Comparing OMI-based and EPA AQS in situ NO2 trends: Towards understanding surface NOx emission changes

With the improved spatial resolution than earlier instruments and more than ten years of service, tropospheric NO2 retrievals from the Ozone Monitoring Instrument (OMI) have led to many influential studies on the relationships between socioeconomic activities and NOx emissions. Previous studies have shown that the OMI NO2 data show different relative trends compared to in situ measurements. However, the sources of the discrepancies need further investigations. This study 15 focuses on how to appropriately compare relative trends derived from OMI and in situ measurements. We retrieve OMI tropospheric NO2 vertical column densities (VCDs) and obtain the NO2 seasonal trends over the United States, which are compared with coincident in situ surface NO2 measurements from the Air Quality System (AQS) network. The Mann-Kendall method with the Sen’s slope estimator is applied to derive the NO2 seasonal and annual trends for four regions at coincident sites during 2005-2014. The OMI-based NO2 seasonal relative decreasing trends are generally biased low compared to the in 20 situ trends by up to 3.7% yr, except for the underestimation in the Midwest and Northeast during Dec-Jan-Feb (DJF). We improve the OMI retrievals for trend analysis by removing the ocean trend, using the MODerate-resolution Imaging Spectroradiometer (MODIS) albedo data in air mass factor (AMF) calculation. We apply a lightning flash filter to exclude lightning affected data to make proper comparisons. These data processing procedures result in close agreement (within 0.3% yr) between in situ and OMI-based NO2 regional annual relative trends. The remaining discrepancies may result from inherent 25 difference between trends of NO2 tropospheric VCDs and surface concentrations, different spatial sampling of the measurements, chemical non-linearity, and tropospheric NO2 profile changes. We recommend future studies to apply these procedures (ocean trend removal and MODIS albedo update) to ensure the quality of satellite-based NO2 trend analysis and apply the lightning filter in studying surface NOx emission changes using satellite observations and in comparison with the trends derived from in situ NO2 measurements. With these data processing procedures, we derive OMI-based NO2 regional 30 annual relative trends using all available data for the West (-2.0%±0.3 yr), the Midwest (-1.8%±0.4 yr), the Northeast (3.1%±0.5 yr), and the South (-0.9%±0.3 yr). The OMI-based annual mean trend over the contiguous United States is -

1.5%±0.2 yr -1 . It is a factor of 2 lower than that of the AQS in situ data (-3.9%±0.4 yr -1 ); the difference is mainly due to the fact that the locations of AQS sites are concentrated in urban and suburban regions.

Introduction
Nitrogen dioxide (NO 2 ) is an air pollutant. At high concentrations, it aggravates respiratory diseases and can lead to acid rain formation (e.g., Lamsal et al., 2015). It is also a key player to produce another pollutant, ozone (O 3 ), through photochemical 5 reactions in the presence of Volatile Organic Compounds (VOCs) under sunlight. Tropospheric NO 2 is emitted both anthropogenically and naturally (e.g., Gu et al., 2016). Anthropogenic fossil fuel combustions and biomass burnings emit mostly nitrogen monoxide (NO) under high temperature, which is later oxidized by O 3 into NO 2 . Major natural NO 2 sources include lightning and soils.
Surface NO 2 concentrations are regulated by the U.S. Environmental Protection Agency (EPA) through the National 10 Ambient Air Quality Standards (NAAQS). NO 2 is measured routinely at the EPA Air Quality System (AQS) sites (Demerjian, 2000). Although the AQS network continually provides valuable hourly NO 2 measurements, AQS sites are mostly located in urban and suburban regions, leaving large regions of rural areas unmonitored. Satellite data provide a better spatial coverage than the in situ measurements.
Several satellites were launched to monitor tropospheric NO 2 vertical column densities (VCDs), such as the SCanning 15 Imaging Absorption spectroMeter for Atmospheric CHartographY (SCIAMACHY), the Global Ozone Monitoring Experiment-2 (GOME-2), and the Ozone Monitoring Instrument (OMI). For trend analysis, the tropospheric NO 2 products from OMI surpass the others for a relatively high spatial resolution and over one decade of continuous operation (Boersma et al., 2004;Boersma et al., 2011). Thus, OMI NO 2 retrievals are widely applied in NO 2 and NOx emission trend studies (e.g., Lin et al., 2010Lin et al., , 2011Castellanos et al., 2012;Russell et al., 2012;Gu et al., 2013;Lamsal et al., 2015;Tong 20 et al., 2015;Cui et al., 2016;Duncan et al., 2016;de Foy et al., 2016ade Foy et al., , 2016bKrotkov et al., 2016;Liu et al., 2017).  reported that the reduction rates calculated from OMI NO 2 VCDs and AQS surface NO 2 data at eight cities were -35% and -38% from 2005 to 2012, respectively. Lamsal et al. (2015) also found the divergence between the annual trends inferred from the two datasets, i.e. -4.8% yr -1 vs -3.7% yr -1 during 2005-2008, and -1.2% yr -1 vs -2.1% yr -1 during 2010-2013. There are several potential factors attributing to the discrepancies between trends from satellite and ground-based 25 measurements: interferences by the oxidation products of NOx from the chemiluminescent instruments (Lamsal et al., 2008(Lamsal et al., , 2014, the differences of sampling time between OMI (~13:30 local time) and AQS (hourly) measurements , a high sensitivity of NO 2 VCDs to high-altitude NO 2 in contrast to the high sensitivity of surface NO 2 concentrations to surface NOx emissions (Duncan et al., 2013;Lamsal et al., 2015), spatial representativeness of satellite pixels , and high uncertainties of satellite retrievals in clean regions . 30 To understand how various factors and the retrieval procedure affect the differences between the OMI derived trends and those derived from the surface AQS measurements, we utilize a regional 3-D chemistry transport model (CTM), a radiative transfer model (RTM), and the Mann-Kendall method (Mann, 1945;Kendall, 1948) to calculate OMI-based NO 2 seasonal relative trends during Dec-Jan-Feb (DJF), Mar-Apr-May (MAM), Jun-Jul-Aug (JJA), and Sept-Oct-Nov (SON) (Section 2).
We find that two procedures are essential to ensure the quality of trend analysis using OMI tropospheric NO 2 VCDs, including the ocean trend removal and the MODerate-resolution Imaging Spectroradiometer (MODIS) albedo update in calculating the air mass factors (AMFs). The lightning filter (Section 3.1) is necessary for comparing OMI-based and in situ AQS NO 2 trends. 5 With these procedures implemented, the differences between OMI-based and AQS in situ annual relative trends are within 0.3% yr -1 of coincident measurements for all the four regions. Finally, we estimate the OMI-based annual relative trends across the nation in Section 3.2. Conclusions are given in Section 4.

EPA AQS surface NO2 measurements 10
The in situ surface NO 2 measurements from the U.S. EPA AQS network are used in this research. Sites with a continuous measurement gap of more than 50 days are removed and the observations of 140 remaining cites are used (Fig. 1). The AQS chemiluminescent analyzers are equipped with molybdenum converters to measure ambient NO 2 concentrations. These analyzers are known to have high biases, since the converters are not NO 2 specific and they measure some fractions of peroxyacetyl nitrate, nitric acid and organic nitrates (Demerjian, 2000;Lamsal et al., 2008). In addition to chemiluminescent 15 analyzers, several NO 2 specific photolytic instruments were deployed since 2013. By utilizing the data from both chemiluminescent and photolytic measurements at coincident sites during the overpassing time of OMI, we calculate the observed NO 2 concentration ratio between both measurements in Fig. S1 in the Supplement. The ratio peaks at 2.3 in June and decreases to 1.3 in November, indicating that the chemiluminescent analyzers overestimate by 27%-132% than photolytic instruments. This finding is in agreement with Lamsal et al. (2008). We correct the chemiluminescent NO 2 data by the observed 20 ratio assuming that the inter-annual change is small and the high bias of the chemiluminescent measurements is identical at all sites. This correction may contribute to the differences between in situ and OMI based absolute NO 2 trends but do not significantly affect the relative trends (since the correction is canceled out in computing relative trends). In this study, we only examine the relative trends and therefore the analysis results are not affected by the uncertainties in the in situ NO 2 measurement corrections. 25

REAM model
We use a 3-D Regional chEmical trAnsport Model (REAM) in the simulation of NO 2 profiles. REAM has widely been used in atmospheric NO 2 studies, including vertical transport (Choi et al., 2005;Zhao et al., 2009a;Zhang et al., 2016a), emission inversions (Zhao et al., 2009b;Yang et al., 2011;Gu et al., 2013Gu et al., , 2014Gu et al., , 2016, and regional and seasonal variations (Choi et al., 2008a(Choi et al., , 2008b. The model has a horizontal resolution of 36 km with 30 vertical layers in the troposphere, 5 vertical layers 30 in the stratosphere, and a model top of 10 hpa. In this study, the domain of REAM is about 400 km larger on each side than the contiguous United States (CONUS). Meteorology inputs driving transport process are simulated by the Weather Research and Forecasting model (WRF) assimilations constrained by National Centers for Environmental Prediction Climate Forecast System Reanalysis (NCEP CFSR, Saha et al., 2010) 6-hourly products. The KF-eta scheme is used for sub-grid convective transport in WRF (Kain and Fritsch,1993). We run the WRF model with the same resolution as in REAM but with a domain 10 grids larger on each side than that of REAM. REAM updates most of the meteorology inputs every 30 minutes while those 5 related to convective transport and lightning parameterization are updated every 5 minutes. The chemistry mechanism expands that of a global CTM GEOS-Chem (V9-02) with aromatics chemistry (Bey et al., 2001;Liu et al., 2010Liu et al., , 2012aLiu et al., , 2012b. For consistency, the GEOS-Chem (V9-02) simulation with 2°× 2.5° resolution is used to generate initial and boundary conditions for chemical tracers.  Guenther et al., 2012). We parameterize lightning emitted NOx as a function of convective mass flux and Convective Available Potential Energy (CAPE) (Choi et al., 2005).
NOx production per flash is set to 250 moles NO per flash, and the emissions are distributed vertically following the C-shaped profiles by Pickering et al. (1998). For recent model evaluations of REAM with observations, we refer readers to Zhang et al. 15 (2016aZhang et al. 15 ( , 2016b, Cheng et al. (2017), and .

OMI-based NO2 VCDs
We retrieve the tropospheric NO 2 VCDs using the tropospheric slant column densities (SCDs, without destriping) from the Royal Dutch Meteorological Institute (KNMI) Dutch OMI NO 2 product (DOMINO v2, Boersma et al., 2011). OMI onboard the Aura satellite was launched in July 2004 and is still active. OMI overpasses the equator at about 13:30 Local Time (LT) 20 and obtains global coverage with a 2600 km viewing swath spanning 60 rows. It has a ground level spatial resolution up to 13 km x 24 km (at nadir). The spatial extent of the OMI pixels will not affect our analysis as we focus on regional trend analysis.
SCDs are retrieved by matching a modeled spectrum to an observed top-of-atmosphere reflectance with the Differential Optical Absorption Spectroscopy (DOAS) technique within a fitting window of 405-465nm. The stratospheric portion of SCDs are estimated and subsequently removed with a global CTM TM4 with stratospheric ozone assimilation (Dirksen et al., 2011). 25 Deriving tropospheric VCDs from the remaining tropospheric SCDs requires the calculation of AMFs. Being an optically thin gas, tropospheric AMF for NO 2 can be calculated from AMF for each vertical layer ( ) weighted by NO 2 partial VCDs at the corresponding layer ( ) (Boersma et al., 2004), as shown in equation (1).
As the vertical distribution of NO 2 is usually unknown, we typically substitute by an a priori profile ( , ) from a 30 CTM.
is the sensitivity of NO 2 SCD to VCD at a given altitude (Eskes and Boersma, 2003), and is computed using the Double Adding KNMI (DAK) RTM (Boersma et al., 2011). As a result, the retrieved tropospheric NO 2 VCD computation depends on the a priori NO 2 vertical profile, the surface reflectance, the surface pressure, the temperature profile, and the viewing geometry (Boersma et al., 2011). Previous studies have addressed the sources of uncertainties in NO 2 retrievals, including surface reflectance resolutions (Russell et al., 2011;Lin et al., 2014), lightning NOx (Choi et al., 2005a;Martin et al., 2007;Bucsela et al., 2010), a priori CTM uncertainties (Russell et al., 2011;Heckel et al., 2011;Lin et al., 2012;Laughner 5 et al., 2016), surface pressure and reflectance anisotropy in rugged terrain (Zhou et al., 2009), cloud and aerosol radiance (Lin et al., 2014(Lin et al., , 2015, and boundary layer dynamics (Zhang et al., 2016a). In this study, we find that the first two factors are essential in NO 2 VCD trend analysis and we will discuss these in the following sections.
AMFs are derived using the pre-computed altitude-dependent AMF lookup table, which is generated by the DAK RTM. 10 We use the NO 2 profiles from REAM, temperature and pressure from CSFR, viewing geometry and cloud information from DOMINO v2 product. We use the REAM results of 2010 to avoid the uncertainty introduced by yearly variation of NO 2 profiles. The yearly variations of meteorology and anthropogenic emission changes have little impact in polluted areas on trend analysis results using OMI data . We use the surface reflectance from DOMINO v2 product as default (Kleipool et al., 2008), and update it using a surface reflectance product with a higher temporal resolution (Section 15 2.3.2). The derived tropospheric NO 2 VCD relative trends with default surface reflectance are referred as "Standard".

Ocean trend removal
For trend and other analyses of OMI tropospheric VCDs, the data of anomalous pixels must be removed. with our understanding of the row anomaly (Schenkeveld et al., 2017). In addition, the data of coarse spatial resolution from rows 1-5 and rows 56-60 are also excluded, as suggested by Lamsal et al. (2015). The selection of rows 6-25 used in this research is stricter than the data flags in the DOMINO v2 product. Furthermore, we exclude OMI data with cloud fraction > 0.3 to minimize retrieval uncertainties due to clouds and aerosols (Boersma et al., 2011;Lin et al., 2014). 2b) with minimal marine traffic. This trend may reflect the inaccurately simulated stratospheric SCDs (A. Richter, personal communication, 2018) or the increase in the magnitude of the stripes (step-wise SCD variability from one row to another) in time, which originates from the use of a constant (2005-averaged) solar irradiance reference spectrum in the DOAS spectral fits throughout the mission and the weak increase of noise in the OMI radiance measurements (F. Boersma, personal communication, 2017;Zara et al., 2018). Fig. 2a shows that there is a positive annual trend of 1.750.45x10 13 molecules cm -2 30 yr -1 . The ocean trend is insensitive to the region selection in the remote North Pacific (varies within 10%). We only analyze OMI tropospheric column trends over the CONUS for grid cells with 2005-2014 averaged VCDs > 1x10 15 molecules cm -2 , which tends to minimize the effect of the background noise. However, removing this background ocean (absolute) trend has a non-negligible effect in reducing the OMI relative trend (Fig. 1). We treat this trend as a systematic bias. We calculate the contribution from the ocean (absolute) annual trend to SCDs for each year and subtract it from OMI tropospheric NO 2 SCDs uniformly in the following analysis. Since the origin of this trend is not yet clear, the ocean trend removal method may need updates in future studies. We refer to such derived (relative) trend data as "Ocean". An alternative method is to subtract monthly SCD trends of the remote ocean region (Fig. S2 in the Supplement) from the OMI tropospheric SCD data. Although 5 the end results (Fig. S3 in the Supplement) are essentially the same as the annual trend removal method, noises are added to the SCD data, making it more difficult to understand the effects of the MODIS albedo update and the lightning filter (next sections). We therefore choose to use the (absolute) annual trend removal method here.

MODIS albedo update
The albedo data used to calculate the in "Standard" and "Ocean" versions of trend analysis are from the DOMINO v2 10 products, which are the climatology of averaged OMI measurements during 2005-2009 with a spatial resolution of 0.5°×0.5° (Kleipool et al., 2008) and is valid for 440 nm. We recalculate the using the MODIS 16-day MCD43B3 albedo product with 1km spatial resolution, which combines data from both MODIS onboard Aqua and Terra satellites (Schaaf et al., 2002;Tang and Zhang, 2007). Aqua and Terra have an equatorial overpassing time of 13:30 LT and 10:30 LT, respectively. The band 3 (459nm-479nm) is used to match the NO 2 fitting window (405nm-465nm). The albedo is spatially integrated to the 15 geometry of OMI pixels and is temporally interpolated to match OMI overpassing dates. In order to maintain the consistency of the DOMINO retrieval algorithm (Boersma et al., 2011), we only use the MODIS data to improve the temporal variations of albedo data used in the retrieval. We scale the MODIS albedo data such that the mean albedo during 2005-2009 is the same as the OMI climatology at 0.5°×0.5°. We recalculate OMI tropospheric VCDs using the MODIS albedo data as described. We recalculate the relative OMI trend and remove the ocean (absolute) trend (Section 2.3.1). We refer to this version of OMI 20 relative trend data as "MODIS".

Lightning event filter
Over North America, lightning is a major source of NOx in the free troposphere and its simulations in CTMs are uncertain (e.g., Zhao et al., 2009a;Luo et al., 2017). The large temporospatial variations of lightning NOx make it difficult to compute satellite based NO 2 trends by changing the vertical distributions of NO 2 affecting the AMF calculation (e.g., Choi et al., 2008b;25 Lamsal et al., 2010) and the SCD values. Furthermore, accompanying lightning occurrences, the presence of cloud significantly affects the lifetime of NOx and the partitioning of NO 2 to NO in daytime and convective transport exports NOx from the surface and boundary layer to free troposphere, changing surface and column NO 2 concentrations. Given the difficulty to simulate lightning NOx accurately across different years and meteorological effects (vertical mixing) of lightning, we use a lightning filter to remove potential effects of lightning NOx on the basis of the flash rate observations of cloud-to-ground (CG) 30 lightning flash data detected by the National Lightning Detection Network TM (NLDN) (Cummins and Murphy, 2009;Rudlosky and Fuelberg, 2010). NLDN only reports the ground point of a CG lightning flash, while the CG lightning flash can extend horizontally to tens of kilometers. A CG lightning flash can affect the NO 2 retrievals not only in the model grid cell where the CG lightning is located but also the nearby model grid cells. The atmospheric lifetime of NOx in the free troposphere can be up to 1 week. Therefore, we exclude the OMI NO 2 data within a radius of 90 km of the NLDN-reported CG lightning location (about two model grid cells around the grid cell where the CG lightning is located) for a period of 72 hours after the lightning occurrence. Since lightning usually occur along the track of a thunderstorm, the 90 km radius is more a constraint on lightning 5 NOx effects across the track. The extended period of 72 hours is to ensure that we exclude data affected by lightning NOx.  (Fig. 3), large amounts of lightning NOx can be produced by high flash ratios of severe thunderstorms and they can be transported northward from the South to the Northeast (Choi et al., 2005). We therefore further filter OMI NO 2 data in the 10 Northeast on the basis of CG lightning flash rates in the South. If the average CG flash rate in the South exceeds the 95 th percentile value of the NLDN observations, which is 0.035 flash km -2 day -1 (Fig. S4 in the Supplement), we exclude in the analysis the Northeast OMI data in the following 72 hours. Excluding the OMI data based on CG lightning data implicitly removes the data affected by cloud-to-cloud lightning collocated with CG lightning. The lightning filter removes about 2%, 27%, 20%, and 19% of OMI data, which are coincident with AQS data, for the West, the Midwest, the Northeast and the 15 South, respectively. We refer to this version of OMI relative trend data as "Lightning filter".

Results and discussion
We group the analysis results into different regions: (a) West, (b) Midwest, (c) Northeast, and (d) South (Fig. 1), following the regional divisions by the United States Census Bureau. To make a fair comparison between the in situ and OMI-based trends, we only use spatially and temporally coincident in situ and OMI NO 2 observations in Section 3.1. The AQS data are temporally 20 interpolated based on the overpassing time of the available OMI pixels which cover the corresponding AQS sites. Similarly, only OMI pixels covering the corresponding available AQS sites are used. The data from each dataset are then aggregated and averaged on a regional basis into time series to calculate the regional trends.
We apply the Mann-Kendall method with the Sen's slope estimator to calculate the relative trend of NO 2 for each season, i.e. DJF, MAM, JJA, and SON, during 2005-2014. We compute the uncertainties of the trends with the 95 th percentile confidence 25 intervals using the Mann-Kendall method. Note that when we compare in situ and OMI-based trends, the lightning filter also removes in situ NO 2 data, which are coincident with the OMI NO 2 data affected by lightning. This leads to slightly different in situ NO 2 trends between Fig. 4 and Fig. 6 (Section 3.2.3). We first compute the trends using the "Standard" OMI VCD data.
The ocean trend removal, MODIS albedo update, and lightning filter are then added in sequence to compute three different OMI-based NO 2 trends (in addition to "Standard") to compare to the AQS in situ results. A subtlety in the comparison is that 30 the coincident data change when the lightning filter is applied. As a result, the AQS in situ results in this set of comparison differ from those in the other three sets. Fig. 4 shows that both AQS in situ and "Standard" OMI-based seasonal relative trends are negative for all seasons across the regions. OMI-based trends generally underestimate the decreasing trends by up to 3.7% yr -1 (the absolute difference between relative trends) except for the large overestimation in the Midwest and the Northeast regions during DJF. The overestimates in these two regions are 3.0% yr -1 and 1.1% yr -1 , respectively. On average, the differences between OMI-based and in situ 5 seasonal relative trends are 1.6% yr -1 , -0.3% yr -1 , 1.0% yr -1 , and 1.4% yr -1 for the West, the Midwest, the Northeast, and the South regions, respectively. Note that the relative trends are calculated using coincident measurements for the comparisons.

In situ and "Standard" OMI-based trends
The focus of this work is to examine if the differences between AQS in situ and OMI-based trends can be reduced.

Improvement due to ocean trend correction
After removing the ocean trend as discussed in Section 2.3.1, the OMI-based NO 2 decreasing trends are more pronounced as 10 shown in Fig. 4 ("Ocean", blue line) by 0.1-0.9% yr -1 . The regional relative trends have different sensitivities to the ocean trend removal due to different tropospheric VCDs levels. In general, the discrepancies between OMI-based and in situ trends are reduced except for the Midwest and the Northeast regions during DJF, which are already biased low. The averaged differences between OMI-based and in situ seasonal relative trends for the West, the Midwest, the Northeast, and the South regions are 1.2% yr -1 , -1.1% yr -1 , 0.4% yr -1 , and 1.0% yr -1 . Only in the Midwest region, removing the ocean trend enlarges the 15 difference due to the large winter bias.

Improvement due to MODIS albedo update
The adoption of the up-to-date MODIS albedo (Section 2.3.2) greatly reduces the difference of relative trends in the Midwest during DJF from -3.6% yr -1 ("Ocean") to 1.3% yr -1 ("MODIS"), the improvement of DJF trend difference is more moderate from -1.7% to 0.5% (Fig. 4). There are no significant changes of the comparisons in other regions or other seasons. Fig. 5  20 shows the albedo seasonal relative trends for the 4 regions coincident with AQS in situ NO 2 data. The OMI DOMINO v2 incorporates a climatology albedo dataset (Kleipool et al., 2008) with snow/ice albedo adjustment, in which the albedo value is reset to be 0.6 if snow/ice is reported in the NASA Near-real-time Ice and Snow Extent (NISE) dataset (Boersma et al., 2011). The climatology albedo data have no trends. Thus, the trends of albedo from the DOMINO v2 product mainly originate from the yearly variation of NISE detected snow/ice and to a lesser extent the OMI sampling variation. The noticeable seasonal 25 trend of the OMI DOMINO v2 albedo dataset is the 3.9% yr -1 increase in DJF of the Midwest and a smaller DJF increase (1.0%) of the Northeast. In contrast, the MODIS albedo dataset exhibits a smaller positive DJF trend (0.8% yr -1 ), 3.1% yr -1 less than the trend from DOMINO v2, in the Midwest, and a small negative DJF trend (-0.8%) in the Northeast. These differences suggest that using a fixed snow/ice albedo and climatology albedo data are inadequate. The comparison to the AQS data shows that the MODIS albedo update leads to better agreement between satellite and in suit trends in winter in these 30 regions (Fig. 4).

Comparison after lightning event filter
As discussed in Section 2.3.3, lightning NOx affects the retrievals of satellite tropospheric NO 2 VCDs and NO 2 vertical mixing. Fig. 6 shows that the lightning filter significantly reduces the difference between the OMI-based relative trend and that of the AQS data by 0.5-1.4% yr -1 in the Northeast and 0.9-1.3% yr -1 in the South. As a result, the seasonal trend differences are within 0.9% yr -1 in these two regions except during SON. The Northeast is affected by the lightning filter due to lightning in this 5 region and transport of lightning NOx from the South (Section 2.3.3). The lightning filter has little effect on the West and the Midwest. While lightning NOx can be significant during the monsoon season in some regions of the West (Fig. 3), the average tropospheric NO 2 VCDs are usually < 1x10 15 molecules cm -2 and lightning affected regions are therefore excluded in trend analysis.
The effect of lightning filter (Fig. 6) cannot be shown in Fig. 4 because the coincident OMI and AQS data points are fewer 10 after applying the lightning filter. We examine the improvements of ocean trend removal, MODIS albedo update, and data screening with the lightning filter by comparing the differences of different OMI-based seasonal relative trends from the AQS in situ trends in Fig. 7. The previously discussed improvements such as OMI albedo update for the Midwest and the Northeast during DJF are shown. By subtracting the AQS trends, we can now find clear improvements of the lightning filter for the South and the Northeast. There remains seasonal variation of OMI-based trend biases relative to in situ data but the discrepancies of 15 the annual trends after the three discussed procedures are relatively small at 0.3% yr -1 , -0.3% yr -1 , -0.1% yr -1 , and 0.0% yr -1 , in the West, the Midwest, the Northeast, and the South regions (Fig. 1), respectively. The remaining seasonal difference of the trends reflects in part the nonlinear photochemistry (Gu et al., 2013), the effects of NOx emission changes on NO 2 retrievals , different spatial coverages of the two measurements, and the inherent difference between trends of NO 2 tropospheric VCDs and surface concentrations. 20

OMI-based NO2 trends
Table 1 summarizes the regional annual trends of coincident AQS in situ and OMI data. The "Standard" OMI data (following the DOMINO v2 algorithm) tend to show less NO 2 reduction than AQS data. After applying the three data processing procedures discussed in the previous section to the OMI data, the agreement with the AQS trends is within the uncertainties of the trends. While lightning NOx is part of OMI NO 2 observations, we treat the influence of lightning on the OMI 25 tropospheric VCD trend as a bias for comparison purposes in this study. Table 1 shows the effects of data sampling when both AQS and OMI data are analyzed and when the lightning filter is applied.
Without the lightning filter, AQS decreasing trends are stronger than the decreasing trends of OMI data (Fig. 7). The lightning trend in the NLDN data is unclear due in part to the changing instrument sensitivity (Koshak et al., 2015). If lightning NOx is not accounted for in OMI retrieval, tropospheric NO 2 VCDs are overestimated. On the other hand, lightning 30 accompanies low pressure systems which mix the atmosphere vertically and tend to reduce surface NO 2 concentrations when anthropogenic emissions are high such as urban and suburban regions. Therefore, lightning has opposite effects on surface and satellite trends. The low-pressure dilution effect on surface NO 2 concentrations depends on anthropogenic emissions (since the end point of dilution is the background NO 2 value). Therefore, the weaker decreasing surface trends likely reflects a reduction of low-pressure dilution effect. Similarly, as anthropogenic emissions decrease, the positive bias of tropospheric VCDs due to lightning NOx becomes larger, likely resulting in weaker decreasing trends. We consider the lightning effects on surface NO 2 trends to be mostly meteorological driven not by lightning NOx directly (e.g., Ott et al., 2010;Lu et al., 2017) 5 and hence the filtered OMI NO 2 data are likely closer to emission related concentration changes.
One advantage of satellite observations over a surface monitoring network is spatial coverage. The processed OMI data 15 ("Lightning filter") coincident with the AQS data show a national annual trend of -3.9±0.3% yr -1 similar to the AQS in situ trend of -4.1±0.4% yr -1 . Using all data available (Fig. 8, Table 1), the OMI data ("Lightning filter") show a much lower trend of -1.5±0.2%yr -1 , about half of the AQS trend (-3.9±0.4% yr -1 ). Fig. 9 shows that the AQS sites, which are mostly urban and suburban sites, tend to be located in regions with high tropospheric NO 2 VCDs. The OMI decreasing trend is a function of tropospheric NO 2 VCDs, increasing from 0% yr -1 to -6% yr -1 (Fig. 9). The national annual trend is close to the value of clean 20 regions which contribute much more than polluted regions. The larger decrease near the anthropogenic source regions reflect in part the nonlinear photochemistry (Gu et al., 2013) and in part to a stronger influence of NOx sources such as soils in rural regions.

Conclusions
Using data from the DOMINO v2 algorithm, we find that the computed OMI-based seasonal NO 2 (relative) trends 25 underestimate the decreasing trends of the EPA AQS data by up to 3.7% yr -1 . While lightning NOx is part of OMI NO 2 observations, we treat the influence of lightning on the OMI tropospheric VCD trend as a bias for comparison purposes in this study. Furthermore, lightning NOx effects need to be removed when using satellite observations to understand the effects of changing anthropogenic emissions.
In this study, we show that removing the background ocean trend, adopting MODIS albedo data (with better temporospatial 30 resolutions and characterization of snow/ice), and excluding lightning influences can bring OMI tropospheric NO 2 VCD trends in close agreement (within 0.3% yr -1 ) with those of the AQS data. Among the corrections, the background ocean trend removal is not as significant as the latter two. Since the origin of this trend is not yet clear, the ocean trend removal method may need updates in future studies. The remaining differences may result from the inherent differences between trends of NO 2 tropospheric VCDs and surface concentrations, different spatial sampling of the measurements, chemical-nonlinearity, and tropospheric NO 2 profile changes. The largest effects of MODIS albedo update are in winter in Midwest and Northeast and those of lightning filter are in the South and the Northeast. After applying these data processing procedures, the derived OMI-5 based annual regional NO 2 trends change by a factor of > 2 for the South, the Midwest, and the West and seasonal changes can be even larger. We derive OMI-based NO 2 regional annual relative trends using all available data for the West (-2.0%±0.3 yr -1 ), the Midwest (-1.8%±0.4 yr -1 ), the Northeast (-3.1%±0.5 yr -1 ), and the South (-0.9%±0.3 yr -1 ).
The national annual trend of the processed OMI data is -1.5±0.2%yr -1 , about half of the AQS trend (-3.9±0.4% yr -1 ). It reflects that the AQS sites are mostly located in the urban and suburban regions, where OMI data show much larger decreasing 10 trends (up to -6% yr -1 ) than rural regions (down to 0% yr -1 ). The reasons for the dependence of OMI derived trends on tropospheric NO 2 VCDs and the seasonal/regional trend differences are still not completely understood. Further studies are necessary to improve our understanding of these trends. The observation-based lightning filter implemented in this study is preliminary. Incorporating chemical transport modeling may improve this filter. Moreover, the results presented here represent an alternative and indirect way to assess the importance of lightning NOx for National Climate Assessment (NCA) analyses 15 described in Koshak et al. (2015), and Koshak (2017). Inversion studies (e.g., Zhao and Wang, 2009;Gu et al., 2013Gu et al., , 2014Gu et al., , 2016 will be needed to quantify the emission and AMF changes corresponding to the OMI tropospheric NO 2 VCD trends.

Region
Annual relative trends of coincident data (% yr -1 ) Annual relative trends using all data (% yr -1 ) a These data include the three data processing procedures of this study, namely, ocean trend correction, MODIS albedo update, and lightning filter screening.