Detection of Strong NO X Emissions from Fine-scale Reconstruction of the OMI Tropospheric NO 2 Product

: Satellite-retrieved atmospheric NO 2 column products have been widely used in assessing bottom-up NO X inventory emissions emitted from large cities, industrial facilities, and power plants. However, the satellite products fail to quantify strong NO X emissions emitted from the sources less than the satellite’s pixel size, with signiﬁcantly underestimating their emission intensities (smoothing e ﬀ ect). The poor monitoring of the emissions makes it di ﬃ cult to enforce pollution restriction regulations. This study reconstructs the tropospheric NO 2 vertical column density (VCD) of the Ozone Monitoring Instrument (OMI) / Aura (13 × 24 km 2 pixel resolution at nadir) over South Korea to a ﬁne-scale product (grid resolution of 3 × 3 km 2 ) using a conservative spatial downscaling method, and investigates the methodological ﬁdelity in quantifying the major Korean area and point sources that are smaller than the satellite’s pixel size. Multiple high-ﬁdelity air quality models of the Weather Research and Forecast-Chemistry (WRF-Chem) and the Weather Research and Forecast / Community Multiscale Air Quality modeling system (WRF / CMAQ) were used to investigate the downscaling uncertainty in a spatial-weight kernel estimate. The analysis results showed that the ﬁne-scale reconstructed OMI NO 2 VCD revealed the strong NO X emission sources with increasing the atmospheric NO 2 column concentration and enhanced their spatial concentration gradients near the sources, which was accomplished by applying high-resolution modeled spatial-weight kernels to the original OMI NO 2 product. The downscaling uncertainty of the reconstructed OMI NO 2 product was inherent and estimated by 11.1% ± 10.6% at the whole grid cells over South Korea. The smoothing e ﬀ ect of the original OMI NO 2 product was estimated by 31.7% ± 13.1% for the 6 urbanized area sources and 32.2% ± 17.1% for the 13 isolated point sources on an e ﬀ ective spatial resolution that is deﬁned to reduce the downscaling uncertainty. Finally, it was found that the new reconstructed OMI NO 2 product had a potential capability in quantifying NO X emission intensities of the isolated strong point sources with a good correlation of R = 0.87, whereas the original OMI NO 2 product failed not only to identify the point sources, but also to quantify their emission intensities (R = 0.30). Our ﬁndings highlight a potential capability of the ﬁne-scale reconstructed OMI NO 2 product in detecting directly strong NO X emissions, and emphasize the inherent methodological uncertainty in interpreting the reconstructed satellite product at a high-resolution grid scale. than the WRF/CMAQ model during the periods. The NO 2 VCD modeled by the WRF/CMAQ compared better to the satellite-retrieved atmospheric NO 2 column concentrations than the WRF-Chem model. This discrepancy may partly be attributed to the deactivation of aqueous chemistry processes and the configuration of higher vertical grid resolution in the WRF-Chem simulation, but both the models’ NO 2 VCD values were within the 1-sigma error range of the satellite product. The atmospheric NO 2 VCD simulated by the WRF-Chem and WRF/CMAQ models were used to calculate the spatial-weight kernels in the conservative spatial downscaling method. over the region for four different months (60 days) to investigate the downscaling uncertainty in calculating spatial-weight kernels in the conservative spatial downscaling method. NO VCD on high-fidelity and resolution remains a critical limitation in quantifying a strong area and / or point sources and steep gradients on a satellite’s sub-pixel scale. This study reconstructed the KNMI OMI NO 2 VCD (13 × 24 km 2 at nadir) over South Korea on a spatial resolution of 3 × 3 km 2 using a conservative spatial downscaling method to investigate the methodological ﬁdelity in quantifying the atmospheric NO 2 column enhancements of the major area and point sources over the South Korean region. Multiple air quality simulations by the WRF-Chem and WRF / CMAQ models were performed over the South Korea region for four di ﬀ erent months (60 days) to investigate the downscaling uncertainty in calculating spatial-weight kernels in the conservative spatial downscaling method. The results showed that the newly reconstructed OMI NO 2 VCD based on high-ﬁdelity ﬁne-scale air quality modeling revealed clearly the strong area and point sources with enhanced spatial heterogeneity and steep gradients in accordance with their emission intensities. This is a critical weakness in the original OMI NO 2 product that signiﬁcantly underestimates the atmospheric NO 2 column


Introduction
Nitrogen oxides (NO X = NO + NO 2 ) are major criteria pollutants forming photochemical ozone and particulate matters (PM) in the atmosphere, which are emitted from various anthropogenic the aircraft measurement which showed 4-9 times higher NO 2 column densities over the source region than the OMI product.
The spatial smoothing effect embedded in the low-resolution satellite products is a critical barrier in quantifying strong emission sources, especially narrow-isolated sources with a scale of less than a few kilometers. The oversampling approach has been developed and improved in order to enhance spatial and temporal variations of satellite-detected pixel data at finer grid scales [35][36][37][38]. Various gridding algorithms have been applied with different complexities to generate gridded products from satellite-detected orbital signals. The simplest method is based on common spatial interpolation methods such as linear interpolation, spline interpolation, and kriging methods [39]. Meanwhile, the physics-based oversampling approach proposed by Sun et al. [38] is one of the advanced methods, in which a generalized 2-dimensional Gaussian function is used to represent the spatial response function of each pixel detected by a satellite sensor. Many studies have been done based on the gridded products, such as an emission estimate (e.g., [40,41]), an environmental exposure assessment (e.g., [42,43]), and source identification (e.g., [44][45][46]). The oversampling approach is able to generate spatially enhanced satellite products finer than their original pixel resolution by virtue of the temporally averaging multiple pixel data over a target grid, which simultaneously results in decreased temporal resolution. This may be conceptually valid when the concentrations detected by a satellite are assumed to be stationary during the averaging period. Directional composite averaging of chemical plumes following local wind direction in a specific source is a method that can be applied to overcome the limitation [37]. The spatial response function in Sun et al. [38] represents a sensitivity distribution of a satellite's sub-pixel in terms of the sensor's detection ability, and the information about spatial concentration distribution of chemical plumes within a satellite pixel is not taken into account, which is an apparent limitation of the oversampling method. On the other hand, Kim et al. [47] proposed a conservative downscaling method to reconstruct the spatial distribution of concentration within a satellite pixel, for which the spatial-weight kernel is calculated from high-resolution spatial concentration data simulated by a high-fidelity air quality model. Unlike the oversampling methods, the method is able to reconstruct the fine-scale spatial distribution of atmospheric concentration on each satellite pixel without the temporal averaging process. Meanwhile, the downscaled products depend on the fidelity of the air quality model applied to calculate the spatial-weight kernels of the satellite pixels, which causes additional downscaling errors in generating the fine-scale satellite products. Recent studies found the conservative downscaling method useful to represent strong spatial gradients over large source areas on a city scale, comparing well with in-situ measurements [27,30,48]. However, the downscaling errors associated with an air quality model are not investigated in those studies which used a single air quality model. This study aims to investigate the methodological fidelity of the conservative spatial downscaling method in quantifying strong NO X emission sources over South Korea and the downscaling uncertainties associated with inclusion of multiple air quality models. To do this, the OMI NO 2 column data were reconstructed over South Korea using independent high-fidelity air quality models of WRF-Chem and CMAQ.
The manuscript is structured as follows: Section 2 describes the air quality models, the OMI NO 2 columns data, and the conservative spatial downscaling method. Section 3 presents the original and newly reconstructed OMI NO 2 products over South Korea. Further discussed are the methodological uncertainty, the smoothing effect, and the potential capability of the reconstructed OMI NO 2 product in quantifying the major point sources over South Korea. The summary and conclusions follow in Section 4.

Data and Method
2.1. Regional Air Quality Models: WRF-Chem and WRF/CMAQ The regional air quality models of WRF-Chem [49] and CMAQ [50] were applied to produce high-resolution spatial distribution of NO 2 concentration over the South Korean region. The WRF Remote Sens. 2019, 11, 1861 4 of 18 model is a three-dimensional Eulerian meteorological model that solves dynamic/thermodynamic conservation equations of momentum, heat, moisture, and mass, and atmospheric physical processes of radiative transfer, turbulent mixing, surface-atmosphere interaction, and precipitation, thus being capable of simulating multiscale atmospheric phenomena [51]. The WRF-Chem model is an on-line air quality model that includes chemical transformation, source emissions, dry and wet deposition of chemical gaseous and aerosol species within the WRF model. The model simultaneously integrates the meteorological and chemical processes in a model's integration time step, thus is able to efficiently represent complex interactions among those physical and chemical processes. The CMAQ model is another three-dimensional Eulerian air quality model that has been developed primarily for regional air quality forecasts by the U.S. Environmental Protection Agency (EPA). It calculates the evolution of the atmospheric gaseous and aerosol species through representing meteorological and chemical processes. Meteorological fields are independently prepared for the model, and they are subsequently fed to the CMAQ model through a meteorology-chemistry interface processor (MCIP) package. In this study, the WRF model is used to produce meteorological fields required for running the CMAQ model (WRF/CMAQ). Both the air quality models have been widely applied for various meteorological and environmental problems such as air quality forecasts, emissions evaluations, regulatory applications, and scientific investigations (e.g., [9,[52][53][54][55][56][57]).
The WRF-Chem (ver. 3.9.1) and WRF/CMAQ (ver. 3.6.1/ver. 4.7.1) models were configured with three domains to model atmospheric NO 2 column concentrations over South Korea (Figure 1). The outermost domain covered a large East Asia region including China, Japan, and South Korea, with a horizontal grid resolution of 27 km (174 × 128 mesh). The second and third nested domains were configured with a horizontal grid resolution of 9 km (69 × 90 mesh) and 3 km (180 × 225 mesh), respectively, covering the whole South Korean region. The vertical grids were stretched ranging from 16 m above ground level at the lowest grid to~20 km ( conservation equations of momentum, heat, moisture, and mass, and atmospheric physical processes of radiative transfer, turbulent mixing, surface-atmosphere interaction, and precipitation, thus being capable of simulating multiscale atmospheric phenomena [51]. The WRF-Chem model is an on-line air quality model that includes chemical transformation, source emissions, dry and wet deposition of chemical gaseous and aerosol species within the WRF model. The model simultaneously integrates the meteorological and chemical processes in a model's integration time step, thus is able to efficiently represent complex interactions among those physical and chemical processes. The CMAQ model is another three-dimensional Eulerian air quality model that has been developed primarily for regional air quality forecasts by the U.S. Environmental Protection Agency (EPA). It calculates the evolution of the atmospheric gaseous and aerosol species through representing meteorological and chemical processes. Meteorological fields are independently prepared for the model, and they are subsequently fed to the CMAQ model through a meteorology-chemistry interface processor (MCIP) package. In this study, the WRF model is used to produce meteorological fields required for running the CMAQ model (WRF/CMAQ). Both the air quality models have been widely applied for various meteorological and environmental problems such as air quality forecasts, emissions evaluations, regulatory applications, and scientific investigations (e.g., [9,[52][53][54][55][56][57]). The WRF-Chem (ver. 3.9.1) and WRF/CMAQ (ver. 3.6.1/ver. 4.7.1) models were configured with three domains to model atmospheric NO2 column concentrations over South Korea (Figure 1). The outermost domain covered a large East Asia region including China, Japan, and South Korea, with a horizontal grid resolution of 27 km (174 × 128 mesh). The second and third nested domains were configured with a horizontal grid resolution of 9 km (69 × 90 mesh) and 3 km (180 × 225 mesh), respectively, covering the whole South Korean region. The vertical grids were stretched ranging from ~16 m above ground level at the lowest grid to ~20 km (50 hPa) at the domain top. The WRF-Chem model simulates both meteorological and chemical species at 35 sigma levels, while the CMAQ model calculates chemical species at 16 reduced sigma levels for computational efficiency. The MCIP produces the reduced meteorological fields compatible to the CMAQ using the WRF-simulated meteorological fields at 35 sigma levels.  The physical options used in the WRF simulations were identically configured with the Goddard scheme [58] and the rapid radiative transfer model (RRTM) [59] for shortwave and longwave radiation, the YSU scheme [60] for atmospheric boundary layer turbulence, the NoahLSM [61] for surface-atmosphere interactions, and the WSM3 scheme [62] and the Kain-Fritsch scheme [63] for grid-scale microphysics and sub-grid convective cloud parameterization. The exceptions were the Dudhia shortwave radiation scheme [64] and the Grell-3D cumulus parameterization in the WRF-Chem. The WRF-Chem uses the regional atmospheric chemistry mechanism (RACM) [65] and the modal aerosol dynamics model for Europe/Secondary organic aerosol model (MADE/SORGAM) [66,67] for gaseous and aerosol chemical mechanisms, respectively, while the WRF/CMAQ uses the Statewide Air Pollution Research Center, Version 99 (SAPRC-99) [68] and the aerosol module version 5 AERO5 [69]. The Model Inter-Comparison Study for Asia 2010 (MICS Asia 2010) developed for the East Asia air quality project [70,71] was used to produce anthropogenic emissions. The MICS-Asia 2010 emissions were gridded compatible to the model domains using the sparse matrix operator kernel for emissions (SMOKE) [72] emission processing module. The chemical speciation of volatile organic compounds (VOCs) follows the SAPRC-99 for WRF/CMAQ and the RACM for the WRF-Chem, and the chemical conversion between the two mechanisms follows Lee et al. [55]. Biogenic emissions were produced by the model of emissions of gases and aerosols from nature, version 2 (MEGAN 2) [73]. Table 1 summarizes the physical and chemical processes used in simulations of the WRF-Chem and the WRF/CMAQ models. Anthropogenic emission MICS-Asia 2010 [71] Biogenic emission MEGAN-2 [73] The simulations were carried for first 15 days of each January, April, July, and October (total 60 days) in 2015. The meteorological initial and boundary conditions of the outermost domain were obtained from the National Centers for Environmental Prediction-Final Analysis (NCEP FNL) that is a global reanalysis data with a spatial resolution of 1 • × 1 • and a temporal resolution of 6 hrs. The simulated meteorological fields for the outermost domain were nudged to the large-scale reanalyzed meteorological fields of air temperature, humidity, and winds using the 4-dimensional data assimilation (4DDA) technique [74]. The simulations of the nested domains were conducted by a one-way nesting approach.

Satellite Measurement: OMI/Aura tropospheric NO 2 Columns
The OMI tropospheric NO 2 vertical column densities (VCD) retrieved by the Royal Netherlands Meteorological Institute Dutch-OMI-NO2 version 2.0 (KNMI DOMINO v2.0) algorithm [75] were used. The OMI instrument, onboard the NASA Aura satellite, passed the equator at~13:45 local standard time (LST) with a cross-track field of view angle of 114 • , a swath width of 2600 km, and a nadir pixel size of 13 × 24 km 2 . The OMI tropospheric NO 2 VCD product was made through a series of data processing procedures. The KNMI OMI NO 2 slant column densities were calculated by a differential optical absorption spectroscopy (DOAS) technique, from which the tropospheric NO 2 column densities were calculated by separating the stratospheric and tropospheric contribution and subsequently by Remote Sens. 2019, 11, 1861 6 of 18 applying the tropospheric air mass factor that was obtained from the global chemical transport model TM4 and the radiative transfer model DAK [75]. The errors estimated during the retrieval procedures were~0.25 × 10 15 molecules cm -2 in the separation of stratospheric and tropospheric contribution, 1.0 × 10 15 molecules cm -2 in the application of the air mass factor, and~0.7×10 15 molecules cm -2 in the spectral fitting process. Figure 2 shows the frequency distribution of overpassing days in terms of the mean pixel size of tropospheric NO 2 VCD passing over the South Korean region (126 • E-130 • E; 34 • N-8 • N). The tropospheric NO 2 VCD data were available for 33 days (~55%) before the data quality control among 60 days of the simulation period and the average pixel sizes ranged from 312-1653 km 2 in the South Korean region. Approximately 33% of the days had a pixel size close to the satellite's nadir resolution, while approximately 30% had a mean pixel size of greater than 1000 km 2 . Meanwhile, the Korean cities have small urbanized areas of ranging from 10.5-227.5 km 2 when compared to the satellite pixel sizes, and the spatial sizes of strong point sources such as electric power plants, industrial facilities are much smaller than those of the cities. This indicates that fine-scale spatial distribution of the NO X emissions at a sub-pixel scale should be hardly detected in the OMI NO 2 VCD product (shaded in Figure 2). of data processing procedures. The KNMI OMI NO2 slant column densities were calculated by a differential optical absorption spectroscopy (DOAS) technique, from which the tropospheric NO2 column densities were calculated by separating the stratospheric and tropospheric contribution and subsequently by applying the tropospheric air mass factor that was obtained from the global chemical transport model TM4 and the radiative transfer model DAK [75]. The errors estimated during the retrieval procedures were ~0.25 × 10 15 molecules cm -2 in the separation of stratospheric and tropospheric contribution, ~1.0 × 10 15 molecules cm -2 in the application of the air mass factor, and ~0.7×10 15 molecules cm -2 in the spectral fitting process. Figure 2 shows the frequency distribution of overpassing days in terms of the mean pixel size of tropospheric NO2 VCD passing over the South Korean region (126°E-130°E; 34°N-8°N). The tropospheric NO2 VCD data were available for 33 days (~55%) before the data quality control among 60 days of the simulation period and the average pixel sizes ranged from 312-1653 km 2 in the South Korean region. Approximately 33% of the days had a pixel size close to the satellite's nadir resolution, while approximately 30% had a mean pixel size of greater than 1000 km 2 . Meanwhile, the Korean cities have small urbanized areas of ranging from 10.5-227.5 km 2 when compared to the satellite pixel sizes, and the spatial sizes of strong point sources such as electric power plants, industrial facilities are much smaller than those of the cities. This indicates that fine-scale spatial distribution of the NOX emissions at a sub-pixel scale should be hardly detected in the OMI NO2 VCD product (shaded in Figure 2).
The quality control procedure was applied to the OMI tropospheric NO2 VCD data because the satellite products were significantly contaminated in the conditions of high cloud fractions and long slant paths. Previous studies generally adopted the filtering conditions of 15-40% for cloud fractions and 1000 km 2 for slant path lengths according to the data availability in the area of interest (e.g. [10,21,[76][77][78]). This study filtered the NO2 VCD with a cloud fraction of less than 40% and a pixel size of less than 1000 km 2 considering frequent cloudy days. Finally, approximately 20% of the OMI NO2 VCD were used in reconstructing the satellite product.

Conservative Spatial Downscaling Method
The conservative spatial downscaling method proposed by Kim et al. [47] was used, which consists of two-step procedures: Conservative spatial re-gridding and conservative downscaling. The conservative spatial re-gridding procedure is to interpolate the OMI NO2 VCD pixel data to high- The quality control procedure was applied to the OMI tropospheric NO 2 VCD data because the satellite products were significantly contaminated in the conditions of high cloud fractions and long slant paths. Previous studies generally adopted the filtering conditions of 15-40% for cloud fractions and 1000 km 2 for slant path lengths according to the data availability in the area of interest (e.g. [10,21,[76][77][78]). This study filtered the NO 2 VCD with a cloud fraction of less than 40% and a pixel size of less than 1000 km 2 considering frequent cloudy days. Finally, approximately 20% of the OMI NO 2 VCD were used in reconstructing the satellite product.

Conservative Spatial Downscaling Method
The conservative spatial downscaling method proposed by Kim et al. [47] was used, which consists of two-step procedures: Conservative spatial re-gridding and conservative downscaling. The conservative spatial re-gridding procedure is to interpolate the OMI NO 2 VCD pixel data to high-resolution air quality model grids by weighting the overlapped areas between the satellite pixel and the model grids. The resulting gridded OMI NO 2 VCD data (OMI j ) is calculated as follows: where OMI i is the ith raw OMI NO-2 VCD pixel value, A i,j is the area fraction of OMI i overlapped on the jth model grid, and N is the number of the OMI pixels overlapped on the jth model grid. Each model grid meets N i A i,j = 1. For example, when two OMI NO 2 VCD pixels are overlaid on a model grid j, is simply re-gridded to the model grids without any changes in spatial distribution. The conservative downscaling procedure is to calculate the gridded OMI NO 2 VCD data (OMI_DS j ) by changing their spatial distribution according to the modeled spatial distribution as follows: where K i,j is a spatial-weight kernel defined by a ratio of modeled NO 2 VCD to the OMI NO 2 VCD. Here, the spatial-weight kernel was constructed on each ith OMI pixel using high-resolution modeled NO 2 VCD in order to conserve the OMI NO 2 VCD after applying the spatial-weight kernel.
The averaged OMI_DS j on the OMI pixel was identical to the OMI NO 2 VCD pixel value. In this study, the three-dimensional distribution and evolution of the atmospheric NO 2 VCD over South Korea were made by the two independent regional air quality models of the WRF-Chem and WRF/CMAQ so as to produce the spatial-weight kernel K i,j . In the modeled NO 2 VCD calculation, the KNMI OMI averaging kernel (AK) was applied to the model's vertical layers of the atmospheric NO 2 VCD. It has been known that the averaging kernel is sensitive to the retrieved NO 2 VCD [21,78,79].

Area and Point Sources of NO X Emissions in South Korea
Several major cities and isolated point sources with high NO X emissions over South Korea were selected in order to investigate methodological fidelity of the conservative spatial downscaling method in quantifying the emissions from the sources. The sources were selected based on the annual NO X emissions reported in the South Korean emission inventory data of the Clean Air Policy Support System (CAPPS). Figure 3 shows the spatial distribution of the monthly mean NO X emissions over South Korea along with geographic locations of the major sources. resolution air quality model grids by weighting the overlapped areas between the satellite pixel and the model grids. The resulting gridded OMI NO2 VCD data ( ) is calculated as follows: where is the ith raw OMI NO2 VCD pixel value, , is the area fraction of overlapped on the jth model grid, and is the number of the OMI pixels overlapped on the jth model grid. Each model grid meets ∑ , = 1. For example, when two OMI NO2 VCD pixels are overlaid on a model grid j, the gridded OMI NO2 VCD is calculated as ( , + , )/( , + , ). The OMI NO2 VCD is simply re-gridded to the model grids without any changes in spatial distribution.
The conservative downscaling procedure is to calculate the gridded OMI NO2 VCD data ( _ ) by changing their spatial distribution according to the modeled spatial distribution as follows: where , is a spatial-weight kernel defined by a ratio of modeled NO2 VCD to the OMI NO2 VCD.
Here, the spatial-weight kernel was constructed on each ith OMI pixel using high-resolution modeled NO2 VCD in order to conserve the OMI NO2 VCD after applying the spatial-weight kernel. The averaged _ on the OMI pixel was identical to the OMI NO2 VCD pixel value. In this study, the three-dimensional distribution and evolution of the atmospheric NO2 VCD over South Korea were made by the two independent regional air quality models of the WRF-Chem and WRF/CMAQ so as to produce the spatial-weight kernel , . In the modeled NO2 VCD calculation, the KNMI OMI averaging kernel (AK) was applied to the model's vertical layers of the atmospheric NO2 VCD. It has been known that the averaging kernel is sensitive to the retrieved NO2 VCD [21,78,79].

Area and Point Sources of NOX Emissions in South Korea
Several major cities and isolated point sources with high NOX emissions over South Korea were selected in order to investigate methodological fidelity of the conservative spatial downscaling method in quantifying the emissions from the sources. The sources were selected based on the annual NOX emissions reported in the South Korean emission inventory data of the Clean Air Policy Support System (CAPPS).   Figure 4 compares the spatial distributions of the re-gridded OMI NO 2 VCD and the modeled NO 2 VCD by the WRF-Chem and WRF/CMAQ models. The modeled NO 2 VCD were averaged for the satellite's valid pixels applying with the KNMI OMI's averaging kernels. Large source areas are clearly seen with enhanced concentrations both in the OMI-detected and the modeled NO 2 VCD, but distinguishable differences are also apparent. The OMI NO 2 VCD is able to discriminate strong NOx emissions from large urbanized areas, but fails to identify strong point emissions (Figure 4a). Previous studies that showed a similar spatial pattern of the KNMI OMI NO 2 VCD over the South Korea region at different time periods (e.g., [21,80]) considered the nation-wide spatial average of the OMI NO 2 VCD in their analysis. Meanwhile, both the modeled NO 2 VCD show more clearly enhanced signals of the strong point sources as well as the urbanized area sources than the OMI NO 2 product (Figure 4b,c), which is also spatially well correlated with the distribution of NO X emissions in Figure 3. The two models simulate similarly the spatial distributions over the South Korea, but the NO 2 VCD modeled by the WRF-Chem tends to be higher than the WRF/CMAQ. Figure 4 compares the spatial distributions of the re-gridded OMI NO2 VCD and the modeled NO2 VCD by the WRF-Chem and WRF/CMAQ models. The modeled NO2 VCD were averaged for the satellite's valid pixels applying with the KNMI OMI's averaging kernels. Large source areas are clearly seen with enhanced concentrations both in the OMI-detected and the modeled NO2 VCD, but distinguishable differences are also apparent. The OMI NO2 VCD is able to discriminate strong NOx emissions from large urbanized areas, but fails to identify strong point emissions (Figure 4a). Previous studies that showed a similar spatial pattern of the KNMI OMI NO2 VCD over the South Korea region at different time periods (e.g., [21,80]) considered the nation-wide spatial average of the OMI NO2 VCD in their analysis. Meanwhile, both the modeled NO2 VCD show more clearly enhanced signals of the strong point sources as well as the urbanized area sources than the OMI NO2 product (Figures 4b,4c), which is also spatially well correlated with the distribution of NOX emissions in Figure 3. The two models simulate similarly the spatial distributions over the South Korea, but the NO2 VCD modeled by the WRF-Chem tends to be higher than the WRF/CMAQ. Figure 5 compares the OMI-detected and modeled NO2 VCD to evaluate the model performance of the WRF-Chem and WRF/CMAQ in simulating the atmospheric NO2 column concentrations over the South Korean region. The monthly mean OMI NO2 VCD was 3.6 × 10 15 molec. cm -2 in July and 9.0 × 10 15 molec. cm -2 in January, with an annual mean value of 6.7 × 10 15 molec. cm -2 . Han et al. [21] reported a similar seasonal change of the KNMI OMI NO2 VCD ranging from 3.4-6.7 × 10 15 molec. cm -2 over the South Korean region in 2006. Kim et al. [80] reported approximately 4.0 × 10 15 molec. cm -2 in summer and 7.0 × 10 15 molec. cm -2 in winter 2010 in the KNMI OMI NO2 VCD Level 3 product. Meanwhile, the WRF-Chem model simulated the higher NO2 VCD levels by 50-145% than the WRF/CMAQ model during the periods. The NO2 VCD modeled by the WRF/CMAQ compared better to the satellite-retrieved atmospheric NO2 column concentrations than the WRF-Chem model. This discrepancy may partly be attributed to the deactivation of aqueous chemistry processes and the configuration of higher vertical grid resolution in the WRF-Chem simulation, but both the models' NO2 VCD values were within the 1-sigma error range of the satellite product. The atmospheric NO2 VCD simulated by the WRF-Chem and WRF/CMAQ models were used to calculate the spatialweight kernels in the conservative spatial downscaling method.  Figure 5 compares the OMI-detected and modeled NO 2 VCD to evaluate the model performance of the WRF-Chem and WRF/CMAQ in simulating the atmospheric NO 2 column concentrations over the South Korean region. The monthly mean OMI NO 2 VCD was 3.6 × 10 15 molec. cm -2 in July and 9.0 × 10 15 molec. cm -2 in January, with an annual mean value of 6.7 × 10 15 molec. cm -2 . Han et al. [21] reported a similar seasonal change of the KNMI OMI NO 2 VCD ranging from 3.4-6.7 × 10 15 molec. cm -2 over the South Korean region in 2006. Kim et al. [80] reported approximately 4.0 × 10 15 molec. cm -2 in summer and 7.0 × 10 15 molec. cm -2 in winter 2010 in the KNMI OMI NO 2 VCD Level 3 product. Meanwhile, the WRF-Chem model simulated the higher NO 2 VCD levels by 50-145% than the WRF/CMAQ model during the periods. The NO 2 VCD modeled by the WRF/CMAQ compared better to the satellite-retrieved atmospheric NO 2 column concentrations than the WRF-Chem model. This discrepancy may partly be attributed to the deactivation of aqueous chemistry processes and the configuration of higher vertical grid resolution in the WRF-Chem simulation, but both the models' NO 2 VCD values were within the 1-sigma error range of the satellite product. The atmospheric NO 2 VCD simulated by the WRF-Chem and WRF/CMAQ models were used to calculate the spatial-weight kernels in the conservative spatial downscaling method.    Figure 6 presents the reconstructed OMI NO2 VCD over the South Korean region by applying the spatial-weight kernels obtained from the independent air quality models. The newly reconstructed OMI NO2 VCD have a spatial resolution of 3 km in accordance with the models' grid resolution. Figure 7 compares the re-gridded and reconstructed OMI NO2 VCD over a large urbanized area (Seoul metropolitan area) and an area with a few isolated strong point sources. The reconstructed OMI NO2 VCD can reveal the highly urbanized area sources as well as the strong isolated point sources with enhanced spatial gradients and the higher level of column concentrations ( Figure 6), which is well compared to the NOX emissions (Figure 3a). In the Seoul metropolitan area, the OMI NO2 VCD downscaled by the WRF-Chem and WRF/CMAQ had values of 3.27 ± 0.75 × 10 16 molec. cm -2 (maximum 4.53 × 10 16 molec. cm -2 ) and 3.43 ± 0.68×10 16 molec. cm -2 (maximum 4.89 × 10 16 molec. cm -2 ), respectively, whereas the re-gridded OMI NO2 VCD had relatively low values of 2.63 ± 0.40 × 10 16 molec. cm -2 (maximum of 3.33 × 10 16 molec. cm -2 ) (Figures 7a-c). During the Megacity Air Pollution Studies-Seoul (MAPS-Seoul) field campaign conducted in May-June 2015, the atmospheric NO2 VCD detected by the Pandora spectrometer ranged from 1.7-1.9 Dobson units (DU, 1 DU ≈ 2.7 × 10 16 molec. cm -2 ) (4.59-5.13 × 10 16 molec. cm -2 ) during 12-14 LST at a site within the Seoul metropolitan area [81]. The fine-scale reconstructed OMI NO2 VCD products compares better to the surface in-situ measurement than the original OMI NO2 VCD. Meanwhile, the capability of the reconstructed OMI NO2 products are apparent in quantifying the emissions from the strong isolated point sources (Figures 7d-f). The strong point sources with a similar amount of NOX emissions were cement industrial facilities separated with a distance of ~20 km. The reconstructed OMI NO2 VCD by the WRF-Chem and WRF/CMAQ had values of 2.23 ± 0.86×10 16 molec. cm -2 (maximum 4.55 × 10 16 molec. cm -2 ) and 2.13 ± 0.89 × 10 16 molec. cm -2 (maximum 4.35 × 10 16 molec. cm -2 ), respectively. The signals of the strong point sources were clearly separated in the reconstruct OMI NO2 products, whereas the re-gridded original OMI NO2 product failed to identify the point sources showing relatively lower concentration levels of 1.50 ± 0.12 × 10 16 molec. cm -2 (maximum 1.70 × 10 16 molec. cm -2 ).  Figure 6 presents the reconstructed OMI NO 2 VCD over the South Korean region by applying the spatial-weight kernels obtained from the independent air quality models. The newly reconstructed OMI NO 2 VCD have a spatial resolution of 3 km in accordance with the models' grid resolution. Figure 7 compares the re-gridded and reconstructed OMI NO 2 VCD over a large urbanized area (Seoul metropolitan area) and an area with a few isolated strong point sources. The reconstructed OMI NO 2 VCD can reveal the highly urbanized area sources as well as the strong isolated point sources with enhanced spatial gradients and the higher level of column concentrations (Figure 6), which is well compared to the NO X emissions (Figure 3a). In the Seoul metropolitan area, the OMI NO 2 VCD downscaled by the WRF-Chem and WRF/CMAQ had values of 3.27 ± 0.75 × 10 16 molec. cm -2 (maximum 4.53 × 10 16 molec. cm -2 ) and 3.43 ± 0.68×10 16 molec. cm -2 (maximum 4.89 × 10 16 molec. cm -2 ), respectively, whereas the re-gridded OMI NO 2 VCD had relatively low values of 2.63 ± 0.40 × 10 16 molec. cm -2 (maximum of 3.33 × 10 16 molec. cm -2 ) (Figure 7a-c). During the Megacity Air Pollution Studies-Seoul (MAPS-Seoul) field campaign conducted in May-June 2015, the atmospheric NO 2 VCD detected by the Pandora spectrometer ranged from 1.7-1.9 Dobson units (DU, 1 DU ≈ 2.7 × 10 16 molec. cm -2 ) (4.59-5.13 × 10 16 molec. cm -2 ) during 12-14 LST at a site within the Seoul metropolitan area [81]. The fine-scale reconstructed OMI NO 2 VCD products compares better to the surface in-situ measurement than the original OMI NO 2 VCD. Meanwhile, the capability of the reconstructed OMI NO 2 products are apparent in quantifying the emissions from the strong isolated point sources (Figure 7d-f). The strong point sources with a similar amount of NO X emissions were cement industrial facilities separated with a distance of~20 km. The reconstructed OMI NO 2 VCD by the WRF-Chem and WRF/CMAQ had values of 2.23 ± 0.86×10 16 molec. cm -2 (maximum 4.55 × 10 16 molec. cm -2 ) and 2.13 ± 0.89 × 10 16 molec. cm -2 (maximum 4.35 × 10 16 molec. cm -2 ), respectively. The signals of the strong point sources were clearly separated in the reconstruct OMI NO 2 products, whereas the re-gridded original OMI NO 2 product failed to identify the point sources showing relatively lower concentration levels of 1.50 ± 0.12 × 10 16 molec. cm -2 (maximum 1.70 × 10 16 molec. cm -2 ). Figure 8 compares the simulated tropospheric NO 2 VCD and the reconstructed OMI NO 2 VCD by the WRF-Chem and WRF/CMAQ models. The tropospheric NO 2 VCD modeled by the WRF-Chem were spatially well correlated with those by the WRF/CMAQ (R = 0.84), but the WRF-Chem model simulated higher NO 2 VCD than the WRF/CMAQ (Figure 8a). On the other hand, a better spatial correlation (R = 0.97) was found in the reconstructed OMI NO 2 VCD products (Figure 8b). The similarity in the reconstructed OMI products shows that the conservative spatial downscaling method is less sensitive to the concentration level of the modeled NO 2 VCD.   Figure 8 compares the simulated tropospheric NO2 VCD and the reconstructed OMI NO2 VCD by the WRF-Chem and WRF/CMAQ models. The tropospheric NO2 VCD modeled by the WRF-Chem were spatially well correlated with those by the WRF/CMAQ (R = 0.84), but the WRF-Chem model simulated higher NO2 VCD than the WRF/CMAQ (Figure 8a). On the other hand, a better spatial correlation (R = 0.97) was found in the reconstructed OMI NO2 VCD products (Figure 8b). The similarity in the reconstructed OMI products shows that the conservative spatial downscaling method is less sensitive to the concentration level of the modeled NO2 VCD.

Fine-scale Reconstructed OMI NO 2 VCD
The relative differences between the reconstructed OMI NO2 VCD products in Figures 6-8 are interpreted as a methodological uncertainty associated with the conservative spatial downscaling   Figure 8 compares the simulated tropospheric NO2 VCD and the reconstructed OMI NO2 VCD by the WRF-Chem and WRF/CMAQ models. The tropospheric NO2 VCD modeled by the WRF-Chem were spatially well correlated with those by the WRF/CMAQ (R = 0.84), but the WRF-Chem model simulated higher NO2 VCD than the WRF/CMAQ (Figure 8a). On the other hand, a better spatial correlation (R = 0.97) was found in the reconstructed OMI NO2 VCD products (Figure 8b). The similarity in the reconstructed OMI products shows that the conservative spatial downscaling method is less sensitive to the concentration level of the modeled NO2 VCD.
The relative differences between the reconstructed OMI NO2 VCD products in Figures 6-8 are interpreted as a methodological uncertainty associated with the conservative spatial downscaling method. The downscaling uncertainty is inherent, but has not been quantitatively analyzed in method. The downscaling uncertainty is inherent, but has not been quantitatively analyzed in previous studies (e.g., [30,47]). Figure 9 presents the spatial distribution of the reconstructed OMI NO2 VCD averaged by the WRF-Chem and WRF/CMAQ models and the frequency distribution of grid-scale normalized absolute difference (NAD). The NAD is calculated as a downscaling uncertainty metric by It is clear that the averaged OMI NO2 product reveals highly urbanized areas and strong point sources over the South Korea region (Figure 9a). The downscaling uncertainty in the reconstructed OMI NO2 VCD is relatively large near the strong area and point sources such as cities, industrial complexes, and power plants, resulting from a dissimilar shape of the simulated plumes (Figures 6 and 7). The NAD ranged from 11.1% ± 10.6% (maximum 131.2%) for the whole grid cells and 12.0% ± 11.5% (maximum 108.9%) at the high-concentration grid cells of >8 × 10 15 molec. cm -2 . The grid-scale NAD values are less than 10% (20%) at approximately 58% (85%) of the whole grid cells, which is similar in the high-concentration grid cells (Figure 9b). These results indicate that the fine-resolution reconstructed OMI NO2 product has a potential capability in detecting steep concentration gradients found near the strong area and point sources with a reasonable error range compared to the original OMI NO2 VCD (Figure 4a).  The relative differences between the reconstructed OMI NO 2 VCD products in Figures 6-8 are interpreted as a methodological uncertainty associated with the conservative spatial downscaling method. The downscaling uncertainty is inherent, but has not been quantitatively analyzed in previous studies (e.g., [30,47]). Figure 9 presents the spatial distribution of the reconstructed OMI NO 2 VCD averaged by the WRF-Chem and WRF/CMAQ models and the frequency distribution of grid-scale normalized absolute difference (NAD). The NAD is calculated as a downscaling uncertainty metric by OMI_DS WRF−Chem −OMI_DS WRF/CMAQ 0.5(OMI_DS WRF−Chem +OMI_DS WRF/CMAQ ) . It is clear that the averaged OMI NO 2 product reveals highly urbanized areas and strong point sources over the South Korea region (Figure 9a). The downscaling uncertainty in the reconstructed OMI NO 2 VCD is relatively large near the strong area and point sources such as cities, industrial complexes, and power plants, resulting from a dissimilar shape of the simulated plumes (Figures 6 and 7). The NAD ranged from 11.1% ± 10.6% (maximum 131.2%) for the whole grid cells and 12.0% ± 11.5% (maximum 108.9%) at the high-concentration grid cells of >8 × 10 15 molec. cm -2 . The grid-scale NAD values are less than 10% (20%) at approximately 58% (85%) of the whole grid cells, which is similar in the high-concentration grid cells (Figure 9b). These results indicate that the fine-resolution reconstructed OMI NO 2 product has a potential capability in detecting steep concentration gradients found near the strong area and point sources with a reasonable error range compared to the original OMI NO 2 VCD (Figure 4a).
grid-scale normalized absolute difference (NAD). The NAD is calculated as a downscaling uncertainty metric by It is clear that the averaged OMI NO2 product reveals highly urbanized areas and strong point sources over the South Korea region (Figure 9a). The downscaling uncertainty in the reconstructed OMI NO2 VCD is relatively large near the strong area and point sources such as cities, industrial complexes, and power plants, resulting from a dissimilar shape of the simulated plumes (Figures 6 and 7). The NAD ranged from 11.1% ± 10.6% (maximum 131.2%) for the whole grid cells and 12.0% ± 11.5% (maximum 108.9%) at the high-concentration grid cells of >8 × 10 15 molec. cm -2 . The grid-scale NAD values are less than 10% (20%) at approximately 58% (85%) of the whole grid cells, which is similar in the high-concentration grid cells (Figure 9b). These results indicate that the fine-resolution reconstructed OMI NO2 product has a potential capability in detecting steep concentration gradients found near the strong area and point sources with a reasonable error range compared to the original OMI NO2 VCD (Figure 4a).

Effective Spatial Resolution and Detection Capability of Strong NOx Emissions
Despite the strong emissions with enhanced spatial gradients and the concentrations being successfully identified in the reconstructed satellite product, the reconstructed OMI NO 2 VCD had somewhat large downscaling uncertainty at the grid resolution of 3 km. Therefore, the effective spatial resolution was further investigated by changing an averaging box size near the sources to see how much the downscaling uncertainty reduces. The averaging box sizes of the emission sources increased gradually from 3 × 3 grid cells (9 × 9 km 2 ) to 23 × 23 grid cells (69 × 69 km 2 ). Figure 10a,b compare the re-gridded and the reconstructed OMI NO 2 VCD and the associated downscaling uncertainties over the Seoul metropolitan area and an industrial facility site as a function of the averaging box size. For the large urbanized area (Figure 10a), the reconstructed OMI NO 2 VCD had higher concentration levels than the re-gridded satellite product by a maximum of 30.0%. The difference between the re-gridded and reconstructed OMI NO 2 VCD decreases gradually with an averaging box size until the averaging area is large enough to cover the source region. The downscaling uncertainty in NAD also gradually decreases from~7% at 225 km 2 (15 × 15 km 2 ). For the isolated industrial facility site (Figure 10b), the tropospheric NO 2 VCD enhancement in the reconstructed satellite product is apparent by 72.8% at the maximum and decreases rapidly with the averaging area. The estimated NAD is approximately 19% at 9 km 2 (3 × 3 km 2 ) as a maximum value, and it decreases rapidly with an averaging box size. The downscaling uncertainty of each area and point sources estimated in NAD are presented for the 6 urbanized area sources and the 13 isolated point sources in Figure 10c,d. The NAD decreases gradually from 17% to 2%, on average, for the urbanized area sources as the averaging area increases, while it decreases more rapidly from 26% to 2% for the isolated point sources. It is also found that the downscaling uncertainty depends on the sources' characteristics such as the spatial size and distribution of emissions, and the complexities in local meteorology. difference between the re-gridded and reconstructed OMI NO2 VCD decreases gradually with an averaging box size until the averaging area is large enough to cover the source region. The downscaling uncertainty in NAD also gradually decreases from ~7% at 225 km 2 (15 × 15 km 2 ). For the isolated industrial facility site (Figure 10b), the tropospheric NO2 VCD enhancement in the reconstructed satellite product is apparent by 72.8% at the maximum and decreases rapidly with the averaging area. The estimated NAD is approximately 19% at 9 km 2 (3 × 3 km 2 ) as a maximum value, and it decreases rapidly with an averaging box size. The downscaling uncertainty of each area and point sources estimated in NAD are presented for the 6 urbanized area sources and the 13 isolated point sources in Figures 10c and 10d. The NAD decreases gradually from 17% to 2%, on average, for the urbanized area sources as the averaging area increases, while it decreases more rapidly from 26% to 2% for the isolated point sources. It is also found that the downscaling uncertainty depends on the sources' characteristics such as the spatial size and distribution of emissions, and the complexities in local meteorology. The smoothing effect of the original OMI NO 2 product in quantifying the atmospheric NO 2 column enhancements for the 6 urbanized area and the 13 isolated point sources over South Korea was investigated. A large variability in the downscaling uncertainties of the area and point sources leads to consideration of the effective spatial resolution that is determined by an appropriate averaging box size to reduce the downscaling uncertainties in a reasonable level. The NAD values reduced to less than~10% for the urbanized area sources when the averaging area was 4-5 times of the source's aerial size. Meanwhile, the averaging area of 5 × 5 grid cells (15 × 15 km 2 ) for the most isolated point sources led to a significant reduction in NAD ranging within 3-25% except for a few point sources where the downscaling uncertainty still remained as large as 37-40%. Figure 11 compares the re-gridded and the reconstructed OMI NO 2 VCD for the 6 urbanized area sources and the 13 isolated point sources over South Korea on the effective spatial resolution. The reconstructed OMI NO 2 VCD were well correlated with the original product for the urbanized area sources (R = 0.99) and the isolated point sources (R = 0.87). However, the reconstructed OMI NO 2 VCD were higher by 18-54% for the urbanized area sources and by 7-60% for the isolated point sources than the original OMI NO 2 product, which is caused by the inability of the satellite's pixels to resolve the spatial heterogeneity of the sources. In this study, the smoothing effect of the original OMI NO 2 product was evaluated by 31.7% ± 13.1% for the urbanized area sources and 32.2% ± 17.1% for the isolated point sources on the effective spatial resolution.
Finally, a potential capability of the reconstructed OMI NO 2 product in quantifying the NO X emission intensity for isolated strong point sources was investigated. Figure 12 compares the re-gridded and the reconstructed OMI NO 2 VCD of the 13 isolated point sources against the 2015 annual NO X emission of each point source. The NO X emissions are profiled in the Korean national emission inventory database from the direct measurements of stack emissions for the point sources. The reconstructed OMI NO 2 VCD were well correlated with the emission intensity of the point sources (R = 0.87), whereas the original OMI product showed only a weakly positive correlation (R = 0.30). The original OMI product shows significant limitations not only in identifying the strong emission sources with a satellite's sub-pixel scale but also in quantifying the atmospheric NO 2 column enhancements corresponding to the emission intensity. In contrast, the newly reconstructed OMI NO 2 product shows a great potential capability in quantifying the emissions of the isolated point sources within a reasonable range of downscaling uncertainties. product was evaluated by 31.7% ± 13.1% for the urbanized area sources and 32.2% ± 17.1% for the isolated point sources on the effective spatial resolution.
Finally, a potential capability of the reconstructed OMI NO2 product in quantifying the NOX emission intensity for isolated strong point sources was investigated. Figure 12 compares the regridded and the reconstructed OMI NO2 VCD of the 13 isolated point sources against the 2015 annual NOX emission of each point source. The NOX emissions are profiled in the Korean national emission inventory database from the direct measurements of stack emissions for the point sources. The reconstructed OMI NO2 VCD were well correlated with the emission intensity of the point sources (R = 0.87), whereas the original OMI product showed only a weakly positive correlation (R = 0.30). The original OMI product shows significant limitations not only in identifying the strong emission sources with a satellite's sub-pixel scale but also in quantifying the atmospheric NO2 column enhancements corresponding to the emission intensity. In contrast, the newly reconstructed OMI NO2 product shows a great potential capability in quantifying the emissions of the isolated point sources within a reasonable range of downscaling uncertainties.

Summary and Conclusions
Large spatial detection coverage of the satellite-retrieved NO2 VCD is beneficial in evaluating air quality models and bottom-up inventory NOX emissions. However, the satellite's coarse pixel resolution remains a critical limitation in quantifying a strong area and/or point sources and steep gradients on a satellite's sub-pixel scale. This study reconstructed the KNMI OMI NO2 VCD (13 × 24 km 2 at nadir) over South Korea on a spatial resolution of 3 × 3 km 2 using a conservative spatial downscaling method to investigate the methodological fidelity in quantifying the atmospheric NO2 column enhancements of the major area and point sources over the South Korean region. Multiple air quality simulations by the WRF-Chem and WRF/CMAQ models were performed over the South Korea region for four different months (60 days) to investigate the downscaling uncertainty in calculating spatial-weight kernels in the conservative spatial downscaling method.
The results showed that the newly reconstructed OMI NO2 VCD based on high-fidelity finescale air quality modeling revealed clearly the strong area and point sources with enhanced spatial heterogeneity and steep gradients in accordance with their emission intensities. This is a critical weakness in the original OMI NO2 product that significantly underestimates the atmospheric NO2 column enhancements over the area and point sources. The downscaling uncertainty of the reconstructed OMI NO2 product was estimated by 11.1% ± 10.6% at the whole grid cells over South Korea due to the use of multiple air quality models in the spatial-weight kernel estimate, which is not negligible. The smoothing effect in the original OMI NO2 product was evaluated, on average, by 31.7% ± 13.1% for the 6 urbanized area sources and 32.2% ± 17.1% for the 13 isolated point sources on the effective spatial resolution that is defined by an averaging box size of about 4 times of the urban size and 5 × 5 grid cells (15 × 15 km 2 ) for the area and point sources, respectively. Without considering the geometric smoothing effect, the surface emissions are easily misleading in terms of their intensity and steep spatial gradient. Finally, the potential capability of the reconstructed OMI product was tested in quantifying the isolated strong point sources that are listed in the Korean emission inventory with direct monitored annual NOX emissions. This showed that the isolated strong point sources with Figure 12. A comparison of the re-gridded and reconstructed OMI NO 2 VCD of the 13 major isolated point sources in South Korea against their annual NO X emissions. The vertical bar denotes the downscaling uncertainty estimated for the fine-scale reconstructed OMI NO 2 product.

Summary and Conclusions
Large spatial detection coverage of the satellite-retrieved NO 2 VCD is beneficial in evaluating air quality models and bottom-up inventory NO X emissions. However, the satellite's coarse pixel resolution remains a critical limitation in quantifying a strong area and/or point sources and steep gradients on a satellite's sub-pixel scale. This study reconstructed the KNMI OMI NO 2 VCD (13 × 24 km 2 at nadir) over South Korea on a spatial resolution of 3 × 3 km 2 using a conservative spatial downscaling method to investigate the methodological fidelity in quantifying the atmospheric NO 2 column enhancements of the major area and point sources over the South Korean region. Multiple air quality simulations by the WRF-Chem and WRF/CMAQ models were performed over the South Korea region for four different months (60 days) to investigate the downscaling uncertainty in calculating spatial-weight kernels in the conservative spatial downscaling method.
The results showed that the newly reconstructed OMI NO 2 VCD based on high-fidelity fine-scale air quality modeling revealed clearly the strong area and point sources with enhanced spatial heterogeneity and steep gradients in accordance with their emission intensities. This is a critical weakness in the original OMI NO 2 product that significantly underestimates the atmospheric NO 2 column enhancements over the area and point sources. The downscaling uncertainty of the reconstructed OMI NO 2 product was estimated by 11.1% ± 10.6% at the whole grid cells over South Korea due to the use of multiple air quality models in the spatial-weight kernel estimate, which is not negligible. The smoothing effect in the original OMI NO 2 product was evaluated, on average, by 31.7% ± 13.1% for the 6 urbanized area sources and 32.2% ± 17.1% for the 13 isolated point sources on the effective spatial resolution that is defined by an averaging box size of about 4 times of the urban size and 5 × 5 grid cells (15 × 15 km 2 ) for the area and point sources, respectively. Without considering the geometric smoothing effect, the surface emissions are easily misleading in terms of their intensity and steep spatial gradient. Finally, the potential capability of the reconstructed OMI product was tested in quantifying the isolated strong point sources that are listed in the Korean emission inventory with direct monitored annual NO X emissions. This showed that the isolated strong point sources with different emission intensities are clearly identified in the reconstructed OMI NO 2 VCD with a good correlation of 0.87. In contrast, the original OMI NO 2 product totally fails to identify the strong point sources and quantify their emission intensities, showing a weakly positive correlation with the emission intensity (R = 0.30).
This study highlights a potential capability of the fine-scale reconstructed OMI NO 2 product in quantifying directly strong NO X emissions and their steep spatial gradients and thus in assessing bottom-up inventory emissions for isolated areas and/or point sources that have a spatial scale of less than the satellite's pixel resolution. It also emphasizes that the methodological uncertainty associated with a spatial-weight kernel estimate should be considered cautiously in interpreting the reconstructed satellite product at a high-resolution grid scale.
The traditional oversampling methods are more beneficial in identifying hidden NO X emissions, but the spatial concentration distribution may be poorly represented due to a relatively coarse spatial and temporal resolution of the product. In contrast, the conservative downscaling method is better in representing fine-scale spatial distribution of identified NO X emissions on a finer spatial and temporal resolution of product. However, the application may be limited to detect unknown sources. Therefore, the oversampling and conservative downscaling methods can be used complementarily.
More research is needed to understand the downscaling uncertainty associated with a priori inventory emissions used in the air quality models, which is another uncertainty factor in the conservative downscaling method. In addition, further validation of the fine-scale reconstructed OMI NO 2 product can be useful if high-resolution in-situ measurements and/or remote sensing data are available.