What caused the extreme CO concentrations during the 2017 high pollution episode in India?

The TROPOspheric Monitoring Instrument (TROPOMI), launched 13 October 2017, measures carbon monoxide (CO) concentrations in the Earth’s atmosphere since early November 2017. In the first measurements, TROPOMI was able to measure CO concentrations of the high pollution event in India of November 2017. In this paper we studied the extent of the pollution in India, comparing the TROPOMI CO with modelled data from the Weather Research and Forecast model (WRF) to identify the most important sources contributing to the high pollution, both at ground-level and in the total column. We 5 investigated the period between 11 and 19 November 2017. We found that residential and commercial combustion was a much more important source of CO pollution than the post-monsoon crop burning during this period, which is in contrast to what media suggested and some studies on aerosol emissions found. Also, the high pollution was not limited to Delhi and its direct neighbourhood but the accumulation of pollution extended over the whole Indo-Gangetic Plain (IGP) due to the unfavourable weather conditions in combination with extensive emissions. From the TROPOMI data and WRF simulations, we observed 10 a build-up of CO during 11-14 November and a decline in CO after the 15 of November. The meteorological situation, characterized by low wind speeds and shallow atmospheric boundary layers, was most likely the primary explanation for the temporal accumulation and subsequent dispersion of regionally emitted CO in the atmosphere, emphasizing the important role of atmospheric dynamics. Due to its rapidly growing population and economy, India is expected to encounter similar pollution events more often in future post-monsoon and winter seasons unless significant policy measures are taken to reduce residential 15 and commercial emissions.

Indo-Gangetic Plain (IGP) of India in support of future pollution mitigation efforts. With WRF we (4) also study the role of meteorology in the accumulation and spreading of CO.
The data and methods section describes the datasets that are used and the setup of the WRF model. In the results section, CO levels measured by TROPOMI over Southeast Asia and by ground-level pollution measurement stations are compared with WRF data. The model is used also to attribute the high total column average mixing ratios over India to specific emission 5 categories as presented in section 4. In this section also the role of meteorological conditions is discussed as well as the results of sensitivity tests on CO chemistry in the model.

TROPOMI
TROPOMI has a shortwave infrared spectrometer module, from which the total column average mixing ratio (XCO) is retrieved 10 using the measured radiance around 2.3 µm. Due to its high spatial and temporal resolution, TROPOMI is able to observe global CO vertical columns on a daily basis (Landgraf et al., 2016).
We used data from 14 orbits of TROPOMI retrieved between 11 and 19 November 2017 that covered the Northern part of India. As in the study of Borsdorff et al. (2018a) on the first TROPOMI CO results, we used XCO values that were retrieved using the operational algorithm SICOR (Landgraf et al., 2016). TROPOMI data were filtered for clear sky observations, and 15 cloudy sky observations with a cloud top height < 5000 m and an aerosol optical thickness >0. 5. Borsdorff et al. (2018c) found that including low-level cloud data increased the amount of available measurements, while hardly affecting the ability to measure relatively small scale sources by applying the SICOR algorithm to data from the SCanning Imaging Absorption SpectroMeter for Atmospheric CHartographY (SCIAMACHY).
We removed the two most westward pixels of every swath, which suffer from a not yet resolved performance issue (Borsdorff 20 et al., 2018a). The first validation study showed that the TROPOMI data is in good agreement with CAMS data with a global mean difference of +3.2% and a Pearson correlation coefficient of 0.97 (Borsdorff et al., 2018b). Moreover, only a small mean bias of 6 ppb, with a standard deviation of 3.9 and 2.4 ppb for respectively clear and cloudy skies has been found compared to ground-based total column measurements of TCCON (Total Carbon Column Observing Network). The signal-to-noise ratio of TROPOMI is high compared to previous satellite instruments retrieving CO (Borsdorff et al., 2018a). 25 The TROPOMI averaging kernel (AK) provides information on the vertical sensitivity of the satellite instrument for each single retrieved CO column (Borsdorff et al., 2014). The relationship between the reported CO vertical profile (C retr and the true CO profile (C true ) is given by Eq. 1. In this equation C retr is the retrieved CO profile and C prior the a priori CO profile.
According to Borsdorff et al. (2014): In this study, we compare the CO columns from TROPOMI, derived from C retr , with the modelled columns from WRF. To make a fair comparison between the TROPOMI CO columns and the modeled CO columns, the AK has been applied in the same way to the modeled CO vertical profile (Eq. 1), by replacing C true with the modelled profiles.

CAMS
The Copernicus Atmosphere Monitoring Service (CAMS, https://atmosphere.copernicus.eu) provides data on air quality in 5 6-hourly time intervals at a global resolution of 0.25°x0.25°. The CAMS CO reanalysis product is derived from the output of a week. In this research we used the CAMS CO reanalysis products at various pressure levels and the total column product (http://apps.ecmwf.int/datasets/data/cams-nrealtime/levtype=sfc/).

WRF
To model XCO and ground concentrations at high spatial resolution we used WRF version 3.8.1 (http://www.wrf-model.org/) with the Advanced Research WRF core (ARW). WRF is a numerical non-hydrostatic model developed at the National Centers 20 for Environmental Prediction (NCEP). It has several choices of physical parameterizations, allowing application of the model to a large range of spatial scales (Grell et al., 2005). Our model domain of 2900 km by 2010 km is over the northern part of India and parts of Pakistan, Nepal, China and Bangladesh, including parts of the Himalaya mountain range (see Fig. 1a). Our model employed a 10x10 km 2 resolution and 29 vertical eta levels, and used the Mellor-Yamada-Janjic (MYJ) planetary boundary scheme (Janjic, 1994), the Unified Noah land surface model for surface physics (Ek et al., 2003;Tewari et al., 2004), and the 25 Dudhia scheme (Dudhia, 1989) and the Rapid Radiative Transfer Method (RRTM) for short-wave and long-wave radiation (Mlawer et al., 1997). Cloud physics are solved with the Grell-Freitas cumulus physics ensemble scheme (Grell and Freitas, 2014).
Our boundary and input meteorological conditions, on 6-hourly basis, were based on ECMWF reanalysis data, similar to the CAMS model. WRF calculates its own meteorology in between these 6-hourly time steps and nudges towards the 30 meteorological boundary conditions every 6 hours. The boundary conditions for CO were from the CAMS CO data on pressure levels, interpolated to the WRF model levels.   Deposition and chemical production from Volatile Organic Compounds (VOCs) are not included in our base setup. The deposition process is slow compared to transport of CO out of the model domain, and direct CO sources over the highly populated (IGP) of Northern India are much larger than the indirect source from VOC oxidation.
However, in a sensitivity simulation (see Section 4.3) we accounted for the chemical reaction between the Hydroxyl radical (OH) and CO using the JPL recommended temperature and pressure dependent reaction rate (Burkholder et al., 2015). Carbon 5 monoxide production from the oxidation of methane and other VOCs are included in this simulation as well. In this chemistry simulation, we used the CO production from the TM5-4DVAR system (Krol et al., 2013) and the corresponding OH climatology based on Spivakovsky et al. (2000) and scaled by 0.92 (Huijnen et al., 2016(Huijnen et al., , 2010Krol et al., 2013)

Ground-level measurements
The central pollution control board (CPCB) of India measures the air quality at several stations in India (http://cpcb.nic.in/automatic-10 monitoring-data/). All the samples are taken at ground-level and are made available as fifteen minute averages. We only used stations here with CO measurements available between the 15 th of October and the 20 th of November. To obtain measurements representative of the urban background, we excluded stations near large roads showing large CO enhancements.
This selection is needed for a meaningful comparison to WRF simulations at 10x10 km 2 using MACCity emissions at only 0.5°x0.5°resolution. In Fig. 2 all stations used for comparison with WRF are listed.

Comparing WRF with TROPOMI and ground-level measurements
As outlined before (section 2.1), the averaging kernel was applied to the WRF data using Eq. 1. Both WRF and TROPOMI data were averaged on a 0.25°x0.25°grid to make the comparison less sensitive to local outliers in the data. We also averaged over several days of data, concentrating on two periods: 11-14 November 2017, and 15-19 November 2017, in order to obtain  outer city). This distinction was used to investigate differences in the source signature of CO inside and outside of cities.

TROPOMI and CAMS over South-East Asia
In some of the first TROPOMI observations collected in the first half of November 2017, the northern part of India, more 10 specifically, the IGP, stood out by its high XCO values (see Fig. 3). XCO values were even significantly higher over the IGP than over any region of South-East Asia, even higher than over China. This is remarkable, since in earlier studies, China used to be the most polluted region of the world (e.g., Baldasano et al., 2003;Kan et al., 2012). On the other hand, China has recently been active in reducing air polluting emissions, including CO (Zheng et al., 2018), while in India, emissions continued to increase over the past years (Krotkov et al., 2016).

15
It was estimated that China reduced its CO emissions by 23% between 2013 and 2017 (Zheng et al., 2018). India only took its first steps to improve the air quality in December 2017 by implementing the National Clean Air Program (NCAP), i.e., after the high pollution event studied in this paper. This makes it plausible that the New Delhi region was more polluted in this

Comparing WRF to TROPOMI
We compared our WRF results with the TROPOMI data, and found that WRF could reasonably well reproduce the high XCO values spread over the whole IGP during 11-14 November 2017 and the lower XCO values during 15-19 November. Fig. 5 shows that both modelled and TROPOMI retrieved XCO are very high in the north-west of India (Fig. 5,

CO columns over and outside of the Indo-Gangetic Plain
The XCO levels measured by TROPOMI and modelled by WRF are clearly enhanced during 11-14 of November over the IGP compared to more southerly regions of India (Non-IGP, Fig. 6). The IGP CO total columns are on average 30 ppb higher than then over non-IGP regions (see Fig. 1b for areas of IGP and non-IGP). When we average over 15-19 November, this difference between the IGP and the non-IGP mostly disappears; the column average XCO over the Indo-Gangetic Plains is now lowered 5 from 162 to 129 ppb for TROPOMI and from 152 ppb to 124 ppb for WRF, while the non-IGP XCO only slightly decreased for TROPOMI (124 to 118 ppb) and remained nearly equal at 129 ppb for WRF. A WRF simulation based on MACCity without GFAS (green bars), shows the same XCO pattern. Since the emissions of MACCity are not changing day-by-day, the difference between the periods is solely caused by different meteorological conditions (see also section 4.2).  boundary layer depth. The CO concentrations generally reach lower levels after the 16 th of November (Fig. 7). The WRF model largely follows the CO enhancement and reduction pattern, although the diurnal cycle seems delayed by 3 hours compared to the ground-based measurements. This might be due to the hourly time profiles that were used for the emissions, which were derived for Europe (van der Gon et al., 2011), but do account for the local time shift. In Fig. 7, we zoom in on November 11-20, the days for which also TROPOMI data are available. Averaged time series are shown of the measurements collected at stations 5 in the provinces of Delhi, Punjab, and Uttar Pradesh and the corresponding averaged WRF concentrations. The stations inside the cities (Fig. 7a) show a clear reduction in mixing ratio during the latter half of this period (1050 ppb, 15-19 November), compared to the first half (1700 ppb, 11-14 Nov). The observed reduction, which we observed also in TROPOMI XCO, is reproduced by WRF (1400 to 880 ppb). At locations outside cities (Fig. 7b) this pattern is less pronounced, both in WRF and in the measurements. WRF largely follows the measured CO mixing ratios, but slightly underestimates the CO values after the 10 16 th of November. WRF shows enhanced XCO during the 15 th and 16 th of November, which is not observed.

Comparing WRF to ground-level measurements
To further investigate the origin of the XCO variations, the contribution of different emission categories in WRF is shown in Fig. 8. We show here the inner-city stations, as these are the areas were most people live, but the picture is not very different for outer-city stations (see also Table 2, section 4.1). As can be seen, the surface concentrations are much less sensitive to the background CO (black) compared to the total column mixing ratios. On all days the category residential and commercial 15 combustion contributes most to the total CO concentration (on average 67% for ground-level and 35% for the total column including the background). Other large contributors are industrial processes and combustion and traffic. Surprisingly, we find a rather small contribution from fires to the total mixing ratio of 1-2% in our simulation with MACCity and standard GFAS emissions (see section 4.1, Table 2). Even with strongly enhanced GFAS emissions, the contribution remains on average within 20%. The larger XCO measured at inner-city stations -compared to the stations outside the cities-also point to large 20 contributions from urban emissions.

Discussion
We found XCO values of over 200 ppb in substantial parts of Northern India, in both the TROPOMI and model simulations.
From the satellite data and total column WRF mixing ratios, it is clear that CO is not only enhanced directly around Delhi, but over the whole Indo-Gangetic Plain, with very high values west of Delhi.
25 The background CO is however rather constant, and day-by-day variations in XCO are caused by residential and commercial combustion, similar to what we observed at ground-level.
At the measurements stations that we considered, except for Punjab at ground-level, only a minor contribution from fire was found of 1% to 2%, both for total column and ground-level CO (see Table 2). At ground-level in Punjab the average and maximum contributions were 6% and 44%, respectively over the whole modelled period of 1 October to 19 November. In the 5 11-19 November period, the maximum contribution of biomass burning to the ground-level contribution there was 23%, with an average of 2% (see Table 2).
There are strong indications that GFAS might severely underestimate the fire emissions (Mota and Wooster, 2018). Cusworth et al. (2018) concluded in their recent paper on biomass burning in India that the resolution of the MODIS satellite instrument, on which GFAS fire emissions are partly based, misses many small fires. In addition, thick smoke from fires might lead to an 10 underestimation of fire emissions from GFAS, as MODIS might identify these as clouds, as was found in a recent study over Indonesia (Huijnen et al., 2016). The results of increasing the fire emissions by a factor 5 -10 in WRF are shown in Figs. 5 and 6. Adding biomass-burning emissions in the WRF simulation does not lead to a higher spatial correlation between WRF and TROPOMI but CO levels get closer to TROPOMI values in the 11-14 November period, so it might be that the GFAS fire emissions were indeed underestimated in this period. However, the mixing ratios during the 15-19 November period are 15 overestimated with respect to TROPOMI when higher GFAS emissions are assumed (see Figs 5,6). Alternatively, MACCity already explains a very large part of the observed CO levels, and increasing the MACCity emissions by 20% gives rather comparable results to increasing the GFAS fire emissions by a factor 5-10 for the total columns (Fig. 6) In this paper, however, we assume that the emissions of MACCity do not grossly overestimate CO emissions over the IGP.
Compared to TROPOMI and the amount of emissions that might come from fires based on GFAS and GFED, this assumption seems legitimate. When comparing the total emissions of MACCity to the EdgarV4.3.1 emission database of the most recent

Meteorological conditions 10
In general, the post-monsoon and winter season are the seasons with the worst air quality in the IGP. The photochemical loss is low and other meteorological variables, e.g., the absence of rain and low wind speeds, contribute to high levels of pollution.
For November 2017 we identified meteorological conditions as the most important reason why the CO mixing ratios at ground-level and in the total column increased as observed. Although not extreme, the meteorological conditions were favourable for the accumulation of air pollutants. The wind speeds near the surface were low for several weeks: < 2.5 m/s at 10 15 meter height, limiting the advection of CO away from the sources (Fig. 9). The temperatures were relatively low, decreasing from 22°C to 16°C from 1 Nov to 19 Nov, thus limiting vertical convection. The planetary boundary layer heights were low with daily averages between 350 and 580 m, diagnosed from WRF, while the air pressure (around 990 hPa) and relative humidity (up to 70%) were relatively high (Fig. 9). The most important changes that we found in meteorological parameters around the 15 th of November when the CO concentrations started decreasing, are in the wind speed, the wind direction, the relative 20 humidity and the boundary layer height. The wind speeds clearly increased after the 15 th of November and the wind direction changed from a north-westerly direction to easterly winds in this period of the highest CO concentrations and the start of the ventilation. The relative humidity (RH) went up from 55% to 70% on November 15 and decreased afterwards to 45%. The boundary layer was highest on the 18 th of November (580 meter, see Fig. 9) but we found that more to the north-west of the IGP, closer to the Himalayas, boundary layer heights were also exceeding the height of this mountain ridge on November 14 25 and 15 (not shown). The highest CO values around Delhi were found during 13-16 November, so just before the winds were turning and increasing. In our WRF simulation, the most important contributors to the decrease in CO were both the ventilation of the IGP with clean air from the Himalayas followed by advection of the pollution to the south-east, which took place over all days after 15 November, and the outflow of CO towards the Nord-West, around the Himalayas, in the upper troposphere.
We could clearly observe this outflow of CO in the upper layers of WRF on November 14 and 15 and it was also visible in 30 the TROPOMI measurements on the same days (see also : Borsdorff et al., 2018a). The emissions of MACCity that went into the WRF simulation with only MACCity emissions were the same every day of November, which means that the increase and decrease in CO levels in the MACCity-only run (Figs. 6, 7) were due to the meteorological conditions. TROPOMI showed very high levels of XCO (>280 ppb) over Northern India during the high pollution event in India in November 2017. TROPOMI captured the spatial pattern of the pollution, covering not only Delhi, but rather the whole IGP.
November is in the post-monsoon crop burning season, and media and scientific papers pointed to emissions from crop residue burning as the main reason for the high pollution levels over the IGP (Jha, 2017;Vadrevu et al., 2011;Liu et al., 2018;Cusworth 5 et al., 2018).
In this study, we analysed two consecutive periods in November: 11-14 November with the highest CO levels and 15-19 November, when CO levels decreased. High CO levels and a subsequent drop in CO were observed by TROPOMI, in groundlevel measurements, and in our WRF simulations. The meteorological situation, characterized by low wind speeds and shallow atmospheric boundary layers, was most likely the primary explanation for the temporal accumulation of regionally emitted 10 CO in the atmosphere. The increase in wind speed and change of wind direction around 14 November led to the subsequent dispersion. The dominant role of meteorology, rather than emission variations, is supported by the fact that the WRF simulations that used constant emissions during the whole period, showed a similar temporal dependence, including decreasing CO levels after the 15 th of November.
After analysing the contribution of specific emission sectors to the simulated and observed CO levels over India, we conclude 15 that residential and commercial combustion explain the largest fraction of the high CO pollution over the IGP. Biomass burning only plays a minor role in the CO enhancement: on average 1-2% at ground-level, and only 1% to the total column pollution level. In earlier studies, it was found that the GFAS biomass burning data, used in our analyses, likely underestimate the actual emissions of CO (Mota and Wooster, 2018; Cusworth et al., 2018;Huijnen et al., 2016). The comparison of TROPOMI data with our WRF simulations, based on MACCity and GFAS data, confirms that CO emissions are underestimated in the 11-14 20 November period. The difference could be accounted for by increasing the GFAS emissions to 500%-1000% of the value in GFAS, a rather large increase compared to the findings of last named studies. In that case, the contribution of biomass burning to the observed pollution levels becomes more important: in the order of 5%-20%, but it would still remain smaller than the contribution of urban CO emissions. Therefore, unless urban MACCity emissions are largely overestimated and GFAS emissions are underestimated even more, which we consider a less likely scenario, the contribution of urban CO emissions is 25 the most important contributor to the CO pollution inside and out of the cities. These findings have important implications for emission mitigation efforts to avoid extreme pollution levels over the IGP during the post-monsoon period.
Our results have implications for ongoing winter time pollution mitigation efforts in India. Meteorology is found to be key driver of the extreme pollution episodes, which cannot be altered cheaply with the current state of geoengineering technologies.
Hence, to mitigate the pollution, reducing the largest CO emission sources (residential and commercial combustion) remains 30 the best solution short-term and long-term.
Data availability. Data used in this study can be found under ftp://ftp.sron.nl/open-access-data/