Figure 1 shows the flow chart of the POMINO–GEMS retrieval algorithm. There are two essential steps. The first is to calculate tropospheric NO2 SCDs on an hourly basis, through the fusion of total SCDs from the official GEMS v1.0 L2 NO2 product, total SCDs and stratospheric VCDs from the TROPOMI Product Algorithm Laboratory (PAL) v2.3.1 L2 NO2 product and diurnal variations in the stratospheric NO2 from the GEOS-CF v1 product. We then calculate tropospheric NO2 AMFs to convert SCDs to VCDs.

In order to validate satellite NO2 products with surface concentration measurements from MEE, we convert tropospheric NO2 VCDs from satellite products on a 0.05∘ × 0.05∘ grid to surface NO2 mass concentrations using GEOS-Chem-simulated NO2 vertical profiles and the box heights of the lowest model layer (Eq. 9). 9Csurf=VCDtropSAT×RGC×MN×HGC×2

In Eq. (9), Csurf represents the estimated surface NO2 mass concentration (in µgm-3), VCDtropSAT the satellite tropospheric VCD (in molec.m-2), RGC the GEOS-Chem-simulated hourly ratio of NO2 sub-column in the lowest layer to the total tropospheric column, M the NO2 molar mass (in µgmol-1), N the Avogadro constant and HGC the box height of the lowest layer (in m). The thickness of the lowest layer of GEOS-Chem (about 130 m) is too large for the layer average NO2 mass concentration to represent that near the ground (M. Liu et al., 2018); thus, the derived concentration is multiplied by a factor of 2 to roughly account for the vertical gradient from the height of the ground instrument to the center of the model layer. However, the constant correction factor of 2 neglects the diurnal variation in the NO2 vertical gradient, which is related to the diurnal variation in the planetary boundary layer (PBL) heights. This issue is discussed in detail in Sect. 3.4.

We use ground-based MAX-DOAS NO2 measurements, together with POMINO–TROPOMI v1.2.2, OMNO2 v4 and GOME-2 GDP 4.8 NO2 products, to validate the POMINO–GEMS retrieval results. The types, geolocations and observation times of MAX-DOAS stations are summarized in Table S3, and the location of each site is shown in Fig. S6. Details of each site are described in Sect. S2. Kanaya et al. (2014) and Hendrick et al. (2014) have discussed the error in MAX-DOAS NO2 retrieval, where uncertainties from a priori aerosol and NO2 profiles are the largest source by 10 %–14 %, and the total retrieval uncertainty is typically 12 %–17 %.

To ensure sampling consistency in time, we average all valid MAX-DOAS measurements within each observation period of GEMS (i.e., 30 min) for hourly comparison and within ± 1.5 h of the TROPOMI, OMI and GOME-2 overpass times for daily comparison. Following the procedures in previous studies (Lin et al., 2014; Liu et al., 2020), we exclude all matched MAX-DOAS data for which the standard deviation exceeds 20 % of the mean value to minimize the influence of local events. To ensure sampling consistency in space, we select valid satellite pixels within 5 km of MAX-DOAS sites for POMINO–GEMS and POMINO–TROPOMI v1.2.2, 25 km for OMNO2 v4 and 50 km for GOME-2 GDP 4.8 and then conduct spatial averaging. The Grubbs statistical test, which is used to detect outliers in a univariate dataset assumed to exhibit normal distribution (Grubbs, 1950), is performed to exclude outliers in both MAX-DOAS and satellite data before comparison. Only one data pair from the Fudan University site is identified as an outlier and removed (Fig. S7), and we have 1348 matched hourly data pairs in total.

We use tropospheric NO2 VCDs from mobile car MAX-DOAS measurements performed by the Chinese Academy of Meteorological Sciences (CAMS) in the Three Rivers source region in July 2021 (Cheng et al., 2023). The Three Rivers source region is on the northeastern Tibetan Plateau in western China, which is isolated from massive anthropogenic activities and hence a good place for observations of atmospheric compositions in the background atmosphere. The field campaign lasted from 18 to 30 July 2021 and included four closed-loop journeys, beginning from the meteorological bureau of the city of Xining (the capital of Qinghai Province) to the meteorological bureau of Dari county of the Guoluo Tibetan Autonomous Prefecture and then from the meteorological bureau of Dari county to the meteorological bureau of the Yushu Tibetan Autonomous Prefecture, and finally back to Xining city (Fig. S6). The spectral analysis of the measurement spectra in the fitting window of 400–434 nm was implemented with the DOAS method. A sequential Fraunhofer reference spectrum (FRS) is used to derive NO2 differential slant column densities (DSCDs), which are then converted to VCDs by adopting the geometric approximation method. The errors are estimated to be less than 20 % at high altitudes. More detailed descriptions of instrumentation, field campaign and data retrieval are in Cheng et al. (2023).

We average all valid mobile car MAX-DOAS measurements within each observation period of GEMS in each 0.05∘ × 0.05∘ grid cell to ensure spatiotemporal consistency. Over relatively clean areas with little human influence and biomass burning, such as the Three Rivers source region, a large portion of NO2 is located in the middle and upper troposphere, which is not accounted for in the mobile car data via such a DSCD-based retrieval method. Indeed, Cheng et al. (2023) showed that the official TROPOMI NO2 VCDs are higher than mobile car data by about 40 %. Considering that the diurnal variation in the middle and upper tropospheric NO2 is much smaller than that in the lower troposphere, we focus on the correlation of NO2 diurnal variation between POMINO–GEMS and mobile car MAX-DOAS data.

We use hourly surface NO2 mass concentration measurements from the MEE air quality monitoring network. By 2021, more than 2000 MEE stations across China had been established, providing hourly observations for NO2 and five other air pollutants. Most stations are in urban or suburban areas.

The spatial distribution of all MEE sites in the GEMS FOV is shown in Fig. S8a and that of MEE sites over urban, suburban and rural regions is shown in Fig. S8b–d, respectively. The classification of sites is based on Tencent user location data, with a horizontal resolution of 0.05∘ × 0.05∘ for every 0.5 s from 31 August to 30 September 2021 (Fig. S8e), as adopted from previous work (Kong et al., 2022). Here, urban MEE sites are defined as places where the mean location request times is larger than 50 times per second, suburban sites refer to 5–50 times per second, and rural sites refer to fewer than 5 times per second. The number of sites for urban, suburban and rural sites is 808, 554 and 71, respectively.

At MEE sites, molybdenum-catalyzed conversion from NO2 to NO and subsequent chemiluminescence measurement of NO is done to estimate NO2 concentrations. The heated molybdenum catalyst has low chemical selectivity, leading to strong interference from other oxidized nitrogen species such as nitric acid (HNO3) and peroxyacetyl nitrate (PAN). Therefore, MEE data tend to overestimate the actual NO2 concentrations, with the extent of overestimation at about 10 %–50 % (Boersma et al., 2009; M. Liu et al., 2018). The overestimation is dependent on the oxidation level of NOx but is currently unclear for each site and hour.

To compare with satellite-derived surface NO2 concentration data, we average over all valid MEE sites in each 0.05∘ × 0.05∘ grid cell to generate gridded MEE NO2 data for each hour. To ensure sampling consistency for each day, we average MEE observations at two consecutive hours to match GEMS hourly observations – for example, we match the mean value of MEE NO2 concentrations at 13:00–14:00 and 14:00–15:00 local solar time (LST) with the GEMS NO2 at 13:45–14:15 LST. We also match MEE observations over the periods 13:00–14:00 LST, with TROPOMI-derived and OMI-derived surface NO2, and 9:00–10:00 LST, with GOME-2-derived surface NO2.

Figure 4 shows mean POMINO–GEMS tropospheric NO2 VCDs at each hour on a 0.05∘ × 0.05∘ grid in JJA 2021. High values of tropospheric NO2 columns (> 10 × 1015 molec.cm-2) are evident over populous regions such as South Korea, central and eastern China and northern India. Clear hotspot signals reveal intense NOx emissions over city clusters such as Beijing–Tianjin–Hebei (BTH), the Yangtze River Delta (YRD), the Pearl River Delta (PRD) and the Seoul metropolitan area (SMA), as well as isolated megacities such as Osaka and Nagoya in Japan, Chengdu and Ürümqi in China and New Delhi in India. Tropospheric NO2 VCDs are much lower (< 1 × 1015 molec.cm-2) over most of western China and the open ocean, due to low anthropogenic and natural emissions.

Figure 5a–c present NO2 VCDs in the morning, noon and afternoon in JJA 2021 for eastern China. Data are averaged at 22:45–01:45 UTC (06:45–09:45 Beijing Time, BJT), 02:45–04:45 UTC (10:45–12:45 BJT) and 05:45–07:45 UTC (13:45–15:45 BJT) to represent the morning, noon and afternoon, respectively. In the morning (Fig. 5a), there are clear city signals with high-NO2 values, reflecting abundant NOx emissions from traffic. The spatial gradients of NO2 from urban centers to outskirts are very strong. However, these spatial gradients are greatly reduced in the noon and afternoon (Fig. 5b and c). For example, the differences in the tropospheric NO2 VCDs between the urban center of Xi'an (108.93∘ N, 34.27∘ E) and its surrounding areas (within 50 km) are reduced from about 8×1015 molec.cm-2 in the morning to about 4×1015 molec.cm-2 at noon and then to below 2×1015 molec.cm-2 in the afternoon. This is likely due to chemical loss of traffic-associated NO2, increased emissions from other sectors (e.g., industry) and/or enhanced horizontal transport smearing the spatial gradient.

Over western China with low tropospheric NO2 VCDs (Fig. 5d–f), there is a gradual increase in the tropospheric NO2 by about 1 × 1015 molec.cm-2 from the early morning to noon. This increase is likely dominated by biogenic NOx emissions that are sensitive to sunshine intensity and surface temperature (Kong et al., 2023; Weng et al., 2020). Future studies are needed to understand the exact causes.

Figure 6 shows the diurnal variation in the POMINO–GEMS tropospheric NO2 VCDs over six different region groups in the GEMS FOV. The six groups are defined based on the levels of mean POMINO–GEMS tropospheric NO2 VCDs at 12:00 LST in JJA 2021 (VCD12:00LST), and their spatial distributions are also shown in each panel. We convert the observation time from UTC to LST for each time zone in this domain (+5 time zone at 70–82.5∘ E; +6 time zone at 82.5–97.5∘ E; +7 time zone at 97.5–112.5∘ E; +8 time zone at 112.5–127.5∘ E; +9 time zone at 127.5–140∘ E) and show the NO2 diurnal variations in each time zone with different colors. For low-NO2 situations (VCD12:00LST≤2×1015 molec.cm-2), NO2 grow in the morning in the +5 and +6 time zones but not in other time zones. Over high-NO2 situations (VCD12:00LST>8×1015 molec.cm-2; in cities and suburban areas), NO2 in all time zones exhibits a minimum around noontime and a morning peak at 09:00–10:00 LST, which is consistent with previous findings for specific polluted locations (Boersma et al., 2008, 2009; J. Li et al., 2021; Ghude et al., 2020; Herman et al., 2019; Biswas and Mahajan, 2021). In all groups and time zones, tropospheric NO2 VCDs grow from noon to the afternoon.

The NO2 diurnal variations are related to multiple driving factors. Different sources with distinctive diurnal patterns dominate the NOx emissions over different regions. Lightning and biogenic activities are the major emission sources over low-NO2 land areas, and they tend to intensify with temperature and radiation in the daytime. Anthropogenic emissions are dominant over polluted cities and suburban areas, where the traffic emissions tend to peak in the mid-morning and late afternoon (Jing et al., 2016; Y.-H. Liu et al., 2018; Naiudomthum et al., 2022). In addition, the photochemistry plays an important role. NO2 is in chemical balance with NO, and the ratio of NO2 and NO depends on radiation, ozone and peroxyl radicals. NOx is oxidized to nitric acid and organic nitrates by radicals in the daytime, the level of which depends on radiation, ozone and volatile organic compounds. Thus, the lifetime of NO2 reaches the minimum value around noon, i.e., a few hours in summer. Furthermore, atmospheric transport also affects the diurnal variation in the NO2 at high-value places (e.g., cities) and their surroundings. Further studies are needed to determine the exact causes of NO2 diurnal variations at individual places.

Figure 7a and b show the POMINO–GEMS and POMINO–TROPOMI v1.2.2 tropospheric NO2 VCDs, respectively, on a 0.05∘ × 0.05∘ grid averaged over JJA 2021. Cloud screening is implemented based on the CRFs from each product. To ensure temporal compatibility, matching between hourly GEMS observations and the TROPOMI data at the closest observation time is done for each day. Overall, POMINO–GEMS agrees well with POMINO–TROPOMI with a spatial correlation coefficient of 0.98, a linear regression slope of 1.18 and a small positive NMB of 4.9 % (Fig. 7c). Regionally, POMINO–GEMS VCDs are higher than those of POMINO–TROPOMI v1.2.2 over eastern China, most of India and the northwestern GEMS FOV but smaller over western China and the oceans (Fig. 7a and b; see Fig. S9c and d for plots showing the differences). These differences are related to tropospheric NO2 AMFs and SCDs. A detailed discussion is given in Sect. S3.

Figure 7d–f and g–i show the comparison results of POMINO–GEMS tropospheric NO2 VCDs with OMNO2 v4 on a 0.25∘ × 0.25∘ grid and GOME-2 GDP 4.8 on a 0.5∘ × 0.5∘ grid averaged over JJA 2021, respectively. POMINO–GEMS NO2 VCDs exhibit good spatial consistency with the two independent products (R=0.87 and 0.83), although with slightly lower values than OMNO2 v4 (by 16.8 %) and GOME-2 GDP 4.8 (by 1.5 %). These VCD differences are expected, considering the differences in the retrieval algorithm. For example, the POMINO–GEMS algorithm implements explicit aerosol corrections in the radiative transfer calculation, while OMNO2 v4 and GOME-2 GDP 4.8 treat aerosols as “effective clouds”. POMINO–GEMS accounts for the anisotropy of surface reflectance by adopting MODIS BRDF coefficients, whereas OMNO2 v4 and GOME-2 GDP 4.8 use geometry-dependent and regular LER, respectively. The horizontal resolution of a priori NO2 profiles in POMINO–GEMS is 25 km (and interpolated to 2.5 km), 1∘ × 1.25∘ in OMNO2 v4 and 1.875∘ × 1.875∘ in GOME-2 GDP 4.8 (Valks et al., 2019a; Lamsal et al., 2021).

Based on comparisons with POMINO–TROPOMI v1.2.2, OMNO2 v4 and GOME-2 GDP 4.8 NO2 VCDs, we conclude that POMINO–GEMS NO2 columns show good agreement with LEO satellite data, with values lower by 20 % at most.

The scatterplot in Fig. 8a compares POMINO–GEMS tropospheric NO2 VCDs in JJA 2021 at all GEMS observation hours with matched ground-based MAX-DOAS measurements at nine sites. POMINO–GEMS correlates with MAX-DOAS (R=0.66), with a small negative bias (NMB = -11.1 %). The linear regression shows a slope of 0.51 and intercept of 3.34×1015 molec.cm-2, reflecting underestimation of POMINO–GEMS tropospheric NO2 VCDs on high-NO2 days.

Figure 8b and c further use MAX-DOAS measurements to evaluate POMINO–GEMS and POMINO–TROPOMI v1.2.2 tropospheric NO2 VCDs at the overpass time of TROPOMI. In Fig. 8b, POMINO–GEMS data at 13:45–14:15 LST are used to match the overpass time of TROPOMI. The POMINO–TROPOMI product is evaluated in the context of understanding the relative performance of POMINO–GEMS. Each data point represents a day. Figure 8b and c show that the day-to-day variability in the MAX-DOAS measurements is well captured by POMINO–TROPOMI v1.2.2 (R=0.83) but less so by POMINO–GEMS (R=0.65). Linear regression results show an underestimate of tropospheric NO2 VCDs in POMINO–TROPOMI v1.2.2 (NMB = -18.1 %), as also found in previous studies (Liu et al., 2020). POMINO–GEMS exhibits a small bias (NMB = -9.0 %), but the station-dependent performance is apparent. At the two remote sites of Fukue and Cape Hedo with low NO2, POMINO–GEMS NO2 columns are higher than those of MAX-DOAS measurements. At the other sites, the data pairs are more scattered and located both above and below the 1:1 line, resulting in a small NMB.

Figure 8d shows the NMBs and correlation coefficients of POMINO–GEMS NO2 VCDs relative to ground-based MAX-DOAS data at each hour. The negative NMBs reach a maximum of about 20 % at 10:00 LST and decrease to less than 10 % in the afternoon. The correlation coefficients are modest or high (0.45–0.73) at all hours.

Figure 9 compares the diurnal variation in the tropospheric NO2 VCDs between POMINO–GEMS and MAX-DOAS at eight stations. At each site, NO2 values are averaged in JJA 2021 at each hour for comparison, and the number of valid days for each hour is also shown. The Cape Hedo site is not included because there are few valid MAX-DOAS data points at each hour. Figure 9a–f show that at the urban and suburban sites, MAX-DOAS NO2 (black lines) peaks in the mid-to-late morning, declines towards the minimum values at noon around 13:00 LST and then gradually increases in the afternoon. A strong correlation of NO2 diurnal variation between POMINO–GEMS (solid red lines) and MAX-DOAS is found at Xuzhou (R=0.82), Hefei (R=0.96), Fudan University (R=0.84), Nanhui (R=0.79) and Xianghe (R=0.94). At the Dianshan Lake site, POMINO–GEMS NO2 columns increase, but MAX-DOAS data decrease from 08:00 to 09:00 LST, resulting in a lower correlation coefficient (R=0.60). At Chongming and Fukue sites, MAX-DOAS NO2 shows a peak in the morning without an evident increase in the early afternoon, but this diurnal pattern is not fully captured by POMINO–GEMS. At Fukue, POMINO–GEMS NO2 exhibits abrupt changes at 12:00 and 13:00 LST due to few valid data points.

In addition, comparison of POMINO–GEMS diurnal variation with NO2 data from GOME-2 in the morning and OMI and TROPOMI in the early afternoon shows good agreement at Hefei, Nanhui, Dianshan Lake, Chongming and Fukue sites. The differences between POMINO–GEMS and MAX-DOAS NO2 VCDs are comparable to or smaller than those between LEO satellite and MAX-DOAS NO2 VCDs.

As we use TROPOMI total NO2 SCDs to correct those of GEMS, this may influence the NO2 diurnal variation in the original GEMS observations. Thus we also compare MAX-DOAS data with re-calculated POMINO–GEMS tropospheric NO2 VCDs without correction in total SCDs (dashed red lines in Fig. 9). Compared to our default POMINO–GEMS data (with correction), excluding the correction leads to lower diurnal correlation coefficients at Xuzhou, Hefei, Fudan University, Nanhui and Dianshan Lake but higher correlation coefficients at Xianghe, Chongming and Fukue. Excluding the correction increases the NMB at three sites but decreases the NMB at five sites. We conclude that at these eight sites (in the eastern areas), no significant influence on the diurnal variation in the POMINO–GEMS tropospheric NO2 VCDs is brought in through TROPOMI-based correction for total NO2 SCDs.

Figure 10 compares the diurnal variations between POMINO–GEMS and mobile car MAX-DOAS tropospheric NO2 VCD data in the Three Rivers source region on the Tibetan Plateau. Results of POMINO–GEMS with and without total SCD correction are shown in the solid and dashed red lines, respectively. Mobile car MAX-DOAS data show an evident decrease in the tropospheric NO2 VCDs from the morning to noon, with little change thereafter. Such NO2 diurnal patterns reflect the spatial and temporal variations in the tropospheric NO2 along the driving route. The high-NO2 values with large standard deviation at 09:00 BJT are due to enhanced pollution and variability in the morning when the mobile car is in or near Xining city. The NO2 diurnal variations in the POMINO–GEMS with correction correlate well with those of mobile car MAX-DOAS data (R=0.81). In contrast, POMINO–GEMS without total SCD correction exhibits much poorer correlation with mobile car MAX-DOAS data, due to the erroneous increase in the afternoon.

Overall, the validation results with independent ground-based and mobile car MAX-DOAS measurements provide confidence on the general characteristics of POMINO–GEMS NO2 diurnal variations.

The scatterplot in Fig. 11a compares surface NO2 concentrations derived from POMINO–GEMS with MEE measurements at all hours. POMINO–GEMS-derived surface NO2 concentrations show good agreement with MEE measurements in terms of spatiotemporal correlation (R=0.78) and bias (NMB = -26.3 %) but are higher than those of MEE at some high-value situations, which mainly occur over the YRD region (Fig. S14). These differences reflect errors in POMINO–GEMS NO2 VCDs themselves, errors in the conversion process from tropospheric NO2 VCDs to surface concentrations, and errors in the MEE data (due to potential contamination by nitric acid and organic nitrates; M. Liu et al., 2018).

Figure 11b and c show the validation results for satellite-derived surface NO2 concentrations with MEE measurements at the overpass time of TROPOMI (i.e., early afternoon). Here, each data pair denotes a MEE site. POMINO–GEMS results at 13:45–14:15 LST are used to match the overpass time of TROPOMI data. Overall, both satellite-based datasets show good spatial correlation with MEE measurements (R=0.63 and 0.61). POMINO–GEMS exhibits higher linear regression slope (0.50) with a smaller NMB (-48.0 %). The values of satellite data are lower than those from MEE, especially in the afternoon (Fig. 11d). This is in part because of the aforementioned contamination issues in MEE data, which becomes more severe in the afternoon as the air ages throughout the daytime.

Figure 12a examines the diurnal variation in the surface NO2 concentrations averaged over JJA 2021 at all sites. The MEE data show a smooth and monotonic decline from the early morning to the early afternoon, with a slight increase beginning at 15:00 LST. This diurnal pattern differs from those seen in ground-based MAX-DOAS VCD data (Fig. 9), due to the difference in sampling size between MEE and MAX-DOAS, the diurnal variation in the NO2 vertical distribution that affects the relationship between surface and columnar NO2 and the insensitivity of NO2 columns to changes in PBL heights. POMINO–GEMS-derived surface NO2 concentrations show similar diurnal variations to those of MEE (R=0.97), although with a peak at 10:00 LST and a gradual increase beginning at 14:00 LST. The discrepancies between POMINO–GEMS and MEE surface NO2 concentrations at different hours are likely caused by the assumed constant correction factor of 2 to account for the vertical gradient of NO2 from the height of the ground instrument to the center of the first model layer (Sect. 2.2). In the morning when the PBL is low, most NO2 molecules are near the ground, and the vertical gradient of NO2 over polluted regions is the largest in the daytime, so the factor of 2 may lead to the underestimation of derived surface NO2 concentrations. In contrast, in the afternoon, the PBL mixing is much stronger, and the vertical gradient of NO2 is much smaller; thus, the factor of 2 may lead to overestimated surface NO2 concentrations. Note that the consistency between POMINO–GEMS and MEE data does not depend on the total SCD correction (Table S4).

To quantify the influences of the diurnal variation in the hourly column-to-surface ratio from GEOS-Chem simulations, we compare the MEE measurements with POMINO–GEMS-derived surface NO2 concentrations using daily column-to-surface ratio (Fig. S15). As expected, POMINO–GEMS-derived NO2 concentrations show a similar diurnal variation to the tropospheric NO2 VCDs, with two peaks in the mid-morning and afternoon and a minimum at noon. The temporal correlation coefficient with MEE is only about 0.23. Thus, it is more reasonable to use an hourly ratio for comparison with MEE measurements, as done in our study.

To further test the reliability of our VCD-to-surface-concentration conversion method (Eq. 9), we apply the same method to MAX-DOAS NO2 VCDs and compare the resulting surface NO2 concentrations with MEE data. As shown in Fig. S16, the diurnal variation in the MAX-DOAS-derived surface NO2 concentrations correlates well with that of MEE measurements (R=0.96), which supports our conversion method.

Figure 12b–d show the comparison of NO2 diurnal variations for different groups of MEE sites. The diurnal variations in the POMINO–GEMS-derived surface NO2 concentrations show similar characteristics over urban, suburban and rural regions, and all results correlate well with those of MEE data. Meanwhile, surface NO2 concentrations derived from LEO satellite observations also agree well with those of POMINO–GEMS, except GOME-2-GDP-4.8-derived surface NO2 concentrations, which are lower than those of POMINO–GEMS by about 30 %–40 %. We conclude that the validation with extensive MEE measurements presents a promising performance of the POMINO–GEMS retrievals, especially the great agreement of the POMINO–GEMS NO2 diurnal variation with MEE data over urban, suburban and rural regions.

Total retrieval errors for POMINO–GEMS tropospheric NO2 VCDs are derived from the calculations of total SCDs, stratospheric SCDs and tropospheric AMFs. Spatial and temporal averaging across GEMS pixels greatly reduces the random errors but hardly affects the systematic errors. Here, we provide a preliminary estimate of POMINO–GEMS errors for the summertime retrieval discussed above.

As described in Sect. 2, we calculate hourly total SCDs based on the original GEMS SCD data and daily TROPOMI-guided corrections. According to the GEMS ATBD of the NO2 retrieval algorithm, the SCD errors from the DOAS method are < 5.65 % at high-NO2 conditions (NO2 VCD >1×1015 molec.cm-2; Park et al., 2020). The NO2 SCD errors in TROPOMI are reported to be 0.5–0.6 ×1015 molec.cm-2 (10 % in a relative sense; Van Geffen et al., 2022a). Given the assumption we made in adjusting GEMS total SCDs to match TROPOMI values, we tentatively estimate the error in our corrected total SCD data to be 0.5–0.7 ×1015 molec.cm-2 (10 % in a relative sense) for most regions and 0.9×1015 molec.cm-2 (20 %–30 %) at the edge of the northwestern GEMS FOV.

In constructing the stratospheric NO2 SCDs, the stratospheric VCDs are taken from TROPOMI PAL v2.3.1, scaled based on GEOS-CF v1 stratospheric NO2 to account for diurnal variation and then applied with geometric AMFs. We assign a constant error of 0.2×1015 molec.cm-2 (5 %–10 %) to our hourly stratospheric SCDs, which is the same as the value for TROPOMI (Van Geffen et al., 2022a). Few studies have assessed the accuracy of stratospheric NO2 and its diurnal variation from GEOS-CF data (Knowland et al., 2022), but our comparison between GEOS-CF and TROPOMI shows great consistency (Sect. 2.1.5). As most of the errors in total SCDs are absorbed in the stratosphere–troposphere separation step (Van Geffen et al., 2015), the errors in tropospheric SCDs should be 10 %–30 %, depending on different cases, with higher relative biases in cleaner situations.

Tropospheric AMF calculations are the dominant error source for the retrieved tropospheric NO2 VCDs over polluted regions. According to Liu et al. (2020), the AMF errors caused by uncertainty in surface reflectance are about 10 %, and errors induced by uncertainties in aerosol parameters are about 10 % in clean regions and 20 % for heavily polluted situations. We further assume that the O2–O2 cloud retrieval algorithm introduces another error at the 10 % level to the NO2 AMFs. The uncertainty in a priori NO2 vertical profiles is estimated to cause an AMF error of 10 % (Liu et al., 2020). Yang et al. (2023) suggested that the NO2 profiles from GEOS-Chem (version 13.3.4) might contain incorrect timing of PBL mixing growth in the morning and thus introduce a relative root mean square error of 7.6 % and NMB of 2.7 % in AMF; however, this error could be greatly dampened by averaging over a long time period. The free tropospheric NO2 bias in GEOS-Chem NO2 profiles might also contribute to the retrieval errors, especially over remote regions. Adding these errors in quadrature leads to the overall AMF errors for POMINO–GEMS at 20 %–40 %.

The overall uncertainty in POMINO–GEMS tropospheric NO2 VCDs is estimated by adding in quadrature the errors in tropospheric NO2 SCDs and AMFs, which is when these errors are expressed in the relative sense. For remote regions with low tropospheric NO2 abundances, the overall retrieval uncertainties can reach 30 %–50 % and are dominated by errors in tropospheric SCDs. For regions with abundant tropospheric NO2, the uncertainties in the retrieved tropospheric VCDs are dominated by the AMF errors and are estimated to be about 20 %–30 %.

As shown in Figs. 8d and 11d, the maximum negative NMB of POMINO–GEMS tropospheric NO2 VCDs relative to ground-based MAX-DOAS data is about 20 % in the mid-morning, and the NMB of POMINO–GEMS-derived surface NO2 concentrations to MEE measurements is -30 % on average. Thus, our estimated error magnitude is supported by the independent ground-based MAX-DOAS and MEE data.