2 observations in other seas.


Introduction
Strong emissions of gases and particulate matter by ships affect the composition of the marine boundary layer, with important consequences for climate change, air quality and public health. Because hardly any regulations for the maritime sector exist in international waters, ships are still allowed to burn marine heavy fuel, resulting in substantial emissions of black carbon (BC), sulfur dioxide (SO 2 ), nitrogen oxides (NO x = NO + NO 2 ), and volatile organic compounds (VOCs) (e.g. Eyring et al., 2010). NO x and SO 2 emissions from ships are relatively high compared to emissions from other transport sectors because marine heavy fuel is high in sulfur content and is combusted at high temperatures without reduction technologies (Eyring et al., 2005). Recently, new legislation has been proposed that sets limits on technology used in new ships to reduce sulfur and nitrogen oxide emissions in so-called sulfur emission control areas (SECAs) and NO x emissions control areas (NECAs). In Europe, SECAs have been in effect since 2006 and 2007 for the North Sea and the Baltic Sea, respectively, and a NECA is planned for the Baltic Sea for 2016. The North American coastal waters are designated as both a SECA and NECA since 2012 (IMO, 2009). Stringent emission limits for ships in these seas will be enforced in several steps, but the International Maritime Organization (IMO) is currently considering delaying the implementation of tighter limits in NECAs. NO x emissions lead to O 3 and particulate matter formation, detrimental to air quality in the densely populated coastal regions close to ship lanes. Furthermore, O 3 influences the hydroxyl radical (OH) concentrations that determine the lifetime of methane (CH 4 ) (Lawrence and Crutzen, 1999), thereby influencing its global radiative forcing (IPCC, 2007). Sulfate  ) formed by oxidation of sulfur dioxide (SO 2 ) is the dominant aerosol emitted from ships and has a negative radiative forcing (due to its efficient reflection of sunlight). Aerosols originating from ships can also have an indirect (negative) effect on radiative forcing by altering the properties of clouds (e.g. Schreier et al., 2007;Lauer et al., 2007). Because of their important effect on both air quality and climate, ship emissions have received increasing attention over the past years. Previous studies (e.g. Corbett et al., 2007;Eyring et al., 2010;Paxian et al., 2010) proposed that global ship NO x emissions amount to 3.0-10.4 Tg N per year (15-35 % of global anthropogenic NO x emissions). However, as individual measurements of ship emissions are sparse, and knowledge of activity, technology and global fleet is limited, these bottom-up inventories suffer from large uncertainties, making it difficult to assess the efficacy of reducing ship emissions in order to mitigate the effects of air pollution and climate change.
The magnitude and geographic location of ship NO x emissions can be constrained by using high-spatial-resolution satellite observations of NO 2 columns. Previous studies have demonstrated this concept, where satellite-observed NO 2 columns, in combination with simulations from a chemistry transport model (CTM), were used to constrain NO x emissions. For example, Martin et al. (2003) used observations from the Global Ozone Monitoring Experiment (GOME) instrument to scale global NO x emissions in the GEOS-Chem CTM. Different instruments have since then been used to constrain various source categories: anthropogenic emissions (e.g. Martin et al., 2006), soil NO x (e.g. Jaeglé et al., 2005) and lightning NO x (e.g. Boersma et al., 2005). Recently, Wang et al. (2012) used high-resolution NO 2 columns from the Ozone Monitoring Instrument (OMI) to detect newly built power plants in China. Satellite observations have also provided information on ship pollution. Beirle et al. (2004) used GOME measurements of tropospheric NO 2 columns to estimate emissions in the ship lane from Sri Lanka to Indonesia. Several ship tracks have been identified in global maps of satellite-observed NO 2 columns from the SCanning Imaging Absorption spectroMeter for Atmospheric CartograpHY (SCIAMACHY) and compared to an emission inventory by Richter et al. (2004). Marmer et al. (2009) took advantage of the higher spatial and temporal resolution of OMI NO 2 observations of the ship track in the Mediterranean Sea to assess several emissions inventories. A trend in NO 2 columns over four ship tracks in Europe and Asia, following the rhythm of global economic activity, was shown by de Ruyter de Wildt et al. (2012) using GOME, SCIAMACHY, OMI and GOME-2 observations. Franke et al. (2009) used a combination of observed NO 2 columns of SCIAMACHY, GOME and GOME-2 instruments and modelled columns using the ECHAM5/MESSy1 CTM to evaluate ship NO x emissions in the ship track from Sri Lanka to Indonesia. They concluded that their modelled NO 2 columns were in good agreement with observed columns for 2002-2007. However, as with most CTMs, their model neglected the in-plume chemistry of ship emissions by instantly diluting the emissions over the model grid cell. This results in an overestimation of modelled NO x concentrations (e.g. Kasibhatla et al., 2000;Davis et al., 2001;Vinken et al., 2011) and therefore in too low ship NO x emissions estimates. Various methods have been proposed in the literature to account for the non-linear chemistry during the first stages of ship plume expansion (for a review, see Paoli et al., 2011). Here we use our recently developed method using a plumein-grid approach (Vinken et al., 2011) that accounts for the non-linear in-plume chemistry in the GEOS-Chem global 3-D CTM. Compared to instant dilution of ship emissions, this method leads to lower (simulated) background NO x concentrations over the North Atlantic Ocean by up to 60 %.
In this study we focus on providing top-down constraints on ship NO x emissions by comparing modelled and satelliteobserved NO 2 columns for four major ship routes in Europe (the Mediterranean Sea, the Bay of Biscay, the Baltic Sea and the North Sea). This is the first time that NO 2 pollution has been evaluated using satellite measurements over the Bay of Biscay, the Baltic Sea and the North Sea. These observed columns are compared to NO 2 columns simulated with the nested version of the GEOS-Chem CTM. The high-resolution (1/2 • × 2/3 • ) of the nested version of GEOS-Chem is capable of resolving major ship tracks in Europe, and improves the localization of emissions. We run the European nested version of GEOS-Chem at a 1/2 • × 2/3 • resolution for 2005-2006 using the plume-in-grid treatment of ship NO x emissions introduced in Vinken et al. (2011). Using the combination of the high-resolution modelled columns and OMI-observed columns, we present for the first time spacebased, seasonal and annual constraints on ship NO x emissions in four major European ship routes for 2005-2006.  (Wang et al., 2004;Chen et al., 2009;Zhang et al., 2012) is operated at 1/2 • × 2/3 • resolution with 47 vertical layers, and a transport and chemistry time step of 10 and 20 min, respectively. The boundary conditions are updated every 3 h using global simulations from the 2 • × 2.5 • parent model (one-way nesting). Both the nested and global simulations are driven by GEOS-5 assimilated meteorological observations from the NASA Global Modeling and Assimilation Office (GMAO). The nested-grid domain extends from 30 to 70 • N and 30 • W to 50 • E. The lowermost layer of the model has an approximate depth of 120 m and the vertical extent of the model is 80 km. GEOS-Chem has a detailed simulation of ozone-NO x -hydrocarbon-aerosol chemistry, as recently described and discussed by Mao et al. (2010) and Lin et al. (2012). The reactive uptake coefficient γ N 2 O 5 for N 2 O 5 on aerosols is from Macintyre and Evans (2010), with a resulting annual mean value for 2005 of γ N 2 O 5 in surface air over our domain of 0.004, at the high end of recently measured values (0. 0005-0.006;Brown et al., 2009;Bertram et al., 2009). We performed a spin-up of 1 yr (2004) and simulations for [2005][2006]. Daily simulated tropospheric NO 2 columns corresponding to the satellite overpass time (between 13:00 and 15:00 LT) were averaged. To ensure consistency with the satellite observations, only days with valid satellite observations (see next section) were included.
Global anthropogenic emissions are from the Emission Database for Global Atmospheric Research (Olivier and Berdowski, 2001). Over Europe these are replaced with the Monitoring and Evaluation of the Long-range Transmission of Air Pollutants in Europe (EMEP) inventory (Vestreng et al., 2007). We replaced ship emissions (NO x , SO 2 and CO) from the EMEP inventory with a combination of the EMEP and (global) AMVER-ICOADS  inventories, as emissions were misplaced in the EMEP inventory (see discussion in Sect. 3.1). NO x emissions from soils are included based on the parametrization of Yienger and Levy II (1995) and Wang et al. (1998). Furthermore, lightning (Sauvage et al., 2007), biomass burning (van der Werf et al., 2006), biofuel (Yevich andLogan, 2003) and aircraft (Baughcum et al., 1996) NO x emissions are included in the model. An overview of the total NO x emissions over Europe for [2005][2006] used in this study is given in Table 1. Anthropogenic sources (7.2 Tg N yr −1 ; including aircraft, biofuel, fertilizer use and ships) account for 87 % of the total NO x emissions in 2005. Natural emissions (biomass burning, lightning and soil) peak in summer, accounting for 25 % of total European NO x emissions in July and August 2005.
We use the plume-in-grid approach developed by Vinken et al. (2011) to take into account non-linear chemistry occurring in ship plumes immediately after emission. In that approach a Gaussian plume model with chemistry has been used to construct a look-up table (LUT) that contains the fraction of NO x remaining and (net) O 3 produced in 5 h of plume expansion after emission as a function of several environmental parameters. Here we extend this method in two ways. First, we limit the chemical aging time in the expanding plume model to 2.5 h, as the plume typically grows to the size of the model grid cell (1/2 • × 2/3 • ) within this time. We store the fraction of NO x remaining and (net) O 3 produced after 2.5 h of plume expansion in a LUT and multiply GEOS-Chem NO x emissions with this fraction of NO x remaining upon release in the model grid cell. The resulting NO x concentrations in (GEOS-Chem) grid cells with ship emissions can be considered as background concentrations, i.e. representative for NO x concentrations after a ship passed by 2.5 h earlier. Because we will compare the simulated columns with satellite-observed NO 2 columns, which also include   The orange to red colours in the third bar represent the contribution to the column from emissions from the last 2.5 h, which are added to the background column to provide model output consistent with a satellite snapshot at 13:30 LT. As reference, the tropospheric NO 2 columns for this location without ship emissions (blue) and when using instant dilution (purple) are also given.
contributions from emissions that occurred during the past 2.5 h, we need to take these into account as well. Therefore, we extended our plume-in-grid approach solely for the model output between 13:00 and 15:00 LT. The continuous plume between release and 2.5 h prior was discretized in three parts (emissions between release and 0.5 h, between 0.5 and 1.5 h, and between 1.5 and 2.5 h). We multiply the emissions in these periods by the fraction of NO x remaining after 15 min, 1 h and 2 h of plume expansion, respectively (Fig. 1a). These emissions are added to the NO x concentration in the grid cell ( Fig. 1b) and the resulting tropospheric columns are stored in order to allow for a fair comparison between the satellite observations (sensitive to all recent emissions) and GEOS-Chem. Note that the ship emissions from the last 2.5 h are not propagated in the regular model chemistry, but are only accounted for to provide a model snapshot that is representative of what the satellite observes. To evaluate tropospheric NO 2 simulated with this extended plume-in-grid method, we compared against the conventional approach of instantly diluting NO x emissions over a grid cell. Our simulation with the plume-in-grid approach (and accounting for fresh emissions) leads to a higher column than instant dilution (Fig. 1b). This is because just after initial release in the plume, OH concentrations are much suppressed, resulting in a longer NO 2 lifetime (during the first part of expansion), and as a result a higher NO 2 column. However, this is not always the case. Depending on ambi-ent conditions (e.g. background NO x and O 3 concentrations), the fraction of NO x remaining (and hence the NO 2 column) can be either higher ( Fig. 2a) or lower ( Fig. 2b) than predicted by instant dilution. For the first hours of expansion, OH is suppressed in the plume, and the fraction of NO x remaining for the expanding plume (solid line) is higher compared to instant dilution of emissions (dashed line). If NO x and O 3 concentrations are relatively high (0.6 and 60 ppbv, respectively; Fig. 2a), the expanding plume simulated by the PARANOX Gaussian plume model (Vinken et al., 2011) results in a higher fraction of NO x remaining compared to instant dilution. For relatively clean ambient conditions -NO x and O 3 concentrations of 0.15 and 39 ppbv, respectively ( Fig. 2b) -the expanding plume simulation has a lower fraction of NO x remaining compared to instant dilution, as in this case the higher NO x concentrations in the plume lead to efficient OH formation and a shorter NO x lifetime. Figure 2b is consistent with a recent study by Valin et al. (2011), who showed, using a 2-D plume model, NO 2 columns for a 2×2 km 2 (i.e. "plume" size) simulation could be higher (by up to 35 %) compared to 48 × 48 km 2 (instant dilution scale) resolution simulations for a moderate emission strength. Furthermore, we note that horizontal transport of emissions in the 2.5 h of expansion is not included in Fig. 1b, while for instant dilution there is transport out of the grid cell in this time, resulting in a lower column. The effect of this transport on our final constraints is minor, as we use long temporal averages and use enhanced NO 2 from ship emissions extended over multiple grid cells. Although the plume-in-grid approach presented in this work will probably introduce additional model errors, it takes into account non-linear chemistry in expanding ship emission plumes, allowing for an appropriate comparison of satellite observations and model simulations of aged pollution plumes.

Ozone Monitoring Instrument
The Dutch-Finish Ozone Monitoring Instrument (OMI) is a nadir-viewing solar backscatter imaging spectrograph aboard the Aura satellite, measuring in the range 264 to 504 nm (Levelt et al., 2006). Aura was launched in sunsynchronous polar orbit on 15 July 2004 with a local Equator-crossing time of 13:40 h. OMI measurements have a spatial resolution up to 13 km × 24 km for nadir pixels and provide daily global coverage. We use the tropospheric NO 2 vertical column densities from Dutch OMI tropospheric NO 2 (DOMINO) v2.0 product (available from the Tropospheric Emissions Monitoring Internet Service (TEMIS); http://www.temis.nl). The DOMINO v2.0 product includes improvements in the radiative transfer modelling, high-resolution surface albedo climatology, better a priori TM4 NO 2 vertical profiles and high-resolution terrain height (Hains et al., 2010;Boersma et al., 2011). The uncertainty in OMI-observed columns due to spectral fitting is 0.7 × 10 15 molecules cm −2 and dominates the overall retrieval error over remote, unpolluted areas (Boersma et al., 2007). Errors arising from incorrect assumptions on surface albedo, aerosols, clouds or the NO 2 vertical profile dominate the overall retrieval error over polluted regions (Boersma et al., 2004). The total error budget for DOMINO v2.0 is estimated to be 1.0 × 10 15 molecules cm −2 + 25 % . Recently, Irie et al. (2012) showed that only a small bias (−10 ± 14 %) exists between DOMINO v2.0 retrievals and in situ multi-axis differential optical absorption spectroscopy (MAX-DOAS) observations at several sites in Japan and China.
The first step of retrievals on OMI data yields slant columns: the integrated abundance of NO 2 along the average photon path through the atmosphere to the instrument. These slant columns are converted to (tropospheric) vertical column densities (VCDs) using a (tropospheric) air mass factor (AMF). This AMF, and hence the retrieved tropospheric NO 2 VCD, is sensitive to the a priori vertical NO 2 profile. In the DOMINO v2.0 retrieval (from now on called DOMINO2), NO 2 vertical profiles simulated by TM4 (Dentener et al., 2003) are used. These vertical profiles have a native spatial resolution of 2 • × 3 • , which is improved upon by spatial interpolation to the OMI pixel centre. Here we replace these a priori vertical NO 2 profiles with profiles from GEOS-Chem nested-grid simulations (1/2 • ×2/3 • horizontal resolution) for the same day and (overpass) time of the OMI measurement, and calculate new tropospheric AMFs. Application of these new tropospheric AMFs results in a new data set of tropospheric VCDs (from now on called DOMINO2_GC), allowing a consistent comparison of the OMI-observed columns with GEOS-Chem-simulated columns, because the vertical distribution assumed in the retrieval is now the same as predicted by the model. The effect of these high-resolution GEOS-Chem profiles will be discussed in the next section.
We exclude clouded situations and snow-or ice-covered pixels to limit retrieval errors by filtering pixels with cloud radiance fraction above 0.5 and surface albedo above 0.2. Fig. 3. Annually averaged tropospheric NO 2 columns on a 1/2 • × 2/3 • resolution for OMI (DOMINO2_GC, top) and GEOS-Chem (bottom). Pixels with cloud radiance fraction above 0.5 and surface albedo above 0.2 are excluded to reduce retrieval errors. Furthermore the outer two pixels on each side of the swath are excluded. OMI pixels are regridded to the GEOS-Chem nested horizontal grid (1/2 • ×2/3 • ), requiring a grid cell coverage of over 75 % and more than three observations per grid cell. The dashed rectangle indicates the area over which spatial averages of OMI and GEOS-Chem are compared in Sect. 2.3.
The effective cloud fraction is obtained from the OMI O 2 -O 2 retrieval (OMCLDO2) (Acarreta et al., 2004;Sneep et al., 2008), and OMI surface albedos are taken from Kleipool et al. (2008). We removed the outer two (large) pixels on each side of the swath to reduce spatial smearing due to viewing geometry. OMI pixels are regridded to the GEOS-Chem nested horizontal grid (1/2 • ×2/3 • ), requiring a grid cell coverage of over 75 % and that there are more than three observations per monthly/seasonal average. A 1 yr averaged map of OMI tropospheric NO 2 columns is shown in Fig. 3 for 2005.

Evaluation of GEOS-Chem and OMI tropospheric NO 2 columns
We compare GEOS-Chem simulations of tropospheric NO 2 columns with OMI-observed columns (both DOMINO2 and DOMINO2_GC) for 2005 ( Figs. 3 and 4). Spatial patterns between DOMINO2_GC and GEOS-Chem for these annual averages over the entire domain are highly consistent (R 2 = 0.90, n = 9270), but the simulated columns are lower than DOMINO2_GC over urban and industrial areas. Although the observations show stronger seasonal variation for 2005-2006 (Fig. 4), the temporal correlation between GEOS-Chem and DOMINO2_GC monthly means (R 2 = 0.95) is remarkably strong, with a root-mean-square error (RMSE) of 0.24 × 10 15 molec cm −2 . For GEOS-Chem and DOMINO2 the temporal correlation is R 2 = 0.89, with an RMSE of 0.45 × 10 15 molecules cm −2 . Huijnen et al. (2010) found a spatial correlation of R = 0.8 (n = 6000) between an ensemble median of regional air quality models and OMI NO 2 observations (DOMINO v1.0.2) over the same domain as in this study for . Furthermore, Huijnen et al. (2010 found that the ensemble median underestimates NO 2 columns by up to 50 % in summer, with only a small bias in winter. The GEOS-Chem model agrees better to OMI observations, reflecting the improved (lower) uptake coefficient for N 2 O 5 on aerosols (Sect. 2.1) and DOMINO NO 2 retrievals (v2.0).
Differences between DOMINO2_GC and DOMINO2 ( Fig. 5) arise from the different a priori NO 2 profiles used in the air mass factor (AMF) calculation. GEOS-Chem NO 2 profiles differ in three ways from the original TM4 profiles: (1) different emissions over the domain, (2) higher spatial resolution and (3) a different CTM (e.g. different vertical mixing and chemical lifetime). Different emissions most likely dominate the changes in AMFs. TM4 used emissions from the POET project (Precursors of Ozone and their Effects on the Troposphere) for the year 1997 (Olivier et al., 2003), which amount to 8.2 Tg N yr −1 for Europe. In GEOS-Chem we use EMEP (Vestreng et al., 2007) emissions, which amount to 6.3 Tg N yr −1 for 2005. Lower emissions lead to lower concentrations in the a priori profiles (Fig. S1), and hence higher AMFs (Fig. S2), resulting in lower tropospheric NO 2 columns (e.g. Martin et al., 2003;Boersma et al., 2004), as can be observed by the reduction in NO 2 columns over western Europe in Fig. 5. In contrast, increased emissions in eastern Europe lead to increased OMI NO 2 columns here. This dependence of AMFs on a priori emissions in the retrieval profiles was earlier found by Barkley et al. (2012) for HCHO vertical columns. The effect of the higher resolution of the GEOS-Chem profiles can be observed near large cities (e.g. Barcelona), where NO 2 columns increase due to better localized emissions in the NO 2 profiles.
OMI NO 2 retrievals using GEOS-Chem NO 2 profiles in the AMF calculation are on average 10 % lower than the original DOMINO2 retrievals using TM4 a priori NO 2 profiles. The new DOMINO2_GC retrievals and GEOS-Chem now agree to within 7 %, with largest differences in winter months. The wintertime underestimation of GEOS-Chem shown in Fig. 4 possibly reflects a too short NO 2 chemical lifetime in GEOS-Chem, as suggested by a number of recent studies on reaction rate updates (e.g. Stavrakou et al.,   Fig. 3. Selection of OMI observations follows the same criteria as Fig. 3.

Fig. 5.
Differences between annually averaged OMI NO 2 columns for 2005 on a 1/2 • × 2/3 • resolution using the DOMINO2_GC retrieval (with a priori NO 2 profiles based on GEOS-Chem 1/2 • × 2/3 • simulations) and DOMINO2 retrieval (with a priori profiles based on TM4 2 • × 3 • simulations). Selection of OMI observations follows the same criteria as Fig. 3. Differences between these two retrievals arise from different a priori NO 2 profiles, as discussed in Sect. 2.3. The AMFs for these retrievals are shown in Fig. S2. 2013; Zhang et al., 2012). In particular, the value for the uptake coefficient γ N 2 O 5 in GEOS-Chem is high compared to recent laboratory and field estimates of this value (e.g. Brown et al., 2009;Mollner et al., 2010;Henderson et al., 2012;Butkovskaya et al., 2007Butkovskaya et al., , 2009. We conclude that the nested GEOS-Chem CTM is in close agreement with OMIobserved NO 2 columns over Europe. Differences in wintertime between OMI and GEOS-Chem are unlikely to influ-ence our results because of our filter criteria (discussed in the next section).

Selection of OMI NO 2 observations
Figure 3 (top) shows two ship tracks in the annually averaged OMI NO 2 observations: one in the Bay of Biscay and one in the Mediterranean Sea. Emission inventories (see next section) suggest that there are also busy ship routes in the North Sea. However, detection of ship tracks in this sea presents significant challenges, since influence from outflow of pollution from land often prohibits a clear view of the pollution from ships. To be able to detect ship emissions in this sea, we inspect daily OMI observations and screen out all days with measurements that are significantly affected by continental outflow (e.g. Fig. 6a). Furthermore, we filter for days that have cloud-free observations (cloud radiance fraction < 0.5) over at least 90% of the ship track area (Fig. 6b), as scattered clouds prevent the unambiguous detection of pollution from ships. In addition to this cloud filter, we also exclude strong negative NO 2 columns (< −0.5 × 10 15 molecules cm −2 ). Using these two criteria we can now identify two additional ship tracks in European seas (the Baltic Sea and the North Sea) (Fig. 6c).

Ship emission inventories
In the GEOS-Chem CTM two (recent) ship emission inventories can be used for Europe: the European regional EMEP inventory (Vestreng et al., 2007) and the global AMVER-ICOADS inventory . The left panels of Fig. 7 show the spatial distribution of European ship NO x Dashed boxes indicate the ship track areas that were used to calculate constraints in this study, and ship tracks were averaged along the white lines in these boxes. emissions, with totals for the EMEP and AMVER-ICOADS inventory for 2005 of 1.1 and 0.8 Tg N, respectively. Differences between these inventories arise from the use of different methodologies (e.g. using fuel consumption versus shipping activity) and spatial allocation. The AMVER-ICOADS inventory is based on international fuel statistics, which do not include fuel consumed for domestic traffic, whereas EMEP does include these domestic emissions. Therefore AMVER-ICOADS underestimates ship emissions over inland seas and coastal zones with significant domestic ship traffic, like the (eastern part of the) Mediterranean Sea (Marmer et al., 2009). The spatial allocation in EMEP is based on the distance each ship covers between ports, information provided by the Lloyd's Register of Shipping (Vestreng, 2003). In the AMVER-ICOADS inventory the spatial allocation is taken from actual ship locations reported to the Automated Mutual-Assistance Vessel Rescue System (AMVER) and International Comprehensive Ocean-Atmosphere Data Set (ICOADS). Comparison with annual OMI NO 2 observations (Fig. 7b) for the eastern part of the Mediterranean Sea shows that the AMVER-ICOADS inventory (Fig. 7c) simulates the ship track closer to the observed tracks. The location of the EMEP emissions is misplaced by up to 150 km (too close to Crete, Fig. 7a). However, as the EMEP inventory does include domestic ship traffic, we combine both inventories and generate a new ship emission inventory (Fig. 7d). This inventory is based on EMEP emission totals, and AMVER-ICOADS emission locations for the eastern part of the Mediterranean Sea. In the following, we will use the combined EMEP & AMVER-ICOADS database as the a-priori emission inventory in our simulations.

Sensitivity of GEOS-Chem NO 2 columns to NO x emissions
Within pollution plumes, the NO x lifetime is influenced by the local NO x concentration. This is because oxidation losses depend on in-plume OH availability, which is determined by the local concentration of pollution, e.g. the NO x concentration itself. We need to account for this non-linear feedback when changing NO x emissions based on observed (changes in) NO 2 columns. Lamsal et al. (2011) introduced a dimensionless factor β, which represents the (modelled) local sensitivity of NO 2 column changes to NO x emissions changes. β is computed by changing NO x emissions by a fixed prescribed percentage and evaluating the local (relative) change in NO 2 column: where E represents the NO x emissions, N GC the simulated tropospheric NO 2 column, E the change in NO low NO 2 concentrations, reflecting efficient OH production and a lower NO x lifetime following an increase in emissions. In polluted areas, β tends to be less than 1 as an increase in NO x will consume OH and increase the NO x lifetime. Recently, Lu and Streets (2012) showed a decrease of β values from about 2 to 0.7 over Indian power plants during the period 1996-2010, following a dramatic increase in NO x emissions. Their study illustrates the strong variability of β, and the need to determine β for a realistic emission strength. We calculate β from seasonal mean NO 2 columns (over the entire ship track as defined in Fig. 6) for the Mediterranean Sea, and from annual mean columns for the Bay of Biscay, the North Sea and the Baltic Sea (filtered by the criteria of Sect. 2.4). As we expect our ship emissions to change by more than 15 % we follow a two-step approach to calculate β values, different from Lamsal et al. (2011). First we run our model with emissions perturbed by the relative difference of observed and simulated NO 2 columns (ignoring the non-linear feedback of emission on simulated columns). From the results of these simulations we calculate β values for our ship tracks and apply these to derive new top-down emissions. We find β values of 0.3-0.9 (Table 2), indicating that emission changes lead to substantial changes in NO 2 columns over the ship lanes. This is expected, as ship emissions are released following our plume-in-grid approach, and NO x concentrations will be relatively high in the expanding plume, a situation comparable to release of NO x in a polluted area. Differences in β are driven by the magnitude of emissions changes and local chemical regime. For example, in the Baltic Sea we impose strong emissions perturbations, and the resulting β values are small (comparable to a decreasing β from ∼ 2 to ∼ 0.7 found in Lu and Streets (2012) for increasing emissions of power plants). Emissions changes for the North Sea and the Mediterranean Sea are similar; however the calculated β values (∼ 0.6 for the North Sea, ∼ 0.8-0.9 for the Mediterranean Sea) indicate different chemical regimes. This is consistent with Fig. S1 in Lamsal et al. (2011), which indicates that β values are lower for wintertime and polluted areas, and higher β values correspond with summer and less polluted areas.

Sensitivity of OMI NO 2 columns to a priori (GEOS-Chem) NO 2 columns
OMI tropospheric NO 2 columns depend on the a priori vertical NO 2 profile. In this study we replaced the TM4 profiles used in the DOMINO v2.0 retrieval with high-resolution GEOS-Chem-simulated NO 2 profiles (Sect. 2.2, leading to OMI NO 2 retrievals 10 % lower than original DOMINO2). As a result of constraining ship emissions in GEOS-Chem, the retrieved OMI NO 2 columns will also change in response to updated a priori NO 2 profiles over the shipping lanes. We quantify the effect of changing GEOS-Chem NO 2 columns  Fig. 6c. Emission strengths are summed over the constrained ship tracks as indicated in Fig. 9d. (NO x emissions) on OMI NO 2 columns by introducing a dimensionless factor γ : where N GC represents the simulated tropospheric NO 2 column corresponding to the a priori profile shape used in the retrieval, N OMI the retrieved OMI tropospheric NO 2 column, N GC the change in simulated NO 2 column as a result of changing emissions, and N OMI the change in retrieved NO 2 column because of the changed a priori NO 2 profile. A γ value of zero would indicate no sensitivity of OMI NO 2 columns to changing GEOS-Chem columns. We calculate γ values in the same way as β values using results of a model run with perturbed emissions. γ is found to be always smaller than 1 (Table 2), indicating that the relative change in OMI NO 2 column from a priori is always smaller than the relative change in a priori GEOS-Chem columns. The γ factor not equal to zero illustrates that OMI retrievals are never completely independent of a priori assumptions, and this factor takes into account how the changing profile shape influences our retrieval. Figure 6c shows a map of OMI tropospheric NO 2 columns for 2005, gridded on the GEOS-Chem horizontal resolution (1/2 • × 2/3 • ). The conventional approach (Martin et al., 2003;Lamsal et al., 2011) to estimate top-down NO x emissions (E top down ) for the four major ship routes that are visible in the Baltic Sea, the North Sea, the Bay of Biscay and the Mediterranean Sea would be to use the relative difference of observed and simulated columns (using the a priori emissions, indicated by the subscript 1) over these ship routes in combination with the modelled sensitivity β 1 to scale the a priori emissions:

Space-based constraints on ship emissions
In this study, we determine β 1 by perturbing the a priori ship emissions by the relative difference of observed and simulated columns: where N GC,2 represents the GEOS-Chem-simulated column after perturbing the a priori ship emissions by the factor ((N OMI,1 − N GC,1 )/N GC,1 ). γ 1 was determined following Eq. 2 by means of where N OMI,2 represents the OMI tropospheric NO 2 column retrieved with the NO 2 profile of the perturbed GEOS-Chem simulation (simulation 2). Replacing the NO 2 profile in the OMI NO 2 retrieval is comparable to replacing the TM4 a priori profiles by GEOS-Chem profiles in the DOMINO2_GC retrieval (see Sect. 2.3 and the Supplement for a discussion of the effect). We now extend the approach by Lamsal et al. (2011) (Eq. 3) by also taking into account the sensitivity of the OMI retrievals to changes in a priori NO 2 profiles. We do so by adding a term to Eq. 3 that accounts for the response of the retrieval to the changed emissions and modifies the overall scaling of the a priori inventory by ((N OMI,1 − N GC,1 )/N GC,1 ) · γ 1 · β 1 .
The new top-down NO x emissions inventory we then obtain via E top down = E a priori + N OMI,1 − N GC,1 N GC,1 · β 1 · E a priori (4) The second term on the right-hand side of Eq. (4) describes the scaling of the a priori emissions required to match the original observed (N OMI,1 ) and simulated (N GC,1 ) NO 2 columns. The third term on the right-hand side may be interpreted as a necessary enhancement of the second term: an increase in a priori emissions by the factor ((N OMI,1 − N GC,1 )/N GC,1 ) · β 1 would lead to more pronounced a priori NO 2 profiles that in turn lead to lower AMFs and higher OMI-retrieved NO 2 columns by a factor ((N OMI,1 − N GC,1 )/N GC,1 ) · γ 1 . In practice, the third term is always smaller than the second term, but still leads to substantial emissions enhancements of up to 35 % on top of the second term (i.e. the approach of (Lamsal et al., 2011)). We found that for the Baltic Sea, and in some seasons for the Mediterranean Sea, emission changes were large and an additional iteration was needed. For this additional iteration we applied Eq. (4) again, but now using the top-down emissions (calculated using Eq. 4) as a priori emissions and also basing the OMI and GEOS-Chem columns on these topdown emissions (this additional iteration is described in the Supplement). For the Baltic Sea, the North Sea and the Bay of Biscay we provide annual constraints based on annual averages of filtered days in 2005 and 2006 (see Sect. 2.4 for selection criteria). For the Mediterranean Sea, more days satisfying our criteria are available, allowing us to provide seasonal constraints for 2005 and 2006.
Using the seasonally and annually averaged tropospheric NO 2 columns, four along-ship-track averages were created by averaging the areas in Fig. 6c over the longitude for the Mediterranean Sea and the Baltic Sea, and over the latitude for the North Sea and the Bay of Biscay (Fig. 8). For these averages, these grid cells were rotated and interpolated along the ship track location. A linear background was fitted to these averages as indicated in Fig. S3 and subtracted from the cross sections. We use this background correction for both OMI and GEOS-Chem, ensuring consistency in the comparison. Using a simulation without ship emissions to determine the contribution of ships could lead to interpretation errors due to non-linearities in the NO x chemistry. The averages clearly show enhanced columns relative to the background (of up to 0.6 × 10 15 molecules cm −2 ) over the ship routes. OMI observations indicate that NO x emissions in GEOS-Chem are too low over the Baltic Sea and the Bay of Biscay, while emissions are too high over the North Sea and the Mediterranean Sea. Furthermore, the cross section for the Bay of Biscay (Fig. 8, left bottom) shows that emissions in GEOS-Chem are located too far to the east by 0.5 • (or ∼ 50 km), compared to the OMI observations. As our inversion does not correct the location of the emissions, we shifted the emissions prior to the inversion in the (combined) inventory to match the OMI location (see next section).

Top-down NO x ship emissions
We proceed and determine the relative difference of the area under the OMI-observed and GEOS-Chem-modelled tropospheric NO 2 cross sections in Fig. 8. We use this relative difference ((N OMI,1 −N GC,1 )/N GC,1 ) to provide constraints following Eq. (4) for sections of the four ship tracks indicated in Fig. 9. We apply these constraints to the much longer tracks shown in Fig. 9d  strong changes in emissions in some ship tracks and strong non-linearities in the NO x chemistry, an additional iteration was needed to match the modelled columns to the simulated ones (see Table S1). This second iteration results in a close match (within 10 %) between observed and simulated column enhancements relative to background for all four ship tracks. Figure 9 shows the results of our constraints on the ship NO x emissions. The OMI top-down total ship NO x emissions amount to 0.96 Tg N for 2005 (1.0 Tg N for 2006), a reduction of 15 % (11 %) compared to the EMEP inventory. The new inventory (Fig. 9b) is more coherent than the EMEP inventory (Fig. 7d), as sudden stepwise jumps in NO x emissions, for instance from the Bay of Biscay to the ship lane west of Portugal, no longer appear. Emission totals for the four constrained European ship tracks are given in Table 2; these show that emissions in the Mediterranean Sea ship track are strongly reduced (to 0.13 Tg N) and closely match AMVER-ICOADS emissions (0.12 Tg N) for 2005 and 2006. Also the strong emissions in the North Sea track are reduced to 0.05 Tg N for 2005 (35 % lower than EMEP) and are more consistent with emission strengths in the Baltic Sea. Figure 9c shows that the AMVER-ICOADS inventory underestimates emissions in the North Sea and the Baltic Sea, and also illustrates that this inventory does not take into account domestic ship traffic in coastal zones. Our top-down emissions are still strong in the English Channel, but our satellite observations do not provide constraints for this area. We note that the TNO-MACC European emission inventory (Denier van der Gon et al., 2011) provides high-resolution ship emissions in the North Sea using ship location (automatic identification system) data (Jalkanen et al., 2009). OMI top-down emissions for the Baltic Sea and the Bay of Biscay ship tracks are 0.05 Tg N for 2005 (131 % higher than EMEP) and 0.08 Tg N for 2005 (128 % higher than EMEP), respectively. The constraints found in this study are also applied to the SO x and CO emission inventories, assuming that the scaling in NO x emissions is due to increased/decreased activity, and not due to changes in emissions factors.
Using simulations with our new top-down emission inventory, four along-ship-track averages of observed and simulated NO 2 columns were created (Fig. 10). Compared to Fig. 8, simulated columns now closely match (within 10 %) observed columns. Figure 10 also shows the emission strength averaged along the ship tracks. Emission strengths are similar in the Bay of Biscay and the Baltic Sea, but NO 2 column enhancements over ship lanes differ. NO 2 column enhancements are a factor of 3-4 lower over the Mediterranean Sea compared to the Baltic Sea and the Bay of Biscay, while emissions are only 50 % lower, indicating different photo-chemical regimes (longer NO 2 lifetimes) in different seas in Europe.

Error of the top-down emission inventory
Our approach to derive new top-down ship NO x emissions is sensitive to errors in both the satellite-observed and simulated NO 2 columns. Important systematic errors in the simulations have been addressed by taking into account the effects of non-linear chemistry in the expanding ship plume and by changing the location of emissions in the top-down inventory for the Mediterranean Sea and the Bay of Biscay. However, other systematic model errors may still persist, related to errors in the NO 2 chemistry (and hence lifetime), and (vertical) transport (e.g. Lin et al., 2012;Stavrakou et al., 2013). Our plume-in-grid method may also introduce errors in our approach, and we cautiously estimate these to be of the order of 15 %. OMI NO 2 observations also have systematic errors, due to errors in the AMFs or stratospheric correction. The error in an OMI-observed NO 2 column is estimated to be 1 × 10 15 molecules cm −2 +25 %, and consists of a systematic and a random part Irie et al., 2012). We reduce random errors in OMI observations by averaging, but might also introduce an additional error in the background correction of our along-ship-track averages (Figs. 8 and 10), resulting in an estimate of OMI-related errors of 20 %. Effects of tighter cloud filtering (cloud radiance fraction < 0.25) are found to result in changes smaller than 10% for the Baltic Sea, and negligible differences for the Mediterranean Sea, the Bay of Biscay and the North Sea . This cloud-related uncertainty is included in the OMI-related error and estimated to be less than 10 %, in agreement with earlier results found for shipping NO 2 columns by Franke et al. (2009). Assuming the errors in OMI, GEOS-Chem and PARANOX to be largely uncorrelated, and using simple error propagation, we estimate the total systematic component of the errors in our approach to be 29 %. Apart from systematic errors in our model or observations, there are also random errors resulting from our approach. These errors arise from statistical errors in the averaging of the columns and determination of the relative difference between (backgroundcorrected) observed and simulated columns. We used the sample standard error to calculate the statistical error in our averages. This error is lowest (30 %) in summer months for the Mediterranean Sea, when a large number of observations are available. The statistical error for the North Sea track is largest (70 %), as NO 2 columns will show more variation due to close proximity to polluted areas. Overall we estimate an error of 40-60 % on our OMI top-down ship NO x emissions inventory.

Conclusions
We developed a method to constrain a large fraction (39 %) of European ship NO x emissions using OMI tropospheric NO 2 columns and the nested-grid GEOS-Chem CTM in the Baltic Sea, the North Sea, the Bay of Biscay and the Mediterranean Sea. Our method allows for a direct comparison between satellite-observed and simulated NO 2 columns (using a model snapshot). First, we updated our previously developed plume-in-grid approach, which accounts for nonlinear chemistry in expanding ship plumes for GEOS-Chem (1/2 • × 2/3 • resolution). OMI showed that ship emissions were misplaced in the Mediterranean Sea and the Bay of Biscay, and we made sure ship emissions were released at the correct location in the CTM. These updates reduce systematic errors in our simulations. Second, we ensured consistency between retrieval and modelling by replacing TM4 a priori NO 2 profiles by high-resolution GEOS-Chem profiles. To reduce influence of continental pollution, and improve detection of ship tracks, we screened out observations that were affected by outflow. Furthermore we only included days that have cloud-free observations over the entire ship track. As a result, we are able to identify ship tracks in four European seas (the Baltic Sea, the North Sea, the Bay of Biscay and the Mediterranean Sea). This is the first time that ship tracks in the Baltic Sea, the North Sea and the Bay of Biscay have been observed in OMI tropospheric NO 2 columns and used to constrain ship NO x emissions.
We use the relative difference between observed and simulated NO 2 columns to provide constraints on ship NO x emissions, and explicitly account for sensitivities to changing emissions in the model and satellite retrieval. The (nonlinear) sensitivity of simulated NO 2 columns to changing NO x emissions (the so-called β factor) found in this study is in the range 0.3-0.9, indicating that emission changes lead to substantial changes in NO 2 columns over ship lanes. We also account for the (non-linear) sensitivity of satellite observations to changing a priori NO 2 profiles. Although the effect of this sensitivity might be minor for small emission changes, the effect on observed NO 2 columns can be significant for large changes in NO x emissions (up to 87 % of GEOS-Chem column change). Our findings stress the need for consistent information in the satellite retrieval and the model, as satellite derived vertical columns are never fully independent of model information (i.e. vertical NO 2 profiles).
Emissions in the main ship track of the Mediterranean Sea in our top-down inventory (0.13 Tg N) closely match the emissions strength of the AMVER-ICOADS inventory (0.12 Tg N) for 2005, and emissions in the Bay of Biscay and the North Sea appear more coherent with emissions in surrounding seas. Our results indicate that Mediterranean Sea emissions in the EMEP inventory are too high (by 60 %), which could have important consequences for local air quality simulations. Future work could focus on the effect of these reduced and relocated emissions on air quality. In the North Sea ship track, our total top-down emissions amount to 0.05 Tg N for 2005 (35 % lower than EMEP). OMI top-down emissions for the Baltic Sea and the Bay of Biscay ship tracks are 0.05 Tg N (131 % higher than EMEP) and 0.08 Tg N (128 % higher than EMEP) for 2005, respectively. Our topdown emission inventory (0.96 Tg N for 2005, 1.0 Tg N for 2006) is about 11-15 % lower than the (regional) EMEP ship emission inventory (1.1 Tg N), and in closer agreement with the AMVER-ICOADS global emission inventory (0.8 Tg N).
Our study provides a framework for future studies to constrain ship NO x emissions using satellite NO 2 observations. This may be particularly valuable given the paucity of measurements of ship pollution over open waters and the upcoming emission control measures. Future work will focus on expanding the analysis to more years, providing an OMIconstrained top-down ship NO x emissions inventory for use in CTMs. Including observations of additional satellite instruments could also be explored in the future in order to reduce systematic and random errors in the top-down emissions.