Improved monitoring of shipping NO 2 with TROPOMI: decreasing NO x emissions in European seas during the COVID-19 pandemic

. TROPOMI measurements of tropospheric NO 2 columns provide powerful information on emissions of air pollution by ships on open sea. This information is potentially useful for authorities to help determine the (non-)compliance of ships with increasingly stringent NO x emission regulations. We ﬁnd that the information quality is improved further by recent upgrades in the TROPOMI cloud retrieval and an optimal data selection. We show that the superior spatial resolution of TROPOMI allows the detection of several lanes of NO 2 pollution ranging from the Aegean Sea near Greece to the Skagerrak in Scandinavia, 5 which have not been detected with other satellite instruments before. Additionally, we demonstrate that under conditions of sun glint TROPOMI’s vertical sensitivity to NO 2 in the marine boundary layer increases by up to 60%. The beneﬁts of sun glint are most prominent under clear-sky situations when sea surface winds are low, but slightly above zero ( ± 2 m/s). Beyond spatial resolution and sun glint, we examine for the ﬁrst time the impact of the recently improved cloud algorithm on the TROPOMI NO 2 retrieval quality, both over sea and over land. We ﬁnd that the new FRESCO+wide algorithm leads to 50 hPa lower cloud 10 pressures, correcting a known high bias, and produces 1-4 · 10 15 molec · cm − 2 higher retrieved NO 2 columns, thereby at least partially correcting for the previously reported low bias in the TROPOMI NO 2 product. By training an artiﬁcial neural network on the 4 available periods with standard and FRESCO+wide test-retrievals, we develop a historic, consistent TROPOMI NO 2 data set spanning the years 2019 and 2020. This improved data set shows stronger (35-75%) and sharper (10-35%) shipping NO 2 signals compared to co-sampled measurements from OMI. We apply our improved data set to investigate the impact 15 of the COVID-19 pandemic on ship NO 2 pollution over European seas and ﬁnd indications that NO x emissions from ships reduced by 10-20% during the beginning of the COVID-19 pandemic in 2020. The reductions in ship NO 2 pollution start in March-April 2020, in line with changes in shipping activity inferred from Automatic Identiﬁcation System (AIS) data on ship location, speed and engine


Introduction
Emissions of nitrogen oxides (NO x =NO+NO 2 ) have several primary and secondary effects on air quality, human health and the environment. NO x is a toxic gas itself (WHO, 2003) and contributes to the formation of secondary pollutants and ozone.
Ozone close to the Earth's surface is a toxic pollutant which can lead to respiratory problems and has negative effects on plant growth and crop yield (e.g. Wang and Mauzerall (2004)). NO x also contributes to acid deposition and eutrophication, harming 25 sensitive ecosystems (European Environment Agency, 2019).
The international shipping sector is a strong source of NO x and other air pollutants to the atmosphere (eg. Eyring et al. (2010)).
Previous studies suggest that international shipping makes up for annual emissions of 2.0-10.4 TgN (Crippa et al., 2018;Eyring et al., 2010;Johansson et al., 2017), or 15-35% of total anthropogenic NO x emissions worldwide. While cars, the power sector and industry have shown substantial reductions in their emissions over the last 10-20 years in Europe and the United States 30 (Curier et al., 2014;Hassler et al., 2016), NO x emissions from shipping activity have increased (De Ruyter de Wildt et al., 2012;Boersma et al., 2015) and the number of ship movements and ship size is expected to keep increasing in the future (Eyring et al., 2005;UNCTAD, 2019). Shipping-related air pollution emissions are estimated to lead to 60,000 premature deaths annually, especially in coastal regions (Corbett et al., 2007;Marais et al., 2015).
To mitigate these and other harmful impacts, more stringent regulations on NO x emissions for ships have been implemented  For new regulations to be effective, monitoring and verification of ship emissions is required. Traditional compliance monitoring includes national authorities conducting on-board checks of engine certificates and keel-laying date. This is not a direct verification of emissions and can only be done for a limited number of vessels. Other methods are on-board measurements at the ship's exhaust pipe (e.g. Agrawal et al. (2008)) or downwind measurements of emission plumes using sniffer techniques or DOAS (Differential Optical Absorption Spectroscopy) measurements (e.g. Lack et al. (2009) ;Berg et al. (2012); McLaren et al. 45 (2012); Pirjola et al. (2014); Seyler et al. (2019)). Modern techniques also include airborne platforms such as helicopters, small aircrafts (Mellqvist and Conde, 2021;Chen et al., 2005b) or drones (Van Roy and Scheldeman, 2016). While these methods do not require inspectors to board the vessel, they require proximity to the ships monitored and are thus less fit-for-purpose when a large number of ships is to be checked, or on open sea away from land. For the above reasons monitoring by satellite remote sensing offers a promising alternative. 50 Satellite instruments have observed enhancements of NO 2 column densities over major shipping routes, e.g. from GOME (Beirle et al., 2004), SCIAMACHY (Richter et al., 2004) and OMI (Vinken et al., 2014b;Marmer et al., 2009). These satellite measurements have recently been continued with new observations from the TROPOMI (TROPOspheric Monitoring Instrument) sensor. With a pixel size of 3.5x5.5 km 2 TROPOMI provides a spatially more resolved evaluation of NO 2 pollution patterns compared to its predecessors GOME (40x320 km 2 ), SCIAMCHY (30x60 km 2 ) and OMI (13x24 km 2 ). Indeed, previous studies 55 demonstrated TROPOMI's capability to pinpoint emissions from the mining industry (Griffin et al., 2019), emissions patterns within cities Lorente et al., 2019), emissions along a gas pipeline in Siberia (van der A et al., 2020) and even from individual ships in the Mediterranean Sea (Georgoulias et al., 2020). Ding et al. (2020) used TROPOMI NO 2 columns and inverse modelling to show NO x emission reductions during the COVID-19 lockdown over urban centers and regions with strong maritime transport. 60 While the aforementioned studies demonstrate the large potential of TROPOMI and its high resolution, retrieval problems remain. Validation studies (e.g. Griffin et al. (2019); ) suggest a 15%-40% low bias in TROPOMI troposheric vertical NO 2 (N v,trop ) columns relative to independent in-situ and MAX-DOAS measurements. Cloud properties present one of the leading sources of uncertainty in trace gas retrieval from space (Boersma et al., 2004;Lorente et al., 2017) and cloud heights used until (and including) v1.3 of the operational TROPOMI retrieval algorithm have been suggested to be 65 biased low . To address this bias in cloud heights, the Royal Dutch Meteorological Institute (KNMI) recently updated the FRESCO+ cloud retrieval by widening the spectral window, which is supposed to improve the sensitivity to low clouds.
The here presented study presents and assesses the impact of steps towards an improved monitoring of shipping NO 2 with TROPOMI. First, we demonstrate TROPOMI's capability to detect ship emissions applying a typical data selection and com-70 pare it to OMI's. We examine previous suggestions of improved retrieval sensitivity over sun glint scenes (Georgoulias et al., 2020). Additionally, we evaluate the new FRESCO+wide cloud pressure retrieval in and its impact on the TROPOMI NO 2 columns in v1.4/2.1 of the operational TROPOMI NO 2 algorithm. Based on our findings, we create a data set of historical TROPOMI NO 2 columns consistent with the v1.4 data allowing for otherwise challenging trend analysis. We conclude with an application of our findings to quantify the effects of the COVID-19 pandemic on ship pollution, an unique opportunity to 75 assess the relationship between the anticipated emission reductions and observed NO 2 columns.

TROPOMI and OMI NO 2 column measurements
The European TROPOMI (Veefkind et al., 2012) is on board the Sentinel-5-Precursor launched in October 2017. TROPOMI has a push-broom design with a 2-D detector, which measures back-scattered radiation from the Earth's atmosphere for viewing Combined, the area of the smallest TROPOMI pixel is 19 km 2 , while it is 325 km 2 for OMI, a factor of 17 improvement in spatial resolution. Both instruments are in a sun-synchronous ascending orbit and have an equator overpass time of about 13:30 hrs local time.

90
To retrieve tropospheric NO 2 columns, TROPOMI uses a 3-step retrieval approach based on the DOAS (Differential Optical Absorption Spectroscopy, Platt and Stutz (2008)) technique: first the slant column density (N s ) is retrieved by spectral fitting of a modeled reflectance spectrum to the observed reflectances in the 405-465 nm window (van Geffen et al., 2021;Van Geffen et al., 2020;Zara et al., 2018). In the second step, data assimilation in the global chemistry Transport Model 5 (TM5-MP) results in vertical NO 2 profiles that are then used to separate the stratospheric and tropospheric contribution to the 95 slant columns (Van Geffen et al., 2020;Dirksen et al., 2011). In the last step, Air Mass Factors (AMFs) are calculated (Lorente et al., 2017) to translate the N s into vertical column densities (N v ). The AMF is calculated using the DAK radiative transfer model (de Haan et al., 1987;Stammes, 2001), and accounts for the viewing and solar geometry as well as surface properties and cloud effects. Cloud height information is retrieved with TROPOMI's FRESCO+ cloud algorithm (driven by the 761 and 765 nm O 2 absorption depth), and cloud fraction from the reflectance levels within the 405-465 nm NO 2 fitting window. Other 100 input parameters to the TROPOMI AMF calculation are the surface albedo climatology (Kleipool et al. (2008), 0.5°x0.5°), a priori NO 2 profiles simulated with TM5-MP (Williams et al. (2017), 1°x1°) and terrein height from Global 3 km Digital Elevation Model (DEM_3KM).
The retrieval of tropospheric NO 2 columns (N v,trop ) from OMI  proceeds along the same lines, and is therefore similar in many aspects. On the other hand, especially spatial resolution, signal-to-noise and the retrieval of cloud 105 properties differ as highlighted in Table 1.
Clouds have several relevant effects on NO 2 retrieval. Clouds shield the lower part of the atmosphere which is most influenced by anthropogenic emissions including those from shipping. Therefore, data users are typically advised to consider scenes with cloud radiance fractions below 50% (Eskes et al., 2019). Initial validation of TROPOMI NO 2 v1.2/1.3 pointed out that the FRESCO+ algorithm retrieves cloud heights close to the surface heights, leading to overestimations in the TROPOMI NO 2 110 AMFs, and, consequently underestimations of the tropospheric NO 2 columns . Accurate knowledge of cloud fraction and height is key for high quality trace gas column retrievals (e.g. Boersma et al. (2004); Van Geffen et al.
(2021)). A detailed description of the TROPOMI and OMI cloud algorithms and recent updates therein is given in the following subsection. 115 2.2 Improved TROPOMI FRESCO+, OMI and VIIRS cloud retrievals FRESCO+ (Fast Retrieval Scheme for Clouds from the Oxygen A band) retrieves cloud pressures from the relative depth of O 2 -A band measurements (Koelemeijer et al., 2001;Wang et al., 2008) using three spectral windows at 758-759 nm (continuum, no absorption), 760-761 nm (strong absorption) and 765-766 nm (moderate absorption). In the algorithm, clouds are assumed to be Lambertian reflectors with a fixed albedo of 0.8, consistent with assumptions for the NO 2 AMF calculation. The surface 120 albedo assumed in the cloud pressure retrieval is from the GOME-2 minimum LER climatology at 758 & 772 nm (Tilstra  , 2017), which is a potential source of uncertainty in the cloud pressure retrieval as the resolution and overpass time of GOME-2 is different from TROPOMI. FRESCO+ has been compared to other cloud data sets by Compernolle et al. (2021), who reported on tendencies in FRESCO+ to overestimate cloud pressures.
To address the high-bias in TROPOMI FRESCO+ cloud pressures, a new version of the FRESCO+ algorithm was introduced 125 and implemented in the operational NO 2 retrieval with the introduction of TROPOMI v1.4 in December 2020. This version, called FRESCO+wide, uses a wider spectral window for the cloud retrieval (765-770 nm, see Table 1), which includes the flank of the absorption band, where oxygen absorption is weaker than in the center of the O 2 -A band (761 and 765 nm). Adding weaker O 2 absorption features improves the sensitivity to clouds low in the atmosphere. This is not possible from the strong O 2 absorption at 761 nm, which is so close to saturation that it becomes difficult to use its absorption depth in order to distinguish 130 between bright reflecting layers at the Earth's surface and reflecting surfaces in the lower atmosphere.
Prior to the implementation of FRESCO+wide in the operational TROPOMI NO 2 retrieval in December 2020, KNMI produced 4 periods with TROPOMI NO 2 test data based on FRESCO+wide, the so-called diagnostic data set 2B (DDS-2B retrieval derives a cloud top temperature using an optimal estimation approach in the thermal infrared spectral bands M14-M16 (8.5-12.3 µm). In a subsequent step, these cloud top temperatures are converted to cloud pressures using Numerical Weather Prediction temperature profiles (Heidinger and Li, 2017). In addition to the cloud top pressure, we use the VIIRS cloud optical 150 thickness (COT) to generate (effective) cloud fractions that can be compared directly to the TROPOMI cloud fractions. First, we derive a geometrical cloud fraction by calculating the share of cloudy VIIRS pixels per grid cell. Then, we translate this geometrical cloud fraction f c,geo into a effective cloud fraction f c,ef f using: with a c the cloud albedo. The cloud albedo is calculated from the VIIRS COT and a previously established empirical relation-155 ship between cloud optical thickness and cloud albedo for liquid water clouds (Buriez, 2005;Boersma et al., 2016) To evaluate the improvements in the FRESCO+wide retrieval, we compare daily gridded, co-sampled cloud data from (partly) 1 In addition to improved cloud parameters, TROPOMI v2.1 data has improved further through a better calibration of level-1 spectra, especially in the treatment of outliers and saturation (Ludewig et al., 2020), and through improvements in the NO 2 algorithm itself (Van Geffen et al., 2021). Version v2.1 is only used for production of the DDS-2B test data, not for publicly released data. Version v2.2, available publicly as of July 2021, is essentially the same as v2.1. 2 The cloud albedo ac for liquid water clouds can be expressed as a 6th order polynomial of the VIIRS cloud optical thickness (τ ) as ac cloudy pixels seen by TROPOMI (FRESCO+ and FRESCO+wide), OMI and VIIRS over parts of the Mediterranean Sea (37.0°N-41.25°N, 2.0°W-8.0°W), the Bay of Biscay (43.5°N-47.5°N,10.0°E-3.0°E) and Northwestern Europe (50.0°N-53.0°N, 4.0°W-9.0°W). These regions represent different surface types (land and ocean), climatological conditions and pollution levels. 160 We define partly cloudy pixels as all pixels with an effective cloud fraction f c ≥ 0.05. For TROPOMI we additionally apply sufficient quality of retrieval (qa ≥ 0.5) and a pressure difference between surface pressure and cloud pressure of at least 7 hPa.
The last filter is applied to filter out 'ghost' clouds coming from sun glint viewing geometries (see Sec. 2.3 below). For OMI, we use the OMCLDO2 cloud properties and take only pixels with solar and viewing zenith angle smaller than 80 • into account.
As Eq. 1 is valid for liquid water clouds only, we select liquid water clouds, and reject ice clouds, as indicated by the VIIRS 165 cloud water phase. Around 25-30% of VIIRS pixels are missed due to this filter.

Sun glint in the TROPOMI NO 2 retrieval
The term sun glint refers to particular satellite viewing geometries, under which the ocean acts as a mirror by reflecting sun light directly to the satellite instrument. In the TROPOMI data product pixels that are potentially in sun glint mode are identified based on the combination of their solar and viewing zenith and azimuth angles. The sun glint condition is fulfilled when the 170 scattering angle Θ is smaller than a threshold angle Θ max : with θ and θ 0 the solar and viewing zenith angles and φ and φ 0 the solar and viewing azimuth angles, respectively (see Supplementary Figure S1). For the TROPOMI data products the maximum threshold angle has been set at 30°. Smaller angles are used before, e.g. for SCIAMACHY and GOME-2 (Loots et al., 2017). The TROPOMI algorithm treats the enhanced albedo 175 as a partially cloudy scene with the cloud pressure located at or close to the sea surface.

Relationship between NO x emissions and columns
When studying NO 2 columns to investigate emission trends, the non-linearity of NO x chemistry needs to be taken into account. For example, the lifetime of NO x depends on the background O 3 level, the available sun light and NO x concentrations themselves (Jacob, 1999). We use a (modeled) β factor to express the sensitivity of relative NO 2 column changes to changes 180 in the relative emission strength following the approach in Vinken et al. (2014a) with where ∆E/E represents the imposed relative change in NO x emission flux and ∆N/N the relative change in subsequently simulated tropospheric NO 2 columns. Here we use beta values from (Vinken et al., 2014b) modeled with GEOS-Chem at 0.5°·0.67°and accounting for plume-in-grid chemistry. These beta values have a similar spatial resolution as the spatially 185 averaged TROPOMI NO 2 signals from ships (see Fig. 9). As we are interested in European Seas only, we average β in the area E ∝ L 2 · v 3 as used e.g. in Georgoulias et al. (2020). For the areas and times under study, ship specific data was available only for 50% (Gibraltar) and 70% (Eastern Mediterranean) of the ships.

Results
We start with demonstrating TROPOMI's capabilities to detect shipping NO 2 applying established data selection criteria. Next, we show steps to optimize monitoring of ship emissions making use of sun glint (Sec. 3.2) and recent improvements in the 205 cloud retrieval (Sec. 3.3) and compare the improved TROPOMI data to OMI data in Sec. 3.4. We end with an application of our findings to quantify NO 2 emissions reductions from shipping due to COVID in 2020 in Sec. 3.5.

Detection of NO 2 pollution over European shipping lanes
TROPOMI detects unprecedented spatial detail in shipping NO 2 over busy shipping routes. Fig. 1 shows the summertime mean (May-September 2019) NO 2 columns from TROPOMI and OMI averaged to a common 0.0625°x0.0625°grid as well as  and selected only grid cells over the Mediterranean Sea. This was done to ensure comparable meteorological and chemical conditions. Next, we binned the data by emission strength in bins of 0.05 · 10 −10 kg m −2 s −1 . A reduced major axis regression of all bins with more than 10 entries lead to the correlation coefficients given above. Corresponding scatter plots can be found in Fig. C1. The y-axis intercept of 1.07 (1.05) ·10 15 molec·cm −2 for TROPOMI (OMI) represents the mean background NO 2 230 column over the summertime Mediterranean. Other emission bin sizes lead to slightly different but comparable regression results.
Besides the higher resolution of the TROPOMI instrument, TROPOMI N v,trop thus have a comparable spatial correlation with emission inventories when compared to OMI's. The distinct shipping lanes visible in Fig. 1 and B1 visualize TROPOMI's unprecedented capabilities to detect shipping NO 2 .

Sun glint
For situations of sun glint (see Sec. 2.3) the usually dark ocean appears bright in the TROPOMI data, leading to a strong increase in the effective scene albedo with decreasing scattering angle as shown in Fig. 3(a). Figure 3(b) shows that the increase in scene albedo leads to substantially higher vertical sensitivities, as diagnosed by the averaging kernels (AK) in the operational TROPOMI NO 2 product. The sensitivity increased most in the lowest vertical layer, where the kernel values are 240 on average ≈60% higher for sun glint compared to non sun glint circumstances (0.44 vs 0.28). Increased albedo generally enhances a satellite sensor's sensitivity to NO 2 concentrations in the lower atmosphere (e.g. Eskes and Boersma (2003)), and sun glint scenes have been tentatively used previously to attribute shipping plumes to individual ships in the Mediterranean Sea (Georgoulias et al., 2020). The scene albedo and vertical sensitivity can be further increased by focusing on scenes with low-moderate wind speeds (≈ 245 2 m/s) as wind-induced waves are expected to change the reflectivity. Fig. 4(a) shows the relationship between effective scene albedo and wind speed for scenes with small scattering angles Θ ≤ 15 • . For very low wind speeds the mean scene albedo is almost as small as for non sun glint scenes and smaller than for all other wind speeds. For wind speeds between 1.5 and 2.0 m/s we find an effective scene albedo of almost 0.25, which is approximately double compared to the average for these scattering angles and more than 5 times as high as for non sun glint scenes. For higher wind speeds the scene albedo decreases to around 250 0.10. In Fig. 4(b) the effect on the averaging kernel profile is shown. As expected low wind speeds lead to the smallest AK in the lower atmosphere, whereas wind speeds between 1.5 and 2.0 m/s show the largest AKs close to the sea surface. This relationship can be understood in terms of wind-induced sea surface roughness (Cox and Munk, 1956). Both very low and strong winds limit the probability that a scattering angle Θ ≤ Θ max leads to sun glint effects at the sensor: For very low wind speeds, the sea surface is effectively flat, leading to sun glint only for very small scattering angles Θ Θ max , whereas for  Table 2.

Cloud properties 270
Here we evaluate TROPOMI's capability to retrieve realistic cloud parameters retrieved from the 405-465 nm continuum reflectances and effective cloud pressures from the O 2 -A band (Table 1), addressing recent improvements in the FRESCO+ algorithm to avoid overestimated cloud pressures . These improvements in cloud retrievals lead to  and super sun glint (Θ ≤ 20°) over the Central Mediterranean north of Libya in June-July-August 2018 (see gray rectangle in Fig. 1(c)). an inconsistency in the tropospheric NO 2 column record.

Cloud fractions 275
We find that improved TROPOMI cloud fractions are of sufficient quality to support the TROPOMI NO 2 AMF calculation.
They show good correlation to independent data such as from OMI and VIIRS. TROPOMI v1.2 and v2.1 cloud fractions are very similar with the new v2.1 cloud fractions being slightly smaller. More details can be found in Appendix D1.

Cloud pressure
FRESCO+wide cloud pressures are a clear improvement over the FRESCO+ data used in v1.2/1.3. Figure 6 shows a compari-280 son of gridded, co-sampled cloud pressure distributions from TROPOMI v1.2 (FRESCO+), TROPOMI v2.1 (FRESCO+wide), OMI QA4ECV and VIIRS over the Bay of Biscay between 1 and 7 July 2018. As expected, the improved TROPOMI v2.1 cloud  to VIIRS with a difference of 2 hPa relative to VIIRS. We find similar agreement between TROPOMI and independent data over the Mediterrenean Sea and northwestern Europe as shown in Table D2. FRESCO+wide cloud pressures agree best but remain higher than VIIRS in the median (both FRESCO cloud pressure distributions show a larger tail towards low pressures compared to VIIRS, possibly caused by filtering for liquid water clouds in VIIRS) . This is in line with expectations as VIIRS's infrared cloud retrieval is mostly sensitive to the cloud top (Platnick et al., 2017), whereas FRESCO's O 2 -A band retrieval is 290 more sensitive to the center of a cloud (e.g. Sneep et al. (2008)). Around 25-30% of VIIRS cloud retrievals in the areas studied here are ice water clouds and therefore not included in the analysis. As these clouds appear at higher altitudes, improved cloud pressures have only little influence on the NO 2 columns (see Sec. 3.3.3).

Effect of improved cloud pressure on TROPOMI NO 2 columns
The improved cloud pressures lead to increases of NO 2 columns of up to 40% depending on area and season. The left panel of 295 Fig. 7 shows the change in tropospheric NO 2 columns as a function of cloud pressure over the Bay of Biscay and northwestern Europe in Summer. We see that NO 2 columns increase most for locations that had the highest original v1.2 cloud pressures, and that the improvements are strongest when cloud pressures are reduced most (light blue dots). The increase over the Bay of Biscay is smaller (up to 0.1 · 10 15 molec·cm −2 ) than over northwestern Europe (up to 1.0 · 10 15 molec·cm −2 ), reflecting the higher pollution levels over the mainland. We see similar patterns with stronger improvements in Winter, as shown in the 300 right panel of Fig. 7  most polluted seas such as the English Channel and shipping lanes. We find a stronger impact of the improved cloud pressures in the winter season, reflecting that NO 2 pollution is confined in a thinner marine boundary layer in that season.

Comparison of TROPOMI and OMI NO 2 columns in shipping lanes
TROPOMI detects a more pronounced and narrower region of ship NO 2 pollution than OMI. On average, TROPOMI v2.1p 320 detects 45% higher peak NO 2 values than OMI. TROPOMI data allow the attribution of 14% more NO 2 to shipping lane enhancements, over 23% narrower shipping lanes. To quantitatively compare TROPOMI's capability to detect NO 2 over shipping lanes under different measurements conditions and compare it to OMI's, we created average NO 2 cross sections over busy shipping lanes. We studied NO 2 enhancements in summer 2019 (June-August) over shipping lanes in the Bay of Biscay, from Sicily to the Suez Canal, and East of Gibraltar, the regions visually defined in Fig. 1(c). First, we defined the location of the shipping 325 lanes according to the emission data shown in Fig. 1(c). Then, we calculated the average NO 2 columns along the shipping lane and parallel to it, taking care to exclude NO 2 columns measured over land. In that way we created an average cross section of NO 2 over shipping lanes. In the last step, we performed a background correction by subtracting a linear NO 2 background to isolate the NO 2 enhancements caused by shipping. The orbital data was gridded to regular grids of 0.0625°x0.0625°and 0.125°x0.125°resolution for TROPOMI and OMI, respectively. For TROPOMI only pixels with qa > 0.75 were taken into 330 account. For OMI, a consistent filtering was applied, including maximal solar and viewing zenith angles of 80°and maximal cloud radiance fractions of 0.5. The resulting cross sections are shown in Fig. 9. Table 4 summarizes the peak value, the area under the curve (i.e. the total NO 2 attributed to shipping) and the full width at half maximum (FWHM) for the three shipping lanes. It should be noted that the grid used for OMI is 2x coarser than the one used for TROPOMI. Gridding TROPOMI to the coarser grid used for OMI only changes the results slightly, indicating that the improved spatial resolution of TROPOMI   As already seen in Fig. 8, the v2.1p data set shows slightly higher NO 2 compared to the TROPOMI v1.2/v1.3 data, especially in the center of the lane while background NO 2 is less affected by the correction. The impact of the DNN is larger in winter than in summer as discussed before.

340
For the Bay of Biscay it is also apparent that the NO 2 peak is shifted to the East for all data sets. As the location is defined by an emission inventory based on AIS data (and therefore real ship location), this is likely an effect of dominant westerly winds.
We conclude that TROPOMI provides a significant improvement for the detection of shipping NO 2 with sharper and more pronounced shipping lanes in seasonal averages. The improved v2.1p TROPOMI data increase the signal further.

345
Emissions proxies derived from AIS data and from TROPOMI NO 2 suggest emission reductions from shipping in 2020 compared to 2019 as depicted in Fig. 10 (c & f). While in the first three months of 2020 the ship emissions were generally higher compared to 2019, both emission proxies show reductions starting in April and lasting until the end of the year. This reduction can be linked to the COVID-19 pandemic, which led to economic lockdowns in many countries of the world. Europe had its most stringent measures in Spring and Autumn 2020.

350
We created daily 0.0625·0.0625°maps of TROPOMI data, using v2.1p NO 2 columns as described in Section 3.3.3 with qa ≥ 0.75. We calculate the area under the cross section as a measure for shipping NO 2 for monthly mean NO 2 columns for the shipping lanes of Gibraltar and Mediterranean defined in Fig. 1(c). Monthly TROPOMI shipping NO 2 for 2019 and 2020 can be seen in Supplementary Figure S2(c). Figure 10(d) shows the relative change in shipping NO 2 from 2019 to 2020 in the Strait of Gibraltar. Using β values and the approach described Section 2.4 and shown in Fig. 10(e), we arrive at the TROPOMI 355 based relative change in emission changes shown in Fig. 10(f).
The uncertainty in our top-down NO x emission changes follows from : (i) The sensitivity of TROPOMI shipping NO 2 to the area of study (σ area = 5%), (ii) the inter-year differences on monthly averaged NO 2 columns over the areas of study caused by meteorology and (iii) the combined spatial and temporal spread of β caused by differences in the chemical regime caused e.g. by differences in atmospheric composition and radiation (σ β = 0.15). Fig. 10(d) and Fig. F1(d) show σ area , in panels (e) 360 uncertainties (ii) and (iii) are used while for panels (e) a full error propagation of all uncertainties listed above was performed.
A full discussion on the uncertainty estimates can be found in Supplement 4.
Additionally, we used AIS data to calculate an AIS based emission proxy as described in Sec. 2.5. We filtered for days with TROPOMI coverage of at least 50% of each study area. AIS data indicates that the number of ships passing per month through the Strait of Gibraltar has reduced from March 2020 onwards relative to 2019 ( Fig. 10(a) and Supplementary Figure S2(a)).

365
The average speed of the ships passing through the shipping lanes is lower between May-September 2020 compared to the same period in 2019 as well ( Fig. 10(b) and Supplementary Figure S2(b)). This is in agreement with a study by (Millefiori et al., 2020)  TROPOMI data and the simplified emission proxy. However, this is not expected to lead to a systematic bias.
We used tropospheric NO 2 column observations from the TROPOMI sensor to optimally monitor ship NO 2 pollution and study 390 the changes in ship NO x emissions over European seas in 2019-2020. Satellite observations of tropospheric NO 2 columns provide valuable information on ship air pollution over open seas, which can be used to inform compliance monitoring by flag states and national authorities. We evaluated the high-resolution TROPOMI NO 2 retrievals for its potential to better detect ship NO 2 pollution. In European waters alone, TROPOMI finds 6 new lanes with enhanced NO 2 ranging from the Aegean Sea to the Skagerrak between Denmark and Norway, which are not detected by OMI, and which have not previously been reported in 395 the literature. These newly found lanes of pollution coincide with busy sailing routes and bottom-up emission proxies.
To better understand the recent detection of an individual ship's NO 2 plume under conditions of sun glint, we examined how sun glint viewing geometries affect subsequent steps in the TROPOMI retrieval procedure. We find that sun glint drives higher apparent scene reflectivity, which enhances the signal strength from spectral fitting of NO 2 columns along the average light path by 20-30% over clear-sky shipping lanes. In such situations, the vertical sensitivity to NO 2 within the marine boundary 400 layer increases by up to 60%. This effect is especially strong when sea surface wind speeds are low, but non-zero. When winds are strong, the wash causes sunlight to be reflected in other directions than directly towards the satellite, leading to little gain in vertical sensitivity. We find that the TROPOMI NO 2 algorithm accounts for these effects, so that data within and outside of sun glint geometries can be used with confidence. Nevertheless, our work clearly indicates that optimal spectral fitting can be accomplished for small scattering angles (<15 • ) and sea surface wind speeds of 1.5-3 m/s. Although selecting a subset 405 fulfilling these sampling criteria reduces the amount of available data sharply, our findings indicate that sun glint conditions are beneficial for quantifying previously undetectable small NO x emissions sources over open sea, and holding promise for also detecting other trace gases with UV/Vis satellite instruments over water, where surface reflectivity and vertical sensitivity is generally small.
In November 2020, KNMI implemented an improved FRESCO+ cloud retrieval called FRESCO+wide in the operational 410 TROPOMI NO 2 algorithm. We find here that this new FRESCO+wide cloud retrieval provides some 50 hPa lower cloud pressures which agree better with coinciding cloud top heights from the VIIRS sensor than the standard FRESCO+. We show that the improved cloud pressures lead to a more realistic description of vertical sensitivities in the TROPOMI NO 2 algorithm, and at least partly address the known low bias in the tropospheric NO 2 product prior to November 2020, thus not only solving a known issue in the TROPOMI NO 2 retrieval but also increasing signal strength. We then trained a neural network on a lim-415 ited data set of simultaneously available standard and improved cloud and NO 2 retrievals. Based on 4 different training sets, the neural network learned the statistical relationship between standard FRESCO+ cloud pressures and other parameters and the new tropospheric NO 2 columns. We used the neural network to predict updated NO 2 columns for the entire 2019-2020 TROPOMI NO 2 record. The neural network predicts a general increase in tropospheric NO 2 columns. Increases are particularly strong (up to 4·10 15 molec·cm −2 ) in the most polluted regions of Europe in wintertime. Our predicted (v2.1p) TROPOMI We compared changes in our v2.1p TROPOMI NO 2 columns between 2019 and 2020 to changes in the number of ships, their speed and their size obtained from AIS data in the main European traffic lanes. From April 2020 onwards, TROPOMI observes 25% less NO 2 pollution than in the year before, in step with a 10% reduction in the number of ships and a 5% speed reduction 425 relative to 2019. Accounting for non-linearity in local NO x chemistry, we infer an average 20% reduction in top-down NO x emissions in the Strait of Gibraltar from ships during months in which COVID-measures were in force in Europe, and global mobility decreased as a result of the pandemic. For future research, a full chemical transport modelling of AIS-based emissions and strict co-sampling of AIS and TROPOMI data can help understanding the observed differences in top-down and bottom-up emission changes and reduce the error margins. 430 We showed that TROPOMI is a superior instrument to analyze relatively small enhancements in NO 2 pollution over dark   where N s,trop is the tropospheric slant column density which can be calculated from the TROPOMI files using where M , N s , and N v mean air mass factor, slant column density, and vertical column density, respectively. The subscripts trop, tot, and strat indicate troposhperic, total, and stratospheric columns, respectively. M geo can be calculated using the solar zenith angle θ and the viewing zenith angle θ 0 as M geo = 1/cos(θ) + 1/cos(θ 0 ). The resulting tropospheric column is shown in Fig. B1.   v2.1 shows high correlation (R 2 =0.66) and somewhat lower cloud fractions (-11%) compared to the co-sampled effective VIIRS cloud fractions. TROPOMI cloud fractions are especially lower for partly cloud-covered scenes, possibly resulting from biased surface albedo's assumed in the TROPOMI retrieval (from the GOME-2 climatology at 0.5 • resolution, see Table 1).
We find similar high correlation and small differences between TROPOMI and independent data over the Mediterrenean Sea and Northwestern Europe as shown in Table D1. An artificial Neural Network allows us to predict v2.1 columns for the full TROPOMI mission period up to December 2020.
We trained the artificial Deep Neural Network (DNN) using the Python package Keras (Chollet and others, 2015) with three 470 hidden layers. We divided the combined v1.2 and v2.1 data sets in 3 random subsets for training (60%), validation (20%), and testing (20%). The input parameters to predict TROPOMI (pseudo) v2.1 NO 2 columns are N v,v1.2 , M trop , f cl , p cl , all viewing geometry parameters, surface albedo, and the qa value (all from v1.2). The DNN was then trained to minimize the mean absolute difference between the predicted and actually retrieved v2.1 NO 2 columns from the training set. This means our prediction does not use FRESCO+wide cloud pressures for dates outside the training set period. Rather, the DNN has been 475 trained to predict new NO 2 columns based on the old FRESCO+ cloud pressures and other parameters. Our DNN application succeeds in reducing the mean difference between the predicted and retrieved v2.1 NO 2 columns to < 0.01 · 10 15 molec·cm −2 (original v2.1 -v1.2 mean difference was 0.12 · 10 15 molec·cm −2 ) over the 3 areas of study during the 4 periods, suggesting considerable skill in the DNN approach. Our improved data set consists of the original L2 TROPOMI NETCDF files with the predicted change in troposhperic NO 2 columns as additional variable.

480
To show that DNN is capable of capturing seasonal variations in NO 2 corrections and, more broadly, that we can use a generic DNN to correct historic TROPOMI v1.2 data, we train a DNN based on 3 seasons (Summer, Winter, and Spring) and tested its predicted NO 2 columns against actually retrieved v2.1 data in Autumn. This analysis is done for the 3 testing areas defined in 3.3. After application of DNN, the mean discrepancy between predicted and retrieved v2.1 NO 2 columns reduces to < 0.01 · 10 15 molec·cm −2 (original mean discrepancy: 0.09 · 10 15 molec·cm −2 ) and R 2 improved from 0.82 to 0.97.

485
Appendix F: Ship NO x emission reductions during the COVID-19 pandemic Figure F1. shipping. Error bars represent the propagated uncertainties in TROPOMI shipping NO2, beta and differences in meteorological conditions between 2019 and 2020 (see discussion in the text).