Nitrogen oxides concentration and emission change detection during COVID-19 restrictions in North India

COVID-19 related restrictions lowered particulate matter and trace gas concentrations across cities around the world, providing a natural opportunity to study effects of anthropogenic activities on emissions of air pollutants. In this paper, the impact of sudden suspension of human activities on air pollution was analyzed by studying the change in satellite retrieved NO2 concentrations and top-down NOx emission over the urban and rural areas around Delhi. NO2 was chosen for being the most indicative of emission intensity due to its short lifetime of the order of a few hours in the planetary boundary layer. We present a robust temporal comparison of Ozone Monitoring Instrument (OMI) retrieved NO2 column density during the lockdown with the counterfactual baseline concentrations, extrapolated from the long-term trend and seasonal cycle components of NO2 using observations during 2015 to 2019. NO2 concentration in the urban area of Delhi experienced an anomalous relative change ranging from 60.0% decline during the Phase 1 of lockdown (March 25–April 13, 2020) to 3.4% during the post-lockdown Phase 5. In contrast, we find no substantial reduction in NO2 concentrations over the rural areas. To segregate the impact of the lockdown from the meteorology, weekly top-down NOx emissions were estimated from high-resolution TROPOspheric Monitoring Instrument (TROPOMI) retrieved NO2 by accounting for horizontal advection derived from the steady state continuity equation. NOx emissions from urban Delhi and power plants exhibited a mean decline of 72.2% and 53.4% respectively in Phase 1 compared to the pre-lockdown business-as-usual phase. Emission estimates over urban areas and power-plants showed a good correlation with activity reports, suggesting the applicability of this approach for studying emission changes. A higher anomaly in emission estimates suggests that comparison of only concentration change, without accounting for the dynamical and photochemical conditions, may mislead evaluation of lockdown impact. Our results shall also have a broader impact for optimizing bottom-up emission inventories.

Starting during week 19, zonification of lockdown areas was performed as "red", indicating the presence of infection hotspots, "orange" indicating some infection, and "green" with no infections. Red zones were further demarcated as buffer and containment zones. Construction activities and traffic movement was relaxed for green and orange zone while red zones remained in lockdown (Ministry of Home Affairs 2020a

Meteorological variability between 2020 and 2019
One of the reasons why comparison of concentration between 2020 and 2019 alone could be misleading is because interannual meteorological variability plays an important role in determining NO2 concentration, particularly wind speeds as they regulate horizontal advection 1 . A 10% difference in wind speeds around 6.5m/s can cause a difference in NO2 column concentration by 10%, especially in winter season 1 . We chose 80m for the analysis because 80m is a typical height of the "surface boundary layer" in the planetary boundary layer. As the temperature profile shows diurnal cycle in a layer below 200m or 300m height, and 80m height is just the central height in the layer, where wind is not so strongly affected by the surface.
Mean weekly trend of zonal (u)and meridional wind speed (v) at 80m height, well as the planet boundary layer height (pblh), as it regulates the intra-day vertical diffusion, is shown in Figure S2. Compared to their values in 2019, during the Phase 1 (starting week 13), u was higher by 92% (mean 3.88 m/s), v was higher by 61% (mean -2.49 m/s). During weeks 17 to 19, u was slower by 70% (mean 1.3m/s). Concurrently, compared to 2019, pblh was about 28% higher in the eight weeks preceding the lockdown and lower by about 13% in the eight weeks beginning with the lockdown. The variation in pblh is more important for surface observation than the satellite retrievals which are column based.

Trend of OMI retrieved NO2
Although OMI was designed for daily retrievals, since June 2007 it has suffered a so-called 'rowanomaly' where its radiance measurements in some field of views have been affected by obstruction 2 . As a consequence of "row anomaly", OMI is missing data for almost half of its observation range, so the world-wide coverage is currently achieved in about two-days 3 . In this study, we used the product, 'OMNO2d' version 3, from NASA's Giovanni portal wherein the row anomaly flagged pixel and image with more than 30% cloud-fractions have been filtered already. To further analyze the impact of such filtering, we checked the weekly frequency of valid daily retrievals over a sample location, (urban Delhi: 76.875°E, 28.375°N -77.625°E, 28.875°N) for 6 years (2015 to 2020) as shown in Error! Reference source not found. (a). Retrievals were the lowest in January (weeks 1 to 4, median: 3 /week) followed by February (weeks 4 to 8, median: 4/week) and December (weeks 48 to 52, median: 4/week). For all other months the median weekly frequency of daily retrievals was higher than 4. A frequency 5/week was judged to sufficiently represent the weekly concentration. Based on the lockdown phases. we were interested mainly the retrievals made during the first six months of a year. Coincidently, the lockdown was initiated from week 13 to 27, where the median weekly retrieval count is at least 5/week and the inferences made from OMI retrievals in these weeks could be used confidently. However inferences made for preceding weeks, especially those of January (weeks 1 to 4) could have higher uncertainties. The weekly mean NO2 concentrations at urban (Delhi) and rural (Fatehabad) for the weeks 1 to 27 are shown in Figure S4 (a) and (b). The seasonal is visually trend is similar to what has been reported earlier based on ground studies 4 . The relatively large standard deviation in the weeks (weeks 1 to 8) could be due to low number of retrievals in those weeks.

Comparison between OMI and TROPOMI tropospheric NO2
It has been pointed out earlier that OMI retrievals are systematically lower than the groundbased Multi-AXis Differential Optical Absorption Spectroscopy (MAX-DOAS) column measurements in strongly polluted areas and vice versa in low concentration areas 5 . TROPOMI retrievals have a tendency to slightly overestimate MAX-DOAS columns by 7% over clean regions (below 2 × 10 15 molec/cm 2 ) and underestimate by 36% over polluted regions (between 3 − 14 × 10 15 molec/cm 2 ), wherein low NO2 column values are better reproduced than high NO2 column values 6 . The bias is thus within the mission requirement of 50%. Global comparison between OMI and TROPOMI has revealed that due to the difference in pixel resolution, representativeness remains a major issue: OMI's larger ground pixels, particularly near the swath-edges different amounts of NO2 and of cloud cover 6 . Over polluted areas TROPOMI is lower than OMI's, with differences varying from a few to -40% (in the winter) while in very clean regions TROPOMI has higher values by about 20% 6 . Over India, the difference between OMI and TROPOMI, the bias of TROPOMI is − 0.2 ± 0.8 × 10 15 (1σ) molec/cm 2 , which is − 6% ± 21% in relative value 7 .
A comparison between the two products was performed over the polluted urban region (Delhi, mean annual TROPMI concentration in 2019: 5.7 × 10 15 molec/cm 2 ) and the clean rural location (Fatehabad, mean annual TROPOMI concentration in 2019: 2.1 × 10 15 molec/cm 2 ) for the period Jan. 2019 to Jun. 2020. The difference between the weekly mean of TROPOMI and OMI columns in 2019 was analyzed. Both the products showed a similar seasonality at urban and rural locations ( Figure  S5). Out of the 52 weeks of 2019, over urban (rural) areas 32 (36) weeks had at least 4 daily retrievals devoid of cloud or row-anomaly. Over the urban area, mean weekly difference between TROPOMI and OMI was 0.3 ± 1.3 × 10 15 molec/cm 2 (6.4% ± 23.1%) , and was biased high during high pollution (regimes greater than 4.8 × 10 15 molec/cm 2 ), which often occurred during the winter months. Additionally, the difference and its deviation was lower (0.2±0.7 × 10 15 molec/cm 2 ) in the weeks when at least 4 daily valid retrievals existed. Over the rural region, TROPOMI was consistently lower than OMI by a mean value of 0.9 ± 0.8 × 10 15 molec/cm 2 (39.9% ± 39.1%). The difference between the two products was smaller (0.7 × 10 15 molec/cm 2 ) in the weeks with at least 4 daily retrievals.
urban rural Figure S5. Weekly mean 2019 NO2 concentrations from TROPOMI and OMI sensors over an urban and rural region.
Furthermore, the relative difference in the concentrations between 2020 and 2019 from TROPOMI and OMI was also analyzed over the two locations (shown in Figure S6). The seasonal trend of the relative difference shows a greater consistency over the urban area compared to the rural area. Over the urban area, the mean relative difference (anomaly) in the pre-lockdown weeks (week 1 to week 12) of TROPOMI and OMI was −0.5 × 10 15 molec/cm 2 (6.5%) and 0.1 × 10 15 molec/ cm 2 (13.5%) respectively, while in the lockdown weeks (week 13 to 27) was −2.2 × 10 15 molec/cm 2 (-38.1%) and −1.9 × 10 15 molec/cm 2 (−34.9%) respectively. Over the rural area, the mean relative difference (anomaly) of TROPOMI and OMI in the pre-lockdown weeks was 0.2 × 10 15 molec/ cm 2 (19.9%) and 0.3 × 10 15 molec/cm 2 (22.3%) respectively, while in the lockdown weeks it stood at −0.4 × 10 15 molec/cm 2 (-15.9%) and −0.3 × 10 15 molec/cm 2 (−11.9%) This shows that although the difference of TROPOMI and OMI with regards to their relative difference or the anomaly was higher during the pre-lockdown months, the difference was lower during the lockdown period. It suggests that despite the underestimation in TROPOMI columns compared to the OMI columns, the calculated anomaly from both the sensors is quite similar and specially in lockdown weeks it is within 5% of each other.

Biomass-fire
Wheat crop-residue burning in the region is an annual activity that takes place around the month May and June 8 . Figure S7 shows the monthly crop-residue events from the NASA's Fire Information for Resource Management System (FIRMS).

Top-down NOx emission estimation
The mean top-down NOx emission rate inventory for the BAU is shown in Figure S10. The total emission rate in the 300×600 km 2 domain was 18.34 kg/sec that corresponds to 1585.16 metric tons/day (by assuming constant diurnal emission rate). In 2020, the 60×60 km 2 region around central Delhi (not including Dadri power-plant) alone accounted for 2.78 kg/sec or 240.19 tons/day NOx emission. The same region in 2019 emitted 3.37 kg/sec or 291.16 tons/day. Spatially local peaks of the emission were found within 3 pixels (10 km) of the emission sources such as power-plants [PM1]. Emission clusters corresponding to strongly emitting power-plants, factories, highways are identifiable during BAU. Especially highways that radiate from Delhi have higher NOx emission due to vehicular traffic as well as industrial clusters along highways.
Bottom-up and regional NOx emission inventories are often found to have been overestimated  30 . A reason for a higher emission estimation could be that the above-mentioned bottom-up studies report annual averages that did not consider seasonality of emissions. For example, power-failure leads to a greater use of diesel generators in Delhi during summers, hence higher diesel generator related emissions are during summer than winter. Non-consideration of such seasonal variation in emission may partially explain top-down emission estimation being lower than bottom-up emission inventory.

Emission uncertainties
Top-down emissions were calculated via E = S + D = = / + ∇( ). Due to the uncertainties in the datasets used and assumptions, the emission estimates have uncertainties.
Tropospheric NO2 retrievals suffer uncertainty in slant column density (due to measurement noise and spectral fitting errors), stratospheric slant column (due to error in separating stratospheric and tropospheric NO2) and tropospheric AMF (due to model parameter errors such as assumed profile shape 14 ). The trace gas vertical profile is needed to derive VCD by separating AMF from SCD. Column uncertainty due to AMF is about 30% 14 . If the satellite retrieval assumed NO2 height profile has a smaller aerosol fraction close to surface compared to the true profile, then tropospheric AMF will be overestimated and correspondingly the retrieved tropospheric NO2 VCD will be underestimated 14 . Shaiganfar et al., (2011) compared OMI NO2 VCD with MAX-DOAS observations over Delhi and found that OMI retrievals underestimate high concentrations. Over highly polluted regions tropospheric NO2 VCDs are partially underestimated due to shielding of emitted NO2 by aerosols 5 . The random errors are reduced by temporal and spatial averaging, while a major part of systematic errors is expected to cancel out through the difference and ratio in the defined anomaly metric.
Estimating D (divergence) and S (sink) is complicated mainly by the variation in wind fields and chemical transformation 9,15 respectively, in addition to the tropospheric NO2 retrieval. As long as the wind has a constant speed and direction and these parameters are known with a high certainty, the steady-state assumption can be applied.
[PM3]However slow winds which change directions with time and space complicate emission estimation as such scenarios have high uncertainty (relative to wind speed) and the sudden change in wind direction breaks the steady-state assumption. However, the error due to non-stationary state is smaller near point sources, which is further diminished by taking a multi-temporal mean. The NCEP modeled winds have an uncertainty of about 20% in speed and 14 o in direction 16 . Based on the sensitivity analysis by Beirle et al. 9 , the speed and direction uncertainty would lead to an uncertainty of 15% and a systematic low bias of 3% respectively in D. Another problem is that while the wind speeds and directions change with altitude, the vertical profile of the trace gas itself is not well-known. For example, the near-surface emissions (from vehicles) and chimney-stack emissions are injected at different altitude and their dispersion is subjected to different wind speeds. The vertical levels chosen for the wind fields result in an uncertainty of 10% in the corresponding D 9 . The total uncertainty due to the wind fields is the quadratic sum of 15% due to wind speed and 10% due to the plume height, amounting to about 20%.
Chemical reaction rates may also add uncertainties in the assumed lifetime and the constant Leighton ratio. As stated earlier NO2/NOx ratio is lower when close to the freshly emitting source in space and time, which increases in aged plumes after chemical conversion 17 . Near strong emission sources emitted NO may not be quickly converted to NO2 if the NO mixing ratios locally exceed those of ozone. The NO/NO2 steady state is completely achieved only after ambient air has mixed with the emitted plume 5 . Turbulence in wind speed may enhance mixing and conversion of NO to NO2 near the source or it may disperse NO downwind before it gets converted to NO2 17 . We assumed a constant L (NOx/NO2) of 1.32 neglecting possible changes by day of the week. We assessed the uncertainty in assuming a constant ratio using NO and NO2 vertical column densities modeled from March to May, 2018 over Asia as part of another project, 'Japan's Study for Reference Air Quality Modeling' (J-STREAM) 18 . As shown in Figure S13, L is lowest (~1.3) over polluted regions like urban areas (Delhi) and over a power-plant (Harduaganj) and is relatively high (~1.4), slightly decreasing between March to May. Standard deviation over polluted regions is 0.06 or about 5%.
We assumed that chemical loss of NOx is first-order given by lifetime varying from 7 hours in January to 4 hours in July 19 . Such a variation only partially accounts for the changes in due to seasonality in actinic flux but neglects the changes due to different ozone concentrations on weekends 20 or during the lockdown 21 and dilution by wind. Uncertainty in this assumption was studied using the method of lifetime fit in which the downwind decay of NO2 from strong NOx emission sources was used for estimating the lifetime. This method was applied to randomly chosen date for each month between January to June over power-plants and urban areas. The lifetime so calculated were within 1.5 hours of the assumed the lifetime. The associated uncertainties in were therefore 30%.
Overall, the uncertainties were added in quadrature and summarized in Table S3. Uncertainties in D dominate over area-and point-based emission such as urban areas and power-plants while uncertainties in S dominate over other regions such as rural-areas.