TROPOMI NO2 in the United States: A Detailed Look at the Annual Averages, Weekly Cycles, Effects of Temperature, and Correlation With Surface NO2 Concentrations

Abstract Observing the spatial heterogeneities of NO2 air pollution is an important first step in quantifying NOX emissions and exposures. This study investigates the capabilities of the Tropospheric Monitoring Instrument (TROPOMI) in observing the spatial and temporal patterns of NO2 pollution in the continental United States. The unprecedented sensitivity of the sensor can differentiate the fine‐scale spatial heterogeneities in urban areas, such as emissions related to airport/shipping operations and high traffic, and the relatively small emission sources in rural areas, such as power plants and mining operations. We then examine NO2 columns by day‐of‐the‐week and find that Saturday and Sunday concentrations are 16% and 24% lower respectively, than during weekdays. We also analyze the correlation of daily maximum 2‐m temperatures and NO2 column amounts and find that NO2 is larger on the hottest days (>32°C) as compared to warm days (26°C–32°C), which is in contrast to a general decrease in NO2 with increasing temperature at moderate temperatures. Finally, we demonstrate that a linear regression fit of 2019 annual TROPOMI NO2 data to annual surface‐level concentrations yields relatively strong correlation (R 2 = 0.66). These new developments make TROPOMI NO2 satellite data advantageous for policymakers and public health officials, who request information at high spatial resolution and short timescales, in order to assess, devise, and evaluate regulations.

and premature mortality (Bell, 2004;Bell et al., 2006). HNO 3 often represents the final chemical state of NO 2 in the atmosphere and when deposited, agitates the equilibrium of our ecosystems due to its acidic properties (Burns et al., 2016). NO 2 can also participate in a series of reactions to create particulate nitrate (NO 3 − ), a component of fine particulate matter less than 2.5 microns in diameter (PM 2.5 ), which is the leading cause of mortality due to air pollution (Cohen et al., 2017).
TROPOMI's smallest pixel size (3.5 × 7.2 km 2 at nadir, reduced to 3.5 × 5.6 km 2 at nadir on 6 August 2019) and enhanced sensitivity are significant improvements when compared to previous satellite instruments (Veefkind et al., 2012). NO 2 is unique due to its relatively short photochemical lifetime which varies from 2 to 5 h during the summer daytime (Beirle, Boersma, et al., 2011;de Foy, B., Wilkins, J. L., et al., 2014;Valin et al., 2013) to 12-24 h during winter (Shah et al., 2020). As a result, tropospheric NO 2 concentrations are strongly correlated with local NO X emissions, which are often anthropogenic in origin.
Initial NO 2 measurements from TROPOMI show the complex spatial heterogeneities of NO 2 pollution with more refined resolution than any instrument before it Ialongo, Virta, et al., 2020). In particular, the smaller pixel sizes aid researchers in differentiating pollution sources within a single metropolitan area such as isolating signals from airports and individual highways (Judd, Al-Saadi, Janz, et al., 2019). These small-scale pixel sizes also show better agreement with the spatial features suggested by ground-based measurements (Ialongo, Virta, et al., 2020;Judd, Al-Saadi, Janz, et al., 2019). In particular, modeling studies have shown that matching the city-wide NO 2 column to 10% accuracy requires a spatial resolution of at least 4 km (Valin, Russell, Hudman, & Cohen, 2011)-the approximate spatial resolution of TROPOMI. Robust high-spatial resolution estimates are also critical inputs to those trying to quantify the surface-level NO 2 exposures (Geddes, Martin, et al., 2016;Lamsal et al., 2008;Larkin et al., 2017).
The improved spatial resolution and instrument sensitivity also allow for shorter temporal averaging ranges (days to months) to gain the similar spatial structure it would normally take >1 year to gather (Beirle, Borger, et al., 2019;Dix et al., 2020;Goldberg, Lu, Streets, et al., 2019;Lorente, Boersma, et al., 2019). As a result, it is easier to gain insight on the short-term variations of NO X pollution when using TROPOMI, which can be especially helpful for those trying to quantify intra-annual changes in NO X emissions (F. Liu et al., 2020).
In this paper, we exploit TROPOMI's small pixel sizes and enhanced instrument sensitivity to analyze spatial and temporal features of NO X columns in the continental United States on annual, seasonal, weekly, and daily timescales. For example, using only a short temporal range of data, we can now answer such questions as: • Which location within each U.S. state has the worst NO 2 air pollution? • How does the NO X emissions cycle vary by day of the week? • How does temperature affect column NO 2 amounts? • How well can we infer surface-level concentrations from satellite data?
While older sensors (e.g., OMI) provided insight into some of these questions, early sensors lacked the same sensitivity and required longer oversampling times. Therefore, answers illuminated by TROPOMI provide a "clarity" that has not been seen before.

TROPOMI NO 2
TROPOMI was launched by the European Space Agency for the European Union's Copernicus Sentinel 5 Precursor (S5p) satellite mission on 13 October 2017. The satellite follows a sun-synchronous, low-earth (825 km) orbit with an equator overpass time of approximately 13:30 local solar time (Veefkind et al., 2012). TROPOMI measures total column amounts of several trace gases in the Ultraviolet-Visible-Near Infrared-Shortwave Infrared spectral regions (VanGeffen et al., 2019). This instrument is characterized as a passive optical satellite sensor due to its reliance on solar UV-Visible radiation to gather measurements. At nadir, pixel sizes are 3.5 × 7 km 2 (reduced to 3.5 × 5.6 km 2 on 6 August 2019) with little variation in pixel sizes across the 2,600 km swath. The instrument observes the swath approximately once every second and orbits the Earth in about 100 min, resulting in daily global coverage.
Using a differential optical absorption spectroscopy technique on the radiance measurements in the 405-465 nm spectral window, the top-of-atmosphere spectral radiances can be converted into slant column amounts of NO 2 between the sensor and the Earth's surface . In two additional steps, the slant column quantity can be converted into a tropospheric vertical column content. In the first step, the stratospheric portion of the column (the amount above approximately 12 km in altitude) is subtracted using a global model estimate that is refined using data assimilation (Boersma, Eskes, & Brinksma, 2004). In a second step, the slant tropospheric column is converted to a vertical column using a quantity known as the air mass factor (AMF). The AMF is the most uncertain quantity in the retrieval algorithm (Lorente, Folkert Boersma, et al., 2017) and is a function of the surface reflectance, the NO 2 vertical profile, and scattering in the atmosphere among other factors. Using accurate and high-resolution data (spatially and temporally) as inputs in calculating the AMF can significantly reduce the overall errors of the AMF (S. Choi et al., 2019;Goldberg et al., 2017;Lamsal, 2020;Laughner, Zare, & Cohen, 2016, Laughner, Zhu, & Cohen, 2019Lin et al., 2015;M. Liu et al., 2019;Russell, Perring, et al., 2011;Zhao et al., 2020) and thus the tropospheric vertical column content.
Operationally, the TM5-MP model (1 × 1° resolution) is used to provide the NO 2 vertical shape profile, and the climatological Lambertian Equivalent Reflectivity (0.5 × 0.5° resolution) (Kleipool et al., 2008) is used to provide the surface reflectivities. The operational AMF calculation does not explicitly account for aerosol absorption or scattering effects, which are partially accounted for in the effective cloud radiance fraction (Chimot et al., 2016). There is already some evidence that the current TROPOMI operational NO 2 product may have a low bias of 20%-40% in urban areas; much of this bias may be attributed to the AMF (Judd, Al-Saadi, Szykman, et al. 2020; Verhoelst et al., 2020). While the operational product does have larger uncertainties in the tropospheric column contents than a product with higher spatial resolution inputs, we limit our analysis to relative trends, which dramatically reduces this uncertainty.

Re-gridding
For our analysis we re-grid the operational TROPOMI tropospheric vertical column NO 2 , with native pixels of approximately 3.5 × 7 km 2 , to a newly defined 0.01° × 0.01° grid (approximately 1 × 1 km 2 ) centered over the continental United States (CONUS; corner points: SW: 24.5°N, 124.75°W; NE: 49.5°N, 66.75°W). Before re-gridding, the data are filtered so as to use only the highest quality measurements (quality assurance flag (QA_flag) > 0.75). By restricting to this QA value, we are removing mostly cloudy scenes (cloud radiance fraction > 0.5) and observations over snow-ice. Once the re-gridding has been completed, the data are averaged over varying timeframes as discussed in the results section.

Other Data sets
Additionally, we use three complementary products in some sections of our analysis. We compare tropospheric vertical column information from TROPOMI to the same quantity from the NASA OMI NO 2 version 4 product in a qualitative sense. Only OMI pixels with cloud fractions < 0.3, surface albedo < 0.3, and not affected by the "row anomaly" are included. When filtering TROPOMI data based on temperature, we use the maximum daily hourly 2-m temperature from the ERA5 re-analysis. To downscale the ERA5 re-analysis, which is provided at 0.25° × 0.25°, we spatially interpolate maximum daily hourly 2-m temperature to 0.01° × 0.01° using bilinear interpolation. For that reason, the heat-urban island effect and any microscale meteorology features (e.g., sea breezes) will not be accounted for, but these effects should be minor for our particular analysis, which groups temperatures in ∼5°C intervals. Figure 1 depicts the 2019 CONUS annual average of TROPOMI and OMI tropospheric vertical column NO 2 compared to averages over monthly, weekly, and daily timeframes.

TROPOMI NO 2 in CONUS
This example illustrates how shorter timeframes compare to the annual average in both magnitude and clarity. In the single daily snapshot (20 September 2019), there are wide sections that are missing due to cloud coverage. Missing data in the OMI NO 2 snapshot is much more widespread than TROPOMI due to the "row anomaly." which obstructs a portion of OMI's field of view. In the areas that do have coverage, values can be a factor of five different than the annual average, but the spatial heterogeneities are generally captured. When oversampling over a one-week period (16-22 September 2019), the TROPOMI image quickly starts to resemble the annual average with some differences in magnitude due to meteorological factors, such as temperature (which will be discussed later), but the OMI image is still very noisy. A monthly oversampled image essentially captures the same spatial heterogeneities as the annual average, but with magnitude differences due to meteorology. In most scenarios, a one-month average should be considered the minimum amount of oversampling time needed for TROPOMI to properly capture spatial heterogeneities, while for OMI ∼12 months of data is needed in order to properly capture spatial heterogeneities. It should be noted that September was specifically chosen for this analysis due to its propensity to have both less cloud coverage and snow cover than other months. If oversampling during winter months (i.e., December-March), which tend to have fewer ideal conditions for satellite retrievals of trace gases, oversampling times will need to be longer to achieve similar clarity. In a qualitative sense, OMI yields larger values than TROPOMI in most areas (rural and urban alike). This is consistent with other literature, which shows OMI yielding larger values than TROPOMI (Wang et al., 2020) and a low bias in TROPOMI in U.S. urban areas (Judd, Al-Saadi, Szykman, et al., 2020).
When visually inspecting the CONUS TROPOMI NO 2 average during the initial 20 months of the TRO-POMI record (1 May 2018-31 Dec 2019) (Figure 2), we now start to see clear spatial heterogeneities across the domain. The largest U.S. cities can be seen, and their magnitudes can be compared to each other (results further discussed later).
Equally important, smaller sources of NO 2 pollution can now be observed, and they are not spatially smeared into the background NO 2 concentration. For example, when magnifying the western United States (Figure 3  TROPOMI data are especially powerful in analyzing local variations in NO 2 pollution as compared to predecessor instruments. In Figure 4, we zoom into five different U.S. states, and in Table 1 we provide the largest NO 2 values in each state; note that in Figure 4 we use a colorbar that is not linear in order to better differentiate urban versus rural values. In Figure 5, we zoom into six different U.S. cities. In each instance, the oversampled TROPOMI NO 2 images exhibit features that match known NO X emissions patterns. The larger NO 2 values correlate very well to the interstate network, population density, and industrial activity hubs (such as manufacturing facilities, airports, and shipping ports). For example, in the image of Maryland, the largest value is observed at the Baltimore Harbor, which is a confluence of several major highways, a large shipping port, the city incinerator, and many industrial facilities. Similarly, the largest values in Chicago exist along the I-55 corridor which has a high traffic volume and a high-density of industrial facilities, with secondary maxima at the O'Hare International airport and the U.S. Steel Corp operations in East Chicago, Indiana. In Los Angeles, the spatial pattern matches the basin outline very well, with the largest values between downtown Los Angeles and the Long Beach Shipping Port. In Houston, Texas the largest values are nearest to the petrochemical refining facilities east of town. For all cases, TROPOMI can accurately quantify the relative relationship between the largest sources of NO X emissions and NO 2 concentrations.

Day-of-the-Week Relationships
A common use of oversampled satellite data is in investigating the weekly cycle of NO X emissions. In Figure 6, we show the weekly pattern of NO 2 across CONUS for three different days of the week as well as the full weekly cycle in seven U.S. cities; we selected U.S. metropolitan areas that were both large and representative of geographic diversity. In all cities, the NO 2 appears to be approximately equivalent amongst all weekdays with some minor exceptions. NO 2 pollution is 2.5% larger on Tuesday than a typical weekday, while Mondays and Fridays have 1.4% and 1.3% lower NO 2 pollution than a typical weekday. On Saturdays, NO 2 is 16% lower than the weekday averages, and on Sundays 24% lower. Standard errors of the mean for each city are shown in Table S1, and are approximately 10% for any given city, and approximately 4% when all cities are aggregated together. This means that NO 2 changes on weekends-including the differences between Saturdays and Sundays-are statistically significant, but the difference between weekdays are not yet statistically significant. As more TROPOMI NO 2 data are acquired over time, these standard errors of the mean will decrease, and we might be able to deduce statistically significant changes between individual weekdays. The weekend changes calculated here (16% drop on Saturdays, 24% drop on Sundays) are less dramatic than previously reported weekend changes (30%-60% drops) in the 2005-2013 timeframe (de Foy, Russell, Valin, et al., 2010;Valin et al., 2014). There are two explanations for the flattening of the weekday-weekend cycle: 1.) as overall emissions are decreasing, the NO 2 lifetime in many cities is increasing (Stavrakou, Müller, Bauwens, et al., 2020)    pronounced weekday-weekend emissions pattern, are continually representing a smaller fraction of NO X emissions over time (Dallmann & Harley, 2010;McDonald et al., 2012).
When analyzing the weekday/weekend differences, there should be some consideration for the difference in traffic patterns and general activity between weekends and weekdays. On weekends, traffic counts generally peak in the early afternoon, while on weekdays traffic counts peak in the evening, with a secondary peak in the early morning (de Foy, 2018). Since the satellite observation is acquired in the early afternoon, we suggest that the 24-h averaged NO X emissions difference between weekdays and weekends may be even greater than implied by the satellite data. The soon-to-be-launched TEMPO instrument, a geostationary satellite, will hopefully be able to better quantify the morning and evening differences of NO X emissions (Chance et al., 2019;Penn & Holloway, 2020;Zoogman et al., 2017).

Hot Versus Warm Days
In Figure 7, we show the variation in column NO 2 as a function of the daily maximum 2-m temperature. Due to varying climates across the United States, most cities do not have values for all temperature bins. In general, as temperatures increase, NO 2 decreases; this is primarily driven by j(NO 2 ) which increases with stronger sunlight. When temperatures are >32°C, we observe a leveling with increasing temperature. This may be related to increasing anthropogenic NO X emissions (Abel et al., 2017;He et al., 2013) at high temperatures despite a shorter NO 2 lifetime. This may also be driven by biogenic or natural causes, such as GOLDBERG ET AL.   the faster dissociation of peroxy-acyl nitrates (PANs) or increased soil NO X emissions (Rasool et al., 2019;Romer et al., 2018) at hot temperatures. The latter reasons are likely causing rural areas to observe increases in NO 2 as temperatures warm above 32°C. The temperature-driven stabilization of NO 2 at very high temperatures appears to hold for all cities except Chicago. Standard errors of the mean for each city are shown in Table S2, and are approximately 7% for any given city on warm/hot temperature days (>20°C), and approximately 2%-3% when all cities are aggregated together. This means that the NO 2 decreases with increasing temperature as well as the NO 2 increases on the hottest days are statistically significant in most areas. It should be noted that there are cross-correlations with increased temperature such as a lower solar zenith angle (which affects photolysis rates of chemical species and the satellite viewing geometry), larger biogenic volatile organic compound (BVOC) emissions in forested areas (which affects the NO 2 lifetime), and higher total water columns (which affects wet deposition and introduces an increased spectral interference). Apportionment of the effects of natural versus anthropogenic sources contributing to NO 2 increases in urban areas on the hottest days will be the subject of future research using model simulations.

Relationship With Surface NO 2 Concentrations
To understand how well TROPOMI NO 2 , without any adjustment, captures surface-level concentrations, we compare the 2019 annual TROPOMI NO 2 average to 24-h annual average EPA AQS monitor data. The surface-level concentrations from the EPA AQS network are known to have a high instrument bias (Dickerson et al., 2019) and thus referred to as NO 2 * hereafter. In Figure 8, we show a scatterplot between 2019 annual averages of oversampled TROPOMI NO 2 and AQS surface-level NO 2 * concentrations. For our analysis, we restrict our fit to monitoring sites that are not "near-road." The EPA requests certain states to site "near-road" NO 2 monitors, which are requested to be within 20 m of a major highway; we do not expect TROPOMI observations to capture this very fine spatial gradient, and are therefore not considered in our fit. Figure 8 demonstrates that there is a strong correlation (R 2 = 0.66) between a linear fit and monitoring sites considered to be "not near-road." which suggests that many (but not all) of the spatial heterogeneities observed by TROPOMI over long time intervals (e.g., year) are real and not an artifact of the processing algorithms. We are encouraged to see that a simple linear fit is able to capture near-surface NO 2 variability well. In order to better estimate surface-level concentrations, TROPOMI NO 2 data should be merged with a model simulation (Cooper, Martin, McLinden, & Brook, 2020) and/or land-use characteristics (Bechle et al., 2015;Beloconi & Vounatsou, 2020;Di et al., 2019;Larkin et al., 2017).

Conclusions
This study investigates the capabilities of the TROPOMI in observing the spatial and temporal patterns of NO 2 pollution in the continental United States (CONUS). Here, we demonstrate that TROPOMI can capture fine-scale spatial heterogeneities in urban areas, such as emissions related to airport/shipping operations and high traffic; this type of spatial precision cannot be matched by predecessor satellite instruments over short timescales (<1 year). We find that Saturday and Sunday concentrations are 16% and 24% lower respectively than during weekdays, with the caveat that diurnal emissions patterns vary among weekdays and weekends. We also analyze the effects of hot temperatures (>32°C) on NO 2 column amounts and find GOLDBERG ET AL.
10.1029/2020EF001665 9 of 16 that column NO 2 is generally larger on the hottest days as compared to warm days (26°C-32°C). Finally, we compare column NO 2 with surface-level NO 2 estimates and find relatively good correlation (R 2 = 0.66).
For this work, we rely on the operational TROPOMI NO 2 algorithm, which underestimates tropospheric vertical column NO 2 in urban areas. Previous studies suggest that this underestimate is due to the AMF and ∼5km pixel size which cannot resolve street-level variations in concentrations (Goldberg, Lu, Streets, et al., 2019;Griffin et al., 2019;Judd, Al-Saadi, Szykman, et al., 2020;Judd, Al-Saadi, Janz, et al., 2019;Zhao et al., 2020); investigating the effects of the AMF bias on trends as well as investigating the effects of the pixels sizes will be the subject of future work. Also, there may be a clear-sky bias (Geddes, Murphy, et al., 2012) associated with any satellite retrieval, but the general spatial heterogeneities of NO 2 pollution should be similar amongst all types of weather conditions when averaged over long timeframes. Lastly, interpreting results from polar-orbiting satellite instruments, such as TROPOMI, should be made with some caution due to the mid-day only data collection time. Work quantifying this bias has shown GOLDBERG ET AL.
10.1029/2020EF001665 10 of 16 Figure 5. Same data shown in Figure 2, but now zoomed into six different U.S. cities. Color bar has been adjusted to better differentiate spatial heterogeneities on a local scale.  that NO 2 column measurements are lower and incrementally more spatially homogeneous in the afternoon than during the morning (Chong et al., 2018;Fishman et al., 2008;Herman et al., 2019;Knepp et al., 2015;Penn & Holloway, 2020;Tzortziou et al., 2015); it is likely that data from geostationary platforms such as TEMPO (Zoogman et al., 2017), GEMS (W. J. Choi, 2018), and Sentinel 4 (Timmermans et al., 2019), will be able to provide further insight on this time-of-day bias.
Because TROPOMI can observe and measure NO 2 increases attributed to relatively small sources, future work should be able to quantify emissions from small sources (e.g., industrial activities, ship plumes, small wildfires) that had previously gone undetected from predecessor spacebased instruments. Furthermore, due to the instrument's excellent stability, precision, and spatial resolution, it is no longer necessary to average over 6+ months of data to gain a clear depiction of regional NO 2 abundances; instead monthly, weekly or even daily aggregations could suffice for many purposes. The examples presented here demonstrate how TROPOMI NO 2 satellite data can be advantageous for policymakers requesting information at high spatial resolution and short timescales, in order to assess, devise, and evaluate regulations. Future health impact assessment studies can use the high-spatial resolution capabilities of TROPOMI NO 2 to investigate disparities in traffic-related air pollution exposure and associated health effects between neighborhoods and population sub-groups within cities.