Assessment of air quality in North Korea from satellite observations

was – 4.8 % yr (cid:0) 1 . Overall decreasing trends, except for NO 2 , are likely due to a decline in coal-fired PES. Positive NO 2 trends are consistent with increasing industrial activities. Each pollutant showed consistent patterns of linear trends, even after correcting the influence of transboundary pollution. Flue gas control and biofuel consumption reduction seem necessary to improve North Korea ’ s air quality.


Introduction
Indoor and outdoor air pollution in the Democratic People's Republic of Korea (hereafter referred to as North Korea-NK) was responsible for 255.4 deaths per 100,000 people in 2019, the highest rate worldwide (WHO, 2022).Since NK shares borders with China and the Republic of Korea (hereafter, South Korea-SK), air quality in these countries affects one another (Bae et al., 2018;J. Choi et al., 2019;Kim et al., 2013Kim et al., , 2016Kim et al., ,2019;;Yeo and Kim, 2019).Inventories from 2013 and 2015 estimated that emissions in NK were 1.5 and 2.7 times those of SK for carbon monoxide (CO) and particulate matter with aerodynamic diameter ≤ 2.5 µm (PM 2.5 ), respectively (Bae et al., 2018).Considering that the primary energy supply (PES) in NK corresponded to only 3.0-3.9% of SK's in 2013-2015, emissions of CO and PM 2.5 per unit energy supply could be tens of times larger in NK (Statistics Korea, 2019;Yeo and Kim, 2019).Moreover, NK emitted even larger amounts of nitrogen oxides (NO x ), sulfur dioxide (SO 2 ), and carbon dioxide (CO 2 ) per PES than SK in 2008, by factors of 3.9, 7.7, and 2.1, respectively (Kim and Kim, 2019).Overall, a view shared by previous studies is that NK has had a high level of air pollutant emissions per energy consumption.
Despite the severity of air pollution in NK suggested by emission inventories, observation-based assessments have been limited because the amounts of available ground-based air quality measurements are extremely small.There are very few examples of published data: international reports (e.g., UNEP, 2012) or proxy measurements such as visibility (Yeo et al., 2019).To this end, satellite data have recently been used to investigate the air quality in NK.For instance, space-borne observations of nitrogen dioxide (NO 2 ) enabled the estimation of NK's NO x emissions and assessment of bottom-up inventories (Han et al., 2020;Jung et al., 2022;Kim et al., 2014;Mijling et al., 2013).Also, satellitederived PM 2.5 concentrations allowed for analysis of the regional variability of aerosol pollution in NK (Yeo and Kim, 2019;Yeo et al., 2019).Moreover, satellite ammonia observations exposed an emission hotspot in NK (Anju city) (Van Damme et al., 2018).To date, however, satellite products have been employed only within the context of a single pollutant to characterize NK's air quality.A comprehensive air quality assessment using satellite observations is hardly found, except NK was briefly mentioned while covering a broader area in East Asia (e.g., Qu et al., 2022).To better understand the air quality in NK and develop mitigation measures, it is necessary to analyze trends of multiple species in accordance with changes in energy consumption (Kim and Kim, 2019).
The present study aimed to assess urban-to national-scale air quality in NK using multiple satellite observations for the first time.Pollutants of interest were NO 2 , SO 2 , CO, and aerosols, whose amounts are strongly associated with anthropogenic activities.Tropospheric NO 2 is produced predominantly via fossil fuel combustion, and to a lesser degree, by biomass burning, soil emissions, and lightning (Conrad, 1996;Ridley et al., 1996;Van Vuuren et al., 2011).Atmospheric SO 2 levels are also largely derived from anthropogenic sources, including fossil fuel combustion, smelters, and oil refineries (Fioletov et al., 2016;Smith et al., 2011).CO originates predominantly from biomass burning and the incomplete combustion of fossil fuels during industrial activities, vehicle use, and domestic heating (Zhong et al., 2017).Ambient aerosols, the key contributor to various adverse effects of pollution on human health (Apte et al., 2015;Wu et al., 2018), are either emitted directly from anthropogenic and natural sources or produced during chemical reactions of their precursors (e.g., NO 2 , SO 2 , and ammonia).
Satellite observations of all these pollutants have accumulated over North Korea for multiple years or even decades, allowing for long-term trend analysis.Besides, recently proposed oversampling methods facilitated the space-borne detection of urban-scale pollution (Pommier et al., 2013;Sun et al., 2018;Zhu et al., 2017).Exploiting these advances, we explored pollution hotspots in NK and analyzed trends for 2005-2018 by oversampling NO 2 and SO 2 columns from the Ozone Monitoring Instrument (OMI) and CO columns from the Measurements of Pollution in the Troposphere (MOPITT).For aerosol pollution investigation, we used 6 km resolution datasets of aerosol optical depth (AOD) for 2011-2018 and PM 2.5 concentrations for 2015-2018 from the Geostationary Ocean Color Imager (GOCI).After characterizing the spatiotemporal variability of air pollution in NK, we suggest possible contributors to the air quality changes based on previous studies and official national statistics for NK.
To put the severity of NK's air pollution into perspective, this study compared statistics between NK and SK, including pollutant amounts and PES.The areas of interest for SK were (1) the entirety of SK territory and (2) Seoul, the capital.Satellite datasets were sampled using the same methods for NK and SK.By aggregating the comparison results between NK and SK for all pollutants of interest, we comprehensively assess NK's air quality.

OMI NASA NO 2 and SO 2 products
The OMI instrument is a nadir-viewing ultraviolet (UV)-visible (VIS) spectrometer launched in 2004 onboard the Aura satellite (Levelt et al., 2006).OMI has 60 cross-track fields of view (FOVs), with spatial resolutions of ~ 13 km × 24 km at nadir and ~ 28 km × 150 km at the swath edges.The overpass time of OMI is 13:30-14:00 local time (LT), and its ~ 2600 km wide swath typically provides 1-2 overpasses per day across the Korean Peninsula.Examples of OMI observations over NK can also be found in previous studies targeting a broad domain in East Asia (Han et al., 2011;Han, 2020;Lee et al., 2014;Souri et al., 2017).Here, OMI data were acquired from January 2005 to December 2018.We excluded OMI observations affected by the so-called "row anomaly," characterized by abnormally low radiance readings found in certain cross-track FOVs since 2007 (Schenkeveld et al., 2017).
To assess the changes in vertical column densities (VCDs) of NO 2 over NK, we employed the NASA OMI NO 2 standard product (OMNO2) (Krotkov et al., 2019a).Recent validation studies found high correlations in East Asia by comparing the OMNO2 product with ground-based in-situ measurements (r = 0.78, Cai et al., 2022) and multi-axis differential optical absorption spectroscopy (MAX-DOAS) observations (r = 0.88, Cheng et al., 2019).We employed the most recent version (4.0) of the product for analyses (Lamsal et al., 2021).Tropospheric VCD (TVCD) data were used for analyzing anthropogenic NO 2 pollution.Only those observations with cloud radiance fractions < 0.3, solar zenith angles < 70 • , and good data quality (as stated by the quality flags, 'VcdQualityFlags' = 0) were used here (Krotkov et al., 2019b).We investigated spatiotemporal variations in SO 2 VCDs over NK using planetary boundary layer (PBL) data from the NASA OMI SO 2 product (OMSO2), which were derived using principal component analysis (PCA) (Krotkov et al., 2020;Li et al., 2020a).Long-term validation of the OMSO2 product in East Asia showed good agreements with ground-based in-situ measurements (r = 0.62, Si et al., 2019) and MAX-DOAS observations (r 2 = 0.54, Tian et al., 2018).We used the latest version of the product in this study (version 2; Li et al., 2020b).As recommended by the data source, the variable ColumnAmountSO2 was used for the analyses related to anthropogenic SO 2 pollution.SO 2 observations from rows 4 to 54 (zero-based) of the OMI charge-coupled device (CCD) detector were used here, except for those from the descending node of the Aura satellite.Observations with radiative cloud fractions > 0.3, solar zenith angles > 65 • , and AMFs < 0.3 were excluded, along with the SO 2 VCDs from pixels with snow or ice.To focus on anthropogenically-driven SO 2 concentrations, we masked transient volcanic SO 2 clouds by excluding OMI swaths, whose 99.9 percentile of SO 2 VCDs within 123-131 • E, 33-43 • N exceeded 9.5 DU, adopting the approach of Krotkov et al. (2016).This approach removed 8 swaths, corresponding to < 1 % of those used in the present study.

MOPITT CO product
MOPITT was launched in December 1999 onboard the NASA Terra satellite, with local equator crossing times of ~ 10:30 and ~ 22:30 (Edwards et al., 2004).MOPITT observations over NK can also be found in previous studies analyzing CO emissions and trends in East Asia (e.g., Zheng et al., 2018).For CO analyses in this study, Level-2 data were used from the MOPITT-derived CO (NIR and thermal IR radiances) V008 product (MOP02J_8), providing vertical profiles of volume mixing ratios (VMRs) and total VCDs at 22 km × 22 km horizontal resolution (MOPITT Algorithm Development Team, 2018; NASA/LARC/SD/ASDC, 2000).This study employed the total VCD values instead of surface VMRs, to account for different injection heights of CO into the air (Kim et al., 2019).Validation of total CO VCDs from this product over the Korean Peninsula showed a bias of -1.80 × 10 16 molecules cm − 2 and a standard deviation of 2.23 × 10 17 molecules cm − 2 (Deeter et al., 2022), which are 1-2 orders of magnitude smaller than typical CO VCDs in NK (see Fig. 2f).All pixels with solar zenith angles of > 80 • were excluded in this study.

GOCI AOD product
The GOCI instrument was launched onboard the Communication, Ocean, and Meteorological Satellites (COMS) in 2010 as the first ocean color imager operating from a geostationary orbit (Choi et al., 2012).GOCI measured hourly backscattered sunlight from 00:30 to 07:30 UTC at a 500 m resolution within a 2,500 km × 2,500 km domain centered at 36 • N and 130 • E. Aerosol optical properties (AOPs) retrieved from GOCI at 6 km resolution by the Yonsei aerosol retrieval (YAER) algorithm showed comparable accuracy to that of ground-based and other satellite-based observations (M.Choi et al., 2016Choi et al., , 2018;;M. Choi et al., 2019;Lee et al., 2010).The GOCI YAER version 2 product was used here, which has previously been effectively employed for assessing air quality across the Korean Peninsula (Lee et al., 2019(Lee et al., , 2021)).This study used the GOCI AOD product from March 2011 (the earliest month available) to December 2018.GOCI AODs at 550 nm showed an excellent agreement H. Chong et al. with ground-based observations from the Aerosol Robotic Network (AERONET) (r = 0.92) from March 2011 to February 2016 over land (Choi et al., 2018).The observed wavelength of AOD is omitted hereafter, as only 550 nm was assessed in this study.

Estimating surface PM 2.5 concentrations
Surface PM 2.5 concentrations were estimated using the random forest (RF; Breiman, 2001) algorithm at a native GOCI AOP resolution of 6 km.We constructed the RF model using GOCI AOP products, meteorological and land variables, ancillary data, and ground-based PM 2.5 measurements.The GOCI AOP products included AOD, fine mode fraction at 550 nm, and surface reflectance at 660 nm.The meteorological variables were from the European Center for Medium-Range Weather Forecast (ECMWF) reanalysis data (ERA5; Hersbach et al., 2020) and comprised temperature, pressure, wind speed, boundary layer height, and relative humidity (RH).The RH was calculated using atmospheric and dew point temperatures at the surface by applying the August-Roche-Magnus formula.For the land variables, low and high vegetation covers were used.Following previous studies, the cosine (Feng et al., 2015) and sine (Park et al., 2019) of the day of the year were employed as ancillary data.The spatial domain of the PM 2.5 estimates was 118-132 • E, 32-44 • N.For hourly surface PM 2.5 information, Airkorea data (https://www.airkorea.or.kr/eng) were employed for SK, the China National Urban Air Quality Real-time Publishing Platform data (https://106.37.208.233:20035/) for China, and the Ministry of the Environment (https://www.nies.go.jp/igreen/) with the Atmospheric Environmental Regional Observation System (AEROS) data (https ://soramame.taiki.go.jp) for Japan.
The probability distributions of PM 2.5 concentrations are typically skewed low; thus, low values are more frequently sampled while constructing the training set for RF, which can lead to the underestimation of satellite-derived values.To address this issue, we oversampled high PM 2.5 concentrations in training by applying the synthetic minority oversampling technique (SMOTE) (Chawla et al., 2002).
PM 2.5 concentrations were estimated on an annual basis, and results were evaluated using a 10-fold cross-validation approach (Rodríguez et al., 2010).The estimates only cover 2015-2018, as official data of ground-based PM 2.5 concentration in SK have existed since 2015.The corresponding annual coefficients of determination (R 2 ) for 2015-2018 were 0.714, 0.699, 0.690, and 0.683, respectively.The validation results and the mean PM 2.5 concentrations over NK for the entire 4-year period are shown in Fig. S1.

Oversampling satellite observations
To generate high-resolution gridded data for OMI NO 2 and SO 2 VCDs, we applied an oversampling approach known as the tessellation method.This approach assumes each OMI FOV to be a quadrilateral on the earth's surface and calculates an area-and error-weighted average VCD for each grid cell (Eq.( 1)): where C(i) is an OMI VCD at FOV i, σ(i) represents its precision, S(i) is the area of FOV i, A(i, j) denotes the overlapping area between FOV i and grid cell j, and C(j) is the resultant average VCD at grid cell j.More detailed descriptions of this oversampling method can be found in the studies of Zhu et al. (2017) and Sun et al. (2018).The longitudes and latitudes of the four corners of an OMI FOV were defined by the OMI pixel corner product (OMPIXCOR version 3) (Kurosu and Celarier, 2010).Tessellation was performed on 0.05 • × 0.05 • grid cells for both OMNO2 and OMSO2 products.For the OMSO2 product, unity was input for σ(i) in Eq. ( 1) as pixel-level precision was not provided.
The MOPITT datasets did not include the pixel corner (vertex) information in the variable list; therefore, an approach known as "pixel averaging" (Fioletov et al., 2011) or "point oversampling" (Sun et al., 2018) was applied to oversample CO VCDs (Pommier et al., 2013).This approach produces gridded data by averaging the CO VCDs from pixels located within a particular radius from each grid.The present study employed a 28 km radius to oversample MOPITT CO VCDs onto 0.05 • × 0.05 • grids.
As the 6 km spatial resolution of GOCI AOD data is adequate for urban/regional-scale air quality analysis (Choi et al., 2018), this data was employed without oversampling.

Time series analysis
This study assumed that the time series of monthly satellite observations at each grid was composed of three subcomponents (Eq.( 2)) (Bhardwaj et al., 2019;Weatherhead et al., 1998): where t represents time (month), α(t) is a seasonal component, β is a linear annual trend component, and ε(t) is a residual component.Here, the α term was derived by combining a constant and sinusoidal harmonic series: where γ 0 , γ 1k , and γ 2k are constant coefficients.The removal of the second term (the sine and cosine terms combined) on the right-hand side of Eq. ( 3) is referred to as "deseasonalization" in this study.The uncertainty of the linear trend σ β is estimated according to Eq. ( 4) (Weatherhead et al., 1998): where σ ε is the standard deviation of the residuals ε(t), l is the number of years, and δ is the autocorrelation coefficient of ε(t) calculated with a 1month lag.If the ratio |β/σ β | ≥ 2, the trend β is significant at the 95 % confidence level.

Changes in atmospheric NO 2 , SO 2 , and CO pollution from 2005 to 2018
Comparisons of average pollution levels between 2005-2011 and 2012-2018 indicated that over the entire NK territory, TVCDs of NO 2 increased, while total VCDs of SO 2 and CO decreased (Fig. 1).Oversampling satellite observations onto 0.05 • × 0.05 • grids enabled the detection of pollution hotspots, particularly for NO 2 and SO 2 , which have shorter lifetimes than CO.The VCDs of these three gaseous species showed overlapping high values in the western region, an area of high population density and low terrain elevations (Fig. S2) that is affected by transboundary pollution (Fig. S3).Pyongyang, Nampo, Pukchang, and Munchon were selected as the cities of interest for the analyses due to their elevated NO 2 , SO 2 , and CO VCDs (Fig. 1).
Relatively large NO 2 and CO VCDs in the capital Pyongyang can be partly attributed to vehicular emissions, as this is the most populated region in NK (Fig. S2a).Also, Pyongyang has the largest industrial district in NK, which comprises thermal power plants, steel mills, and machine plants (Dormels, 2014), contributing to the high NO 2 and SO 2 VCD values.Nampo is a port city and a national hub of the shipbuilding and smelting industries; thus, its large VCDs of NO 2 , SO 2 , and CO are likely associated with industrial emissions, including shipyards and steel mills (Dormels, 2014).The Pukchang Thermal Power Plant, the most substantial thermal power plant in NK (Dormels, 2014;Yeo and Kim, 2018), presumably contribute to the large local NO 2 and SO 2 VCDs in Pukchang.The city of Munchon belongs to the Wonsan industrial district along the eastern coastal area of NK (Dormels, 2014); thus, the high local NO 2 and SO 2 VCD values are likely caused by smelters and other industrial-facility emissions.The relative changes in average NO 2 TVCDs from 2005-2011 to 2012-2018 were + 26, +19, +8, and + 25 % for Pyongyang, Nampo, Pukchang, and Munchon, respectively, all of which indicated NO 2 increases.Meanwhile, SO 2 VCD changes corresponded to -34, -45, -43, and -33 %, respectively, and total CO VCDs decreased by 3-4 % in all four cities.We also found consistent hotspots and percentage changes for NO 2 and SO 2 VCDs by applying the "physical oversampling" method (Sun et al., 2018) (see Supplementary data), even though it resulted in lower values of 7-year mean VCDs than tessellation in four selected cities (Figs.S4 and S5).This consistency corroborates the air pollution changes in NK discussed above.
Further assessments of NK's air quality presented below were performed by constructing monthly mean VCD datasets.Satellite observations were oversampled every month using the tessellation method, and linear trends were calculated for the entire 14-year period (Fig. 2).Calculation results are listed in Table 1, along with the 14-year average VCDs.Negative monthly SO 2 VCDs were derived when the averaged data point number was insufficient to smooth out negative daily VCDs that were possibly generated by SO 2 concentrations below the OMI detection limit (Fig. 2).Such negative VCDs were not excluded from the trend estimation for their implication for low SO 2 concentrations.
The NO 2 TVCD averaged for the entire NK territory for 2005-2018 was 31 % of that observed in SK (Table 1).Among the cities examined in NK, Pukchang showed the largest mean NO 2 TVCD, likely due to the emissions from the Pukchang Thermal Power Plant.The relative linear trends of NO 2 TVCDs over the selected four cities in NK ranged between + 2.1-4.1 % yr − 1 , and rates over the entire territory showed a similar positive trend of + 3.6 ± 0.6 % yr − 1 (Fig. 2d).To put the severity of NO 2 pollution into perspective, we compared TVCDs between NK and SK in association with their respective PES values (Table 1).From 2005 to
The average SO 2 VCD in NK for 2005-2018 corresponded to 81 % of that recorded in SK over the same period (Table 1).The SO 2 VCD ratio of NK to SK was significantly higher than NO 2 , likely due to extensive coal burning in NK.The International Energy Agency (IEA) reported that coal-fired energy generally accounts for 60-80 % of PES in NK (IEA, 2022).The largest mean SO 2 VCD among the four NK cities appeared in Pukchang, likely resulting from a combination of power plant emissions and transboundary pollution.Contrary to NO 2 TVCDs, SO 2 VCDs showed negative trends over all cities analyzed in NK (Fig. 2e).The national average SO 2 VCDs for NK also showed a negative trend of -4.4 ± 0.7 % yr − 1 ; however, SO 2 VCDs per unit PES were still larger in NK than in SK.The average ratio of SO 2 VCD per unit PES between NK and SK was 17.8 from 2005 to 2018 (even higher than NO 2 ).Severe SO 2 pollution in NK is possibly due to the lack of desulfurization of fuels and flue gases.
The average CO VCD ratio between NK and SK was 97 %, the largest among the gaseous species of interest (Table 1).On average, NK experienced a 20.6 times larger CO VCD per unit PES than SK from 2005 to 2018.The most severe CO pollution in NK was recorded in Nampo.The large CO VCDs in NK can be attributed to transboundary pollution and domestic emissions from coal and biofuel combustion.The pollution levels declined from 2005 to 2018, but the rates were slow; the CO VCDs decreased over the four selected cities and the entirety of NK at rates between -0.6 and -0.3 % yr − 1 (Fig. 2f).The CO VCDs slowly decreased by 0.5 ± 0.1 % yr − 1 also over the reference sector in the Yellow Sea (see Fig. 1g for the definition of the reference sector), implying the continuous effects of long-range transport from China on CO pollution over NK (Table 1; Fig. S6c).Domestically, previous studies found that residential biomass burning was responsible for a significant proportion of CO emissions in NK (Kim et al., 2011;Kim and Kim, 2019), although estimates of local biofuel consumption vary greatly by data sources (IEA, 2022;Streets et al., 2003;UNEP, 2012).For example, Streets et al. (2003) estimated that up to 40 % of NK's CO emissions originated from biofuel combustion in 2000.
According to the major statistical indicators published by the SK's government (Statistics Korea, 2018, 2019), coal-fired PES in NK decreased during 2005-2018 at a rate of -4.7 % yr − 1 (Fig. 3b and c).Considering the strong coal dependence of NK's industry (Kim and Kim, 2019;Yeo and Kim, 2019), one may expect that local NO 2 and SO 2 VCDs would display similar decreasing patterns across the study period.Indeed, SO 2 VCDs decreased, but NO 2 TVCDs showed positive trends (Fig. 3a and b).Mijling et al. (2013) and Lee et al. (2014) suggested that transboundary pollution from China affected NK's NO 2 pollution.To examine pollution changes in NK driven only by domestic factors, we attempted to correct the influence of transboundary pollution by removing the dependency of NO 2 TVCDs in NK on those observed in the reference sector in the Yellow Sea (see Supplementary data).This reference sector was selected for two reasons: (1) it has minimal local emissions (generally restricted to those from ships), and (2) air masses from China typically travel to this region before arriving at the Korean Peninsula (Bhardwaj et al., 2019;Lee et al., 2021) (see Fig. S7).However, the positive NO 2 trends were preserved even after correcting for the effects of transboundary pollution (Figs.S6d and S8a).
Moreover, seasonal mean NO 2 TVCDs revealed several features implying that domestic factors maintained a significant influence on the positive trends separately from the transboundary pollution (Fig. S9).First, NK's NO 2 TVCDs in spring (MAM) indicated a positive trend of 3.7 × 10 13 molecules cm − 2 yr − 1 , significantly larger than the spring trend over the Yellow Sea (0.8 × 10 13 molecules cm − 2 yr − 1 ).Second, the maximum positive trend in NK appeared in autumn (SON), while the most rapid increase over the Yellow Sea was observed in winter (DJF).Third, the NO 2 TVCDs in NK increased more rapidly than over the Yellow Sea in summer (JJA).Fourth, NO 2 TVCDs in NK were generally the largest in spring (MAM), while maximum values over the Yellow Sea were typically seen in winter (DJF).
We examined if soil emissions were significant factors in the increasing NO 2 trends in NK, based on a previous study suggesting that NO 2 TVCDs typically peaked in summer in regions where soil emissions dominated (van der A et al., 2008).Seasonal analysis of the maximum NO 2 TVCD at each grid pixel in NK showed that TVCDs increased more rapidly at grids where maximum TVCDs were observed outside of the summer season (Fig. S10).Furthermore, grids with maximum NO 2 values in JJA did not account for a large fraction in NK; thus, we concluded that soil emissions were not the primary source of nationalscale increases of NO 2 in NK.Additionally, the negative trends of CO VCDs imply that biomass burning was not a significant NO 2 source either (see Figs. 2f, 3c, and S6c).Kim and Kim (2019) estimated that in NK, from 2000 to 2008, industrial activities, transportation, and power generation contributed to NO x emissions by 41.7 %, 38.4 %, and 12.7 %, respectively.Therefore, as the candidate variables to explain the increasing NO 2 patterns, we selected the 'index of all industrial production,' 'number of registered motor vehicles,' and 'electrical energy supply via thermal power generation' from the statistical indicators generated by the SK's government (Statistics Korea, 2018) (Fig. 3a).Since these statistics were derived at the national scale, only the NO 2 TVCDs averaged over the entirety of NK were analyzed here.Fig. 3a shows that thermal power generation fluctuated during the study period, and its changing patterns appeared weakly associated with the observed NO 2 TVCD variability.The number of vehicles registered in NK increased, but the trend (1.3 % yr − 1 ) was insufficient to account for the NO 2 TVCD increases alone.Conversely, the statistics reported an increase in industrial production in NK, even though coal-fired PES in NK decreased during the study period (Statistics Korea, 2018, 2019).This report implies the possibility of growth in coalindependent industrial activities and/or the transition of coal usage from other sectors into industries requiring a reduced coal supply.
However, investigating which of the two scenarios applies to NK was beyond the scope of the present study.Overall, we deduced that NK's changes in industrial production could be responsible for the increasing NO 2 concentrations.This finding is consistent with a previous study that estimated that domestic NO x emissions in NK increased by 6 % yr − 1 from 2007 to 2011 (Mijling et al., 2013).On urban scales, NO 2 TVCDs in industrial cities such as Pyongyang, Nampo, and Pukchang increased more rapidly than the average over the entire NK territory (Fig. 2d), supporting that industrial activities are important factors accounting for the increasing NO 2 trends in NK.
Since coal combustion is the major emission source of SO 2 , 'primary energy supply by coal' was selected as the lone explanatory variable of SO 2 pollution changes.The correlation coefficient between changes in the coal-fired PES and the SO 2 VCDs was 0.74 (Fig. 3b).The same explanatory variable was used to assess CO change, revealing a notably lower correlation coefficient (0.53) (Fig. 3c).As mentioned, the slighter decrease in CO VCD compared to SO 2 VCD can either be attributed to the influence of residential biomass burning or the longer lifetime of CO.
As briefly mentioned while interpreting positive NO 2 trends, transboundary pollution could contribute to the changes in NK's pollutant amounts.We attempted to remove the dependence of NK's pollution on long-range transport by applying the same correction method to each pollutant (see Supplementary data).It is worth noting that the VCDs preserved the signs of their linear trends in every region of interest in NK after correction, without decreasing patterns flipped to increasing or vice versa (See Figs. 2 and S8).On a national scale, the relative change rates of SO 2 and CO VCDs only reduced by 0.9 and 0.1 % yr − 1 , respectively, and the NO 2 trend showed the same rate after correction.Postcorrection VCDs of each pollutant showed the strongest association with the same explanatory variable as before correction, only with slightly changed correlation coefficients for SO 2 and CO (Fig. S8).These analyses with the transport-corrected VCDs support our findings that the domestic factors contributed to NK's pollution changes.

Changes in AOD and PM 2.5 concentration from March 2011 to December 2018
The average AOD over NK from March 2011 to December 2018 was 0.26 ± 0.11, notably similar to that over SK for the same period (0.28 ± 0.12) (Table 1).By normalizing the annual mean AOD with the PES, we found that NK experienced an average of 24.3 times larger AOD per unit energy supply than SK over the analysis period.
Fig. 4 shows the biennial means of AODs observed from GOCI during 2011-2018.Similar to VCDs of the gaseous pollutants, high AOD values appeared over the western parts of NK (particularly in the four cities analyzed), implying that aerosol pollution was associated with longrange transport and anthropogenic activities (especially coal combustion).Meanwhile, decreasing AODs over the analysis period were observed on a national scale.
To further investigate AOD, annual mean values were analyzed for Pyongyang, Nampo, Pukchang, Munchon, and at the national scale (Fig. 5a).Nampo maintained the highest AOD values, likely resulting from the large population, industrial activities, biofuel combustion, and transboundary transport.Notably, all the regions of interest showed negative AOD trends (Fig. 5a; Table 1).A relative linear decrease at the national scale was -4.8 ± 1.0 % yr − 1 , and those from the selected cities ranged between -5.9 % yr − 1 to -3.4 % yr − 1 .
We found significant correlation coefficients between deseasonalized monthly AOD time series of the Yellow Sea and most grid pixels in NK (~81 %) from 2011 to 2018, with a national average of 0.44 (Fig. S11a).The correlation coefficients were higher in the western parts of NK, indicating a stronger influence of transboundary pollution.For example, Pyongyang showed a correlation coefficient of 0.55.Slope values derived from linear regression between the Yellow Sea and NK time series were also higher in the western parts, supporting that those regions were more sensitive to transboundary pollution (Fig. S11b).There was a longitudinal gradient in NK's 2011-2018 mean AOD field between 124 • E and 129 • E with higher values in the west.We found that correcting the influence of transboundary pollution makes this gradient weaker and AODs lower (Fig. S11d).The differences between the meridional averages of pre-and post-correction AODs in NK showed a maximum of 0.086 (27.7 %) at 125.25 • E and decreased eastwards up to  128.75 • E, being evidence of a higher level of transboundary pollution in the western parts.Overall, urban-scale enhancements were preserved in the western parts of the transport-corrected AOD fields but with a weakened national-scale longitudinal gradient (Fig. S11c).
We estimated that transported amounts of aerosols were reduced by time, implying that the observed negative AOD trends in NK partly occurred due to a decline in transboundary aerosol pollution (Table S1).Indeed, the change rates of AODs appeared to be slower after correcting the transport influence (see Figs. 5a and S12a).Consistent with the gaseous pollutants, however, the post-correction AODs preserved the signs of trends, showing decreasing patterns in every region of interest in NK (Fig. S12a).The negative trends of post-correction AODs support the contributions of domestic factors to the decreasing AODs in NK.
To identify possible explanatory variables, we used the statistical indicators of NK generated by the SK's government (Statistics Korea, 2018, 2019), similarly to the gaseous pollutants.As with SO 2 and CO, we deduced that the 'primary energy supply by coal' was the most influential in describing the decreasing AOD trends.Fig. 5b shows the temporal changes in the coal-fired energy supply and AOD in NK from 2011 to 2018.Notably, the changing patterns of AODs in Nampo and Pukchang reflected the coal-fired PES variations (Fig. 5b), and consistent features were also found in transport-corrected AODs (Fig. S12b).This is reasonable as the air quality in these two industrial cities could heavily depend on local variations in coal consumption.Biofuel combustion is another potential factor substantially influencing AOD values (Kim et al., 2019); however, the quantification of biofuel consumption is marked by substantial uncertainty.
Estimating aerosol mass concentrations assists in intuitive assessments of aerosol amounts.GOCI estimates indicated that annual mean PM 2.5 concentrations over NK from 2015 to 2018 were 43.5, 40.0, 41.1, and 42.7 μg m − 3 (Fig. 5c).The highest PM 2.5 concentrations appeared in Pyongyang, with corresponding annual estimates of 55.7, 50.4, 45.4, and 47.2 μg m − 3 , respectively (Fig. 5c).These values may be accompanied by uncertainty since no ground-based measurement data were available in NK for input into the RF algorithm; however, the estimates agreed well with the mean population exposure to PM 2.5 provided by the OECD (48.0,43.5,43.5,and 45.1 μg m − 3 over the entirety of NK, and 55.4, 50.5, 49.7, and 52.3 μg m − 3 for Pyongyang from 2015 to 2018, respectively) (OECD, 2020).Similar to the findings reported by the OECD, the estimates here showed declining trends (Fig. 5c).Accordingly, the decrease in PM 2.5 concentrations was presumably driven by a reduction in coal-fired PES and a decline in transboundary aerosol pollution.

Discussion
Dividing NK's pollutant amounts by those of SK provides a useful perspective to understand NK's air quality better.Let the ratio of NK to SK be an indicator of the pollution level.The first and second highest levels were found for CO and aerosols, whose amounts in NK were larger or close to those in SK.Since they have long lifetimes in the atmosphere, their high concentrations partly resulted from long-range transport.However, considering the common emission sources of CO and aerosols, their high pollution levels also imply substantial biofuel consumption in NK, aligned with findings from previous studies.The third highest pollution level was found for SO 2 , indicating 81 % of the concentrations in SK during 2005-2018.Aside from the contribution of long-range transport, the high SO 2 pollution suggests insufficient desulfurization of fuels and flue gases from the enormous coal burning in NK, although the coal-fired energy supply showed a decreasing trend.Lastly, NO 2 concentrations were as low as 31 % of those in SK on average from 2005 to 2018.However, it is worth noting that NO 2 is the only species showing positive trends.The positive NO 2 trends imply that NK increased industrial activities to combat the declining economy under decreasing energy supply conditions.

Conclusions
All pollutants of interest in this study showed decreasing trends except NO 2 in NK, and the most probable cause is the decline of the coalfired energy supply.The signs of linear trends were even preserved after correcting the transboundary pollution effects.However, it is noticeable that NK still experienced significantly larger amounts of air pollutants per energy supply compared to SK from 2005 to 2018, even with the prevailing negative trends.This finding suggests low removal efficiencies and high emission coefficients in NK.Reduction of biofuel combustion seems necessary to lower CO and aerosol concentrations, and flue gas control is an essential measure to reduce SO 2 pollution.
International cooperation would be beneficial in emissions control in NK.Indeed, NK has participated in international projects as well as enacted domestic laws and policies to protect the environment (UNEP, 2012).Yeo and Kim (2018) comprehensively summarized the direction of NK's policy regarding environmental protection.In particular, NK has been actively involved in the Clean Development Mechanism (CDM) (UNFCCC, 2022), hosting projects including hydropower generation and methane emission control (NCCE, 2012).It would be most desirable to assist emission mitigation through the forms of international cooperation favored by the NK's government.
Continuous satellite data accumulation and frequent trend analysis updates are essential to link NK's air pollution with anthropogenic activities.The year 2020 would be particularly of interest in future studies since lockdown measures were globally implemented due to the spread of Coronavirus disease 2019 (COVID-19) in that year.We found that applying the same analysis method to the period 2005-2020 resulted in a linear NO 2 trend of + 2.5 % yr − 1 (Fig. S13a), which is 2.1 % yr − 1 lower than that from 2005-2018 (Fig. 3a), contributed by the changes in 2019 and 2020.Meanwhile, the SO 2 and CO trends showed small differences of only a few tenths of a percent or less between 2005-2018and 2005-2020 (Figs. 3b, 3c, S13b, and S13c) (Figs. 3b, 3c, S13b, and S13c).Since mixed factors likely affected the decline of NK's national average NO 2 TVCDs in 2020, detailed analyses need to be conducted in the future to assess the impact of the COVID-19 lockdown on NK's air quality.
As advanced satellite instruments have continuously been developed, more precise assessments will be possible in the future.For example, the Tropospheric Monitoring Instrument (TROPOMI) (Veefkind et al., 2012) launched in 2017 is now compiling long-term VCDs of NO 2 , SO 2 , and CO with a higher spatial resolution than OMI.Furthermore, the Geostationary Environment Monitoring Spectrometer (GEMS) launched in 2020 has enabled the first space-borne diurnal observations of NO 2 and SO 2 VCDs over Asia (Kim et al., 2020).Such observations will contribute significantly to expanding the understanding of NK's air quality and its interactions with that in other countries in East Asia, including SK and China.

Fig. 1 .
Fig. 1.Changes in NO 2 TVCDs, SO 2 VCDs, and CO VCDs across North Korea.Panels (a-c) indicate the averages for 2005-2011, and panels (d-f) represent those for 2012-2018.Panel (g) shows the cities of interest: Pyongyang, Nampo, Pukchang, and Munchon, in addition to the reference sector (123.5-124.5 • E, 38.0-39.6 • N) over the Yellow Sea.Panel (h) shows NO 2 , SO 2 , and CO VCD values averaged for 2005-2011 and 2012-2018 in the selected cities. (For interpretation of the references to color in this figure legend, the reader is referred to the web version of this article.)

Fig. 2 .
Fig. 2. Time series of NO 2 TVCDs, SO 2 VCDs, and CO VCDs in South and North Korea from 2005 to 2018.Colors correspond to different regions.Thin, dashed lines indicate monthly data, and thick lines represent linear regression fits.Inset values represent the absolute (relative) trends.Absolute trends are in 10 15 , 10 15 , and 10 18 molecules cm − 2 yr − 1 for NO 2 , SO 2 , and CO, respectively.Note that panels (a) and (d) have different y-axis ranges.

Fig. 3 .
Fig. 3. Time series of annual mean (a) NO 2 TVCDs, (b) SO 2 VCDs, and (c) CO VCDs across North Korea.Changes in explanatory variables are overlaid.Shading indicates standard deviations of the VCDs across North Korea.Insets present the relative trend values and correlation coefficients between the VCDs and explanatory variables.For NO 2 , the correlation coefficient was calculated using the index of all industrial production.

Fig. 5 .
Fig. 5. Interannual changes in AOD and PM 2.5 concentrations.(a) Annual mean AODs, with inset values depicting relative trends for each region of interest; (b) AODs for entire North Korea, Nampo, and Pukchang, where green shading indicates standard deviations across the entirety of North Korea.The change in the primary explanatory variable (i.e., primary energy supply by coal) is also presented.(c) Annnual mean PM 2.5 concentrations, with inset values depicting relative trends for each region of interest.(For interpretation of the references to color in this figure legend, the reader is referred to the web version of this article.)