Inferring ground-level nitrogen dioxide concentrations at fine spatial resolution applied to the TROPOMI satellite instrument

Satellite-based estimates of ground-level nitrogen dioxide (NO2) concentrations are useful for understanding links between air quality and health. A longstanding question has been why prior satellite-derived surface NO2 concentrations are biased low with respect to ground-based measurements. In this work we demonstrate that these biases are due to both the coarse resolution of previous satellite NO2 products and inaccuracies in vertical mixing assumptions used to convert satellite-observed tropospheric columns to surface concentrations. We develop an algorithm that now allows for different mixing assumptions to be used based on observed NO2 conditions. We then apply this algorithm to observations from the TROPOMI satellite instrument, which has been providing NO2 column observations at an unprecedented spatial resolution for over a year. This new product achieves estimates of ground-level NO2 with greater accuracy and higher resolution compared to previous satellite-based estimates from OMI. These comparisons also show that TROPOMI-inferred surface NO2 concentrations from our updated algorithm have higher correlation and lower bias than those found using TROPOMI and the prior algorithm. TROPOMI-inferred estimates of the population exposed to NO2 conditions exceeding health standards are at least three times higher than for OMI-inferred estimates. These developments provide an exciting opportunity for air quality monitoring.


Introduction
Air pollution is a leading risk factor for premature mortality, as exposure to air pollution is associated with increases in heart disease, stroke, cancer, and other respiratory and cardiovascular diseases (Cohen et al 2017). Nitrogen dioxide (NO 2 ) is a major contributor to poor air quality and exposure to NO 2 has been associated with increased rates of asthma incidence (Anenberg et al 2018, Achakulwisut et al 2019, lung cancer (Hamra et al 2015), and overall mortality (Burnett et al 2004, Brook et al 2007, Crouse et al 2015. NO 2 is often used as an indicator of air pollution from traffic and other combustion sources (Brook et al 2007, Levy et al 2014, Achakulwisut et al 2019. Additionally, NO 2 is a precursor to aerosol and ozone production, which are also important pollutants (Cohen et al 2017, Stieb et al 2019. Quantifying the impacts of NO 2 on health requires accurate monitoring of NO 2 concentrations and their spatial variability.
The distribution of NO 2 has substantial spatial gradients due to its short lifetime and the spatial variability of emission sources. This inhomogeneity inhibits constructing accurate NO 2 fields by in situ monitoring alone. Satellite instruments provide NO 2 observations with greater spatial coverage. Previous studies have shown that surface NO 2 concentrations can be inferred from satellite-retrieved vertical column densities using a chemical transport model to relate the retrieved column to ground-level concentrations (Lamsal et al 2010, Kharol et al 2015, Geddes et al 2016, Gu et al 2017. Recent studies have used machine learning techniques to convert column abundances to ground-level concentrations (Chen et al 2019, De Hoogh et al 2019, Beloconi and Vounatsou 2020, Di et al 2020, Qin et al 2020. These satellite-based ground-level concentration estimates have been useful in health impact studies as an estimate of pollution exposure (Anenberg et al 2018) or as an input to land use regression models used to estimate exposure at finer spatial resolution (Larkin et al 2017, Achakulwisut et al 2019. Previous evaluations indicate that satelliteinferred surface concentrations correlate well with in situ measurements but were biased low (Geddes et al 2016, Gu et al 2017 for reasons that have not been resolved. The contribution of biases in columnar retrievals to this surface bias has largely been attributed to issues related to horizontal resolution, either from coarse resolution a priori NO 2 profiles used in the retrieval (e.g. Heckel et al 2011 or from using point ground monitors to evaluate satellite observations with coarse pixel resolution (Kharol et al 2015, Judd et al 2019. These same factors may be exacerbated when surface concentrations are inferred from columnar retrievals. Comparisons with ground monitors and aircraft indicate that there is substantial variability in NO 2 concentrations within the coarse pixel size of prior satellite observations, causing coarse resolution satellite measurements to underestimate peaks in the NO 2 field by a factor of two or more (Broccardo et al 2018, Judd et al 2019Lamsal et al 2017). Fine resolution NO 2 profiles have been shown to improve NO 2 retrieval accuracy , Mclinden et al 2014, Laughner et al 2016, Goldberg et al 2017, Ialongo et al 2020, Liu et al 2020. Fine spatial resolution observations also reduce the likelihood of a satellite observation being affected by clouds that obscure surface-level NO 2 from observation and hinder the ability to resolve fine-scale spatial structure of the surface NO 2 field.
Application of algorithmic advances to the TROPOspheric Monitoring Instrument (TROPOMI) provides an opportunity to build upon these previous studies to re-examine the bias with surface concentrations. The spatial resolution of TROPOMI data used here (3.5 × 7 km 2 ) is roughly a factor of 13 finer than its predecessor Ozone Monitoring Instrument (OMI) (13 × 24 km 2 ). Preliminary studies have shown that TROPOMI is capable of observing pollution at unprecedented scales (Hu et al 2018, Alvarado et al 2020  This paper presents a new algorithm for inferring ground level NO 2 concentrations that improves upon previous methods by updating assumptions regarding vertical mixing in the boundary layer and allowing the satellite-retrieved column densities to inform the relation of column abundances to ground level concentrations. The high spatial resolution of TROPOMI NO 2 column observations, together with these algorithmic developments, addresses the bias in satellite-inferred surface concentrations and reveals unprecedented fine-scale information in estimated ground-level NO 2 concentrations.

Satellite NO 2 data
In this study we use retrieved NO 2 column densities from both TROPOMI (Veefkind et al 2012, Van Geffen et al 2020 and OMI (Levelt et al 2006) satellite instruments from July 2018-June 2019. Both instruments have spectrometers measuring in the UV-Vis spectral bands, and are on sun-synchronous orbits with local overpass times around 1:30 P.M. NO 2 column retrievals from TROPOMI use the DOMINO retrieval method (Boersma et al 2011(Boersma et al , 2018. NO 2 retrievals from OMI are from the NASA Standard Product version 3 (Krotkov et al 2017). Only observations with retrieved cloud fractions less than 0.1 are used. The important difference between the two satellite products for this study is their spatial resolution; while both TROPOMI and OMI have a 2600 km swath width, TROPOMI (3.5 × 7 km 2 ) has 450 across-track pixels while OMI (13 × 24 km 2 ) has 60. Additionally, a partial blockage of the OMI field of view, known as the row anomaly, prevents quality observations from many pixels and further hinders OMI's observation density (http://projects.knmi.nl/omi/research/product/ rowanomaly-background.php).

GEOS-Chem
The GEOS-Chem chemical transport model version 12.3.2 (www.geos-chem.org) is used here. GEOS-Chem contains a detailed HO x -NO x -VOC-O 3 -aerosol chemical mechanism (Bey et al 2001, Park et al 2004 and is driven by Goddard Earth Observing System Forward Processing (GEOS-FP) assimilated meteorological data. We use a nested grid with 0.25 • × 0.3125 • resolution over North America (9.75 • -60 • N, 60 • -130 • W) with dynamic boundary conditions at 2 • × 2.5 • resolution. The nested grid values are averaged to coarser resolutions for testing the effect of resolution on inferred surface concentrations.
Anthropogenic NO x emissions are from the Community Emissions Data System (Hoesly et al 2018) with regional overwrites where more detailed information is available over Canada (Air Pollutant Emission Inventory) and the US, (National Emissions Inventory 2011, Travis et al 2016. Biomass burning emissions are from the GFED4.1 inventory (Van der Werf et al 2010). Lightning NO x emissions are described by Murray et al (2012). Soil NO x emissions are described by Hudman et al (2012). Of importance to this work is the capability of GEOS-Chem to translate column abundances to surface concentrations, which requires accurate representation of the NO 2 profile; past evaluations demonstrate that GEOS-Chem simulated NO 2 profiles are consistent with aircraft observations (Lin andMcelroy 2010, Travis et al 2016).

In situ NO 2 observations
We use hourly surface NO 2 measurements from the US Environmental Protection Agency Air Quality System (US-EPA AQS, https://aqs.epa.gov/ aqsweb/documents/data_mart_welcome.html) over the continental US and Environment and Climate Change Canada's National Air Pollution Surveillance Program (NAPS, http://maps-cartes.ec.gc.ca/rnspanaps/data.aspx) over Canada to evaluate the satellite products. We calculate annual mean concentrations at each of the 625 sites by averaging observations between 13:00 h and 15:00 h corresponding to the satellite overpass times. NAPS measurements from 2017 are used as these are the most recent available dates. Following Lamsal et al (2008) we use a correction factor derived from GEOS-Chem to correct for the known overestimate in regulatory measurements of NO 2 concentrations due to interference of other reactive nitrogen species (i.e. peroxyacetyl nitrate (PAN), nitric acid (HNO 3 ), and organic nitrates): (1)

Improving the air mass factor calculation
A significant source of systematic error in satellite NO 2 retrievals is the a priori assumed NO 2 profile used in calculating the air mass factor (AMF) that converts line-of-sight slant columns to vertical column densities (Lorente et al 2017;Boersma et al 2018). Replacing the a priori used in the retrievals with profiles from fine-resolution models has been shown to improve retrieved vertical column densities ( Inconsistencies between a priori profiles may introduce spurious differences when comparing two different satellite products, or when using a model to relate surface concentrations to vertical columns (Boersma et al 2016). For these reasons we remove the influence of the a priori profile in both the TROPOMI and OMI AMFs and replace them with GEOS-Chem vertical profiles using the method outlined in Lamsal et al (2010): where Ω o is the updated vertical column, Ω s is the observed slant column, M is the retrieval AMF, A k is the retrieval averaging kernel, and n G k is the normalized tropospheric NO 2 profile from the GEOS-Chem simulation. We use the Interactive Multi-sensor Snow and Ice Mapping System (IMS) (Helfrich et al 2007) 4 km product to identify snow-covered pixels, which are then omitted from this analysis. IMS is used here in addition to the snow identification used in the TRO-POMI and OMI retrievals as it has been shown to more accurately identify snow cover (Cooper et al 2018).

Surface estimation
We estimate satellite-based surface NO 2 concentrations at both the 0.25 • × 0.3125 • (≈28 × 28 km 2 at 35 • N) resolution of the GEOS-Chem nested simulation (i.e. 'moderate' resolution) and at a sub-modelgrid resolution of 0.025 • × 0.03125 • (i.e. 'fine' resolution, ≈2.8 × 2.8 km 2 at 35 • N). Satellite vertical column densities are gridded daily using an areaweighted oversampling approach (Spurr 2003). We infer fine-resolution surface concentrations (S s o ) by further developing the method outlined in Kharol et al (2015) and Lamsal et al (2008) that accounts for subgrid variation in the column to surface relationship to correct for biases related to vertical mixing assumptions.
Surface concentrations (S) can be inferred from satellite-retrieved vertical column abundances (Ω) using a simulated surface-to-column conversion factor from a chemical transport model: The superscript o signifies quantities based on satellite observations (otherwise quantity is modeled). Satellite-based surface concentrations at a submodel-grid resolution could be similarly calculated: where the subscript s represents sub-grid resolution values (model resolution values otherwise), however S s and Ω s are unknown. Following Lamsal et al (2008) we use the satellite-observed sub-grid variability to infer sub-grid variability in the simulated tropospheric column where Ω o is the average satellite-retrieved tropospheric column within the model grid box. We assume that the tropospheric column consists of two parts: where Ω up is the upper portion of the column that is well mixed both vertically and horizontally across the model grid box (such that Ω up = Ω s up ), and Ω low is the lower portion of the column that is well mixed vertically but has sub-grid horizontal variability due to local surface emissions. Using these definitions, Equation (6) can be expressed for sub-pixel columns as: Applying the sub-pixel tropospheric column variability from the satellite observations via equation (5) into equation (7) gives: We assume that air density is constant within a grid box and treat the ratio of surface concentration to lower partial column as the same at sub-grid and grid resolution, i.e.

S s
Substituting equations (5), (8), and (9) into equation (4) gives An important decision is selecting how to distinguish the horizontally well-mixed upper partial column from the lower partial column (which has sub-grid variability). Previous studies have defined the lower partial column as the being within the boundary layer and the upper column as the free troposphere (e.g. Lamsal et al 2010, Kharol et al 2015. This has been shown to well represent the spatial distribution as the free troposphere is generally well mixed and has less horizontal variability compared to the boundary layer. However, this definition assumes that the boundary layer NO 2 profile is uniform vertically, which is not the case when there are large emission sources that produce significant surface enhancements, and leads to an underestimation of surface NO 2 concentrations (Zhang et al 2016). This bias has been noted in evaluations of previous satelliteinferred surface concentration estimates (Kharol et al 2015).
Alternatively, one could define the lower column as the surface-level partial column (i.e. the lowest model grid box, centered at ∼50 m altitude) and the upper column as the remaining column above (i.e. model level 2 to tropopause). This is likely a better assumption over large emission sources and can better attribute local peaks in NO 2 columns to surface enhancements, reducing the underestimations in the previous method. However, assuming all the sub-grid variability occurs at the surface ignores potential subgrid variability in the lower boundary layer away from the emission source, artificially inflating the surface level variability and producing an unrealistic sub-grid spatial distribution.
Of these two possibilities, assuming a uniform free troposphere and uniform boundary layer NO 2 (S o s,BL ) is more reasonable in most cases and produces a more realistic spatial distribution, while assuming a well mixed surface-level partial column (S o s,surf ) is more accurate over large emission sources. Therefore, we calculate satellite-inferred surface concentrations using both methods, and define a scaling factor χ in each model grid box as the ratio of the maximum values from each method: The maximum value in each grid box represents the value where assuming a uniform surface-level partial NO 2 column is likely the more accurate assumption. Therefore, this scaling factor quantifies the magnitude of the bias that occurs from incorrectly assuming a uniform boundary layer NO 2 over large emission sources. One can then apply this scaling factor to the original method over polluted regions where a surface enhancement is expected: In relatively clean regions, the mixed boundary layer assumption without the scaling factor is preferred: We define clean regions as those with annual mean TROPOMI tropospheric columns less than 10 15 mol cm −2 and polluted regions as those with annual mean columns greater than 11 × 10 15 mol cm −2 . This allows the satelliteretrieved column abundance to inform the shape of the sub-grid vertical profile. We use a linear interpolation between the two methods for regions that fall between the clean and polluted thresholds.
To summarize, sub-grid surface concentrations are given by: The 10 15 and 11 × 10 15 mol cm −2 cut-off values were chosen following a sensitivity test where values ranging from 1 to 20 × 15 mol cm −2 were tested. The selected values yielded the best agreement between satellite-inferred concentrations and the in situ surface measurements. Typical values of κ are 1 in clean regions, 1.5-2 in most cities, and 2.5-3 in major cities like New York and Los Angeles. Setting κ equal to unity makes equation (14) equivalent to the method used in Lamsal et al (2008) and Kharol et al (2015).
We use GEOS-Chem to correct for sampling biases in the satellite records due to persistent cloudy periods or surface snow cover by sampling the GEOS-Chem simulated surface concentrations to match the satellite (S sampled ), and using the ratio of the sampled mean to the true annual mean (S annual ): Figure 1 shows surface concentrations inferred from OMI and TROPOMI at both moderate (≈28 × 28 km 2 at 35 • N) and fine (≈2.8 × 2.8 km 2 at 35 • N) resolution. NO 2 concentrations from both instruments exhibit enhancements over urban and industrial areas. TROPOMI-inferred concentrations are typically larger than those from OMI at moderate resolution (mean fractional bias 13%). At fine resolution, TROPOMI-inferred concentrations are higher over cities and large emission sources. The advantage of the smaller TROPOMI pixel size is apparent at fine resolution, as the TROPOMI-inferred concentrations at this resolution display less noise and reveal more fine-scale features in the NO 2 field than those from OMI. Area-weighted and population-weighted mean NO 2 concentrations over North America for both instruments are inset in figure 1. TROPOMI-inferred population-weighed concentrations are 41%-91% higher than those inferred from OMI, with potential implications for health impact assessments. Figure 2 shows fine resolution surface concentrations inferred from TROPOMI and OMI on a regional scale. While OMI can observe enhancements surrounding major cities, the TROPOMI-inferred surface concentrations show greater fine-detail structure (i.e. more well-defined enhancements over Indianapolis and Columbus, distinguishing Edmonton/Red Deer/Calgary) as well as signatures from smaller cities not captured by OMI (i.e. Orlando, Saskatoon, Winnipeg). TROPOMI-inferred surface concentrations also show signatures from major industrial corridors that cannot be observed in the OMI-inferred NO 2 field (i.e. I-75 and -81 from Atlanta through eastern Tennessee, I-35 from Dallas to Austin). TROPOMI-inferred surface concentrations also show a greater enhancement over regions with significant oil and gas production (i.e. the Canadian oil sands region, eastern Wyoming oil fields). Table 1 summarizes the impact that resolution has on the likelihood of observing clear-sky conditions that are ideal for surface NO 2 sensitivity. We define clear pixels as those with retrieved cloud fractions less than 0.1, and the fraction of observed area that is clear as the fraction of observed area consisting of pixels with cloud fractions <0.1. Pixels affected by the OMI row anomaly are not included. The fine resolution of TROPOMI allows for a greater fraction of clear pixels and a greater observed clear area overall. These clear-sky conditions have greater sensitivity to near-surface NO 2 and thus improve the quality of surface NO 2 inferences. Figure 3 shows a scatterplot comparing satelliteinferred surface concentrations to the ground-based measurements over the US and Canada. TROPOMI has a higher correlation and lower bias compared to the in situ observations at both resolutions. The new algorithm developed here (figure 3(c)) improves upon the prior algorithm of Lamsal et al (2008) (figure 3(b)). The supplementary material (available online at stacks.iop.org/ERL/15/104013/mmedia) shows that the quality of these estimates persists at seasonal scales. Figure 4 examines the effect of a priori resolution on TROPOMI-inferred surface concentrations. The prior method of Lamsal et al (2008) is sensitive to model resolution, with agreement with in situ observations deteriorating as profile resolution increases. The method developed here allows for the TROPOMI column observations to inform sub-grid profile information, including adjusting limits on w at each resolution (upper w limit = 11, 10, 9, and 4 × 10 15 mol cm −2 for panels (A) through (D)). This allows for consistency with surface measurements at all resolutions tested here. Figure 5 shows estimates of the population exposed to different NO 2 concentrations across North America as a function of different instruments and algorithms. TROPOMI-inferred surface concentrations are higher than those inferred from OMI, with TROPOMI indicating nearly three times as many people across North America experience exposures that exceed the Canadian Ambient Air Quality Standard annual mean NO 2 exposure of 17 ppb than indicated by OMI. TROPOMIbased estimates indicate that 9 million people live  in North American regions that exceed the World Health Organisation annual mean guideline of 40 ppb and 3 million exceeding the US EPA guideline of 53 ppb, while OMI-inferred concentrations suggest no exceedances at these levels. In contrast, exposures calculated using the prior algorithm of Lamsal et al (2008) suggest no exceedances for either TROPOMI or OMI.

Discussion and conclusion
A longstanding question has been why previous satellite-derived surface NO 2 concentrations underestimated ground level measurements. Here we developed improved algorithms to derive surface NO 2 from columnar satellite observations and find that the biases observed in previous studies cannot be corrected by increasing satellite resolution alone, as they were partially due to inaccurate assumptions regarding vertical mixing within the boundary layer.
The new algorithm developed here allows for different vertical mixing assumptions to be made based on satellite-observed NO 2 conditions, substantially reducing these biases. Improvements in spatial resolution achieved by TROPOMI have greatly improved the ability to infer surface concentrations from satellite NO 2 columns compared to its predecessor OMI. TRO-POMI provides a more detailed surface NO 2 field across North America, allowing for identification of features that were previously difficult to observe, such as smaller cities and industrial corridors along highways. TROPOMI-inferred surface concentrations also have better agreement with surface observations at both resolutions. Fine resolution estimates from TROPOMI have significantly better agreement than those from OMI found here as well as in previous studies, owing both to TROPOMI's greater resolution and improvements to the method of inferring surface concentration from column abundances developed here. The finer spatial resolution of TROPOMI also allows for more clear-sky observations and a larger percentage of clear-sky observed area compared to OMI, which allows for increased sampling of surface NO 2 concentrations. Population-weighted concentrations estimated here from TROPOMI at 0.025 • × 0.03125 • resolution are up to twice as high as similar estimates from OMI, demonstrating how fine resolution maps of NO 2 that more accurately describe surface NO 2 variability can impact exposure estimates and health studies. Land use regression models used to produce very high (sub-km) resolution surface NO 2 distributions for epidemiological studies will also likely benefit from having a higher resolution satellite-inferred surface concentration data set as an input.
Past studies have shown the importance of using a priori NO 2 profiles from high resolution models in AMF calculations for improving both the magnitude and spatial distribution of satellite retrieved vertical column densities , Mclinden et al 2014, Laughner et al 2016, Goldberg et al 2017, Ialongo et al 2020, Liu et al 2020. High resolution model information is also important when interpreting these satellite observations, particularly for NO 2 due to the nonlinearity of NO x chemistry (Valin et al 2013, Goldberg et al 2019. Model resolutions of 4-12 km 2 are likely needed to fully resolve nonlinear chemistry effects on NO 2 concentrations when estimating NO x emissions . However, we find good agreement between in situ measurements and our TROPOMI-inferred surface concentrations based on simulated information at ∼28 × 28 km 2 , and tests indicate that the algorithm developed here is largely insensitive to simulation resolution. This robustness to model resolution is due to the new development of using TROPOMI columns to inform the vertical mixing assumptions used in the algorithm.
Comparisons performed here indicate that fine resolution TROPOMI-inferred surface concentrations agree well with surface observations for annual means and in winter, spring, and autumn. Monitor placement within the grid used to average satellite values combined with increased sub-grid spatial variability in the NO 2 field during summer can also contribute to this bias, even at the relatively fine resolution of TROPOMI (Kharol et al 2015, Judd et al 2019. Changes to TROPOMI instrument settings made in August 2019 have improved its spatial resolution for future observations, with along-track pixel size reduced from 7 km to 5.5 km. This will allow for even greater detail in observing NO 2 plumes. Further improvement to satellite-inferred surface NO 2 concentrations can be expected from future satellite instruments that promise even higher spatial resolutions, such as the geostationary constellation of TEMPO (Zoogman et al 2017), Sentinal-4/UVN (Ingmann et al 2012), and GEMS (Kim 2012). These fine resolution satellite products can greatly improve air quality monitoring from space and will allow for inferring surface NO 2 concentrations at the fine resolution needed for improving studies on health impacts.