OMI NO 2 column densities over North American urban cities : the e ff ect of satellite footprint resolution

Introduction Conclusions References Tables Figures


Introduction
Tropospheric nitrogen dioxide (NO 2 ) is an important component of urban atmospheric chemistry.It is one of the major pollutants affecting humans and the biosphere (Chauhan et al., 2003;Kampa and Castanas, 2008), and works as an important precursor in tropospheric ozone chemistry and aerosol formation.Continuous monitoring of tropospheric NO 2 is important for understanding urban air quality and changes in anthropogenic emissions.NO 2 is also used as an important indicator for traffic and urbanization (Rijnders et al., 2001;Ross et al., 2006;Studinicka et al., 1997).
NO 2 plumes from urban anthropogenic sources, especially from point and mobile sources, usually have a fine structure, as small as a few hundred meters and as large as 10-20 km, as reported in comparisons of column NO 2 based on in situ observations and modeled calculations (Heue et al., 2008;Valin et al., 2011;Ryerson et al., 2013).Heue et al. (2008) used an airborne instrument based on imaging Differential Optical Absorption Spectroscopy (iDOAS) to build a two-dimensional (2-D) distribution model of urban plumes.By comparing NO 2 column densities over the industrialized South African Highveld with OMI and SCIAMACHY measurements, they demonstrated that iDOAS shows strong enhancements close to industrial areas, 4-9 times higher than measurements from OMI and SCIAMACHY.Previous studies have demonstrated that modeled ozone production depends strongly on the spatial scale of the modeling grid due to the nonlinear dependence of ozone production on NO x concentration (e.g., Cohan et al., 2006;Gillani and Pleim, 1996;Liang and Jacobson, 2000;Sillman et al., 1990); therefore, an accurate comparison of urban NO 2 plumes in fine scale is crucial for understanding surface ozone chemistry and air pollution over urban cities.Using 1-D and 2-D models, Valin et al. (2011) computed the resolution-dependent bias in the predicted NO 2 column, demonstrating large negative biases over large sources and positive biases over small sources at coarse model resolution.
The inhomogeneity of urban NO 2 plumes within the scale of satellite footprint pixels is of rising interest as satellitebased measurements are being compared with fine-scale modeling (Beirle et al., 2004(Beirle et al., , 2011;;Hilboll et al., 2013).Richter et al. (2005) showed that there are considerable differences between GOME and SCIAMACHY observations for locations with steep gradients in the tropospheric NO 2 columns; on the other hand, these observations agree very well over large areas of relatively homogeneous NO 2 signals.Hilboll et al. (2013) argued that these effects result from spatial smoothing that differs depending on the ground resolution of the instruments; therefore, the inherent spatial heterogeneity of the NO x fields must be considered when studying them over small, localized areas.Hilboll et al. (2013) also presented approaches to account for instrumental differences while preserving individual instruments' spatial resolutions.In comparing GOME and SCIAMACHY, they used an explicit climatological correction factor to convolve GOME pixels (40 km × 320 km) with better-resolution SCIAMACHY (30 km × 60 km) data, producing a combined data set for studying long-term trends.In this study, we try to investigate and to quantify the uncertainty resulting from the geometry of OMI satellitebased Nitrogen dioxide vertical column density (NO 2 VCD) measurements by comparing these data to a fine-scale regional quality model.First, a pseudo-OMI data set is built from the outputs of fine-scale model simulations, and then these results are compared to model data in order to quantify the impact from pure differences in geometry.Second, we extend the basic concept of Hilboll et al. (2013) to apply spatial-distribution information from the fine-scale model to the OMI measurements, and demonstrate how the new approach adjusts the original OMI measurements.Satellite and model data are described in Sect. 2. Construction of pseudo-OMI data and the quantification of the impact of pixel geometry are discussed in Sect.3. In Sect.4, the downscaling approach is discussed; Sect. 5 concludes and discusses the implications of findings for emission policy decision-making.

OMI
We utilized OMI tropospheric NO 2 VCD data, retrieved by the Royal Netherlands Meteorological Institute (KNMI).The OMI instrument, onboard NASA's Earth Observing System Aura satellite, is a nadir-viewing imaging spectrograph measuring backscattered solar radiation with a measuring wavelength ranging from 270 to 500 nm and with a spectral resolution of about 0.5 nm.Its telescope has a 114 • viewing angle, which corresponds to a 2600 km wide swath on the surface.In its normal global operation mode, its pixel size is 13 km (along) × 24 km (across) at nadir, which can be reduced to 13 km × 12 km in zoom mode (Levelt et al., 2006).Data were downloaded from the Tropospheric Emission Monitoring Internet Service (TEMIS; http://www.temis.nl/airpollution/no2.html) of the European Space Agency (ESA).DOMINO version 2.0 retrieval based on the Differential Optical Absorption Spectroscopy (DOAS) technique was used for the study.We disregarded data pixels with cloud fractions over 40 % or other contaminated pixels using quality flags.Details on the NO 2 column retrieval algorithms and error analysis are described in Boersma et al. (2004Boersma et al. ( , 2007)).

NAQFC
The US National Air Quality Forecast Capability (NAQFC) provides daily, ground-level ozone predictions using the Weather Forecasting and Research non-hydrostatic mesoscale model (WRF-NMM) and Community Multi-scale Air Quality (CMAQ) framework across the Contiguous United States (CONUS) with a 12 km resolution domain (Chai et al., 2013;Eder et al., 2009).In our analysis, we used the experimental version of NAQFC, which uses WRF-NMM with B-grid (NMMB) as a meteorological driver and the Carbon Bond (CB05) chemical mechanism.Meteorological data are processed using the PREMAQ, which is a special version of the Meteorology-Chemistry Interface Processor (MCIP) designed for the NAQFC system.lected during September 2013.The blue line indicates size distribution counts for each 50 km 2 bin, while the red line indicates the cumulative distribution of the OMI pixel sizes.
The size distribution has high occurrences near 300 km 2 , as expected from the OMI's resolution at the nadir (that is, 13 × 24 = 312).However, many pixels still have larger sizes; around half of total pixels are larger than 500 km 2 , and 20 % of total pixels are larger even than 1000 km 2 .Geographical coverage rapidly increases with pixel size, so deciding a threshold for footprint-pixel sizes and available coverage may present a serious dilemma.
Figure 2 shows the relationship between OMI footprintpixel size and actual geographical coverage over the CONUS.With 1 July 2011 data, 25 % of OMI pixel sizes are less than 342 km 2 , and they cover 1.4 % of the CONUS domain.CONUS coverage changes to 11.5, 24.0, and 58.8 % when 50, 75, and 100 % of OMI pixels are used, respectively.Using only finer data may provide detailed information, but they represent only a small part of all the data.If we also use coarser-resolution data, they provide more coverage but tend to be biased over areas with spatial gradient, as discussed in the previously mentioned studies (Hilboll et al., 2013).We therefore estimated the theoretical range of biases deriving from this geometric effect by constructing a pseudo-OMI data set out of a fine-scale model.Using the fine-scale regional CMAQ simulations and assuming this model represents a true world, we constructed a data set to mimic OMI instrument measurement of this modeled world.
In order to construct the pseudo-OMI data, we utilized a conservative spatial regridding technique to perform a lossless conversion of gridded modeling outputs into actual OMI footprint pixels.Figure 3 demonstrates the concepts of conservative regridding.The gray grid cells are 12 km grid cells for modeling -zoomed on the Houston region as an example -and the blue lines are actual OMI pixel coverage.The blue, shaped pixel is an example of an actual OMI pixel, while the pink boxes are model grid cells overlaid by the example OMI pixel.The numbers in the grid cells are calculations of the fractional area overlaid by the OMI pixel for each cell using the Sutherland-Hodman polygon-clipping algorithms available from the Interactive Data Language (IDL)-based Geospatial Data Processor (Kim et al., 2013); 0.74 means the OMI pixel covers 74 % of the corresponding grid cell.The pseudo-OMI value for the blue OMI pixel area in Fig. 3 can be estimated as where i and j are indices for the model grid cell and OMI pixel, respectively, and f i,j indicates the fractional area of cell i overlaid by OMI pixel j .Figure 4 compares the spatial distributions of CMAQ NO 2 VCDs (assumed to be a true world) and pseudo-OMI (pOMI) NO 2 VCDs, along with the difference and percentage difference, (pOMI-CMAQ)/CMAQ × 100, over the northeastern US.It is evident that there are prominent differences between the original fine-scale modeled NO 2 VCDs and reconstructed pseudo-OMI distribution, especially over and near urban locations.As expected from the smoothing effects of larger pixel sizes, pOMI shows a slightly smoothed transition from urban cores to suburban, and most of the sharp peaks near small cities are gone in the pOMI distribution.As already mentioned, this is purely a result of geometry.
We can see that, for all the major cities, pOMI underestimates the actual NO 2 VCD values while overestimating at the boundaries of major cities, as clearly seen in the New York, Pittsburgh, Philadelphia, Baltimore, and Washington D.C. areas.This effect is also prominent in locations with small but strong NO x emission sources, such as power plants, or small cities such as Norfolk, VA.It should be noted that these discrepancies result from purely geometric effects deriving from the OMI's designed pixel sizes and are around ±5-10 × 10 15 molec cm 2 , with 20-30 % under-or overestimation biases for major cities and more than 100 % under-or overestimation for local cities like Norfolk and Richmond, VA.In the next section, we introduce a new approach -the conservative downscaling method -to reduce this effect of resolution due to varying OMI footprint-pixel sizes.

OMI NO 2 VCD downscaling
As described in the previous section, urban NO 2 plumes usually have too fine of a spatial structure compared to OMI's measuring footprints.In this section, we introduce a new ap-  proach for adjusting those geometric effects.Downscaling is a common concept in meteorological simulations, used especially in global circulation models to provide initial and boundary conditions for regional models.We use a similar concept, describing a downscaling method in data processing as a special case of spatial regridding that provides further details through the incorporation of additional information into a set of coarse-resolution data.This approach differs from simply increasing the resolution, as the raw, coarse data are restructured using a set of logics, analogous to a regional meteorological model that downscales global meteorology using its own set of physical and thermal field balances.Conceptually, we use a calculation process reversed from that used to construct the pseudo-OMI data set.
Figure 5 graphically depicts the steps of conservative downscaling from OMI pixels.Figure 5a shows actual OMI NO 2 VCD measurements over Los Angeles on 4 May 2010, and Fig. 5b shows the corresponding CMAQ NO 2 VCDs calculated from NAQFC modeling outputs at the same time and location.As readers can easily see, OMI footprint pixels are much bigger (∼ 650 km 2 ) than the CMAQ grid cells (12 × 12 = 144 km 2 ).As a result, an OMI pixel can overlay more than 10 CMAQ grid cells, as demonstrated in Fig. 5b (black box representing the OMI pixel).We collected those CMAQ pixel values and then normalized them so that the total value of each grid cell sums to one.We call this a spatialweighting kernel (Fig. 5c), and we apply this weighting kernel to the original OMI measurement.As a result, we generate a reconstructed OMI pixel with a finer structure but without any loss of original quantity.Summing the reconstructed pixels gives the original OMI pixel measurement.It should be noted that we strictly apply this method conservatively; theoretically, if there are no missing or duplicated pixels, the quantity of the original data is numerically preserved.This method can be summarized as fusing a satellite-measured "quantity" with modeled "spatial information"; the strength of the modeled NO 2 field does not at all affect the result.
As expected, the accuracy of this method indeed depends on the model's performance, especially regarding its wind-field simulation and inputs of emission source locations, so this method clearly has its own limitation.Considering the uncertainties resulting from emission source locations, the air-quality community has had an excellent archive of geographical information about the geophysical locations of emission sources thanks to the efforts of US EPA, although the strengths of these sources are somewhat highly uncertain.As just described, however, the downscaling method is not affected by emission strength, so we do not think that the uncertainty associated with known emission sources is very high.On the other hand, the use of a downscaling method can be limited when there are uncertainties in emission inventory information such as unknown emission sources or removal of known sources.Wind field is important for simulating NO 2 plume transport.With the short lifetime of NO 2 , especially during summer, the spatial distribution of NO 2 plumes is strongly determined by the location of emission sources.Improving information about emission-source locations would somewhat improve the model, but it is more important to note that the downscaling method tends to convert the error characteristics.Near urban cores, OMI's coarse footprint resolution always causes unidirectional, systematic biases, with underestimation near urban cores and overestimation at the urban boundary.Using the downscaling method, these systematic biases from resolution are converted to random bias from wind-field error.Since these biases are random, they may be corrected by averaging over a certain time period, unlike the systematic bias resulting from resolution.

2010 CalNex campaign case
We applied the downscaling technique to compare the OMI and downscaled OMI with aircraft-borne measurements from the California Research at the Nexus of Air Quality and Climate (CalNex) campaign.The CalNex field study was conducted in California from May to July 2010 and focused on atmospheric-pollution and climate-change issues, including an emission inventory, atmospheric transport and dispersion, atmospheric chemical processing, cloud-aerosol interaction, and aerosol radiative effects (Ryerson et al., 2013).Here, we compared NO 2 VCD observations from the campaign's P3 flight with corresponding OMI measurements using both the standard and downscaling methods.More detailed descriptions regarding data preparation and a discussion of the influence of environmental inhomogeneity and urban NO 2 plumes are provided by Judd et al. (2016) Figure 6 shows scatter-plot comparisons between the P3 measurements and OMI NASA standard product (Fig. 6a), OMI KNMI product (Fig. 6b), and OMI KNMI downscaled (Fig. 6c) for 3 days: 4, 7, and 16 May 2010.As reported, the OMI NO 2 VCD tends to underestimate near the Los Angeles urban area.The KNMI retrieval showed a slightly better comparison with slope = 0.73 and R = 0.85, while the downscaled product clearly showed the best agreement with the P3 measurements, R = 0.88 and slope = 1.0.Deviations still remain from a true one-to-one line even with the downscaling method; these are possibly caused by errors in wind-field simulation.We expect these random errors to average out as the amount of available data increases.The downscaling method seems to work even with daily timescale data sets.
Figure 7 compares OMI NO 2 VCD spatial distributions for the original KNMI products with downscaled products for 4 May 2010, the day when the downscaling method gave the most dramatic changes in the spatial distribution.In the original retrieval, OMI pixels were coarse and mostly smoothed out over Los Angeles.However, by applying the downscaling technique, the adjusted OMI data show a shape much closer to the urban boundary and enhanced NO 2 VCD values at the center of Los Angeles, agreeing very well with the P3 aircraft measurements.On 7 May, the downscaling method reproduced several peak values very well but failed to generate a clean spot at the edge of Los Angeles.On 16 May, the changes from downscaling are not dramatic due to generally low NO 2 concentrations due to less urban traffic on Sunday (e.g., the weekend effect), but the downscaling method still showed slight enhancement (shown in supplementary plots).

Comparison with NAQFC
Comparing modeled NO 2 VCDs to satellite-observed NO 2 VCDs has been a popular way to evaluate the NO x emission inventory.Since modeled NO 2 VCDs and satellite NO 2 VCDs have different optical and vertical properties, some researchers have used additional processing to fairly compare satellite and modeled column densities.In this section, we performed vertical and spatial adjustments by applying averaging kernel (AK) information in conjunction with the downscaling technique.First, we compared NAQFC NO 2 VCDs with and without AK to OMI NO 2 VCDs with and without downscaling processing.
The sensitivity of the instrument to tropospheric tracer density is highly height dependent.Since the measured tracer profile may have large systematic errors as a result, the retrieved tracer columns should be interpreted with proper additional information (Eskes and Boersma, 2003) of the target species for each layer throughout the atmospheric column (Bucsela et al., 2008) and can be applied to a modeled atmospheric column for a fair comparison with satellite retrievals.For each OMI DOMINO product pixel, 34 layers of AKs are provided.We first converted total AKs to tropospheric AKs, AK trop , by applying the total air mass factor (AMF) and tropospheric AMF, and we then applied AK trop to model layers before vertically integrating, as described by Herron-Thorpe et al. (2010).When multiple OMI pixels overlaid a model grid cell, we conducted the conservative spatial remapping method explained above.Figure 8 compares the monthly averaged NO 2 VCD distributions for CMAQ without and with AK (Fig. 8a and b, respectively) and for OMI NO 2 VCDs without and with downscaling (Fig. 8c and d, respectively).In general, AK-applied CMAQ NO 2 VCDs tend to be slightly lower than CMAQ NO 2 VCDs without AK information.On the other hand, while OMI NO 2 VCDs without downscaling (DS) shows a much smoother pattern, the DS-applied OMI reconstructs the sharp spatial structures near urban areas.DS-applied OMI NO 2 VCDs are evidently able to construct sharp gradients near cities, and especially near mid-size cities.
Figure 9 compares CMAQ and OMI NO 2 VCDs using AK and DS methods together.Figure 9a shows a scatter-plot comparison between CMAQ and OMI NO 2 VCDs at US Environmental Protection Agency Air Quality System (AQS) surface-monitoring site locations during September 2013.In this comparison, CMAQ NO 2 VCDs are much higher compared to OMI NO 2 VCDs, implying that the CMAQ simulation possibly overestimates NO x emissions.Figure 9c compares OMI and CMAQ NO 2 VCD with AK information applied; estimated CMAQ NO 2 VCD is reduced, showing better agreement with OMI NO 2 VCD.Readers may notice that high CMAQ pixels are shifted to the left.On the other hand, applying the DS method to OMI shifts OMI pixels vertically (Fig. 9b).Finally, in Fig. 9d, both AK and DS methods are applied; this comparison shows the best agreement between OMI and CMAQ NO 2 VCD pixels.Its correlation coefficient R = 0.89 and the slope of line fit is 0.59.Clearly, the application of the AK and DS methods not only improved the satellite-model comparison in the high NO 2 concentration range but also significantly improved the comparison in the low NO 2 range (i.e., 0-10 × 10 15 molecules cm −2 ), implying that this method can help interpret NO x emission in major and mid-size cities.We have conducted the same analyses for all summer months in 2013 and 2014, and the results are consistent.The differences in spatial distributions between monthly averaged OMI and CMAQ NO 2 VCDs during September 2013 are shown in Fig. 10.Positive values indicate that CMAQ NO 2 VCDs are higher than OMI VCDs, which should likely be interpreted as an overestimation of the NO x emission inventory used in the CMAQ modeling.The difference between the original OMI and CMAQ NO 2 VCDs show strong positive values over most urban locations (Fig. 10a).Applying AK (Fig. 10b) and DS (Fig. 10c) reduce positive biases for major and middle-to-small cities, showing the best agreement when both AK and DS are included.NO 2 VCDs are still overestimated over major cities -New York, Philadelphia, Detroit, and Chicago -as is expected from the continuous trend of NO x emission reduction, but they are much weaker than in the original comparison.Slight overestimations over Baltimore, Washington D.C., Richmond, and Norfolk have almost disappeared.We also notice broad underestimation of NO 2 VCDs over Pennsylvania and West Virginia, which might be related to recent changes in this region, but detailed analysis is beyond the scope of this study.Another interesting feature is that there are spots of underes-timation over small cities or local power plants; we therefore suspect the DS method slightly overweighted urban emissions due to the lack of soil NO x emissions in the current modeling system.

Conclusions
This study reports that satellite footprint sizes might cause a considerable effect on the measurement of fine-scale urban NO 2 plumes.Comparing OMI NO 2 VCDs over North American urban cities to a 12 km CMAQ simulation from NOAA NAQFC, we found that OMI footprint-pixel sizes are too coarse to resolve urban plumes, resulting in possible underestimation (and overestimation of model NO 2 VCDs) over the urban core and overestimation outside.In order to quantify this effect of resolution, we first conducted a perfect-model experiment.Pseudo-OMI data were constructed using finescale outputs of a model simulation, assuming that the finescale model output is a true measurement.To match the footprint coverage from real OMI pathways, we conducted conservative spatial regridding with the corresponding fine-scale model outputs to generate a set of pseudo OMI pixels.
When compared to the original data, the pseudo-OMI data clearly showed smoothed signals over urban locations, with 20-30 % underestimation over major cities and up to 100 % bias over smaller urban areas.We then introduced conservative downscaling of OMI NO 2 VCDs using spatial information from the fine-scale model to adjust the spatial distribution, also applying averaging kernel (AK) information to adjust the vertical structure.Four-way comparisons were conducted between OMI with and without downscaling and CMAQ with and without AK information.Results show that OMI and CMAQ NO 2 VCDs show the best agreement when both downscaling and AK methods are applied, with correlation coefficient R = 0.89.
These results should be considered when using satellite data in the evaluation of emission inventories and translating these data into decision-making around emission policy.Table 1 shows a summary of the comparisons between OMI and CMAQ NO 2 VCDs described in Figs. 8 and 9.When CMAQ without AK and OMI with DS are compared, the percentage difference is (6.43 − 3.61)/3.61× 100 = 78 %, implying that the current emission inventory likely overestimates NO 2 VCDs.Comparing between OMI with DS and CMAQ without AK or between OMI without DS and CMAQ with AK still implies that the current emission inventory is possibly overestimating.However, when both vertical and spatial profiles are adjusted using the AK and DS methods, a slight underestimation is found, −7 %, in modeled NO 2 VCDs over AQS monitoring locations, implying that the current inventory possibly underestimates emissions.This may represent an important implication for how spatial information should be considered when investigating fine-scale phenomena such as urban NO 2 plumes.
Without question, satellite observations are very useful with their large coverage supplementing sparse surfacemonitoring sites.Interpretation of satellite-based measure-ment, however, should be performed cautiously with consideration of the instrument's characteristics, especially when translating results into policy-making.We expect our current study to provide a reference for the uncertainty of satellite-based information regarding local or regional pollutants, especially until we have the measurement data at a more enhanced resolution that will be provided by future satellites, such as Tropospheric Emissions: Monitoring of Pollution (TEMPO), Tropospheric Monitoring Instrument (TROPOMI), and Geostationary Environmental Monitoring Spectrometer (GEMS).
The Supplement related to this article is available online at doi:10.5194/gmd-9-1111-2016-supplement.

Figure 1 .
Figure 1.Size distribution of OMI pixel footprint (blue) and its cumulative percentile (red) during September 2013.
Emissions are projected to 2012 level using Department of Energy Annual Energy Outlook and EPA Cross-State Air Pollution Rule (CSAPR) from the 2005 National Emission Inventory.Detailed information on the emission is available from Pan et al. (2014) and references within.

Figure 3 .
Figure 3. Calculation of pseudo-OMI (pOMI) data.Blue boxes are actual OMI pixel footprints and the gray cells are 12 km grid cells.Fraction of cells overlapped by an OMI pixel are shown, and pOMI (sky blue) data are estimated by a weighted average of the corresponding grid cells (pink).

Figure 8 .
Figure 8. Spatial distributions of (a) CMAQ NO 2 VCDs without AK and (b) with AK; (c) OMI NO 2 VCDs without downscaling and (d) with downscaling during September 2013.

Figure 9 .
Figure 9.Comparison of OMI and CMAQ NO 2 VCDs for (a) OMI and CMAQ with AK, (b) downscaled OMI and CMAQ with AK, (c) OMI and CMAQ with AK, and (d) downscaled OMI and CMAQ with AK during September 2013.

Figure 10 .
Figure 10.Comparisons of OMI and CMAQ NO 2 VCD spatial distributions in the northeast US region during September 2013.

Table 1 .
Comparison of OMI and CMAQ NO 2 VCD monthly averages (September 2013) at AQS sites.