Fine-Scale Columnar and Surface NO x Concentrations over South Korea: Comparison of Surface Monitors, TROPOMI, CMAQ and CAPSS Inventory

: Fine-scale nitrogen oxide (NO x ) concentrations over South Korea are examined using surface observations, satellite data and high-resolution model simulations based on the latest emission inventory. While accurate information on NO x emissions in South Korea is crucial to understanding regional air quality in the region, consensus on the validation of NO x emissions is lacking. We investigate the spatial and temporal variation in ﬁne-scale NO x emission sources over South Korea. Surface observations and newly available ﬁne-scale satellite data (TROPOspheric Monitoring Instrument; TROPOMI; 3.5 × 7 km 2 ) are compared with the community multiscale air quality (CMAQ) model based on the clean air policy support system (CAPSS) 2016 emission inventory. The results show that the TROPOMI NO 2 column densities agree well with the CMAQ simulations based on CAPSS emissions (e.g., R = 0.96 for June 2018). The surface observations, satellite data and model are consistent in terms of their spatial distribution, the overestimation over the Seoul Metropolitan Area and major point sources; however, the model tends to underestimate the surface concentrations during the cold season.


Introduction
Nitrogen oxides (NO x = NO + NO 2 ) play an important role in tropospheric chemistry as well as in the formation of surface ozone and secondary aerosol formation for particulate matter. Most NO x emissions are released in the form of NO, but they react quickly and are transformed into NO 2 . Thus, NO 2 is a good indicator of NO x emissions, and using NO 2 concentration as a proxy of NO x emissions has been accepted for practical purposes [1][2][3][4][5]. Similarly, the observed concentration of surface NO 2 concentration or vertically integrated column density has been used to estimate the amount of NO x emissions, especially in highly industrialized and urbanized areas such as East Asian cities. As providing accurate information on the intensity and location of emission sources is crucial to improving a chemical transport model, numerous studies have used observational data from surface and space-borne measurements to evaluate a model's performance and update emission inventory information [6,7].

Surface Observations
Hourly observations of surface NO and NO 2 concentrations during 2018 were obtained from the National Institute of Environmental Research (NIER). Data from 305 surface monitoring sites, including 251 urban monitoring sites (urban), 38 roadside sites (roadside) and 17 background sites (rural), were selected after screening data quality and availability. These measurements were taken from chemiluminescence monitors with a molybdenum converter that often overestimated the NO 2 concentration caused by interference from other nitrogen species (NO z ) [42,43]. These interferences were greatest on a relative basis during the afternoon on high-O 3 days when surface NO x concentrations were lowest overall, usually due to high mixing height and/or shortened NO x lifetime by photochemical losses. We did not correct any potential bias caused by such interference in this study.

TROPOMI
TROPOMI, onboard Sentinel-5P satellite, is one of the most recently available instruments used to monitor NO x emissions from many anthropogenic sources in the troposphere. TROPOMI is a hyperspectral spectrometer with a nadir-viewing 108 degree field-of-view, with wavelength coverage over ultraviolet-visible (270 nm to 495 nm), near infrared (675-775 nm) and shortwave infrared (2305-2385 nm) [44,45]. Compared with its precedents (i.e., SCHIAMACHY and OMI), TROPOMI has a superior spatial resolution, with a 7 × 3.5 km resolution at the nadir and more coverage of clear-sky observations thanks to a high signal-to-noise ratio. We used a level 2 product (S5P_L2__NO2) from NASA GES DISC, which is based on the DOMINO algorithm, a differential optical absorption spectroscopy method, a pre-calculated air-mass factor look-up table and a stratospheric and tropospheric contribution separation method based on a data-assimilated chemistry transport model. A quality band, qa_value, was also provided for each pixel (ranging from 0 (poor) to 1 (excellent)), and we filtered out all the pixels with a qa_value less than 0.75 to remove cloud interference and ensure retrieval quality.

Model
Our modeling framework uses the WRF [46] model v3. 3 and community multiscale air quality (CMAQ) [47] model v4.7.1 to simulate the meteorology and chemistry. The meteorology-chemistry interface processor (MCIP) [48] v3.6 and sparse matrix operator kernel emission (SMOKE) are used as a preprocessor and as an emission processing model, respectively. The AERO5 module and statewide air pollution research center version 99 (SAPRC99) [49] are used as the aerosol module and chemical mechanism, based on anthropogenic emissions, from CREATE 2015 (Asia) and [50] CAPSS 2016 (South Korea) [41], and biogenic emissions (Model of Emissions of Gases and Aerosols from Nature; MEGAN) [51]. Natural emissions from dust and fire are not included in the simulations due to their high uncertainty. Although dust and fire emissions do not significantly add to NO x concentration in South Korea, the impact through the boundary condition (e.g., NO x emissions from Siberian fires) could make a partial contribution. Multiscale modeling has been conducted over East Asia in 27-km domain, and over South Korea in 9-km domain. Comparisons with satellite and observations in this study were managed within the 9-km domain. Model configurations are described in Table 1.

CAPSS Emissions
CAPSS, developed by the National Institute of Environmental Research, is an integrated air quality management system designed to support pollutant emission control and air quality management policy in South Korea. CAPSS provides emission information from South Korean emission sources, including point, area, on-road and non-road emission sectors. It uses a multilevel classification of the sources of emissions, with 12 upper-level, 54 intermediate-level, 312 lower-level and 527 detailed-level categories. CO, NO x , SO x , PM 10 and VOCs emissions are available for each upper-level category except for fugitive dust emissions [41]. "Nonindustrial combustion plants", "combustion in manufacturing industries" and "production processes, storage and distribution of fuels" are common in the point and area emission sectors. The point sector also includes "combustion in energy industries" and "waste treatment and disposal", while the area sector includes "other sources and sinks" and "fugitive dust". "Solvent use", "other mobile sources and machinery" and "fugitive dust" belong to the mobile sector. Point source emissions are estimated using both the direct method, based on the real-time measurement of pollutant emissions, and the indirect method, based on emission factors and activity data. Emissions from area sources are estimated using the indirect method. Table 2 summarizes the major source classification code categories and their amount in NO x emissions from CAPSS 2007, 2010, 2013 and 2016. The sector with the highest NO x emissions is road transport. Total mobile sources (which also include additional non-road mobile sources) account for around two-thirds of NO x emissions in South Korea. NO x emissions from South Korean sources were estimated to increase by 13.7% from CAPSS 2013 to CAPSS 2016. While several studies have tried to evaluate the reliability of NO x emissions using satellite measurements, few have fully assessed the CAPSS emission inventory, especially the latest version.

Spatial and Vertical Data Processing
To compare satellite models, conducting a fair comparison by matching spatial resolution and vertical sensitivity is crucial [62][63][64]. In this study, we processed satellite data using oversampling based on the conservative spatial regridding method and applied the average kernel to the modeled layers before integrating into the column density. Oversampling can enhance weak signals of pollution by increasing the frequency of spatial sampling as well as generate high definition data with less noise [65,66]. This method has been used widely to detect signals of anthropogenic emission sources of NO 2 and SO 2 , which usually have a smaller spatial structure than model grid sizes.
For the data processing, we extended the concept of oversampling. Unlike previous approaches that simply increase the spatial sampling frequency, we included the concept of conservative regridding in which all the satellite footprint pixels are treated as polygons that have coverage areas instead of a simple center point. The conservative regridding method reconstructs satellite data pixels into a target domain grid cell by calculating the fractional coverage of raw data pixels within the target grid cell, as described by Kim et al. (2016) [62,63] The concentration of grid cell Cj can be calculated as where i and j are the indices of P (data pixel) and C (grid cell), respectively. f i,j is the overlapping fraction, which sums to 1 (i.e., f i,j = 1 for each grid cell). For all comparisons between model, satellite and surface observations, data were processed to properly match their spatiotemporal characteristics. Satellite data were regridded conservatively, and surface observations were averaged over those points within a domain cell. Temporal coverage is also matched consistently.

Base Model Performance
Before the full analysis, we evaluated the performance of the basic model to demonstrate the capability of the modeling system (i.e., can it reproduce the general features of tropospheric chemistry in the targeted region). Figure S1 presents the time series comparisons of surface PM 2.5 , ozone and NO 2 concentrations over approximately 300 surface monitoring sites in South Korea. The comparisons clearly demonstrate that the model can reproduce the general features of regional and local air quality from both meteorological and chemical perspectives.

Spatial Distribution
We then investigated the spatial distribution of NO x emission sources. Figure 1 compares the geographical distributions of the NO 2 vertical column densities from TROPOMI with the NO x emissions from the CAPSS 2016 emission inventory. The TROPOMI data were oversampled onto 3-km resolution grids using the conservative spatial regridding method and the NO x emission data were obtained from the 9-km simulation data for the CMAQ simulation. As noted earlier, oversampling has been used to detect emission flux signals from a variety of anthropogenic emission sources and CAPSS 2016 captures most potential NO x emission sources well. As expected, the fine resolution of the satellite and the model provide the details of the fine-scale structures of the emission sources. In particular, the highest NO2 column densities are shown in the Seoul metropolitan area (SMA) region (C1) where almost half of South Koreans reside. Mostly due to the high NOx emissions from transportation, the SMA shows high concentrations. In general, both TROPOMI and OMI (not shown) show similar spatial patterns representing the major NOx emission sources over South Korea. Urban cities and industrial facilities are other clear sources of NOx emissions. Busan (C2), Daegu (C3) and Gwangju (C4) are major cities that can be recognized from space-borne observations. Incheon (C5) also has a high concentration of NO2 column density, but is not isolated from emissions from Seoul due to its proximity to the capital. Gwangyang (P1), Ulsan (P2) and Pohang (P3) are other cities with large industrial facilities that are well recognized by TROPOMI. Other cities, including Daejeon (C6) and Sejong (C7), are also highly populated. Several point sources are also noticeable, including Dangjing (P4), Boryung (P5) and Saemangeum (P6). Comparisons of the location of known point sources in South Korea are discussed further in Section 3.4. North Korea also has a noticeable signal at Pyongyang (C8), likely the signals from the two fossil fuel power plants in Pyongyang and East Pyongyang. We confirm that the spatial patterns detected by TROPOMI and those reported in CAPSS 2016 agree to a large degree, implying the high quality As expected, the fine resolution of the satellite and the model provide the details of the fine-scale structures of the emission sources. In particular, the highest NO 2 column densities are shown in the Seoul metropolitan area (SMA) region (C1) where almost half of South Koreans reside. Mostly due to the high NO x emissions from transportation, the SMA shows high concentrations. In general, both TROPOMI and OMI (not shown) show similar spatial patterns representing the major NO x emission sources over South Korea. Urban cities and industrial facilities are other clear sources of NO x emissions. Busan (C2), Daegu (C3) and Gwangju (C4) are major cities that can be recognized from space-borne observations. Incheon (C5) also has a high concentration of NO 2 column density, but is not isolated from emissions from Seoul due to its proximity to the capital. Gwangyang (P1), Ulsan (P2) and Pohang (P3) are other cities with large industrial facilities that are well recognized by TROPOMI. Other cities, including Daejeon (C6) and Sejong (C7), are also highly populated. Several point sources are also noticeable, including Dangjing (P4), Boryung (P5) and Saemangeum (P6). Comparisons of the location of known point sources in South Korea are discussed further in Section 3.4. North Korea also has a noticeable signal at Pyongyang (C8), likely the signals from the two fossil fuel power plants in Pyongyang and East Pyongyang. We confirm that the spatial patterns detected by TROPOMI and those reported in CAPSS 2016 agree to a large degree, implying the high quality of both data sets. Clearly, TROPOMI can detect all the small emission signals we expect from the CAPSS inventory. On the contrary, most CAPSS information is presented as real measurements.
A more complete comparison of emission intensity can be achieved using the satellite observations and model simulations of the NO 2 vertical column densities. Figure 2 shows the spatial comparisons of the NO 2 column densities between CMAQ and TROPOMI during June 2018. The comparisons for the other months are shown in Figure S2. While there are slight differences by month, there is a clear overestimation in the modeled NO 2 column densities over major cities and industrialized locations. The only location with an underestimation is the area near the boundary of Chungbuk, Gangwon and Gyeongbuk, where several cement industries are located (P7 and P8). The potential impact of emissions from the cement industry in this region will be explored in a future study. patterns, we compare the surface observations and the corresponding model output at the monitoring sites with the spatial distribution by converting the point data into a spatial distribution using Krig spatial interpolation ( Figure 3B,D). The patterns of model biases are consistent from these site comparisons as well as the interpolated two-dimensional field comparisons. We see a considerable overestimation over the SMA region, especially over the western SMA. This comparison changes when we compare seasons. However, even for the cold season, we still see a strong overestimation from the major cities, especially in the western SMA, and other cities. Further, we see overestimations for the industries. To summarize, the model tends to overestimate at locations with high NOx emissions as well as in major cities and industries. Most of the underestimation of NO2 concentration is seen in suburban and rural locations. The lack of natural NOx emissions from soil in the model may account for these underestimations. While the similarity in the model biases from TROPOMI and the surface observations is also confirmed from the scatterplot in Figure 4, it shows slightly different patterns for the other months ( Figure S3). For all the months, the NO2 column density bias shows an overprediction, whereas the surface model biases are much weaker or show an underprediction, especially during the cold season. Additional plots are also available in Figures S4 and S5.
While the absolute value of NO2 concentration, both in surface and in columnar, is subject to errors from a variety of sources, the consistency of the model biases from the surface and columnar comparisons sheds light on CMAQ's ability to simulate the tropospheric NOx chemistry. Any inconsistency in this comparison would preclude us from concluding that the discrepancy is due to the retrieval of the satellite data. The model could suffer issues, especially in the simulations of the planetary boundary layer, which seriously affects the ratios between the surface and columnar concentrations.    Figure 2) and the surface observations ( Figure 3) have consistent bias patterns. Clearly, both bias patterns are consistent. To further confirm the spatial patterns, we compare the surface observations and the corresponding model output at the monitoring sites with the spatial distribution by converting the point data into a spatial distribution using Krig spatial interpolation ( Figure 3B,D). The patterns of model biases are consistent from these site comparisons as well as the interpolated two-dimensional field comparisons. We see a considerable overestimation over the SMA region, especially over the western SMA. This comparison changes when we compare seasons. However, even for the cold season, we still see a strong overestimation from the major cities, especially in the western SMA, and other cities. Further, we see overestimations for the industries. To summarize, the model tends to overestimate at locations with high NO x emissions as well as in major cities and industries. Most of the underestimation of NO 2 concentration is seen in suburban and rural locations. The lack of natural NO x emissions from soil in the model may account for these underestimations. While the similarity in the model biases from TROPOMI and the surface observations is also confirmed from the scatterplot in Figure 4, it shows slightly different patterns for the other months ( Figure S3). For all the months, the NO 2 column density bias shows an overprediction, whereas the surface model biases are much weaker or show an underprediction, especially during the cold season. Additional plots are also available in Figures S4 and S5.

Temporal Comparison
The seasonal variation in daily mean NO2 concentration over the 300 surface monitoring sites is shown in Figure 5. Both the observed and the modeled surface NO2 concentration show typical seasonal variation (i.e., high during the cold season and low during the summer season) due to the longer lifetime and increased emissions from anthropogenic sources during the cold season. However, this variation is much higher in the observed than in the modeled outputs. As a result, the modeled concentration shows a slight overestimation during the summer season and general underestimations during the other seasons. The scatterplots in Figure 5b also confirm the same pattern. While they show good statistics in terms of the annual mean (mean bias 20.95 − 21.79 = −0.84, i.e., −3.9%) and correlation coefficient (R = 0.83), there is a clear underestimation at a high concentration with a line fitting slope of 0.81. Figure 6 compares the diurnal variation in NO2 concentration from the surface monitoring sites and model simulations. These comparisons were conducted by the type of monitoring site (urban, roadside and rural). The measurements at rural sites indicate the background concentration. Interestingly, the diurnal variation in NO2 concentration shows varying patterns by monitoring site type and season. The modeled diurnal variations show better agreement for urban sites than roadside sites, which show a clear underestimation during the daytime. The discrepancy at roadside sites might be partially attributed to the model grid resolution, since the spatial distribution of NO2 concentration near strong sources may have a sharp gradient. While the absolute value of NO 2 concentration, both in surface and in columnar, is subject to errors from a variety of sources, the consistency of the model biases from the surface and columnar comparisons sheds light on CMAQ's ability to simulate the tropospheric NO x chemistry. Any inconsistency in this comparison would preclude us from concluding that the discrepancy is due to the retrieval of the satellite data. The model could suffer issues, especially in the simulations of the planetary boundary layer, which seriously affects the ratios between the surface and columnar concentrations.

Temporal Comparison
The seasonal variation in daily mean NO 2 concentration over the 300 surface monitoring sites is shown in Figure 5. Both the observed and the modeled surface NO 2 concentration show typical seasonal variation (i.e., high during the cold season and low during the summer season) due to the longer lifetime and increased emissions from anthropogenic sources during the cold season. However, this variation is much higher in the observed than in the modeled outputs. As a result, the modeled concentration shows a slight overestimation during the summer season and general underestimations during the other seasons. The scatterplots in Figure 5b also confirm the same pattern. While they show good statistics in terms of the annual mean (mean bias 20.95 − 21.79 = −0.84, i.e., −3.9%) and correlation coefficient (R = 0.83), there is a clear underestimation at a high concentration with a line fitting slope of 0.81. Figure 6 compares the diurnal variation in NO 2 concentration from the surface monitoring sites and model simulations. These comparisons were conducted by the type of monitoring site (urban, roadside and rural). The measurements at rural sites indicate the background concentration. Interestingly, the diurnal variation in NO 2 concentration shows varying patterns by monitoring site type and season. The modeled diurnal variations show better agreement for urban sites than roadside sites, which show a clear underestimation during the daytime. The discrepancy at roadside sites might be partially attributed to the model grid resolution, since the spatial distribution of NO 2 concentration near strong sources may have a sharp gradient.

Point Sources
In this section, we summarize the comparison of the model and TROPOMI at the locations of point sources. Figure 7 shows the spatial distribution of NOx emissions from the point sources in CAPSS 2016. The point source NOx emissions in each 9-km modeling cell are summed and grid cells with more than 2000 ton/year NOx emissions are marked with blue circles. The number shown next to each circle is an identifying number (in Figure 8), which is sorted by the order of model biases. Zoom-in plots for congested areas (e.g., Chungnam, Gwangyang and Pohang) are shown in Figure  S6. Figure

Point Sources
In this section, we summarize the comparison of the model and TROPOMI at the locations of point sources. Figure 7 shows the spatial distribution of NO x emissions from the point sources in CAPSS 2016. The point source NO x emissions in each 9-km modeling cell are summed and grid cells with more than 2000 ton/year NO x emissions are marked with blue circles. The number shown next to each circle is an identifying number (in Figure 8), which is sorted by the order of model biases. Zoom-in plots for congested areas (e.g., Chungnam, Gwangyang and Pohang) are shown in Figure S6. Several point sources are noticeable in this point source comparison. We see the overestimation of NO2 concentration at Dangjin (ID25) near a steel factory that has received special attention as a potential emission source owing to highly concentrated aerosols. The point source at Pohang (ID29) is where another steel company is located; this facility has the most overestimated NO2 emission sources compared with the TROPOMI column density. Since it is located in a coastal area, further research focusing on fine-scale models is warranted. In addition, a few point sources show the underestimation of NOx emissions, including the cement factories in Danyang, Yongwol and Jecheon (ID1, -2, -3 and -7).
While the comparison of the modeled column density with satellite retrieval data aimed to evaluate the reliability of emission inventories, this information can be used as guidance on the

Nitric Oxide Comparison
In addition, the ratio between NO and NO2 from emission sources is a major barrier to understanding the characteristics of NOx emissions [67][68][69][70][71][72][73]. The ambient NO2:NOx concentration ratio is often used as an important proxy for NO to NO2 chemistry in air quality models, and most existing models use a simple assumption of the NO to NO2 emission ratio without fully characterizing the important atmospheric chemical and mechanical processes. The model in this study also used a fixed NO:NO2 emission ratio, 9 to 1, to allocate total NOx emissions. Figure 9 summarizes the diurnal variations of the NO2 to NOx concentration ratio for the different locations and seasons. In general, the diurnal variation in the observed NO2:NOx ratio is consistent with previous in situ studies, with the ratios varying from 0.4 to 0.8 and the lowest ratio around morning rush hour [63]. The model also represents diurnal variation well. In all the locations, the model shows a higher NO2:NOx concentration ratio, especially from roadside sites. This suggests important implications on the limitation in modeling NO:NO2 emissions and chemistry. On one hand, the proportions of NO emissions could be higher from actual mobile emission sources. On the other hand, the conversion chemistry from NO to NO2 could have been inaccurate. In addition, modeling resolution also affects the result if the emitted NO emission at nighttime is measured before fully reacting with O3 to form NO2. Indeed, further studies including in situ measurements and finescale modeling studies are warranted. Additional plots for NO comparison are also available in Figures S7,S8.

Discussion on Uncertainties
While comparing modeled concentration to observation is the most common and practical way of evaluating emissions inventory, this comparison also should be done with caution because it can give misleading guidance on pollution control policy [74]. Potential sources of uncertainty, including the measuring interference of commercial instrument [74], satellite retrieval error [75], model resolution, changed NO:NO2 ratios due to diesel vehicle emission control equipment [73], and modeled boundary layer should be carefully considered in the interpretation of NOx emissions Several point sources are noticeable in this point source comparison. We see the overestimation of NO 2 concentration at Dangjin (ID25) near a steel factory that has received special attention as a potential emission source owing to highly concentrated aerosols. The point source at Pohang (ID29) is where another steel company is located; this facility has the most overestimated NO 2 emission sources compared with the TROPOMI column density. Since it is located in a coastal area, further research focusing on fine-scale models is warranted. In addition, a few point sources show the underestimation of NO x emissions, including the cement factories in Danyang, Yongwol and Jecheon (ID1, −2, −3 and −7).
While the comparison of the modeled column density with satellite retrieval data aimed to evaluate the reliability of emission inventories, this information can be used as guidance on the emission inventory but not as evidence to judge it. While we cannot conclude that these model biases are an absolute measure of the overestimated NO x emissions in the CAPSS 2016 emission inventory, we confirm that the Pohang cell is the most overestimated emission inventory information compared with the other point source cells in South Korea. In particular, considering its geographical features (i.e., this Pohang cell does not have nearby area emission sources unlike the point sources in Chungnam), its uncertainty in the overestimated emission assumption is not high.

Nitric Oxide Comparison
In addition, the ratio between NO and NO 2 from emission sources is a major barrier to understanding the characteristics of NO x emissions [67][68][69][70][71][72][73]. The ambient NO 2 :NO x concentration ratio is often used as an important proxy for NO to NO 2 chemistry in air quality models, and most existing models use a simple assumption of the NO to NO 2 emission ratio without fully characterizing the important atmospheric chemical and mechanical processes. The model in this study also used a fixed NO:NO 2 emission ratio, 9 to 1, to allocate total NO x emissions. Figure 9 summarizes the diurnal variations of the NO 2 to NO x concentration ratio for the different locations and seasons. In general, the diurnal variation in the observed NO 2 :NO x ratio is consistent with previous in situ studies, with the ratios varying from 0.4 to 0.8 and the lowest ratio around morning rush hour [63]. The model also represents diurnal variation well. In all the locations, the model shows a higher NO 2 :NO x concentration ratio, especially from roadside sites. This suggests important implications on the limitation in modeling NO:NO 2 emissions and chemistry. On one hand, the proportions of NO emissions could be higher from actual mobile emission sources. On the other hand, the conversion chemistry from NO to NO 2 could have been inaccurate. In addition, modeling resolution also affects the result if the emitted NO emission at nighttime is measured before fully reacting with O 3 to form NO 2 . Indeed, further studies including in situ measurements and fine-scale modeling studies are warranted. Additional plots for NO comparison are also available in Figures S7  and S8.
Atmosphere 2020, 10, x FOR PEER REVIEW 14 of 20 inventory. This is an important issue in urban chemistry, because the magnitude and trend of NOx emissions are still controversial in scientific community [76][77][78][79][80].

Conclusions
Over South Korea, NOx emission sources were examined using fine-scale satellite products, surface observations and a fine-scale modeling system with the latest emission inventory information. Here are the findings of this study: (1) The current South Korean emission inventory, CAPSS 2016, represents the geographical distributions of the NOx emission sources over the country well. It shows good agreement (e.g., R = 0.96 for June 2018) with the TROPOMI NO2 column density distribution;

Discussion on Uncertainties
While comparing modeled concentration to observation is the most common and practical way of evaluating emissions inventory, this comparison also should be done with caution because it can give misleading guidance on pollution control policy [74]. Potential sources of uncertainty, including the measuring interference of commercial instrument [74], satellite retrieval error [75], model resolution, changed NO:NO 2 ratios due to diesel vehicle emission control equipment [73], and modeled boundary layer should be carefully considered in the interpretation of NO x emissions inventory. This is an important issue in urban chemistry, because the magnitude and trend of NO x emissions are still controversial in scientific community [76][77][78][79][80].

Conclusions
Over South Korea, NO x emission sources were examined using fine-scale satellite products, surface observations and a fine-scale modeling system with the latest emission inventory information.
Here are the findings of this study: Prominent underestimations of daytime concentrations at roadside monitors are observed; (5) The modeled NO 2 :NO x ratio is higher than that of observations in all cases, and the largest differences are observed at roadside sites.
From these findings, we conclude that most power plants and industrial facilities were well detected by TROPOMI, implying that it can play a critical role in monitoring emission sources in urban areas. We also confirm that the spatial distribution of TROPOMI is consistent with the surface observations, suggesting that the newly available TROPOMI data can be a useful tool for monitoring fine-scale emission sources. TROPOMI has excellent spatial resolution, and it resolves and detects signals from known point sources in South Korea. We also find that the model usually overestimates both columnar and surface NO 2 concentrations over the SMA and major point sources, especially during the warm season. One notable exception is the underestimation for cement industries, implying that further investigations are required for emissions from those facilities.
While the interpretation of the model comparison is used to evaluate emission inventories, the comparisons shown in this study suggest a complicated interpretation to decide if the CAPSS 2016 NO x emission inventory is accurately estimated. Using such a simple comparison, both NO and NO 2 concentrations are widely underestimated in CAPSS 2016. However, the model overestimates compared with the surface observations in a large number of sites. The key point is that those sites are near the core of strong NOx emission sources, at the center of the SMA and in major industrial point sources. Hence, this mixed result from the NO x concentration comparison fails to provide clear evidence of a NO x emission underestimation or overestimation. It might thus be a premature conclusion to directly link the overestimation of the concentration to the overestimation of emissions, and uncertainty due to the model's limitations should be considered when interpreting the results of our study.