Urban NO_x emissions around the world declined faster than anticipated between 2005 and 2019

OMI is a passive spectrometer launched on the NASA Aura satellite in July 2004 and has been providing global observations of NO₂ atmospheric column densities since 1 October 2004 (Levelt et al 2018). Aura is situated in a polar-orbiting flight path approximately 700 km above the Earth's surface with a Equatorial crossing time of 13:45 local time (Levelt et al 2006). Each day Aura has 14–15 orbits and was designed to have global coverage every day. Since the development of the 'row anomaly' in 2007 (Dobber et al 2008), which obstructs ∼30% of the field of view, it now has global coverage once every 2–3 days. OMI NO₂ slant column densities are derived from backscattered radiance measurements in the 402–465 nm spectral window of the UV–Vis spectrometer (Marchenko et al 2015, Lamsal et al 2021). OMI measures backscatter radiances in a 2600 km swath with a nadir (center of the swath) pixel size of 13 × 24 km². The consistency of the data record over a 16+ year period has allowed for numerous long-term evaluations of trace gas species (McLinden et al 2016, Levelt et al 2018, Liu et al 2018, Abad et al 2019, Shen et al 2019, Silvern et al 2019, Goldberg et al 2019a).

OMI NO₂ data version 4.0 are operationally released by NASA (Lamsal et al 2021) (https://disc.gsfc.nasa.gov/datasets/OMNO2_003/summary). The version 4.0 update includes a high-resolution surface reflectivity product in the calculation of the air mass factor and a recently updated cloud scheme, but still has a low bias of approximately 50% in urban areas when compared to column observations from in situ measurements. We filter the Level 2 OMI tropospheric column NO₂ data to ensure only valid pixels are used. Daily pixels with solar zenith angles ⩾80°, cloud radiance fractions ⩾0.5, and surface albedo ⩾0.3 are removed as well as the five largest pixels at the swath edges (i.e. pixel numbers 1–5 and 56–60). We also remove any pixel flagged by NASA including pixels with missing values and those affected by the row anomaly. The daily data are then re-gridded to a global 0.1° × 0.1° grid.

The uncertainty in any daily measurement in the operational data has been assigned to be approximately 1.0 × 10¹⁵ molecules cm⁻² (Krotkov et al 2017). This equates to roughly a 5%–20% uncertainty over polluted areas. However, because we are oversampling over many days (>100 days), we assume that random errors will cancel due to the large number of observations used (Russell et al 2010). This leaves only the systematic errors, such as the air mass factor bias in urban areas, which we discuss in section 2.4.

We use a top-down inverse statistical modeling technique to derive NO_X emissions from a combination of satellite data and re-analysis meteorology. In this method, all OMI NO₂ data over individual city centers or 'hotspots' are compiled and rotated based on the daily-observed wind direction, so that the oversampled plume is decaying in a single direction. We utilize the 100 m wind speed and direction from the ERA5 re-analysis dataset (Hersbach et al 2020) generated at 0.25° × 0.25°. For each city we use the closest gridded value without interpolation.

This top-down method can only be applied when NO₂ is photochemically active and the NO₂ lifetime is short. Wintertime has more erroneous data due to snow cover and the longer NO₂ lifetime during wintertime yields urban plumes that are much more likely to overlap; both factors cause issues with the statistical fit. Therefore, we only use OMI NO₂ data from May to September in the Northern Hemisphere north of 25° N, November–March in the Southern Hemisphere south of 25° S, and all monthly data in Equatorial regions between 25° N and 25° S. We do not expect any significant systematic biases due to this temporal filtering. We aggregate all daily satellite data into 3 year averages; 36 months of data for tropical regions and 15 months of data for extratropical regions. We choose 3 year averages in lieu of 1 year or a shorter timeframe in order to average out the noise in daily measurements and to account for the row anomaly which causes fewer available measurements in the later time record.

Once all daily plumes have been rotated to be aligned as an effective horizontal plume and averaged together during a three year period (usually 100–600 snapshots), we integrate ±0.5° along the y-axis about the x-axis to compute a one-dimensional line density in units of mass per distance. The line densities, which are parallel to the wind direction, peak near the primary NO_X emissions source and gradually decay downwind as the NO_X is transformed into different chemical species or deposited to the surface.

The line densities are fit to a statistical exponentially modified Gaussian (EMG) model (Beirle et al 2011, Valin et al 2013). This particular method was chosen due its ability to convert NO₂ column information into NO_X emission rates while accounting for meteorological influences and due to the multitude of studies verifying the methodology (De Foy et al 2014, Verstraeten et al 2018, Goldberg et al 2019c). A full description of the method can be found in the supplementary. An illustrative example of the method applied to Paris is shown in figure S1 (available online at stacks.iop.org/ERL/16/115004/mmedia). The five output parameters of the statistical fit are the: NO₂ burden, NO₂ background, decay distance, horizontal location of apparent source (ideally at the origin), and sigma of the Gaussian plume. The NO_X emissions rate from the source can be calculated from the NO₂ burden, decay distance, and NO_X /NO₂ ratio, which is assumed to be 1.33 (Beirle et al 2011, Valin et al 2011). In two final adjustments, the derived NO_X emissions are multiplied by a factor of 1.37 (Goldberg et al 2019c) due to a known low bias in urban areas caused by coarse resolution a priori vertical profile information incorporated in the air mass factor, and then by a factor 0.77 to account for an early afternoon high bias in the emissions rate compared to the 24 h average emissions rate reported by annual inventories, using diurnal allocation factors described in Denier Van Der Gon et al (2011); sensitivity analyses of the early afternoon adjustment factor can be found in the supplementary. A discussion of the uncertainties associated with all multiplicative factors are noted in section 2.4 and the supplementary.

We acquired gridded anthropogenic total NO_X emissions data from five widely used inventories and projections (table 1): the Community Emissions Data System (CEDS) (McDuffie et al 2020), the Emissions Database for Global Atmospheric Research (EDGAR) version 5.0 (Crippa et al 2020), the Evaluating the Climate and Air Quality Impacts of Short-Lived Pollutants (ECLIPSE) version 5a (Klimont et al 2017), the Monitoring Atmospheric Composition and Climate CityZen (MACCity) project (Lamarque et al 2010), and the Shared Socioeconomic Pathways (SSP) projections (Riahi et al 2017). The two SSP scenarios displayed herein represent sustainability (SSP 1–1.9) and continued fossil fuel development (SSP 5–8.5) pathways.

The five inventories and their trends by region are shown in the supplementary (figures S2 and S3). In some cases, sectoral emissions needed to be added together to create total anthropogenic NO_X emission files. The inventories report NO_X emissions as 'equivalent' NO₂, but the NO₂/NO_X ratio may vary by urban area, and can be considered a source of uncertainty. None of the inventories projected to 2020 include the effects of the COVID-19 lockdowns on air pollutant emissions.

All NO_X inventories are re-gridded to a common spatial resolution of 0.05° × 0.05°, while retaining all original values at the coarser spatial resolution. A final emissions output file is created which lists the NO_x emissions within radii of 0.1° to 0.75° at 0.05° increments of each city. We match the satellite-derived emissions to the emissions inventories using the sigma of the Gaussian plume which varies spatiotemporally. We assume that the satellite-derived emissions should be matched to a Gaussian radius of 2-sigma from the city center. The 2-sigma radius for all cities are provided in table S3. Varying the city radius between 1.5-sigma and 3-sigma will affect the magnitude comparison by ±30%, but has a lesser effect on the trend comparison (figure S4).

The total error associated with the magnitude of the top-down vs. bottom-up comparison is calculated to be 53%, and is the sum of the quadrature of seven potential sources of error: the tropospheric vertical column measurement in urban areas (30%), the wind speed and direction (25%), the collocation of the spatial extent between the top-down fit and bottom-up emission inventory (30%), the early afternoon to 24 h conversion emissions rate (10%), the 'clear-sky' bias (10%) which for these purposes is a result of emissions being different on clear-sky days compared to cloudy days, the NO_x /NO₂ ratio (10%) (Kimbrough et al 2017), and the random error of the statistical EMG fit (10%) (De Foy et al 2014). This total uncertainty is comparable to Verstraeten et al (2018), who quantified an uncertainty of 55% using this method with OMI NO₂. For the trend analysis, the total uncertainty is much reduced, since the systematic uncertainties in the emissions are consistent throughout the time period, thus leaving only the random EMG fitting error of roughly 10% (De Foy et al 2014). For further information on this method or the uncertainties associated with this method, please see the discussion in the supplementary or other literature (De Foy et al 2014, Lu et al 2015, Verstraeten et al 2018, Goldberg et al 2019c).

Regional NO₂ temporal trends since 2005 have been well-documented (Duncan et al 2016, Krotkov et al 2016, Georgoulias et al 2019). We update the findings here to include the most recent years of annual data, and to narrow the focus on urban NO_X emissions—instead of NO₂ concentrations. We purposefully exclude 2020 data due to the emission anomalies associated with the COVID-19 lockdowns. In 2005, the global regions with the largest anthropogenic emissions were: eastern United States, western Europe, and east Asia. This is documented by both the 'top-down' OMI NO₂ annual average of tropospheric vertical column NO₂ and the 'bottom-up' CEDS inventory (figure 1; note the non-linear colorbars in each panel).

Zoom In Zoom Out Reset image size

Between 2005 and 2012, NO₂ concentrations dropped dramatically (25%–40%) in North America, western Europe, and Japan in response to stringent policies enacted to reduce NO_X emissions. Conversely, in China, India, and the Middle East, a lack of regional policies controlling NO_X emissions yielded a further increase (+10%–50%) in the NO₂ concentrations as compared to 2005 concentrations. Regional signals in Latin America, Africa, and Southeast Asia are mixed primarily due to the influence of biomass burning in these regions. In other locations, such as Central Asia, Northern Africa, and Oceania regional differences are dominated by natural variability due to sparse anthropogenic activities in these areas.

Between 2012 and 2019, NO₂ concentrations either dropped or held steady in most global regions. The largest decreases during this timeframe were in eastern China. In very few regions, were there large obvious increases. NO₂ changes between 2005 and 2019 at the regional level are displayed in figure S5.

Our top-down OMI NO_X calculation converged for 80 (n = 80) global cities for the 15 year period of interest (2005–2019); 16 of them are shown in figure 2. We first performed the analysis on the 97 C40 cities (www.c40.org/cities), and found that the method generally does not work for metropolitan areas with population sizes of <2 million residents because of OMI's lack of sensitivity to their daily NO₂ plumes. We then expanded our analysis to include all global urban areas with metropolitan area populations exceeding 2 million residents. In total, we performed the statistical fit on 189 global cities. In most cities (167 out of 189), the statistical fit converged in at least one out of the five 3 year periods of interest, but many cities did not have a full temporal range or the statistical fit would yield an unreasonably small effective NO₂ lifetime (<0.5 h) and unusually large sigma (>100 km); these instances were excluded from our analysis because discontinuous estimates are harder to screen for reliability and trend consistency. Our method only works for cities isolated from other large cities within a 200 km radius. Examples of cities in which this method does not work due to insufficient isolation are Beijing, Shanghai, Kinshasa, Amsterdam, Boston, and Washington DC. The comparison between top-down OMI NO_X emissions and the five bottom-up emissions inventories for all 80 cities can be found in figures S6−S15. We anticipated that C40 cities, a group of cities pursuing high ambition climate action, would have larger NO_X reductions but we found no statistical difference between the trends in C40 cities and non-C40 cities since 2009 (figure S16).

Zoom In Zoom Out Reset image size

When the 80 cities are grouped by region, patterns begin to emerge (figure 3). In the United States and Canada, top-down OMI NO_X estimates were available for 14 cities (n = 14), and when combined together, matched the bottom-up inventories in both trend and magnitude to within ±10%; therefore, we assert no consistent bias in the urban NO_X inventories for these two countries. Similarly, excellent agreement was generally found in Australia and New Zealand.

Zoom In Zoom Out Reset image size

In Europe (n = 13) and South Korea/Japan (n = 3), the temporal downward trends of NO_X emissions match to within ±10%, but all the inventories appear to be underestimating the magnitude of NO_X emissions. We have two hypotheses for this magnitude disagreement. This could be indicative of an error in the inventory caused by a large fraction of diesel vehicles in these countries, which are known to have been underestimated in the past (Anenberg et al 2017). Another hypothesis is that this could be indicative of daily lifestyle differences which would present itself in the mid-day to 24 h average conversion. For example, if the activities leading to NO_X emissions in these countries are concentrated in the late morning or early afternoon, and less in the morning or evening due to a heavier reliance on public transit, then the mid-day to 24 h average multiplicative conversion factor should be lower. An additional hypothesis for Europe is that a longer NO₂ lifetime due to Europe's extratropical latitudes is not being fully captured in our method.

In China (n = 6), we observe a broad peak in NO_X emissions in the 2009–2012 timeframe, and subsequent reduction since 2012. Between 2012 and 2018, we calculate that urban NO_X emissions decreased 35%, which was similar in magnitude to the urban NO_X reduction between 2006 and 2012 in the United States and Canada (37%). When comparing to the satellite-based estimates, all bottom-up inventories appear to underestimate the rapid decrease in Chinese NO_X emissions. The CEDS inventory, which uses the regionally-compiled MEIC inventory (Zheng et al 2018) with different spatial downscaling proxies, does capture the urban decreases, but not the extent—between 2012 and 2015, the gridded inventory reports a 7% decrease for the cities considered while the satellite data show a 22% decrease over the same 3 year period. On a national scale, reported Chinese NO_X emissions decreased by 17% in the inventory between 2012 and 2015, but this larger decrease was driven by power plants located in rural locations. These findings are consistent with Zheng et al (2018) who also report that OMI NO₂ downward trends are larger than the regional bottom-up inventory. They documented that downward trends in surface NO₂ concentrations between 2012 and 2017 are smaller than the downward trends in the emissions inventory and satellite data; the ultimate reason for the disagreement is still unknown. However, the difference in this study is that we now account for the NO₂ chemical lifetime, which has been documented to change over time and is responsible for some fraction of urban NO₂ changes (Laughner and Cohen 2019).

In the Middle East (n = 11), all bottom-up inventories suggest a consistent increase in urban NO_X emissions between 2006 and 2018, but the top-down OMI NO_X estimates do not support this. Instead, the satellite measurements indicate that NO_X emissions peaked in 2009, with a slow decrease in the following years. While urban NO_X emissions still appear to be larger in 2018 than in 2006, the top-down estimate suggests only 10% larger as compared to 40%–60% larger as reported by the inventories. This discrepancy appears to be mostly driven by four cities (Dubai, Riyadh, Jeddah, and Karachi), which have shown relatively flat NO_X emission changes between 2005 and 2018. In Latin America (n = 11) and Africa (n = 6), the narrative is similar to the Middle East in that projected upwards NO_X emission trends in the later part of the time record (2012–2018) were in fact steady or downward trends. Scant country-level data exists on emissions or their trends for these regions to inform or validate the global inventories.

Top-down urban NO_X emissions are most uncertain in India (n = 8). There are many reasons for this. First, the satellite measurements in India, especially Delhi, are most affected by aerosol interference as compared to other urban areas around the globe (Vohra et al 2021). Because of this bias, top-down urban NO_X emissions are likely biased low in this region. Satellite measurements in India are also biased by the wet monsoon (limited measurements) and dry monsoon (long-range transport from biomass burning may influence the urban calculation). Further, Indian bottom-up inventories sometimes show unrealistic changes, such as the 31% drop in the EDGAR-reported Delhi NO_X emissions between 2011 and 2012. Sharp changes in the inventory at the urban scale are likely due to the downscaling/disaggregation of national emissions because national trends do not have similar sharp changes (figure S3). With that said, urban area NO₂ trends are noticeably different when compared to the relatively rural, but highly industrial areas of east central India (Chhattisgarh and Jharkhand). In the largest Indian cities (e.g. Delhi, Mumbai, and Kolkata), NO₂ trends are relatively flat between 2005 and 2019, while there have been large increases in east central India (figure S5). This may suggest an issue in the spatial disaggregation of emissions as opposed to an error in the national inventory.

When summing all urban areas in our study, (n = 80), we find that satellite-based measurements show a larger decrease in global urban NO_X emissions than currently reported in the inventories or projections (figure 4). Between 2009 and 2018, the satellite measurements indicate that urban NO_X emissions dropped by 3%–4% yr⁻¹, while the inventories and projections suggest drops generally less than 2% yr⁻¹. The OMI observed NO_X reductions in the 2015–2018 timeframe are most similar to the CEDS inventory (2.2% yr⁻¹) and the SSP 1–1.9 projection (2.0% yr⁻¹) as seen in figure 4. The CEDS inventory captures the recent global decreases best, presumably because it relies more heavily on country-inventory data. ECLIPSE (a projection between 2015 and 2020) and MACCity (a projection throughout the entire timeframe) both show a steady decrease over time (∼1% yr⁻¹), but fail to capture the dramatic drops starting 2012. The EDGAR inventory likewise does not capture the drop starting in 2012. We attribute this to an underestimation of Chinese decreases as well as slower increases in developing nations such as Latin America and Africa (Hickman et al 2021) in the 2012–2018 timeframe.

Zoom In Zoom Out Reset image size