Satellite observations of VCDtrop NO2 were obtained from OMI (2016–2020) and TROPOMI (2019–2020). Surface NO2 observations (2016–2020) at 139 sites across India were from the Central Pollution Control Board (CPCB). The period from 25 March to 3 May each year is defined as the analysis period. Average NO2 levels during the analysis period in 2020 and previous years are referred to as lockdown (LDN) NO2 and business-as-usual (BAU) NO2, respectively. The BAU years for OMI and CPCB are 2016–2019, whereas for TROPOMI the BAU year is 2019 because of the unavailability of earlier observations.

NO2 data were analysed for six geographical regions (north, Indo-Gangetic Plain (IGP), north-west, north-east, central, and south) of India (Fig. S1 in the Supplement). The NO2 changes over various land-use categories (i.e. urban, cropland, and forestland) have been analysed using spatially collocated land-use land cover (LULC) data (NRSC, 2012) and OMI- and TROPOMI-observed VCDtrop NO2. Visible Infrared Imaging Radiometer Suite (VIIRS) fire count data were used to study the fire anomalies during the LDN and other analysis periods.

Air pollutant concentration over a region is governed by emission sources and prevailing meteorological conditions. Meteorological factors (e.g. wind, temperature, radiation, and rainfall) can affect the NO2 concentration (Barré et al., 2020) as well as biogenic emissions (Guenther et al., 2012). The meteorological variations between years can cause ∼ 15 % variations in monthly column NO2 values (Goldberg et al., 2020). However, the NO2 levels are likely to be similar under similar meteorological conditions. Recent studies (e.g. Singh et al., 2020; Navinya et al., 2020; Sharma et al., 2020) have shown that meteorological conditions remained relatively consistent over recent years during the lockdown period and therefore assumed that the changes in the pollution levels during the lockdown are primarily driven by the emission changes. However, it is important to highlight the meteorological differences during the study period to assess the uncertainties associated with meteorological differences.

We used monthly mean ERA-5 reanalysis data (Hersbach et al., 2020) at 0.25∘ × 0.25∘ resolution for March, April, and May for BAU as well as LDN periods at the satellite local overpass time. We considered temperature (T), wind speed (WS), and boundary layer height (BLH) in our analysis. Figure 1a–c show the spatial variation in these quantities during BAU (left column), LDN (middle column) and the calculated difference (LDN-BAU, right column). The probability density function (PDF) using kernel density estimation (KDE) of the meteorological parameters is also shown (Fig. S3) for the BAU (blue) and LDN (red). KDE is a non-parametric way to estimate the PDF. The peak of the distribution shows the most probable value, and the width of the distribution shows the variability. The temperature difference between LDN and BAU shows a slight reduction (∼ 0–3 K range) during the lockdown. Wind speed values also show a reduction (up to 2 m s-1) during the lockdown, although the reduction is mainly seen in certain parts of central India. Reduction in the BLH is also seen in most parts of India. In general, the meteorological parameters during the lockdown were similar. However, the PDF (Fig. S3) during BAU and LDN shows a small reduction (less than 5 %) in temperature and wind speed and ∼ 10 % reduction in BLH. Although small, this weather variability can further add to the variability in the NO2 levels. However, during the lockdown in India, the NO2 change was more sensitive to the emission change than the meteorology variability. Shi et al. (2021) compared the detrended and de-weathered change in NO2 observed over selected cities in India, Europe, China, and the USA. While the reduction in NO2 was highest for Delhi (∼ 50 %), the difference between a detrended and de-weathered change in NO2 observed over Delhi was much smaller (∼ 2 %) as compared to the difference calculated for other cities. This suggests that weather variability did not have much impact on NO2 levels over India and that most of the changes were driven by a change in the anthropogenic emissions.

Forest fires are an important source of surface NO2 and VCDtrop NO2 (Sahu et al., 2015; Yarragunta et al., 2020), depending on the occurrence time and the intensity of fires (Mebust et al., 2011). Also, as the forest fire plumes can be transported longer distances (Alonso-Blanco et al., 2018), forest-fire-related NO2 can contribute to regional and global air pollution. In India, forest fires are prevalent as 36 % of the country's forest cover is prone to frequent fires, of which nearly 10 % is extremely to very highly prone to fires (ISFR, 2019). Long-term satellite-derived fire counts suggest that Indian fire activities typically peak during March–May (Sahu et al., 2015), predominantly over the north, central, and north-east regions (Venkataraman et al., 2006; Ghude et al., 2013). However, the spatial and temporal distribution of fire events is largely heterogeneous (Sahu et al., 2015), meaning an abrupt increase or decrease in fire activity could significantly impact NO2 levels over anomalous regions during the lockdown.

An investigation of fire counts during the 2020 lockdown (LDN analysis period), when compared with the corresponding 2016–2020 average, highlights a substantial decrease over the eastern part of central India and an increase over the western part of central India and the north-east. In Fig. 2a widespread fire activity (counts of 10–50) is shown across India, such as the central region (Madhya Pradesh, Chhattisgarh, Odisha), parts of Andhra Pradesh, the Western Ghats in Maharashtra, and the north-east region (Assam, Meghalaya, Tripura, Mizoram, and Manipur). The fire anomaly during the lockdown (Fig. 2b) shows positive fire counts (5–20) over the north-east region, west of Madhya Pradesh in central India, and scattered locations in South India. The negative fire anomalies (-20 to -5) observed over the central region (Chhattisgarh and Odisha) suggest a decrease in fire activity during the 2020 lockdown period. To minimise the impact of fire emission in our analysis, we have considered the grids with zero fire anomaly to assess the changes in NO2 during the lockdown. By considering the grids with zero fire anomaly, we excluded almost all the grids which have recorded fire activity during the analysis period. However, the impact of long-range transport of forest fire plumes cannot be ignored.

The spatial distribution of VCDtrop NO2 is largely determined by local emission sources; therefore, NO2 hotspots are found over urban regions, thermal power plants, and major industrial corridors. For the Indian subcontinent, maximum NO2 is observed during winter to pre-monsoon (December–May) and minimum NO2 during the monsoon (June–September). Region-specific peaks such as the wintertime peak (December–January) in the IGP are associated with anthropogenic emissions, or the summertime peak (March–April) in central India and north-east India is associated with enhanced biomass burning activities (Ghude et al., 2008, 2013; Hilboll et al., 2017).

We compare the LDN mean VCDtrop NO2 with the BAU mean for OMI and TROPOMI. The spatial distribution of the BAU and LDN VCDtrop NO2 observed by OMI and TROPOMI is shown in Fig. 3a–d. The mean VCDtrop NO2 from the two instruments shows similar spatial distributions during the LDN and BAU analysis period. In BAU years, the NO2 hotspots are seen over the large fossil-fuel-based thermal power plants (∼ 1000 × 1013 molec. cm-2), urban areas (∼ 400–700 × 1013 molec. cm-2), and industrial areas. Scattered sources are also present in western India, covering the industrial corridor of Gujarat and Mumbai, various locations of south India, and densely populated areas (e.g. IGP). The spatial distribution showed significant changes during the lockdown in 2020. The details of absolute and percentage changes are discussed in the subsequent sections.

There is a substantial reduction in VCDtrop NO2 between the LDN and BAU (Fig. 4a and c). A large reduction in the number of hotspots, mainly urban areas, is seen in both OMI and TROPOMI observations. However, hotspots due to coal-based power plants remain during the lockdown as electricity production was continued. Over the NO2 hotspots, there has been an absolute decrease of over 150 × 1013 molec. cm-2 (∼ 250 × 1013 molec. cm-2 over megacities) detected by both OMI and TROPOMI. The rural VCDtrop NO2 has typically reduced by approximately 30–100 × 1013 molec. cm-2, representing a percentage decrease of 30 %–50 % for OMI and 20 %–30 % for TROPOMI (Fig. 4b and d). For urban regions, both OMI and TROPOMI see a decrease of approximately 50 %, but reductions in smaller urban areas are clearly noticeable in the TROPOMI data, given its better spatial resolution. Both instruments observe an increase in VCDtrop NO2 in the north-eastern regions and moderate enhancement over the western and central regions. These enhancements are linked with the biomass burning activities during this period (Fig. 2).

Anthropogenic NOx emissions are typically more localised in urban and industrial centres, while biogenic sources (e.g. soil) are more important in rural regions. OBB activities peak in March–April (Sahu et al., 2015) and represent more sporadic sources. As the lockdown is expected to have reduced urban anthropogenic NOx sources (as shown in Fig. 4), it is important to assess the lockdown impact over the rural regions such as cropland and forestland as well. This section estimates the changes in VCDtrop NO2 over different land types such as cropland, forestland, and urban areas (Fig. S2). Industrial emissions are often part of the urban agglomerates scattered around the city and are part of urban emissions. To minimise the impact of OBB emissions in our analysis, we exclude grids with fire anomalies (Fig. 2) and those containing thermal power plants (Fig. S2d). However, absolute separation of the impact of long-range transportation is beyond the scope of this study.

Figure 7a–b show the relationship of OMI and TROPOMI NO2 with surface NO2 for the BAU and LDN periods, respectively. During BAU, there are reasonable positive correlations between the satellite instruments and the surface sites (OMI: 0.48, 95 % CI 0.33–0.60 and TROPOMI: 0.52, 95 % CI 0.37–0.64). In LDN, these correlations drop to 0.36 (95 % CI 0.20–0.49) and 0.28 (95 % CI 0.12–0.43), respectively. The decrease in the correlation during LDN could be due to the decrease in the signal-to-noise ratio, potentially linked with the primary reduction in urban NO2 levels. We also determined the correlation between satellite- and surface-observed changes during the lockdown (Fig. 7c), finding values of 0.44 (95 % CI 0.28–0.57) for OMI and 0.49 (95 % CI 0.33–0.63) for TROPOMI. This indicates that the lockdown NO2 reductions appear to be present in both measurement types, providing us with confidence in the observed changes detected in this study. The correlation observed over India in this study is lower than that reported for the USA (Lamsal et al., 2015). The low correlation between OMI and surface NO2 has been reported previously by Ghude et al. (2011). While they report the temporal correlation for a single site, our study reports the spatial correlation representing the satellites' ability to capture the spatial heterogeneity. One of the reasons for the lower correlation could be the choice of surface station. Generally, urban background sites are preferred for this kind of analysis. However, the surface NO2 monitoring station type classification is not available for the CPCB sites. Therefore, sites used in the analysis could be potentially impacted by traffic emissions, resulting in lower correlation. Another reason is that in situ measurements are more sensitive to the local emission sources than remotely sensed measurements and therefore have larger variability, resulting in low correlation. Proper classification of the monitoring stations could provide a better assessment of satellite-based observations.

The LDN NO2 percentage change, observed by surface and spatially co-located satellite measurements, is shown in Fig. 8a for various Indian regions. For this comparison, the number of available CPCB surface monitoring stations was 17, 15, 81, 25, and 1 for central, north-west, IGP, south, and north-east regions (north region data not available), respectively. Most of the CPCB stations are in urban areas, so our results reflect changes in predominantly urban-sourced NO2. At all surface sites in all regions, there was a percentage reduction greater than 20 % (Fig. 8a). Satellite observations show a similar trend except for the north-east region, where enhancements are due to forest fires. Both OMI and TROPOMI observed the highest reduction (∼ 50 %) over IGP. A smaller average reduction of ∼ 20 % over central India might be due to the aggregate effect of power plants, forest fires, and prevalent biomass burning activities during this season. While the effect of forest fires can be observed in the column NO2, its impact on the surface NO2 is minimal. For the central, IGP, and south regions, the mean percentage change observed by the surface monitoring station is comparable to that observed by the satellites.

We have inter-compared the percentage change in NO2 observed at the surface and satellite over the major Indian cities (i.e. New Delhi, Chennai, Mumbai, Bangalore, Ahmedabad, Kolkata, and Hyderabad; Fig. 8b). A significant reduction in the range of ∼ 25 %–75 % is observed, consistent in all observational sources used in this study. A similar reduction observed by the satellites over the cities in other parts of the world has been reported (Tobías et al., 2020; Naeger and Murphy, 2020; Kanniah et al., 2020; Huang and Sun, 2020). The satellites observe the largest reduction over Delhi and the smallest over Kolkata. While the observed decline is comparable for cities, Ahmedabad and Kolkata showed smaller declines than observed by ground measurements. Also, the reduction observed at the surface has a larger spatial variability than the one observed from the space. This is potentially linked to the influence of the local emissions which could not be detected by the space-based instruments because of relatively large satellite footprints. The results of percentage change observed by OMI are consistent with the change reported by Pathakoti et al. (2020), although Siddiqui et al. (2020) reported a higher decline of NO2 using TROPOMI. This is because we computed the changes between lockdown and BAU during the same period of the year, whereas Siddiqui et al. (2020) estimated the changes between the pre-lockdown NO2 and the lockdown NO2, which includes the seasonal component of NO2. We have also analysed the changes in VCDtrop NO2 observed by both OMI and TROPOMI for the other major cities (Guttikunda et al., 2019), as shown in Fig. S4. A reduction of over 20 % was observed in most cities except for a few in north-east and central India. Cities showing enhancement or smaller reductions reflect the enhanced fire activities in the north-east and central Indian regions. TROPOMI can capture the reduction over the cities near the fire-prone areas (e.g. Indore and Bhopal) because of its higher spatial resolution.

In this section, we examine the VCDtrop NO2 and population relationship for India except where fire anomalies or large thermal power plants existed. The scatter density plots between VCDtrop NO2 and population density for the BAU and LDN analysis period are shown in Fig. 9 for OMI and TROPOMI. The data were log-transformed to establish the log–log relationship as neither dataset is normally distributed. As the observed changes had negative values, this log transformation was obtained by adding a constant value (Ekwaru and Veugelers, 2018), which was later subtracted when plotting to display the corresponding NO2 values. Both OMI and TROPOMI NO2 show a similar relationship with the population density, with correlations of ∼ 0.65 during the LDN and BAU periods, suggesting a strong dependence upon the population (i.e. anthropogenic emissions). The slopes of the lines in Fig. 9a, b, d, and e show that VCDtrop NO2 follows a power-law scaling with population density (Lamsal et al., 2013). During BAU, the VCDtrop NO2 observed over a grid increased by factors of 100.28=1.9 and 100.20=1.58 for OMI and TROPOMI, respectively, with a 10-fold increase in the population density. The rate of increase of the VCDtrop NO2 during LDN was 100.23=1.7 and 100.16=1.45 times for OMI and TROPOMI, respectively, which was lower than BAU. The correlation during the LDN period was marginally lower than the BAU period. This could be due to a larger reduction in the NO2 levels in the densely populated grids. The changes observed in the VCDtrop NO2 during the LDN (Fig. 9c and f) were negatively correlated (i.e. reduction was positively correlated) with the population density. The linear relation suggests an increase in the reduction with an increase in the population density; however, some grids exhibit enhancements in VCDtrop NO2 due to the local emissions.

In order to link the observed reduction in NO2 levels with the traffic emissions over the urban areas, Fig. 10 shows the 7 d moving average of the daily percentage change observed by OMI, TROPOMI, and CPCB across urban India from 1 March to 31 May 2020 against the Google mobility percentage reduction for three mobility categories: transit stations, workplace, and residential. Transit stations and workplace, proxies for traffic emissions (Forster et al., 2020), show a sharp decline (∼ 70 %) due to the lockdown. The signatures of reduced traffic can be seen even before the start of lockdown in mid-March 2020. The decrease in the workplaces resulted in the enhancement (25 %–30 %) of people at a residential location. The percentage reductions observed by satellites and surface monitoring are consistent with each other and follow the same trend of the workplaces and transit stations. The reductions observed by satellites and surface monitoring are ∼ 20 % lower than the reductions in workplaces and transit stations, which are compensated for by the enhancement in residential emissions. Surface (CPCB) measurements exhibit higher correlation (∼ 0.9 and 0.8, with and without moving average) with the mobility reduction compared to the satellite observation, which has a relatively weaker correlation (∼ 0.8 and 0.5). The positive correlation of NO2 reduction with workplaces and transit stations suggests that the reduction observed over the urban areas was linked with reduced traffic emissions due to travel restrictions for COVID-19 containment. Moreover, the mobility reduction was higher for larger cities as compared to the smaller ones (Fig. S6).

This study has few limitations that need to be considered while interpreting the results. The observed changes in the NO2 levels are the combined effect of changes in the emissions, local meteorology, large-scale dynamics, and non-linear chemistry. The variability in NO2, caused by weather patterns and non-linear chemistry is not included in the present work. Our study does not distinguish the differences in the upwind and downwind transport of plumes originating from urban areas and thermal power plants. Moreover, the estimates can be biased by the forest-fire plumes, which can be transported over a long distance. These limitations warrant a detailed modelling study to quantify the impact of long-range transport of plumes in the drastic reduction of urban emissions. One of the limitations arises due to the unavailability of the surface monitoring classification according to its location and vicinity of the local sources, which restricted a proper assessment of the space-based NO2 observation. To overcome this limitation, proper classification of the monitoring stations (Geiger et al., 2013) based on the environment type and vicinity of the sources will be helpful in air quality assessment.