Inﬂuence of convection on the upper tropospheric O 3 and NO x budget in southeastern China

. Thunderstorms can signiﬁcantly inﬂuence the air composition via strong updraft and lightning nitrogen oxides (LNO x ). In this study, the ozonesondes and TROPOMI nitrogen dioxide (NO 2 ) observations for two cases are combined with model to investigate the effects of typical strong convection on vertical redistribution of air pollutants in Nanjing, southeastern China. The ozonesonde observations show higher O 3 and water vapor mixing ratios in the upper troposphere (UT) after convection, indicating the strong updraft transporting lower-level airmass into the UT, and the possible downward O 3 -rich air 5 near the top of UT over the convective period. During the whole convection life cycle, the UT O 3 production is driven by the chemistry ( > 87 %


Introduction
Convection can transport the surface pollutants and moisture from the planetary boundary layer to the upper troposphere (UT) in a short time, where the gaseous pollutants have a longer lifetime due to the slower reaction rates, except for photolysis, in the colder environment (Dickerson et al., 1987). As trace gases remain for more than 1 week in the UT, they are distributed by the upper-level winds around the globe (Ridley et al., 2004). Meanwhile, the vertical profiles of trace gases are reshaped by the updraft and downdraft on timescales of hours (Huntrieser et al., 2016;Barth et al., 2019). On a global scale, the chemical reactions of transported ozone (O 3 ) and its precursors can increase the amount of UT O 3 (Lawrence et al., 2003;Murray, 2016).
Nitrogen oxides (NO x ), volatile organic compounds (VOCs), and hydrogen oxide radicals (HO x ) are the three main O 3 precursors which can be uplifted or produced by thunderstorms (Bozem et al., 2017). Lightning-produced nitrogen oxides (LNO x ) are the dominant natural source of UT NO x , contributing as much as 35 %-45 % of global freetropospheric ozone Liaskos et al., 2015). Several lightning parameterizations have been developed for chemistry transport and climate models to evaluate the relationship between LNO x and other trace gases (Murray et al., 2012;Gordillo-Vázquez et al., 2019;Luhar et al., 2021). Therefore, the precise estimation of LNO x is crucial for the global O 3 trend and feedback between lightning and climate change (Finney et al., 2016;Chen et al., 2021). Globally, the LNO x is estimated as 2-8 Tg N yr −1 , which is substantially less well quantified than the flash rate .
The large uncertainty in LNO x estimation might be reduced by using cloud-resolving chemistry models in combination with satellite and aircraft observations. It is beneficial to take advantage of sonde and satellite observations for exploring the convection effects, especially for the estimation of the LNO x production efficiency (PE). Recently, many studies have used different satellites to quantify the LNO x PEs (Beirle et al., 2009;Pickering et al., 2016;Zhang et al., 2020). Two main methods have been proposed to distinguish LNO x from the NO 2 background pollution: (1) subtracting the weighted temporal average NO 2 of areas with few flashes before the satellite passing time Bucsela et al., 2019;Allen et al., 2019;Lapierre et al., 2020) and (2) directly using customized lightning air mass factors (AMFs) for each convection event (Beirle et al., 2009;Zhang et al., 2020). Recently, Allen et al. (2021) proved the potential of deriving LNO x PEs by the geostationary lightning instruments (e.g., Lightning Mapping Imager, LMI; Yang et al., 2017;Geostationary Lightning Mapper, GLM;Rudlosky et al., 2019) and NO 2 observations such as Tropospheric Emissions: Monitoring of Pollution (TEMPO; Chance et al., 2019). Furthermore, aircraft observations and chemical models indicate that the transport from the stratosphere to the troposphere can also increase the UT O 3 besides the chemical production from LNO x (Pan et al., 2014). As revealed in the Deep Convective Clouds and Chemistry (DC3) 2012 studies and mesoscale convective system simulations, the compensation of subsidence and differential advection beneath the convective core can lead to the anvil wrapping effects (Huntrieser et al., 2016;Phoenix et al., 2020). The different mechanisms of stratosphere-troposphere exchange and the effects on the tropospheric chemistry have been discussed in Holton et al. (1995) and Stohl (2003).
At present, most aircraft observations and model simulations of convection effects focus on the tropics or the United States (Vaughan et al., 2008;Barth et al., 2019). Little is known about the role of convection in southeastern China (Murray, 2016;Guo et al., 2017), where thunderstorms and lightning have increased significantly because of urbanization during recent decades (Yang and Li, 2014;Pérez-Invernón et al., 2021). This is likely due to the increasing aerosol concentration, which can invigorate storms in a moist and convectively unstable environment (Koren et al., 2008;Rosenfeld et al., 2008;Tao et al., 2012). In this study, we combine ground observations and model simulations to investigate the origin of higher UT O 3 and water vapor mixing ratio (Qv) after convection, and we try to distinguish the contributions of physical processes, chemical reactions, and LNO x . For the first time, the TROPOspheric Monitoring Instrument (TROPOMI; Veefkind et al., 2012) NO 2 observations are used to identify LNO x PEs in southeastern China. Section 2 describes the datasets used with a brief introduction of the cloud-resolved chemistry model and the LNO x retrieval method. Section 3 evaluates the model simulations, and the physical and chemical effects of convection are analyzed in Sects. 4 and 5. We apply new a priori NO 2 profiles into the retrieval algorithm to explore the sensitivity of AMFs to LNO x in Sect. 6. Conclusions are summarized in Sect. 7.

Ozonesonde data
Five ozonesondes were launched from the Nanjing National Reference Climatological Station (31.93 • N, 118.90 • E) on 25 July 2019 and 1 September 2020; both days experienced strong convection. Both pre-convection and duringconvection/post-convection campaigns were designed to investigate the convection effects. The convection and ozonesonde trajectories are illustrated in Fig. 1a and b.
Three Institute of Atmospheric Physics (IAP) ozonesondes were launched near the air mass convection that developed on 25 July 2019. The IAP ozonesonde uses an electrochemical concentration cell (ECC). The complete parameters and performance are described in Zhang et al. (2014). The average bias is less than 0.3 mPa from the surface up to 2.5 km, close to zero below 9 km, and less than 0.5 mPa between 9 and 18 km. The first IAP ozonesonde was launched at 05:35 UTC on a sunny day (23 July), and the other two were at 05:10 UTC (pre-convection) and 06:35 UTC (postconvection) on 25 July. Because of water leakage, the preconvection one lost signal just a few seconds after the release, and instead the ozonesonde launched on 23 July is chosen. Although the time interval is 2 d, the largest relative difference of forecast O 3 profiles above 10 km is usually smaller than 25 % (Fig. S2); therefore the daily variation cannot explain the observed difference of more than 65 %.
The two Vaisala ECC ozonesondes were launched successfully at 23:45 UTC on 31 August (pre-convection) and 06:10 UTC on 1 September (during-convection), following the standard manual to ensure the precision is better than 5 % and the accuracy is within ± (5-10) % below 30 km (Smit et al., 2007). The captured squall line was developing from the convergence of cold air and typhoon Maysak's outer region circulation. Note that the during-convection ozonesonde entered directly into the cloud, providing a unique opportunity of exploring the ozone affected by the convective clouds.

Lightning data
Three lightning datasets were used in this study: the China National Lightning Detection Network (CNLDN; Yang et al., 2015), the Earth Networks Total Lightning Network (ENTLN; Marchand et al., 2019), and the World Wide Lightning Location Network (WWLLN; Rodger et al., 2006). The detection efficiency (DE) of cloud-to-ground (CG) flashes is about 90 % for the CNLDN data in Jiangsu Province (Li et al., 2017a), while ENTLN and WWLLN detect both intracloud (IC) and CG flashes with a specific detection frequency (1 Hz-12 MHz for ENTLN and 3-30 kHz for WWLLN). In the ENTLN data, groups of pulses are classified as a flash if they are within 700 ms and 10 km. Both strokes and lightning flashes composed of one or more strokes are included in the preprocessed data obtained from the ENTLN. The detailed processing algorithm of the WWLLN is given by Rodger et al. (2004). The WWLLN strokes and pulses are combined with ENTLN into one dataset (ENGLN) within 10 km and 0.7 s as mentioned in Virts and Goodman (2020). To increase the lightning data coverage in our study, the CG flashes of ENGLN and CNLDN datasets are combined using spatial and temporal clustering criteria of 10 km and 0.5 s (Zhao et al., 2020). Merging these three datasets should provide a sufficiently high CG flash detection efficiency for this analysis. Because the IC DE of all these lightning data is low in China, we conservatively use the merged CG data with a constant IC / CG ratio of 3 : 1 based on Wu et al. (2016) and Bandholnopparat et al. (2020). IC data will become more accurate if more Chinese total lightning networks, such as Beijing Lightning Network (BLNET; Srivastava et al., 2017), are available to be compared with lightning imaging sensors (Rudlosky and Shea, 2013;Poelman and Schulz, 2020).

TROPOMI data
On 13 October 2017, TROPOMI on the Sentinel-5 Precursor satellite was launched successfully . For the current study, we used the Royal Netherlands Meteorological Institute (KNMI) standard product v2.1 test as input to our LNO x retrieval algorithm. A spike removal is included in the product to better deal with the detector saturation and blooming effects, which enables more valid data over the bright clouds generated by convection (Ludewig et al., 2020;van Geffen et al., 2022). Each pixel includes a slant column density quality flag (no2_scd_flag = 0, see Appendix A) that can be used to get the data without known retrieval errors (Allen et al., 2021;van Geffen et al., 2022).
The official NO 2 columns are retrieved using the nearultraviolet and visible (UV-VIS, 405-465 nm) spectrometer backscattered solar radiation measurements on TROPOMI (van Geffen et al., 2015). The retrieval consists of three main procedures for each measured Level-1b spectrum.
1. Total NO 2 slant column density (SCD) is determined by the DOAS (differential optical absorption spectroscopy) method.
2. The stratospheric and tropospheric SCDs are separated by data assimilation of slant columns in the tracer model, version 5, tailored for the application of satellite retrievals (TM5-MP; Williams et al., 2017).
3. The stratospheric and tropospheric NO 2 vertical column densities (VCDs) are obtained via the air mass factor (AMF) lookup tables (Lorente et al., 2017).
We replaced the tropospheric AMF (AMF trop ) with a new AMF called AMF for LNO x (AMF LNO x ) to derive the tropospheric LNO x vertical column density (VCD LNO x ). The concept of AMF LNO x is inherited from the AMF trop derived by a function of several parameters (solar zenith angle, viewing zenith angle, relative azimuth angle, surface albedo, surface pressure, cloud fraction, cloud height, and a priori trace gas profile). Briefly, the numerator is the modeled tropospheric NO 2 slant column density (SCD tropNO 2 ) and the denominator is the modeled VCD (VCD NO 2 or VCD LNO x ). In detail, these two AMFs can be calculated as follows: (1 − f r ) p tp p surf w clear (p)NO 2 (p) dp +f r p tp p cloud w cloudy (p)NO 2 (p) dp p tp where p surf is the surface pressure, p tp is the tropopause pressure, p cloud is the cloud optical pressure, f r is the cloud radiance fraction in the NO 2 window, w clear and w cloudy are respectively the pressure-dependent scattering weights from Then, their VCD LNO x can be calculated as the VCD NO x subtracted from a tropospheric NO x background. Because our AMF LNO x converts the SCD tropNO 2 to VCD LNO x directly, the additional estimation of background NO 2 is not needed for calculating LNO x PE in Sect. 6.2.

Model simulations
This study uses WRF-Chem version 4.1.4. The initial and boundary conditions of meteorological parameters are provided by the hourly ECMWF atmospheric reanalysis (ERA5) data (Hersbach et al., 2020). Simulations are performed with one-way nesting with 75 vertical levels and a 50 hPa model top. The domain settings are illustrated in Fig. S1. The microphysical processes are computed with the WRF Single-Moment 6-class scheme (WSM6; Hong and Lim, 2006), while the shortwave and longwave radiation is calculated by the Rapid Radiative Transfer Model for GCMs (general circulation models) scheme (RRTMG; Iacono et al., 2008). The land surface processes are simulated by the Noah scheme (Koren et al., 1999). However, we use different planetary boundary layer (PBL) parameterizations to simulate the convection. Specifically, the 2019 case uses the Yonsei University scheme (YSU; , while the quasinormal scale elimination (QNSE;Sukoriansky et al., 2005) is applied to the 2020 case.
The chemical initial and boundary conditions are defined using the output from the Whole Atmosphere Community Climate Model (WACCM; https://www.acom.ucar. edu/waccm/, last access: 22 April 2022). The initial O 3 profile of the 2020 case is replaced by the O 3 profile from the ozonesonde. Anthropogenic emissions are driven by the 2016 Multi-resolution Emission Inventory for China (MEIC) version 1.3 (http://www.meicmodel.org/, last access: 22 April 2022). The Model of Emissions of Gases and Aerosol from Nature (MEGAN; Guenther et al., 2006) is used for biogenic emissions. The chemical mechanism is the Model for Ozone and Related chemical Tracers (MOZART) for gas-phase chemistry and Goddard Chemistry Aerosol Radiation and Transport (GOCART) model for aerosols (Pfister et al., 2011). The photolysis rates are adjusted by the presence of aerosols and clouds using the new TUV (tropospheric ultraviolet and visible) photolysis option with the scaled cloud optical depth (cloud_fraction 1.5 ). Note that the bimodal profile modified from the standard Ott et al. (2010) profile is employed as the vertical distribution of LNO in WRF-Chem (Laughner and Cohen, 2017), while the LNO x parameterization is activated as 500 mol NO per flash . The resulting lightning nitrogen monoxide (LNO) and lightning nitrogen dioxide (LNO 2 ) profiles are defined as the difference of vertical profiles between simulations with and without lightning.
To simulate the convection and LNO x realistically, a lightning data assimilation (LDA) technique is applied to WRF-Chem. The details of the LDA technique can be found in Fierro et al. (2012) and Li et al. (2017b) and are briefly illustrated here. The water vapor mass mixing ratio is increased at constant temperature layers in columns where flashes occur: where Q sat is the water vapor saturation mixing ratio (g kg −1 ), Q g is the graupel mixing ratio (g kg −1 ), and X is the flash rate. In our simulations, the layer between 263.15 and 290.15 K is chosen to let the convection root itself in the PBL quickly as the Q v in the lower troposphere is the deeper layer (Marchand and Fuelberg, 2014;Finney et al., 2016;Li et al., 2017b). Parameter settings follow Li et al. (2017b): A = 0.94, B = 0.2, C = 0.001, and D = 0.25 and α = 2.2. The resampled total lightning flash data are read through the auxiliary input stream of WRF every 10 min. For example, if the beginning time of LDA is 05:00 UTC with a time step of 10 min, all flashes in a specific grid between 05:00 and 05:10 UTC are summed as the contribution during this period. At the next time step, the flashes are classified as the next new group. Therefore, the flash count is the flash rate density (units: flashes per 10 min per dx km per dy km), where dx and dy are the resolutions of model grids in the x and y direction, respectively. This leads to flash rate densities that are the same through all nested domains, as done in Fierro et al. (2012) and Li et al. (2017b), and these are used as the flash counts in WRF-Chem directly following the method employed in CMAQ (Community Multiscale Air Quality; Kang et al., 2019aKang et al., , b, 2020.

Model evaluation
Compared with the radar observations, the simulated initiations of convection are ahead by 60 and 30 min for the 25 July 2019 and 1 September 2020 cases, respectively (Figs. 2 and 3). The lightning assimilations are applied with the same time step forwards. For comparisons below, we choose the matched stages instead of the same time.
The air mass convection on 25 July 2019 was initialized as isolated cells. WRF-Chem reproduces the position and intensity of isolated convections at the initial stage ( Fig. 2a and d). At 05:40 UTC, the cells were presented with the northeastsouthwest orientation, and the column maximum radar reflectivity (CRF) reached 60 dBZ (Fig. 2b), which is stronger than the simulated convection (CRF = 55 dBZ, Fig. 2e). The vertical cross section of simulated radar reflectivity across the core of cells is compared with the observations (Fig. S3). Although there were much missing data caused by the long distance between convection and radar, the horizontal and vertical structures of isolated cells are roughly shown without artificial interpolation. While the simulated 45 dBZ contour extends to 12 km, the observed one only reaches 10 km because of reduced data quality above 10 km.
The squall line on 1 September 2020 was formed in the north, strengthened, and moved towards the observation site (Fig. 3). The strongest convective stage with a CRF of 60 dBZ was at 05:50 UTC, which is the TROPOMI overpass time (Fig. 3b and e). Although the highest level reached was lower than that of the 2019 case, the reflectivity in the lower troposphere (2-8 km) was larger and broader (Fig. S4). Note that the simulated dissipated cells deviate from the radar observation, and this leads to the region for the ozonesonde comparison moving to the west of the station (Fig. 3c and f).
The measured O 3 and Qv profiles at different convection stages are shown in Fig. 1c and d. Generally, the observed UT O 3 and Q v are higher with convection, while the largest enhancements are between 10 and 16 km. However, the 2020 case showed a larger increase in the lower troposphere (2-8 km, LT). Additionally, a two-valley shape of O 3 profile exists in both cases but at different levels: 2 and 8 km for the 2019 air mass and 4 and 10 km for the 2020 squall line. Although the WRF-Chem model tends to underestimate the O 3 concentration in the LT and UT for the 2019 and 2020 cases, respectively, it can reproduce the detailed O 3 structures and provide the opportunity to analyze the mechanisms of convection.
Three possible sources can explain the enhancements of O 3 in the UT: convective transport, chemical production, and O 3 directly produced by lightning. Only the first two factors are discussed in detail in Sect. 4 as lightning O 3 is beyond the scope of this study and still uncertain, as shown by limited observations and model simulations (Morris et al., 2010;Ripoll et al., 2014).

Convection impacts
To evaluate the higher UT O 3 concentration after the convection, the mean vertical profiles of O 3 in the atmospheric regions passed by the ozonesondes are illustrated at three stages of convection: initiation, development, and dissipation ( Fig. 4a and d). The UT O 3 increases continuously during the whole cycle of the 2020 case, but for the 2019 case, it declines at the developing stage because of the uplift of O 3poor air, while later it starts rising again. This phenomenon can be explained by the O 3 vertical cross sections at the developing stages (Fig. 4b and e). For the 2019 case, the cells of low O 3 concentration reach 16 km by the updraft, and then the high O 3 air wrapped behind the convection moves into the region. However, the observed increasing O 3 of the 2020 case is mostly from the vertically transported background O 3 .
To determine the processes causing the differences between the two cases, we analyzed the outputs of mean integrated physical rates (IPRs) from 10 to 14 km during the convective period ( Fig. 4c and f). Generally, the opposite trend of the horizontal advection (advh) and vertical advection (advz) governs the net decrease in the UT O 3 production rate of the 2019 case. Note that the advz contribution is negative between 10 and 11.5 km and positive between 11.5 and 13.8 km. This is due to the uplifted O 3 -poor air and downward O 3 -rich air caused by the stronger updraft compared with the 2020 case. As indicated by the higher Q v after convection (Fig. 1c) and tropopause height (Fig. 4b), the updraft of the 2020 case is not strong enough to wrap the stratospheric O 3 like the mesoscale convective system (Phoenix et al., 2020). While the dynamic processes play an important role in the O 3 production, the positive chemistry contribution cannot be neglected in both cases and leads to the net increase in UT O 3 during the convective period of the 2020 case. Specifically, the chemistry increases O 3 in both cases,

Impact of lightning NO x on the O 3 profile
Furthermore, the IPR outputs including LNO emissions are compared with these excluding LNO to explore the effects of LNO x on O 3 (Table 1). The LNO x reduces the net O 3 production by 25 % during the convective period and life cycle of the 2019 case. The decreased chemistry contribution is less significant (≤ 1 %) for the 2020 case which has a smaller lightning density near the station. Note that the LNO x can certainly enhance the downwind ozone production on the scale of days (Pickering et al., 1996;DeCaria et al., 2005). Therefore, it is necessary to estimate the LNO x PE accurately (discussed later in Sect. 6.2).
Additionally, the convection was divided into three regions by TROPOMI data: fresh lightning, downwind of fresh lightning, and aged lightning (See Sect. 6.1 and Fig. 8 for details). Firstly, the difference of O 3 ( O 3 ) profiles is obtained with different LNO PE assumptions (Fig. 5a-c). In contrast with the net loss of ozone (< 4 ppbv) over all height levels in Ott et al. (2007), the O 3 is mostly positive (< 1 ppbv) between 2 and 5 km and negative (> −3 ppbv) between 5 and 12 km in our cases. The higher PE (700 mol per flash) slightly reduces the O 3 concentration by less than 1 ppbv at all levels compared with the default PE (500 mol per flash), and it even leads to negative O 3 between 2 and 5 km downwind of fresh lightning (Fig. 5b). The maximum O 3 loss is between 8 and 10 km due to the peak of LNO x (up to 2.6 ppbv) introduced in the model.
Then, the integrated reaction rate (IRR) is applied to determine the chemistry mechanism and the effect of LNO on the O 3 variation for two layers in which the O 3 is opposite: 800-500 hPa ( O 3 > 0) and 500-200 hPa ( O 3 < 0). The tropospheric O 3 is mainly controlled by five reaction rate terms (Pickering et al., 1990;Bozem et al., 2014): where k i is the rate coefficient of the reaction between peroxy radicals (R i O 2 ) and NO. The time series of each contribution to O 3 production are illustrated in Fig. 5d-i. Overall, the time series of IRR are more variable for the 2019 case due to the stronger activity as clarified in Sect. 4. The total net chemistry production of O 3 remains positive for both layers. In detail, the reaction between NO and HO 2 always dominates the production, while the oxidation of NO by RO 2 is about 40 %-60 % of that production. The dominant loss of O 3 is the photolysis (described by the reaction of O( 1 D) and H 2 O), while the reaction between O 3 and OH is comparable during the convective period. The lowest contribution to O 3 loss, O 3 + HO 2 → OH + 2O 2 , is reduced during the convection because of the production of LNO which captures the HO 2 reacting with O 3 . Note that although the increase in total IRR induced by LNO x can reach 1.36 × 10 7 and 2.60 × 10 6 molec. cm −3 s −1 in the low layer and high layer over these three regions, respectively, the net O 3 production actually decreases in the high layers ( Fig. 5a-c) due to the combination of dynamic transport and chemical production related to LNO x .

Relation between lightning and TROPOMI products
As the LNO x production estimation from TROPOMI depends upon the SCD tropNO 2 , we compare the SCD tropNO 2 distributions with the observed lightning flashes (Fig. 6a-d).
Although the SCD tropNO 2 over the most active pixels is not valid due to the detector saturation and blooming effect, the nearby or outflow regions still have useful data. While the flashes occurred less than 30 min before the TROPOMI overpass time in the 2019 case, the 2020 case had both fresh and aged LNO 2 (Fig. 6d).
Specifically, the SCD tropNO 2 of the convective pixels (f r ≥ 0.7) is smaller than that in other regions. This is opposite to previous studies of large-scale convective systems with high flash density (Beirle et al., 2009). Four factors could lead to this unexpected result: the cloud top heights, flash counts, flash occurrence time, and background NO 2 . Either the inadequate flash or weak convection could lead to a smaller SCD tropNO 2 over pixels with f r ≈ 1 because TROPOMI can only see the LNO 2 above the clouds. In other words, the polluted NO 2 below the broken or thinner clouds is partially exposed if f r < 1. The sensitivity tests of the WRF-Chem a priori SCD tropNO 2 can help explain this phenomenon clearly (Fig. S5). The pixels of high SCD tropNO 2 with low cloud fraction belong to background NO 2 pollution ( Fig. S5a and e), but the SCD tropNO 2 increased by UT LNO 2 is still visible compared with the lower SCD tropNO 2 without LNO 2 (Fig. S5b-d and f-h).
An LNO PE upper limit of 700 mol NO per flash (Ott et al., 2010) is applied to WRF-Chem for investigating the importance of LNO x for the AMF trop and AMF LNO x calculations. The changes in the retrieved AMFs are examined by replacing profiles in three tropospheric layers independently: middle troposphere (MT, 800 to 400 hPa), up- per troposphere (UT, 400 to 150 hPa), and troposphere (surface to tropopause). Unless otherwise specified, the changes in AMFs are obtained by increasing LNO x in this section. Figure 7 shows that the AMF changes are mostly controlled by the LNO x in the UT layer where the detection sensitivity is high (Beirle et al., 2009;Laughner and Cohen, 2017) and the LNO x production reaches the peak (Fig. S6). While the AMF LNO x decreases by 5 %-40 % for both cases, the changes in AMF trop ( AMF trop ) are regionally specific and can be classified by the lightning activity: fresh lightning (MT AMF trop < −20 %), downwind of fresh lightning (MT AMF trop > 20 %), and aged lightning (UT AMF trop > 20 %). Figure 8a illustrates the relation-ship between p cloud and f r over these three regions. The clouds are higher than 400 hPa (p cloud < 400 hPa), and f r is larger than 0.8 over fresh lightning pixels, but both aged lightning and downwind of fresh lightning areas have clouds lower than 400 hPa. This coincides with the mean cloud pressures in Fig. S6 and explains why UT AMF trop > 20 % exits in Fig. 7bi and biii, indicating the possibility of LNO x estimations over the aged lightning regions (Sect. 6.2).
As can be used to determine which parameter controls AMF trop : enhanced a priori SCD tropNO 2 or a priori VCD tropNO 2 (Fig. 8b-d). Briefly, the dominant one The tropospheric NO 2 slant column density (SCD tropNO 2 , filled color) and cloud optical pressure (line color). These white grid cells stand for missing TROPOMI data (no2_scd_flag = 0), as defined in Appendix A. The solid black border is Jiangsu Province. (c, d) The cloud radiance fraction in the NO 2 window and flashes whose color depends on the occurrence time relative to the TROPOMI overpass time.  belongs to the larger ratio. First, if the LNO 2 is included in the tropospheric layer of a priori NO 2 profiles (Fig. 7ci), the AMF trop decreases over most fresh lightning pixels because of the increased a priori VCD tropNO 2 (Fig. 8b). The situation is opposite for the downwind of the fresh lightning region.
There the AMF trop is larger no matter which layer is chosen ( Fig. 7ai-ci) because the LNO 2 was convected, reached above the cloud top, and led to the larger a priori SCD tropNO 2 (Fig. 8c). Interestingly, for the 2020 case, the AMF trop of aged lightning pixels increases more than 50 % for the UT layer (Fig. 7biii). It demonstrates the important role of advected UT LNO 2 and that the cloud exists as a barrier, causing the difference between a priori SCD tropNO 2 and a priori VCD tropNO 2 (Fig. 8d). Although the difference is smaller than that of the other two regions due to the LNO 2 lifetime, it is still useful for retrieving the LNO x . Moreover, considering the region-specific LNO x effects on AMFs, we need to include the representation of LNO 2 in the TROPOMI NO 2 retrievals better, especially outflow regions. The comparisons (Fig. S6) between the TROPOMI standard NO 2 profiles from TM5-MP and WRF-Chem also illustrate the importance of LNO x , resolved convection transport, and emissions. Aircraft observations of NO and NO 2 will be useful to determine the exact roles (Laughner and Cohen, 2017).

Estimations of LNO x
Satellite observations with fresh convection are usually used to estimate the LNO x PE. However, it is difficult to apply the same method to regions with small convection like the 2019 case because of the pixel saturation of TROPOMI and limited coverage of convective area (Fig. 6c). Instead, we focus on dissipated convection, which is the southern part of the 2020 case, while the northern part has multiple missing values regardless of the adequate flashes and high cloud fractions. As shown in Fig. 9b-c, the time differences between flashes and TROPOMI overpass time are longer than 30 min but shorter than 3 h. Since the lifetime of NO 2 is ∼ 3 h in or near the field of convection (Nault et al., 2016) and ranges from 2 to 12 h depending on the convective location, these pixels can still be used for the LNO x estimation. Equation (5) is applied to determine the mean LNO x PE (mol per flash): where p stands for pixels affected by LNO x , V i (mol m −2 ) is the LNO x vertical column densities (VCD LNO x = SCD tropNO 2 / AMF LNO x ) over pixel i with an area called A i (m 2 ), N is the total number of flashes contributing to VCD LNO x , and the exponential component considers the lifetime of NO x for each flash (F j ). Specifically, t 0 is the time of TROPOMI overpass, t j is the time of the lightning flash, and τ represents the NO x 3 h lifetime near convection. As the dissipated convection produced enough lightning and the UT winds within the storm were blowing from west-northwest to east-southeast (Fig. 9a), the pattern of VCD LNO x can still be clearly identified (dashed rectangle in Fig. 9). Fortunately, there is a low VCD LNO x strip separating the northern and southern convection. With the careful selection, the LNO x PE is estimated as 60 mol NO x per flash. Although there are a few lightning flashes related to the VCD LNO x outside of the region selection, it only affects the LNO x PE by ≈ 2 mol, which is within the uncertainty discussed below.  (2019) and Zhang et al. (2020), the uncertainty of LNO x is determined by the LNO x lifetime, lightning DE, NO / NO 2 ratio, LNO profile, and other sources (summarized in Table 2). The lifetime (τ ) of NO 2 is replaced by 2 and 6 h to evaluate the uncertainty as 27 %, while another uncertainty is also 27 %, which is related to lightning DE by changing the ratio of IC to CG to 2 : 1 and 4 : 1. The uncertainty caused by modeled NO / NO 2 ratios is assumed to be 30 % based on Allen et al. (2019), and the uncertainty related to the LNO profile is 26 % by using the a priori NO 2 profile with 330 and 700 mol NO per flash. The uncertainty associated with the stratospheric vertical column is considered as 7 % by applying a bias of ± 10 14 molec. cm −2 (van Geffen et al., 2022). The uncertainty caused by other possible error sources including systematic errors in slant columns, optical cloud pressure, and NO 2 redistributed by convection is difficult to quantify and is assumed to be 10 % following Allen et al. (2021). Assuming no correlation between errors, the total uncertainty (56 %) is estimated as the square root of the sum of the squares of all individual uncertainties. As a result, the LNO x PE is 60 ± 33 mol NO x per flash. It is less than that of our previous work (90 ± 50 mol NO x per flash) over the continental United States  and at the lower end of 120 ± 65 mol NO x per flash obtained in Allen et al. (2021). Thus, more studies over China are necessary for the estimation of region-dependent LNO x PEs.

Conclusions
Both the 2019 and 2020 cases saw enhanced UT O 3 concentrations according to the ozonesonde observations. As revealed by the WRF-Chem model, the dynamic contribution of O 3 variation in the 2019 case was generated by mixing with the UT O 3 -rich air due to strong updraft, while it was caused by vertical advection of high background O 3 in the 2020 case. The detailed analysis of integrated physical rates shows that the dynamic processes dominate the UT O 3 decrease during the convective stage of both cases. However, in the convection life cycle, the contribution of chemistry reactions to the UT O 3 production is 5-10 times that of dynamics. Moreover, the UT O 3 enhancement of the 2019 case Figure 7. The percent differences of AMFs by replacing the a priori NO 2 profiles at three layers: middle troposphere (left), upper troposphere (middle), and troposphere (right). AMF trop is the comparison of the AMF trop with 500 mol NO per flash relative to 0 mol NO per flash.
AMF LNO x is the comparison of the AMF LNO x with 700 mol NO per flash relative to 500 mol NO per flash. Three regions are annotated: fresh lightning (blue), downwind of fresh lightning (red), and aged lightning (black). Because of the quite large AMF LNO x values in pixels with little lightning, AMF LNO x is shown over pixels where 0 < AMF LNO x < 10. decreases by 40 % if the lightning nitrogen oxides (LNO x ) are included in the model, indicating the importance of the LNO x . Utilizing the outputs of integrated reaction rates improves our understanding of the chemical reactions and effects of lightning nitrogen oxides (LNO x ) on O 3 . While the reaction of NO and HO 2 dominates the production of O 3 , the O 3 loss processes are suppressed because the LNO x reacts with the HO 2 which is the reactant in the production.
The WRF-Chem results are incorporated into the retrieval algorithm to explore how LNO x affects the official TROPOMI products and prove that TROPOMI is useful for LNO x studies over small-scale convection regions. The sen-sitivity tests imply that air mass factors of tropospheric NO 2 are smaller for fresh lightning regions and the opposite for the aged lightning pixels or downwind of fresh lightning. The air mass factors of tropospheric LNO x always decrease with increased LNO x . Since the LNO x affects the variation in tropospheric NO 2 in the outflow regions, better consideration of LNO x is essential for studies of the tropical and mid-latitude regions in summer.
Because the saturation of TROPOMI pixels and blooming effects lead to the failure of detecting LNO x over active convection, we focus on the dissipation regions where aged LNO x still exists. The production efficiency is estimated to  be 60 ± 33 mol LNO x per flash, which is less than but still within the range of previous studies. Although the current results are limited, the technique can be implemented worldwide to region-specific LNO x retrievals. To quantify and refine the LNO x production estimates, TROPOMI data over both active and dissipated convection could provide valuable information.

Appendix A: Flag definitions used in this study
The no2_scd_flag is introduced to make the usage of the NO 2 slant column density (SCD) data easier by gathering information from a few variables into one flag (Table A1). This flag can thus be used for filtering, though with care as it probably does not cover all possible situation. Here "δ" refers to the SCD error (in mol m −2 ), and "pqf" stands for processing quality flag. No SCD value due to saturation limit exceeded, i.e., pqf = 54 0 SCD with δ < 3.3 × 10 −5 and no error reported 1 SCD with δ < 3.3 × 10 −5 and error reported: pqf = 55 2 SCD with δ < 3.3 × 10 −5 and other error reported, e.g., pqf = 41 3 SCD with δ ≥ 3.3 × 10 −5 and no error reported 4 SCD with δ ≥ 3.3 × 10 −5 and error reported: pqf = 55 5 SCD with δ ≥ 3.3 × 10 −5 and other error reported, e.g., pqf = 41 FillValue No SCD due other error (prior to the differential optical absorption spectroscopy fit)

Appendix B: Contributions to ∆AMF trop
In Fig. 8b- , where the LNO x superscript indicates that the a priori variable is calculated with LNO x (500 mol NO per flash), and the noLNO x superscript is without LNO x . The two contributions are derived by taking the logarithms of Eqs. (B1) and (B2) and then subtracting them in Eq. (B3). Here, several abbreviations are defined to simplify the symbols: S is SCD tropNO 2 , V is VCD tropNO 2 , subscript 1 is with LNO x , and 0 is without LNO x .
Therefore, if S 1 S 0 is larger than V 1 V 0 (the solid line is higher than the dashed line in Fig. 8b-d), then AMF 1 is larger than AMF 0 . In other words, these two variables determine how a priori LNO x affects the retrieval of NO 2 .
Author contributions. YY directed the research, and YY, XZ, and RvdA designed the research with feedback from the other coauthors. XZ, RvdA, HE, and JvG developed the retrieval algorithm. XZ modified lightning assimilation code written by YL. JLL provided guidance and supporting data on the ENTLN data. XZ, KC, XK, ZZ, JH, CH, JZ, XY, and HC participated in the field campaigns. XZ performed simulations and analysis with the help of YY, RvdA, XK, JC, CH, and RS. XZ, YY, RvdA, and JLL interpreted the data and discussed the results. XZ drafted the manuscript with comments from the co-authors. RvdA, YY, and JLL edited the manuscript.
Competing interests. The contact author has declared that neither they nor their co-authors have any competing interests.
Disclaimer. Publisher's note: Copernicus Publications remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.