Long-term trends in air quality in major cities in the UK and India: A view from space

Air quality networks in cities can be costly, inconsistent, and typically monitor a few pollutants. Space-based instruments provide global coverage spanning more than a decade to determine trends in air quality, augmenting surface networks. Here we target cities in the UK (London and Birmingham) and India (Delhi and Kanpur) and use observations of nitrogen dioxide 20 (NO 2 ) from the Ozone Monitoring Instrument (OMI), ammonia (NH 3 ) from the Infrared Atmospheric Sounding Interferometer (IASI), formaldehyde (HCHO) from OMI as a proxy for non-methane volatile organic compounds (NMVOCs), and aerosol optical depth (AOD) from the Moderate Resolution Imaging Spectroradiometer (MODIS) for PM 2.5 . We assess the skill of these products at reproducing monthly variability in surface concentrations of air pollutants where available. We find temporal consistency between column and surface NO 2 in cities in the UK and India (R = 0.5-0.7) and NH 3 at two of three rural sites in 25 the UK (R = 0.5-0.7), but not between AOD a significant and substantial (3.1 % a -1 ) increase in PM 2.5 . NO 2 , NH 3 and PM 2.5 decline in London and Birmingham are likely due in large part to emissions controls on vehicles. Trends are significant only for NO 2 and PM 2.5 . Reactive NMVOCs decline in Birmingham, but the trend is not significant. There is a recent (2012-2018) steep (> 9 % a -1 ) increase in reactive NMVOCs in London. The cause for this rapid increase is uncertain, but may reflect increased contribution of oxygenated VOCs from household products, the food and beverage industry, and domestic wood burning, with implications for formation of ozone in 35 a VOC-limited city. that we obtain from the slope of SMA regression of hourly PM 2.5 and PM 10 at 6 sites in Birmingham with both measurements. We use a similar approach as applied to NO 2 to assess AOD. Only surface observations around the satellite overpass (12h00-15h00 LT) and with consistent detrended month-to-month variability (R > 0.5) are retained to obtain city-wide monthly mean PM 2.5 . This results in 20 sites in London for 2009-2018 and 5 sites in Birmingham for 2009-2017. Mean midday city-average 290 PM 2.5 for the period of overlap (2009-2017) is 13.7 µg m -3 in London and 11.3 µg m -3 in Birmingham. MODIS AOD monthly means are estimated for London by averaging the pixels centred within its administrative boundary and for Birmingham within and 6.5 km beyond the administrative boundary, as with OMI NO 2 (Section 3.1). We remove months with < 160 observations; equivalent in spatial coverage to 5 OMI pixels at nadir (the threshold used for OMI). After filtering, 53 % of months are removed for London and 72 % for Birmingham, mostly in winter. Fewer months than OMI are retained, as MODIS uses et al., 2019). The trends from satellite observations of HCHO in megacities obtained by De Smedt et al. (2010) using multiple instruments for 1997-2009 are consistent with ours for Delhi (1.6 ± 0.7 % a -1 ), but opposite for London (-0.4 ± 2.1 % a -1 ). There is a shift in the magnitude of the HCHO trend for London around 2011 (Figure 9) from an increase of 0.3 % a -1 (p-value = 0.9) in 2005-2011 to a rapid increase of 9.3 % a -1 (95% CI: 0-26% a -1 ) in 2012-2018. Visually the data suggest a 440 decline in OMI HCHO in 2005-2011, as in De Smedt et al. (2010), but our trend estimate for 2005-2011 is affected by a limited analysis period and large interannual variability.


Introduction
More than 55 % of people live in urban areas and this is projected to increase to 68 % by 2050 (UN, 2019). Air pollution in cities routinely exceeds levels safe for human health (Landrigan et al., 2018). Regulatory air quality monitoring networks, such as those employed in cities in the UK and India, provide detailed data concerning individual species and specific locations, but are labour intensive to operate and maintain, with potential gaps in spatial coverage and discontinuities hindering longer-term 45 trend discovery. Here we assess the ability to use the long record of satellite observations of atmospheric composition to monitor long-term trends in surface air quality in cities in the UK (London, Birmingham) and India (Delhi, Kanpur) of variable size, at a range of development stages, and with air pollutant concentrations that pose greater risk to health than previously thought (Vodonos et al., 2018;Vohra et al., 2021).

50
Our study focuses on two large cities in the UK (London and Birmingham) and two in India (Delhi and Kanpur). Each is at a different stage of development: London is well developed, Birmingham is undergoing urban renewal, Delhi is experiencing rapid development (Singh and Grover, 2015), and Kanpur is a rapidly industrialising city (World Bank, 2014). Air quality policy is well established in the UK and the rapid decline in regulated air pollutants and their precursors has been monitored since 1970. According to the National Atmospheric Emission Inventory (NAEI), precursor emissions of fine particles with 55 aerodynamic diameter < 2.5 m (PM2.5) decreased in 1970-2017 by 1.5 % a -1 for nitrogen oxides (NOx  NO + NO2), 2.0 % a -1 for sulfur dioxide (SO2), and 1.4 % a -1 for non-methane volatile organic compounds (NMVOCs). Primary PM2.5 emissions decreased by 1.6 % a -1 over the same time period compared to a decline of just 0.2 % a -1 for ammonia (NH3) emissions during 1980-2017(Defra, 2019a. In UK cities, vehicles make a large contribution to air pollution year-round, with seasonal contributions from residential fuelwood burning, agricultural activity, and construction, and sporadic contributions from long-60 range transport of Saharan dust (Fuller et al., 2014;Crilley et al., 2015;2017;Harrison et al., 2018;Ots et al., 2018;Carnell et al., 2019). Despite the decline in emissions, many areas in the UK still exceed the legal annual mean limit of NO2 of 40 g m -3 (Barnes et al., 2018), a threshold that may not adequately protect against health effects of long-term exposure to NO2 (Lyons et al., 2020). Many areas will also exceed the annual mean PM2.5 standard, if updated from 25 to 10 µg m -3 , the WHO guideline (Defra, 2019b). Reported annual mean PM2.5 in 2016, obtained as the surface monitoring network average, is 12 g m -3 for 65 London and 10 g m -3 for Birmingham (WHO, 2018). There is increasing concern over emissions of the important PM2.5 precursor, NH3, as there are no direct controls on the agricultural sector, the dominant NH3 source (Carnell et al., 2019). There has even been a recent increase in NH3 emissions of 1.9 % a -1 in 2013-2017 (Defra, 2019a), attributed to agriculture (Carnell et al., 2019).

70
Air quality policy in India is in its infancy compared to the UK. The first air pollution act was passed in 1981; 30 years after the equivalent in the UK. There has been a steady rollout of European-style (Euro VI) vehicle emission standards, starting with Delhi in 2018 and scaling up to the whole country by 2020 (Govt. of India, 2016). Strict controls on coal-fired power plants have been in place since December 2015, but most power plants are non-compliant (Sugathan et al., 2018). National PM2.5 concentration targets have been set at 20-30 % reductions by 2024 relative to 2017 levels (Govt. of India, 2019), but in 2016 75 measured annual mean PM2.5 in Delhi and Kanpur exceeded the national standard (40 g m -3 ) by about a factor of 4: 143 g m -3 for Delhi; 173 g m -3 for Kanpur (WHO, 2018). In Delhi and Kanpur year-round emissions are dominated by vehicles, construction and household biofuel use in the city and industrial activity and coal combustion nearby (Guttikunda and Jawahar, 2014;Venkataraman et al., 2018). Seasonal enhancements come from intense agricultural fires along the Indo-Gangetic Plain (IGP) north of Delhi, frequent firework festivals, and dust storms originating from the Thar Desert and Arabian Peninsula 80 (Ghosh et al., 2014;Parkhi et al., 2016;Yadav et al., 2017;Cusworth et al., 2018;Liu et al., 2018). Like the UK, the agricultural sector is not directly regulated and intense agricultural activity in the IGP contributes to the largest global NH3 hotspot (Warner et al., 2017;Van Damme et al., 2018;Wang et al., 2019).
Surface monitoring networks in cities in the UK and India needed to evaluate city-wide trends in air pollutant concentrations 85 and precursor emissions can be exceedingly sparse and are often short term. To illustrate this, we show in Figure 1 the coverage of surface sites in the four cities that continuously monitor NO2, the most widely monitored air pollutant in both countries.
There are also diffusion tubes and emerging technologies that measure NO2 at low cost, but these are susceptible to biases Castell et al., 2017) and so are excluded. The points in Figure 1 show sites established and maintained by national agencies, local city councils, and academic institutions. These are coloured by multi-year mean NO2 around the 90 satellite midday overpass (12h00-15h00 local time or LT) for our period of interest (2005)(2006)(2007)(2008)(2009)(2010)(2011)(2012)(2013)(2014)(2015)(2016)(2017)(2018). London has the most extensive surface coverage. There can be more than 100 sites operating simultaneously, but many of these are short-term. Most long-term sites are in central London, and southeast London is devoid of stations. Birmingham has eight monitoring stations, but only two operated for the majority of 2005-2018. There are recently established comprehensive air quality monitoring sites in London and Birmingham, but these started operating in late 2018. More than 40 % of the NO2 monitoring stations in Delhi 95 were established in 2018 and there are concerns over data access and quality . Fewer stations in the four cities monitor PM2.5 than NO2 and measurements of NMVOCs are limited to a few short-term intensive campaigns and long-term sites that only measure light (short-chain) non-methane hydrocarbons. Long-term continuous monitoring of NH3 in the UK is limited to hourly measurements at rural European Monitoring and Evaluation Programme (EMEP) sites ( Figure 1) and monthly measurements at UK Eutrophying and Acidifying Pollutants (UKEAP) Network sites. 100 Satellite observations of atmospheric composition (Earth observations) provide consistent, long records (> 10 years) and global coverage of multiple air pollutants, complementing surface monitoring networks with limited spatial coverage and temporal records (Streets et al., 2013;Duncan et al., 2014). These have been used extensively as constraints on temporal changes in surface concentrations of air pollutants and precursor emissions (Kim et al., 2006;Lamsal et al., 2011;Zhu et al., 2014), but 105 typically just targeting 1-2 pollutants. In this work, we consider Earth observations of NO2, formaldehyde (HCHO), NH3, and aerosol optical depth (AOD). HCHO is a prompt, high-yield, ubiquitous oxidation product of NMVOCs used as a constraint on NMVOCs emissions (Miller et al., 2008;De Smedt et al., 2010;Marais et al., 2012;2014b;2014a). AOD has been used to derive surface concentrations of PM2.5 for global assessment of the impact of air pollution on health (van Donkelaar et al., 2006;2010;Brauer et al., 2016;Anenberg et al., 2019). 110 Here we conduct a systematic evaluation of the ability of satellite observations of NO2, NH3, HCHO and AOD to reproduce temporal variability of surface air pollution in the UK and India before going on to apply these satellite observations to estimate long-term changes in air pollution to assess the efficacy of air quality policies in the four cities of interest.

Space-Based and Surface Air Quality Observations
Earth observations of NO2 and HCHO are from the Ozone Monitoring Instrument (OMI), NH3 from the Infrared Atmospheric Sounding Interferometer (IASI), and AOD from the Moderate Resolution Imaging Spectroradiometer (MODIS). There are also observations of SO2 and the secondary pollutant ozone from OMI, but SO2 is below or close to the detection limit yearround for all cities, except in some months in Delhi, and UV measurements of tropospheric column ozone have limited 120 sensitivity to ozone in the boundary layer (Zoogman et al., 2011). TROPOspheric Monitoring Instrument (TROPOMI) sensitivity to SO2 is 4-fold better than OMI, but the observation record is short (October 2017 launch) (Theys et al., 2019). We use hourly observations of NO2 and PM2.5 from the network of surface sites in the four target cities and NH3 from the rural EMEP sites in the UK, to assess whether satellite observations of NO2, AOD, and NH3 reproduce temporal variability of surface air quality. There are no direct reliable measurements of HCHO in the UK and measurements of NMVOCs are limited 125 to a few sites that only measure light ( C9) hydrocarbons. 3.24° W) (Malley et al., 2015;Walker et al., 2019). Instruments at the Harwell site were relocated to Chilbolton 130 Observatory in 2016, providing the opportunity to assess the satellite data at sites with distinct agricultural activity and anthropogenic influence (Walker et al., 2019). There are also passive NH3 samplers in the UK, but these have coarse temporal (monthly) resolution (Tang et al., 2018) and no temporal correlation (R < 0.1) with a previous version of the IASI NH3 product (Van Damme et al., 2015).

Earth Observations of Air Pollution 155
OMI onboard the NASA Aura satellite, launched in October 2004, has a nadir spatial resolution of 13 km  24 km, a swath width of 2600 km, and passes overhead twice each day. OMI is a UV-visible spectrometer and so only provides daytime observations (13h30 LT). Global coverage was daily in 2005-2009 and is every 2 days thereafter due to the row anomaly (http://projects.knmi.nl/omi/research/product/rowanomaly-background.php). We use the operational NASA OMI Level 2 product of tropospheric column NO2 for 2005-2018 (version 3.0; doi:10.5067/Aura/OMI/DATA2017; last accessed 29 160 February 2020) (Krotkov et al., 2017). Total columns of HCHO are from the Quality Assurance for Essential Climate Variables  (De Smedt et al., 2018). We remove OMI NO2 scenes with cloud radiance fraction  50 %, terrain reflectivity  30 % and solar zenith angle (SZA)  85° (Lamsal et al., 2010) and OMI HCHO scenes with processing errors and processing quality flags not equal to zero (De Smedt et al., 2017). This removes scenes with cloud radiance fraction > 60 % and SZA > 80°. We 165 apply additional filtering to remove scenes with cloud radiance fraction  50 % to be consistent with the threshold applied to OMI NO2. This additional filtering removes 16 % of the data for London, 19 % for Birmingham, 7 % for Delhi, and 8 % for Kanpur.
IASI on the polar sun-synchronous Metop-A satellite, launched in October 2006 is an infrared instrument with a morning 170 (09h30 LT) and nighttime (21h30 LT) overpass. It provides global coverage twice a day with circular 12 km diameter pixels at nadir and a swath width of 2200 km. We use observations for the morning only, when the thermal contrast and sensitivity to the boundary layer is greatest (Clarisse et al., 2010;Van Damme et al., 2014). We use the Level 2 reanalysis product of total column NH3 (version 3R) obtained with consistent meteorology (ERA5) for clear-sky conditions (cloud fraction < 10 %) (Van Damme et al., 2020). The earlier IASI NH3 product version (version 2R) was shown to be consistent with ground-based 175 measurements of total column NH3 at 9 global sites (Dammers et al., 2016).
The MODIS sensor onboard NASA's Aqua satellite, launched in May 2002, has a swath width of 2330 km, crosses the Equator at 13h30 LT and provides near-daily global coverage. We use the Level 2 Collection 6.1 Dark Target daily AOD product at 550 nm and 3 km resolution Wei et al., 2019) (https://ladsweb.modaps.eosdis.nasa.gov/; last accessed 29 180 February 2020). We use only the highest quality AOD data (quality assurance flag of 3) Remer et al., 2013;Gupta et al., 2018).

Consistency between Earth Observations and Surface Air Pollution
Earth observation products retrieve column densities of pollutants throughout the atmospheric column (total for HCHO, AOD and NH3; troposphere for NO2), and are compared in what follows to surface concentrations from the surface monitoring 185 network sites. This is to evaluate whether monthly variability in the column reproduces variability in surface concentrations before going on to use the satellite observations to quantify long-term trends in air pollution in the four cities. The majority of the enhancement in the column, with the exception of events like long-range transport, is near the surface (Fishman et al., 2008;Duncan et al., 2014). Sources of errors in retrieval of HCHO and NO2 column densities include uncertainties in simulated vertical profiles, and presence of clouds and aerosols (Boersma et al., 2004;Lin et al., 2015;Zhu et al., 2016;Silvern et al., 190 2018). Retrieval of NH3 column densities from IASI relies on thermal contrast between the Earth's surface and atmosphere and a sufficiently large training dataset (Whitburn et al., 2016;Van Damme et al., 2017). Errors in retrieval of AOD include uncertainties in aerosol properties and atmospheric conditions in matching simulated and observed top-of-atmosphere radiances from single viewing angle instruments like MODIS (Remer et al., 2005;Levy et al., 2007;. To the extent that errors are random, these are reduced with temporal and spatial averaging . 195 In what follows, city-average OMI NO2 and MODIS AOD are compared to representative city-average surface concentrations of NO2 in all four cities, and PM2.5 in London and Birmingham. IASI NH3 is compared to coincident surface observations of NH3 at UK EMEP sites ( Figure 1).

Assessment of OMI NO2
Data for NO2 in the UK include 152 monitoring sites in London, 8 in Birmingham, 37 in Delhi, and 2 in Kanpur ( Figure 1).
The data we use for London and Birmingham have been independently ratified, but we still find and remove spurious NO2 observations. These include persistent (> 24 hours) low (< 1 µg m -3 ) values that do not exhibit diurnal variability. This occurs at fewer than 10 % of the sites and accounts for at most 1 % of the data at these sites. We identified that NO2 data from DPCC 205 and CPCB (Delhi) and from UPPCB (Kanpur) networks are inconsistently reported in either ppbv or g m -3 . As information on the units of the individual data are not provided, we determine whether NO2 is reported in ppbv or g m -3 by regressing total NOx (reported throughout in ppbv, following the CPCB protocol (CPCB, 2015)) against the sum of the reported NO and NO2. We identify that NO2 reported in ppbv (29 % of DPCC, 10 % of CPCB and 74 % of UPPCB data) populates along the 1:1 line and so we convert these to g m -3 using 1.88 g m -3 ppbv -1 . The same unit inconsistency does not exist for the IMD 210 NO2 data. These are reported throughout in ppbv and so are converted to g m -3 .
We only consider surface observations coincident with the OMI record (2005)(2006)(2007)(2008)(2009)(2010)(2011)(2012)(2013)(2014)(2015)(2016)(2017)(2018), around the satellite overpass (12h00-15h00 LT). We find that NO2 declines at most sites in London (ranging from -0.8 to -3.6 % a -1 ) and Birmingham (-1.1 to -3.8 % a -1 ), with the exception of a few sites influenced by local sources. These include Marylebone Road in central London and 215 Moor Street in Birmingham City Centre. Both are impacted by dense traffic and development projects (Carslaw et al., 2016;Harrison and Beddows, 2017). We find that NO2 increases in Moor Street by 6.8 % a -1 from 2013 to 2017. There are too few long-term sites in Delhi and Kanpur to determine trends at individual sites. We do not filter out sites based on site classification, as this information is not readily available for sites in India. Instead, we remove sites influenced by local effects and not consistent with month-to-month variability representative of the city. This we do by detrending surface NO2 at each site, cross-220 correlating the detrended data for each site, and selecting sites with consistent month-to-month variability (R > 0.5) in the detrended data. The original surface NO2 (including the trend) at the selected sites are then used to obtain city-average monthly mean NO2 for comparison to OMI NO2.
The selected sites are shown as triangles in Figure  We sample satellite observations within the administrative boundaries of the four cities ( Figure 1) to capture the domain that policymakers would target and assess. This is extended a few km beyond the administrative boundary for Birmingham, as otherwise there are too few observations due to frequent clouds and small city size (~300 km 2 ). Error-weighted OMI NO2 monthly means are estimated for individual pixels centred within the administrative boundaries (including 6.5 km beyond for 235 Birmingham). Months with < 5 observations are removed. The number of months retained is 77 % for Birmingham, > 90 % for London, and > 95 % for Delhi and Kanpur. tropospheric column and surface NO2 in locations with large seasonal shifts in temperature and solar insolation include reduced photolysis rates leading to longer NOx lifetime in winter than summer (Boersma et al., 2009;Kenagy et al., 2018;Shah et al., 2020) and a lower mixed layer height in winter than summer contributing to accumulation of pollution. Maximum mixed layer height for London is 900 m in winter compared to 1500 m in summer (Kotthaus and Grimmond, 2018). The slope for Birmingham in winter (0.43 × 10 15 molecules cm -2 (µg m -3 ) -1 ) is steeper than that for non-winter months (0.27 × 10 15 molecules 245 cm -2 (µg m -3 ) -1 ), but the difference is not significant. The surface NO2 measurements are also susceptible to interferences (positive biases) from thermal decomposition of NOx reservoir compounds such as peroxyacetyl nitrates in chemiluminescence instruments that use heated molybdenum catalysts (Dunlea et al., 2007;Reed et al., 2016). The effect is worse in winter than summer in London and Birmingham due to abundance of NOx reservoir compounds in winter (Lamsal et al., 2010). OMI and surface NO2 monthly variability is consistent (R = 0.51-0.71), except for London in winter (R = 0.33). The correlation degrades 250 (R = 0.40 for London, R = 0.54 for Birmingham) if all months are considered. The seasonal dependence of the relationship between satellite and surface NO2 affects the ability to use OMI NO2 to infer seasonality in the underlying NOx emissions. The same consistency in monthly mean OMI and surface NO2 in non-winter months (R  0.6) has also been found over the UK city Leicester (surface area 73 km 2 ) (Kramer et al., 2008). Data for all months are used for Delhi and Kanpur, as there is less variability in mixed layer height in India than the UK. Seasonal mean maximum planetary boundary layer height in Delhi 255 varies from 1200 m in winter to 1400 m during monsoon months (Nakoudi et al., 2019). Month-to-month variability in tropospheric column and surface NO2 (Figure 2) is consistent in Delhi (R = 0.55) and Kanpur (R = 0.52). OMI NO2 exhibits much greater variability for an increment change in surface NO2 in the UK than in India, resulting in order-of-magnitude lower slopes for Delhi and Kanpur (0.033 and 0.039 × 10 15 molecules cm -2 (µg m -3 ) -1 ) than for London and Birmingham (0.35 and 0.27 × 10 15 molecules cm -2 (µg m -3 ) -1 ) (Figure 2). This difference is likely due to a combination of representativeness of surface 260 sites and systematic biases in the OMI NO2 retrieval. In Delhi, the proportion of sites used in Figure 2 that measure the relatively lower concentration range of NO2 (annual mean NO2 < 50 g m -3 ) is just 20 % compared to 74 % for London, leading to a positive bias in city-average surface NO2 in Delhi. In Kanpur, we use only one site located 600 m from a national highway.

Assessment of IASI NH3
Figure 3 compares monthly mean IASI and surface NH3 at the three UK EMEP sites. IASI is sampled up to 20 km around the surface site following the approach of Dammers et al. (2016) and surface observations are sampled around the IASI morning overpass (08h00-11h00 LT) on days with coincident IASI observations. As with NO2, only months with more than 5 270 observations are used. 38 % of months are retained for Auchencorth Moss, 62 % for Harwell and 61 % for Chilbolton Observatory. For the months retained, average NH3 is 1.6 µg nitrogen (N) m -3 for Auchencorth Moss, 2.5 µg N m -3 for Harwell and 6.1 µg N m -3 for Chilbolton Observatory. Chilbolton is southwest of mixed farmland, contributing to levels of NH3 about 3 times higher than at Harwell (Walker et al., 2019). Harwell has more dynamic range in NH3 and stronger correlation (R = 0.69) than the other two sites (R = 0.37 for Auchencorth Moss; R = 0.50 for Chilbolton Observatory). Weak correlation at 275 Auchencorth Moss may be because surface NH3 concentrations are near the instrument detection limit (monthly mean NH3 < 2.0 µg N m -3 ) and also because of low thermal contrast between the surface and overlying atmosphere (Van Damme et al., 2015;Dammers et al., 2016). The slope for Auchencorth Moss (4.02 × 10 15 molecules cm -2 (µg N m -3 ) -1 ) is steeper than the slopes observed at sites with greater surface concentrations of NH3 (Harwell = 2.23 × 10 15 molecules cm -2 (µg N m -3 ) -1 and Chilbolton = 2.07 × 10 15 molecules cm -2 (µg N m -3 ) -1 ). Steeper slopes for sites with relatively low NH3 concentrations is 280 consistent with assessment of earlier IASI NH3 product versions (Van Damme et al., 2015;Dammers et al., 2016). that we obtain from the slope of SMA regression of hourly PM2.5 and PM10 at 6 sites in Birmingham with both measurements.

Assessment of MODIS AOD
We use a similar approach as applied to NO2 to assess AOD. Only surface observations around the satellite overpass (12h00-15h00 LT) and with consistent detrended month-to-month variability (R > 0.5) are retained to obtain city-wide monthly mean means are estimated for London by averaging the pixels centred within its administrative boundary and for Birmingham within and 6.5 km beyond the administrative boundary, as with OMI NO2 (Section 3.1). We remove months with < 160 observations; equivalent in spatial coverage to 5 OMI pixels at nadir (the threshold used for OMI). After filtering, 53 % of months are removed for London and 72 % for Birmingham, mostly in winter. Fewer months than OMI are retained, as MODIS uses stricter 295 cloud filtering. The correlations in Figure 4 are weak (R = 0.34 for London, R = 0.23 for Birmingham) and do not improve if we apply a less strict threshold for the number of observations required to calculate monthly means. The poor correlation may be due to environmental factors that complicate the relationship between AOD and surface PM2.5, such as variability in meteorological conditions, aerosol composition, enhancements in aerosols above the boundary layer, and the aerosol radiative properties (Schaap et al., 2009;van Donkelaar et al., 2016;Shaddick et al., 2018;Sathe et al., 2019). We find that the same 300 assessment is not feasible for Delhi or Kanpur as the record of surface PM2.5 and PM10 in these cities is too short. Birmingham (2009)(2010)(2011)(2012)(2013)(2014)(2015)(2016)(2017) to assess whether the weak correlation in Figure 4 affects agreement in trends of the two quantities.
PM2.5 is longer lived than NO2, so trends in PM2.5 (lifetime order weeks) for the limited number of sites mostly located in 305 central London should be more representative of variability across the city than the surface sites of NO2 (lifetime order hours against conversion to temporary reservoirs). The steeper decline in surface PM2.5 in Birmingham (3.7 % a -1 ) than in London (2.7 % a -1 ) is reproduced in the AOD record (3.7 % a -1 in Birmingham; 2.5 % a -1 in London), although the AOD trends are not significant. In the two UK cities, surface PM2.5 peaks in spring, whereas AOD peaks in the summer, determined from multiyear monthly means (not shown). There are too few PM2.5 measurements in Delhi and Kanpur to compare long-term trends. 310 We compare the MODIS AOD product against ground-truth AOD from AERONET at long-term sites in Kanpur and Chilbolton to assess whether errors in satellite retrieval of AOD contribute to the weak temporal correlation between MODIS AOD and surface PM2.5. Daily AERONET AOD at 550 nm is estimated by interpolation using the second-order polynomial relationship between the logarithmic AOD and logarithmic wavelengths at 440, 500, 675 and 870 nm (Kaufman, 1993;Eck et 315 al., 1999;Levy et al., 2010;Li et al., 2012;Georgoulias et al., 2016). AERONET is sampled 30 minutes around the MODIS overpass and MODIS is sampled 27.5 km around the AERONET site Petrenko et al., 2012;Georgoulias et al., 2016;McPhetres and Aggarwal, 2018). Months with fewer than 160 MODIS observations are removed. Figure 6 compares coincident AOD monthly means from MODIS and AERONET for Kanpur and Chilbolton. Monthly 320 variability in MODIS and AERONET AOD is consistent at both sites (R  0.8). MODIS exhibits no appreciable bias at Kanpur.
There is positive variance (slope = 1.4) at Chilbolton that may result from sensitivity to errors in surface reflectivity at low AOD Bilal et al., 2018) and residual cloud contamination (Wei et al., 2018;. Mhawish et al. (2017) obtained similarly strong correlation (R = 0.8), but positive bias (26 %), of MODIS AOD at Kanpur from an earlier 3 km MODIS AOD product (Collection 6). 325

Air Quality Trends in London, Birmingham, Delhi, and Kanpur
The consistency we find between satellite and ground-based monthly mean city-average NO2 ( Figure 2) and rural NH3 ( Figure   3), and trends in city-average PM2.5 ( Figure 5) supports the use of the satellite record to constraint surface air quality. Variability in NO2, HCHO, and NH3 columns can also be related to precursor emissions of NOx, NMVOCs, and NH3 (Martin et al., 2003;330 Lamsal et al., 2011;Marais et al., 2012;Zhu et al., 2014;Dammers et al., 2019), as their lifetimes against conversion to temporary or permanent sinks are relatively short, varying from 1-12 hours depending on photochemical activity, abundance of pre-existing acidic aerosols, and proximity to large sources (Jones et al., 2009;Richter, 2009;Paulot et al., 2017;Van Damme et al., 2018). We adopt the same sampling approach as used to evaluate OMI NO2. That is, we sample the satellite observations within the city administrative boundaries for London, Delhi and Kanpur, and extend the sampling domain for 335 Birmingham beyond the administrative boundary by 6.5 km for OMI and MODIS and 10 km for IASI.
We apply the Theil-Sen single median estimator to the time series and also test the effect of fitting a non-linear function (Weatherhead et al., 1998;van der A et al., 2006;Pope et al., 2018) to account explicitly for seasonality: is city-average satellite observations for month , is the number of months from the start month (January 2005 for OMI and MODIS, and January 2008 for IASI), and , , and ∅ are fit parameters. is the city-average satellite observations in the start month, is the linear trend, and [ sin( + ∅)] is the seasonal component that includes the amplitude , 345 frequency (fixed to 12 months) and phase shift ∅. We only show the fit in Equation (1) if the trend B is different to that obtained with the Theil-Sen approach. The confidence intervals (CIs) for the Theil-Sen trends are estimated using bootstrap resampling and trends are considered significant for p-value < 0.05, that is, if the 95 % CI range does not intersect zero. Figure 7 shows the time series of monthly means of city-average OMI NO2 in the four cities for 2005-2018. Decline in OMI 350 NO2 in both London and Birmingham is 2.5 % a -1 and is significant. In Delhi, the OMI NO2 increase is 2.0 % a -1 and is significant (p-value = 0.003), whereas the increase in Kanpur of 0.9 % a -1 is not (p-value = 0.06). The relationship between tropospheric column and surface NO2 in London and Birmingham exhibits seasonality (Figure 2). This is in part due to seasonality in mixing depth. We find that excluding the winter months in the time series has only a small effect on the trend.
NO2 should exhibit seasonality in all cities due to seasonal variability in its lifetime and sources (van der A et al., 2008). The 355 fit in Equation (1)  obtained for OMI NO2 in London (-2.5 % a -1 ) is steeper than we estimate with the surface monitoring sites shown as triangles in Figure 1 (1.8 % a -1 for 2005-2018). Most sites are in central London, and NO2 trends in outer London are 1.6 times steeper than in central London (Carslaw et al., 2011). The decline in NO2 in the two UK cities is less than the rate of decline in national NOx emissions (3.8 % a -1 ) for 2005-2017 from the national bottom-up emission inventory (Defra, 2019a). This may reflect a combination of factors. There is less steep decline in NOx emissions in London compared to the national total that may in part 365 be due to discrepancies between real-world and reported diesel NOx emissions (Fontaras et al., 2014), sustained heavy traffic in central London, and an increase in NO2-to-NOx emission ratios dampening decline in NO2 (Grange et al., 2017). There is also weakened sensitivity of the tropospheric column to changes in surface NO2 due to a gradual increase in the relative contribution of the free tropospheric background to the tropospheric column (Silvern et al., 2019). This weakening of the trend in the tropospheric column will likely be less in London than in Birmingham, due to greater local surface emissions in large 370 cities such as London (Zara et al., 2021). The positive trends in Delhi and Kanpur likely reflect a 2-fold increase in vehicle ownership in Delhi (Govt. of Delhi, 2019), rapid industrialisation in Kanpur (Nagar et al., 2019), and limited effect of air quality policies on pollution sources. This is corroborated by NOx emissions compliance failures at more than 50 % of coalfired power plants in Delhi and the surrounding area (Pathania et al., 2018). The lack of trend reversal in Delhi, despite implementation of air quality policies, is consistent with the lack of trend reversal reported by Georgoulias et al. (2019). They 375 used a 21-year record (1996-2017) of multiple space-based sensors to estimate a significant and sustained increase in NO2 of 3.1 % a -1 in Delhi. By the end of 2018, tropospheric column NO2 is similar in London and Delhi (5.7 × 10 15 molecules cm -2 ; Figure 7) but OMI NO2 over India may be biased low, due to the presence of optically thick aerosols (AOD > 0.4; Figure 6) that are not explicitly accounted for in the retrieval (Section 3.1).

380
The direction of the trends for all four cities is consistent with other trend studies, with differences in the absolute size of the trend due to differences in instruments, time periods, and sampling domains. Pope et al. (2018)

observed declines in OMI NO2
for 2005-2015 of 2.3 ± 0.5 × 10 14 molecules cm −2 a −1 for London and 1.1 ± 0.5 × 10 14 molecules cm −2 a −1 for Birmingham. We obtain a similar trend for Birmingham but a steeper decline for London of 2.6 × 10 14 molecules cm −2 a −1 using our sampling domain for 2005-2015, though the difference is not significant. Schneider et al. (2015) obtained less steep and non-significant 385 changes in NO2 in London (-1.7 ± 1.2 % a -1 ) and Delhi (1.4 ± 1.2 % a -1 ) from the SCanning Imaging Absorption spectroMeter for Atmospheric CHartographY (SCIAMACHY) for 2002. Trends in OMI NO2 for 2005from ul-Haq et al. (2015 are similar to ours for Delhi (2.0 % a -1 ) but lower for Kanpur (0.2 % a -1 ). Studies have also combined multiple instruments to derive trends since the mid-1990s. These find decreases in NO2 over London of 0.7 % a -1 for 1996-2006 (van der A et al., 2008) and 1.7 % a -1 for a longer observing period (1996-2011) (Hilboll et al., 2013), and a consistent increase for Delhi of 7.4 390 % a -1 in 1996(van der A et al., 2008) and 1996(Hilboll et al., 2013; much steeper than ours in Figure 7. 15-20 times more in Delhi and Kanpur than in London and Birmingham due to larger emissions of NH3 in the IGP, higher ambient temperatures promoting volatilization of NH3, and greater sensitivity of IASI to NH3 due to greater thermal contrast 395 between the surface and the atmosphere over India (Van Damme et al., 2015;Dammers et al., 2016;Wang et al., 2019). IASI NH3 decreases by 0.1 % a -1 in Kanpur, 0.6 % a -1 in Birmingham and 2.4 % a -1 in London, and increases by 0.5 % a -1 in Delhi.
None of the trends are significant. Measurements of surface NH3 from continuous monitors deployed in Delhi in April 2010 to July 2011 exhibit the same seasonality as IASI NH3, peaking in the monsoon season (July-September) (Singh and Kulshrestha, 2012). We investigated the effect of NH3 seasonality on the trend using Equation (1) (grey solid lines in Figure  400 8). Similar to NO2, all four cities show significant seasonality (p-value < 0.05 for the amplitude of the seasonality, ). The linear trends (grey dashed lines in Figure 8) are more positive than those obtained with Theil-Sen for all four cities, but are still not significant. This leads to a trend reversal in Kanpur (+1.0 % a -1 ) and Birmingham (+2.1 % a -1 ), steeper increase in Delhi (+3.7 % a -1 ), and a less negative trend in London (-0.6 % a -1 ).

405
Relating trends in NH3 columns to trends in NH3 emissions is complicated by partitioning of NH3 to aerosols to form ammonium and dependence of this process on pre-existing aerosols that have declined in abundance across the UK due largely to controls on precursor emissions of SO2 (Vieno et al., 2014). Harwell and Auchencorth Moss include measurements of gasphase NH3 and aerosol-phase ammonium in PM2.5. These exhibit large and distinct seasonality, so we use Equation (1) to estimate changes of -0.096 µg N m -3 a -1 for ammonium and +0.031 µg N m -3 a -1 for NH3 at Auchencorth Moss in 2008-2012 410 and similar changes at Harwell in 2012-2015 of -0.10 µg N m -3 a -1 for ammonium and +0.035 µg N m -3 a -1 for NH3. Only the decline in ammonium at Auchencorth Moss is significant. This suggests the increase in rural NH3 includes contributions from unregulated agricultural emissions and reduced partitioning of NH3 to pre-existing aerosols. The opposite trend (decline) in NH3 in London obtained with Theil-Sen and Equation (1) (Figure 8) may be because decline in local vehicular emissions of NH3 with a shift in catalytic converter technology (Richmond et al., 2020) outweighs the increase in NH3 from waste and 415 domestic combustion (Defra, 2019a), and nearby agriculture (Vieno et al., 2016) and offsets reduced partitioning of NH3 to acidic aerosols with decline in sulfate. The opposite effect would be expected in Delhi due to nation-wide increases in SO2 emissions and sulfate abundance (Klimont et al., 2013;Aas et al., 2019). That is, the increase in NH3 emissions may be steeper than the increase in NH3 columns in Figure 8 due to a corresponding increase in partitioning of NH3 to pre-existing aerosols as these become more abundant. 420 Figure 9 shows the time series of city-average monthly mean OMI HCHO for the four cities for 2005-2018 after removing the background contribution from oxidation of methane and other long-lived VOCs to isolate variability in the column due to reactive NMVOCs (Zhu et al., 2016). A representative background is obtained as monthly mean OMI HCHO over the remote Atlantic Ocean (25-35° N, 35-45° W) for the UK and the remote Indian Ocean (10-20° S, 70-80° E) for India. The non-linear 425 function in Equation (1) is fit to these background HCHO values and used to subtract the background contribution, as in Marais et al. (2012), from the city-average monthly means. OMI HCHO columns from oxidation of reactive NMVOCs in Delhi and Kanpur are almost twice those in London and Birmingham due to a combination of unregulated sources  and high temperatures enhancing emissions of isoprene, a dominant HCHO precursor in India (Surl et al., 2018;Chalilyakunnel et al., 2019). The trends suggest reactive NMVOCs emissions have decreased in Birmingham (1.6 % a -1 ) and 430 increased in London (0.5 % a -1 ), Delhi (1.9 % a -1 ) and Kanpur (1.0 % a -1 ). Only Delhi has a significant trend. According to the UK bottom-up emission inventory, national NMVOCs emissions decreased by 2.4 % a -1 from 2005 to 2017 (Defra, 2019a). This is supported by decline in short-chain hydrocarbons measured at Harwell from 2-3 g m -3 in 2008 to 0.8-445 0.9 g m -3 in 2015. These include hydrocarbons from vegetation (isoprene and monoterpenes) and vehicles (light alkanes and aromatics), but exclude oxygenated VOCs (OVOCs) that in the UK include increasing contributions from domestic combustion, the food and beverage industry, and household products (Defra, 2019a). OVOCs have relatively high HCHO yields (Millet et al., 2006) and VOC concentrations measured during field campaigns in London and cities in India, including Delhi, are dominated by OVOCs (> 60 % in London) (Valach et al., 2014;Sahu et al., 2016;Wang et al., 2020). In London, 450 OVOCs also dominate inferred fluxes of VOCs (Langford et al., 2010) and reactivity of VOCs with the main atmospheric oxidant, OH (Whalley et al., 2016). The rapid increase in HCHO also has implications for ozone air pollution and the radical budget in London, as ozone formation is VOC-limited and HCHO photolysis is the second largest source of hydrogen oxide radicals (HOx  OH + HO2) in London (Whalley et al., 2018). Mean AOD in Delhi and Kanpur is on average 5-6 times more than in London and Birmingham, due to large local anthropogenic emissions, nearby agricultural emissions of PM2.5 and its precursors in the IGP, and long-range transport of desert dust (David et al., 2018). Our results, as absolute AOD trends for London (-0.004 a -1 ) and Birmingham (-0.007 a -1 ) for 460 2005-2018, are similar to trends obtained by Pope et al. (2018) for 2005-2015 (-0.006 a -1 for London; -0.005 a -1 for Birmingham). Our trends for both cities in India are less steep than the increase for Delhi (4.9 % a -1 ) obtained for 2000-2010 with the MODIS 10 km AOD product (Ramachandran et al., 2012) and for Kanpur (10.3 % a -1 ) obtained for 2001-2010 with AERONET AOD at the Kanpur AERONET site (Kaskaoutis et al., 2012). This may reflect a recent dampening of the trend or differences in data products and sampling domain/period. Sulfate from coal-fired power plants in India makes a large 465 contribution to PM2.5 (Weagle et al., 2018) and emissions from these nearly doubled from 2004 to 2015 (Fioletov et al., 2016).

Conclusions
Satellite observations of atmospheric composition provide long-term and consistent global coverage of air pollutants. We assessed the ability of satellite observations of nitrogen dioxide (NO2) and formaldehyde (HCHO) from OMI for 2005-2018, 470 ammonia (NH3) from IASI for 2008-2018, and aerosol optical depth (AOD) from MODIS for 2005-2018 to provide constraints on long-term changes in city-average NO2, reactive NMVOCs, NH3, and PM2.5, respectively in four cities: 2 in the UK (London and Birmingham) and 2 in India (Delhi and Kanpur).
Assessment of satellite observations against ground-based measurements followed careful screening of the in-situ 475 measurements for poor quality data, correcting NO2 data reported in inconsistent units at monitoring sites in Delhi and Kanpur, and removing sites influenced by local sources. OMI NO2 reproduces monthly variability in surface concentrations of NO2 in cities, whereas satellite AOD reproduces trends, but not monthly variability, in PM2.5 in cities. MODIS and AERONET AOD are consistent at long-term monitoring sites in Kanpur and a UK EMEP site in southern England. IASI NH3 is consistent with monthly variability in surface NH3 concentrations at two of three rural UK EMEP sites. There were no appropriate 480 measurements of reactive NMVOCs to compare to OMI HCHO.
According to the long-term record from Earth observations, NO2, PM2.5, and NMVOCs increased in Delhi and Kanpur. There is no reversal in the increase in NO2 or PM2.5 in Delhi or Kanpur, as would be expected from successful implementation of air pollution mitigation measures. In all four cities, the magnitude and direction of trends in NH3 is sensitive to treatment of NH3 485 seasonality and none of the NH3 trends are significant. In London and Birmingham, NO2 and PM2.5 decrease, and HCHO, a proxy for reactive NMVOCs emissions, decreases in Birmingham, but exhibits a recent (2012-2018) sharp (> 9 % a -1 ) increase in London. This may reflect increased emissions of oxygenated VOCs and long-chain hydrocarbons from household products, the food and beverage industry, and residential fuelwood burning. This would have implications for formation of secondary organic aerosols (SOA) contributing to PM2.5, the radical (HOx) budget that includes large contribution from HCHO photolysis, 490 and formation of surface ozone that is VOC-limited in London.

Data availability
Corrected hourly NO2 data for Delhi and Kanpur are available at https://github.com/karnvoh/India-NO2-data. Data from IIT Kanpur can be obtained by contacting SNT (snt@iitk.ac.in). Data for Birmingham not publicly available can be obtained by request from the Birmingham City Council. IASI NH3 data are available at https://iasi.aeris-data.fr/nh3/. 495

Author contributions
KV analysed and interpreted the data and prepared the manuscript, EAM assisted in the writing and provided supervisory guidance, with co-supervision from WJB. LK provided data analysis and usage guidance. SS derived the relationship between hourly PM10 and PM2.5 for Birmingham. Observations are from RS, AG, and SNT for the surface site in Kanpur, and MVD, LC, and PFC for IASI NH3. 500

Competing interests
The authors declare that they have no conflict of interest.