Estimating ground-level PM 2 . 5 in eastern China using aerosol optical depth determined from the GOCI satellite instrument

We determine and interpret fine particulate matter (PM2.5) concentrations in eastern China for January to December 2013 at a horizontal resolution of 6 km from aerosol optical depth (AOD) retrieved from the Korean geostationary ocean color imager (GOCI) satellite instrument. We implement a set of filters to minimize cloud contamination in GOCI AOD. Evaluation of filtered GOCI AOD with AOD from the Aerosol Robotic Network (AERONET) indicates significant agreement with mean fractional bias (MFB) in Beijing of 6.7 % and northern Taiwan of −1.2 %. We use a global chemical transport model (GEOS-Chem) to relate the total column AOD to the near-surface PM2.5. The simulated PM2.5 / AOD ratio exhibits high consistency with ground-based measurements in Taiwan (MFB=−0.52 %) and Beijing (MFB=−8.0 %). We evaluate the satellitederived PM2.5 versus the ground-level PM2.5 in 2013 measured by the China Environmental Monitoring Center. Significant agreement is found between GOCI-derived PM2.5 and in situ observations in both annual averages (r = 0.66, N = 494) and monthly averages (relative RMSE= 18.3 %), indicating GOCI provides valuable data for air quality studies in Northeast Asia. The GEOS-Chem simulated chemical composition of GOCI-derived PM2.5 reveals that secondary inorganics (SO 4 , NO − 3 , NH + 4 ) and organic matter are the most significant components. Biofuel emissions in northern China for heating increase the concentration of organic matter in winter. The population-weighted GOCI-derived PM2.5 over eastern China for 2013 is 53.8 μg m, with 400 million residents in regions that exceed the Interim Target-1 of the World Health Organization. Published by Copernicus Publications on behalf of the European Geosciences Union. 13134 J.-W. Xu et al.: Estimating ground-level PM2.5 in eastern China


Abstract
We determine and interpret fine particulate matter (PM 2.5

) concentrations in eastern
China for January to December 2013 at a horizontal resolution of 6 km from aerosol optical depth (AOD) retrieved from the Korean Geostationary Ocean Color Imager (GOCI) satellite instrument. We implement a set of filters to minimize cloud 5 contamination in GOCI AOD. Evaluation of filtered GOCI AOD with AOD from the Aerosol Robotic Network (AERONET) indicates significant agreement with mean fractional bias (MFB) in Beijing of 6.7% and northern Taiwan of -1.2%. We use a global chemical transport model (GEOS-Chem) to relate the total column AOD to the nearsurface PM 2.5 . The simulated PM 2.5 /AOD ratio exhibits high consistency with ground-10 based measurements in Taiwan (MFB= -0.52%) and Beijing (MFB= -8.0%). We evaluate the satellite-derived PM 2.5 versus the ground-level PM 2.5 in 2013 measured by the China Environmental Monitoring Center. Significant agreement is found between GOCIderived PM 2.5 and in-situ observations in both annual averages (r 2 =0.66, N=494) and monthly averages (relative RMSE=18.3%), indicating GOCI provides valuable data for 15 air quality studies in Northeast Asia. The GEOS-Chem simulated chemical composition of GOCI-derived PM 2.5 reveals that secondary inorganics (SO 4 2-, NO 3 -, NH 4 + ) and organic matter are the most significant components. Biofuel emissions in northern China for heating increase the concentration of organic matter in winter. The populationweighted GOCI-derived PM 2.5 over eastern China for 2013 is 53.8 µg m -3 , with 400 20 million residents in regions that exceed the Interim Target-1 of the World Health Organization.

Introduction
Fine particulate matter with aerodynamic diameter less than 2.5 µm (PM 2.5 ) is a robust indicator of mortality and other negative health effects associated with ambient air pollution (Goldberg et al., 2008;Laden et al., 2006). It is estimated that more than three million people lost their lives prematurely due to PM 2.5 in 2010 (Lim et al., 2012), of which one million occurred in East Asia (Silva et al., 2013). In China, there have already 5 been several episodes with PM 2.5 described as "beyond index" levels. Thus, it is of paramount importance to monitor PM 2.5 concentration across China. Satellite remote sensing has a high potential to monitor PM 2.5 .
Satellite retrievals of aerosol optical depth (AOD), which provide a measure of the amount of light extinction through the atmospheric column due to the presence of 10 aerosols, have long been recognized to relate to ground level PM 2.5 (Wang and Christopher, 2003). Many studies have developed advanced statistical relationships to estimate with high accuracy surface PM 2.5 from satellite AOD (Liu et al., 2009;Kloog et al., 2012;Hu et al., 2013). For example, Ma et al. (2014) estimated PM 2.5 concentrations in China from satellite AOD by developing a national-scale geographically weighted 15 regression model, and found strong agreement (r 2 =0.64) with ground measurements.
In addition to empirical statistical methods, satellite AOD can also be geophysically related to surface PM 2.5 by the use of a chemical transport model to simulate the PM 2.5 to AOD relationship (Liu et al., 2004;van Donkelaar et al., 2010). This approach was first demonstrated using data from the Multiangle Imaging Spectroradiometer (MISR) aboard 20 NASA's Terra satellite over the United States for 2001 (Liu et al., 2004). Van Donkelaar et al. (2006 extended this approach to estimate PM 2.5 from AOD retrieved from both the MODIS (Moderate Resolution Imaging Spectroradiometer) and the MISR satellite instruments, and developed a long-term global estimate of PM 2.5 at a spatial resolution of approximately 10 km x 10 km. Boys et al. (2014) used AOD retrieved from MISR and the SeaWiFS (Sea-viewing Wide Field-of-view Sensor) to produce a 15-year (1998−2012) global trend of ground-level PM 2.5 . These previous studies have proven to 5 be globally effective, but more detailed regional investigation is needed in densely polluted and populated regions like China.
The Geostationary Ocean Color Imager (GOCI) is the first geostationary satellite instrument that offers multi-spectral aerosol optical properties in Northeast Asia (Park et al., 2014). GOCI has a high observation density of 8 retrievals/day (hourly retrievals from 10 09:00 to 16:00 Korean Standard Time) over a location, which exceeds the retrieval density of traditional low-Earth polar-orbiting satellite instruments. Thus, GOCI is promising for more detailed investigations on aerosol properties in highly polluted and populated regions including eastern China.
In this study, we estimate ground-level PM 2.5 in eastern China for 2013 at a horizontal 15 resolution of 6 km by 6 km, by using AOD retrieved from GOCI, coupled with the relationship of PM 2.5 to AOD simulated by a chemical transport model (GEOS-Chem).
Section 2 describes the approach and data. Section 3 evaluates the GOCI AOD, the simulated PM 2.5 to AOD relationship, and the GOCI-derived PM 2.5 using recently available ground-level measurements from the China Environmental Monitoring Center 20 (http://113.108.142:20035/emcpublish/). We also interpret the GOCI-derived PM 2.5 by using the GEOS-Chem model to estimate its chemical composition. Section 4 summarizes the major findings and potential future improvements of the current analysis.

Methods
2.1 Aerosol optical depth from the GOCI satellite instrument GOCI operates onboard the Communication, Ocean, and Meteorology Satellite (COMS) 5 that was launched in 2010 in Korea (Lee et al., 2010). The spatial coverage of GOCI is 2500 km x 2500 km in Northeast Asia, including eastern China, the Korean peninsula and Japan (Kang et al., 2006). GOCI has eight spectral channels for aerosol retrievals, including six visible bands at 412, 443, 490, 555, 660, 680 nm and two near infrared bands at 745 and 865 nm (Park et al., 2014). The Level 2 AOD products are retrieved at a 10 spatial resolution of 6 km by 6 km, using a clear-sky composite method for surface reflectance and a lookup table approach based on AERONET observations (Lee et al., 2010;Lee et al., 2012).
A challenge using GOCI to detect aerosols in the atmosphere is the absence of midinfrared (IR) channels to detect clouds, which means that significant errors could be 15 induced in the estimates of AOD. The operational GOCI products screen clouds based on spatial variability and threshold tests at each 6 km x 6 km pixel in combination with a meteorological imager that has 4 IR channels (at 3.7 µm, 6.7 µm, 10.8 µm, 12 µm wavelengths) at 4 km by 4 km resolution onboard the same satellite (Cho et al., 2006). However, as will be shown here cloud contamination still occurs. Therefore, we apply a 20 set of spatial filters following Hyer et al. (2011) and temporal filters to further eliminate cloud contamination in GOCI AOD. The filters include 1) a buddy check that sets a minimum number of 15 retrievals per 30 km x 30 km grid cell, 2) a local variance check to eliminate grid cells where the coefficient of variation of AOD is larger than 0.5 within the surrounding 5 x 5 grid cells and 3) a diurnal variation check that excludes grid cells with diurnal variation (maximum -minimum) of AOD larger than 0.74 which is the 90 th percentile of diurnal variation of AERONET AOD in Beijing and northern Taiwan for 5 2013. In this study, we use GOCI AOD for Jan-Dec 2013 to derive ground-level PM 2.5 in eastern China.

Aerosol optical depth from AERONET ground-based measurements
The Aerosol Robotic Network (AERONET) is a globally distributed network of CIMEL Sun photometers (Holben et al., 1998) that provide multi-wavelength AOD 10 measurements with a low uncertainty of < 0.02 (Holben et al., 2001). Here we use AERONET Level 1.5 cloud screened data (Smirnov et al., 2000) for Jan-Dec 2013 from 4 stations within the GOCI domain: Beijing, Beijing-CAMS, Taipei_CWB and EPA-NCU.
AERONET Level 2 data for 2013 are not available for some stations discussed in this paper. Thus, we use Level 1.5 for consistency. We compared Level 2 and Level 1.5 data 15 for 2013 for stations that do have Level 2 available, and found Level 1.5 AOD is highly consistent with Level 2 AOD with RMSE of 0.01-0.02 (relative RMSE of 2%-7%).
Criteria for selecting an AERONET station are 1) a PM 2.5 ground monitor has to be located within 10 km and 2) a complete time series of AOD data records for the period of study has to be available. Beijing and Beijing-CAMS stations are located in downtown 20 Beijing, with the closest available PM 2.5 monitors 9.5 km and 7.5 km away, respectively.
However, due to interrupted time series of PM 2.5 records at both these stations, we combine the AERONET AOD from the Beijing and Beijing-CAMS stations and PM 2.5 from the corresponding two in-situ ground-based sites as a "combined Beijing" site.
Taipei_CWB and EPA-NCU stations are located in populated northern Taiwan, with nearly collocated PM 2.5 monitors (< 3 km). We similarly combine the Taipei_CWB and EPA-NCU as "northern Taiwan" site. We use these sites to evaluate GOCI AOD and the relationship between AOD and PM 2.5 simulated by a global chemical transport model. 5

Simulation of the relationship between AOD and PM 2.5 by GEOS-Chem
We use the GEOS---Chem chemical transport model (version 9-01-03; http://geos--chem.org) to calculate the spatiotemporally resolved relationship between ground-level PM 2.5 and satellite-retrieved column AOD.
Our nested GEOS-Chem simulation at 1/2° x 2/3° spatial resolution with 47 vertical 10 levels (14 levels in the lowest 2 km) is driven by assimilated meteorology from the Goddard Earth Observing System (GEOS-5). A global simulation at 2° x 2.5° spatial resolution is used to provide boundary conditions for the nested domain (Wang et al., 2004). We spin up the model for one month before each simulation to remove the effects of initial conditions on the aerosol simulation. 15 GEOS-Chem includes a fully coupled treatment of tropospheric oxidant-aerosol chemistry (Bey et al., 2001;Park et al., 2004). The GEOS-Chem aerosol simulation includes the sulfate-nitrate-ammonium system Pye et al., 2009), primary (Park et al., 2003) and secondary (Henze et al., 2006;Henze et al., 2008;Liao et 20 al., 2007;Fu et al., 2008) organics, mineral dust (Fairlie et al., 2007), and sea salt (Jaegle et al., 2011). We estimate the concentration of organic matter (OM, which includes elements such as hydrogen, oxygen and nitrogen) from the simulated primary organic carbon (OC) using spatially and seasonally resolved values from OMI (Ozone Monitoring Instrument) NO 2 and AMS (Aerosol Mass Spectrometer) measurements following Philip et al. (2014). Gas-aerosol phase partitioning is simulated using the ISORROPIA II thermodynamic scheme (Fountoukis and Nenes, 2007). GEOS-Chem calculates AOD 5 using relative humidity dependent aerosol optical properties following Martin et al. (2003). Dust optics are from Ridley et al. (2012).  (Yienger and Levy, 1995;Wang et al., 1998), lightning NO x (Murray et al., 2012), aircraft NO x (Wang et al., 1998;Stettler et al., 2011), ship SO 2 15 from EDGAR (Olivier et al., 2001) and volcanic SO 2 emissions (Fischer et al., 2011).  We apply GEOS-Chem to simulate daily relationships between ground level PM 2.5 and column AOD, specifically PM 2.5 /AOD. PM 2.5 concentrations are calculated at 35% relative humidity for consistency with in-situ measurements. For consistency with GOCI AOD and PM 2.5 ground-based measurements, we sample the simulated AOD only from hours that GOCI has retrievals (00:00 -07:00 UTC), and calculate the simulated daily PM 2.5 from 24-hour averages as reported for the ground-based PM 2.5 measurements. The simulation period is May 2012 -April 2013 as the GEOS-5 meteorological fields are not available afterward. The mismatch with observations for May-Dec 2013 has the potential 5 to degrade performance, but as will be shown here no clear loss of quality is apparent.

In-situ PM 2.5 measurements
We collect PM 2.5 measurements from 494 monitors to evaluate the GOCI-derived values. http://taqm.epa.gov.tw). The specific instrument (BAMs or TEOMs) used by each monitoring site is unknown. The effective relative humidity of the resultant PM 2.5 20 measurement likely varies diurnally and seasonally as a function of the ambient temperature. Semivolatile losses are expected from the TEOMs. The network design appears to include compliance objectives that may affect monitor placement. Despite these issues, we use the monitoring data to evaluate our satellite-derived PM 2.5 since the monitoring data offer valuable information about ground-level PM 2.5 concentrations. We also collect PM 2.5 measurements from a monitor in Beijing as part of the Surface PARTiculate mAtter Network (SPARTAN; www.spartan-network.org) using a threewavelength nephelometer and an impaction filter sampler (Snider et al., 2015). The 5 SPARTAN, CEMC and TEPA PM 2.5 monitoring data combined with AERONET AOD are used to estimate the empirical relationship between PM 2.5 and AOD, and to further evaluate the relationship simulated by the model.

Statistical Terms
Root mean square error (RMSE), relative root mean square error (rRMSE), mean 10 fractional bias (MFB) and mean fractional error (MFE) are defined as where S i is the satellite-derived value of the parameter in question, O i is the corresponding observed value, and N is the number of observations.

Results and Discussion
3.1 Evaluation of satellite AOD and the simulated relationship between PM 2.5 and AOD Figure 1 shows the effects of our cloud-screening filters on GOCI AOD. The left panel shows GOCI true color images from 5 July 2013 at 10:30 (top) and 11:30 (bottom) 5 Korean Standard Time. The boxes identify challenging regions with thick white cloud, dark cloud-free oceans and grey shading that appears to be thin cloud. The operational GOCI AOD retrievals, shown in the middle panel, correctly exclude thick clouds, but report high AOD for the potentially thin clouds. Although these grey regions could contain aerosol, we err on the side of caution. Application of our additional temporal and 10 spatial cloud filters removes the suspicious pixels from the original GOCI data, as shown in the right panel. Our filters reject 10.3% of all the operational GOCI AOD data investigated in this study. We evaluate the cloud filters further below. GOCI AOD is highly consistent with AERONET observations with MFB of 6.7% in Beijing and -1.2% in Taiwan. GOCI AOD and AERONET AOD are positively skewed at both stations, and the skewness is reduced in GOCI AOD at both stations due to more records for extremely small AOD (< 0.04) in GOCI products. The relatively larger rRMSE between GOCI and AERONET AOD in northern Taiwan may reflect the fewer 20 observations there.
We investigate the filtered diurnal variation of GOCI AOD at the above AERONET stations and find the level of AOD is uniform within a day (e.g. the coefficient of variation in Beijing is 0.1), similar to AERONET observations. The effect of excluding our cloud-screening filters is negligible for coincident comparisons with AERONET since AERONET is already cloud-screened. The exclusion throughout the year as shown in Fig. 2 (top). Snider et al. (2015) interpreted coincident measurements of AOD, PM 2.5 , and nephelometer measurements of aerosol scattering and found that the temporal variation of 5 the PM 2.5 /AOD ratio in Beijing was primarily driven by the vertical profile in aerosol scattering. We examine the seasonal variation in the simulated PM 2.5 /AOD and similarly find that the ratio of ground-level aerosol scatter to columnar AOD contributes most (89%) of the monthly variability in the PM 2.5 /AOD ratio in Beijing.
3.2 Evaluation of ground-level PM 2.5 derived from GOCI AOD 10 provinces where the annual PM 2.5 is almost 100 µg m -3 or more. Prior work has attributed this regional enhancement to high emission rates Zhang et al., 2013) that in part arises from emissions when producing goods for exports (Jiang et al., 2015). based sites. A significant correlation (r 2 =0.66, N=494) with a slope near unity (1.01) is found in the annual scatter plot. The slope remains near unity (0.95-1.01) in seasonal scatter plots. The weaker correlation for all four seasons implies random representativeness differences between point in situ measurements and area-averaged satellite values when data density diminishes. Semivolatile losses from some in situ instruments (TEOMs) might contribute to scatter in winter when nitrate constitutes a larger fraction of PM 2.5 . We focus on more meaningful aggregated measurements. Using 5 the same technique, we also estimated PM 2.5 from MODIS Collection 6 AOD for 2013, and found GOCI-derived PM 2.5 achieves greater consistency than MODIS-derived PM 2.5 when compared with ground-based measurements (slope=1.1, r 2 =0.61). GOCI-derived PM 2.5 also corrects the significant underestimation of PM 2.5 from GEOS-Chem (slope=0.68, r 2 =0.85) when compared with ground measurements. 10  Ye et al., 2003;Tao et al., 2012;Zhang et al., 2013). Zhang et al. (2008) showed consistent seasonal patterns in OM at 18 stations in China, with a winter maximum, and a 20 summer minimum, similar to the seasonality of OM in this work. Zhang et al. (2013) studied the chemical composition of PM 2.5 in Beijing and found the percentage of SIA in PM 2.5 is largest in summer, consistent with our result.
The seasonal variation of PM 2.5 in Fig. 6 is driven by a combination of meteorological conditions, emissions, and nitrate formation. All three processes have greater seasonal variation in the north than south. The mixing height over northeastern China has strong seasonal variation with summer having an average mixing height from GEOS-5 that is 1.9 times higher than in winter. The GEOS-Chem simulation reveals that the increase of 5 OM in winter is primarily driven by biofuel emissions from burning wood, animal waste and agricultural waste (Bond et al., 2004) for heating in eastern China. The spatial distribution of biofuel emission is primarily north of the Yangze River, especially from the North China Plain. The significant contribution from biofuel emissions to the OM concentration in our work is consistent with Bond et al. (2004) who found residential 10 biofuel emissions were responsible for ~70% of OC emissions in China. The increase of NO 3 in winter in Fig. 6 is consistent with prior attribution of the increase of NO 3 in winter to the favorable formation of NH 4 NO 3 at low temperatures ).
Supplemental Figure S1 shows the spatial distribution of PM 2.5 chemical components. Table 1 shows the annual chemical composition of GOCI-derived PM 2.5 in regions 15 outlined in Fig. 3  http://sedac.ciesin.columbia.edu/). Table 1 also provides the population-weighted GOCI-20 derived PM 2.5 for regions outlined in Fig. 3   As shown in Table 1, population-weighted PM 2.5 for eastern China considerably exceeds the Interim Target-1 level of PM 2.5 concentration, especially in Beijing and Shandong regions where the PM 2.5 concentration is almost triple the Interim Target-1 level. These elevated concentrations threaten the health of 433 million inhabitants (Table 1)

in eastern
China who live in regions that exceed this target. 15

Conclusions
We estimated the ground-level concentration of PM 2.5 in eastern China for 2013 using AOD retrieved from the GOCI satellite instrument, coupled with the relationship of AOD to PM 2.5 simulated by a global chemical transport model (GEOS-Chem). GOCI-derived PM 2.5 was compared with in-situ measurements throughout eastern China. 20 We applied a set of filters to GOCI AOD to remove cloud contamination. The filtered GOCI AOD showed significant agreement with AERONET AOD at Beijing and northern Taiwan (MFB of 6.7% to -1.2%). We also evaluated the simulated relationship of PM 2.5 and AOD from GEOS-Chem by using an empirical relationship calculated from nearly collocated ground-based PM 2.5 monitors and AERONET AOD stations. A high degree of consistency was observed between the GEOS-Chem simulation and ground-based measurements with MFB of -0.52% to 8.0%. 5 The GOCI-derived PM 2.5 were highly consistent with in-situ measurements, capturing the similar seasonal and spatial distribution throughout eastern China. The highest PM 2.5 concentrations were found in winter over northern regions. The annual averages of GOCI-derived PM 2.5 were strongly correlated (r 2 =0.66) with surface measurements with a slope near unity (1.01). Monthly comparison of GOCI-derived PM 2.5 with ground-based

Acknowledgements
We are grateful to the GOCI, AERONET, CEMC, TEPA and SPARTAN for providing 15 available data used in this study. Funding for this work was provided by NSERC (Natural

Sciences and Engineering Research Council of Canada) and by an Izaak Walton Killiam
Memorial Scholarship for J.-W. Xu. Computational facilities are partially provided by ACEnet, the regional high performance computing consortium for universities in Atlantic Canada. 20 Favez, O., Cachier, H., Sciare, J. and Le Moullec, Y.: Characterization and contribution to PM2.5 of semi--volatile aerosols in Paris (France), Atmos. Environ., 41(36), 7969-7976, doi:10.1016/j.atmosenv.2007.09.031, 2007 Global budgets of atmospheric glyoxal and methylglyoxal, and implications for formation of secondary organic aerosols, J. Geophys. Res., 113, D15303, doi:10.1029/2007JD009505, 2008  Holben, B., Eck, T., Slutsker, I., Tanre, D., Buis, J., Setzer, A., Vermote, E., Reagan, J., Kaufman, Y., and Nakajima, T.: AERONET-A federated instrument network and data archive for aerosol characterization, Remote Sens. Environ., 66, 1-16, 1998.         is calculated from the average of all grid boxes that contain PM2.5 ground monitors. The chemical composition is calculated by applying the GEOS-Chem simulated mass fraction of PM 2.5 chemical components to GOCI-derived PM 2.5 mass concentration. Aerosol water is associated with each PM 2.5 component according to its hygroscopicity. Error bars represent standard errors. Table 1. Annual PM2.5 concentrations, area-weighted concentrations of chemical composition and affected population of PM 2.5 in regions outlined in Fig. 3 and in overall eastern China (excluding northern Taiwan) for 2013. Aerosol water is not associated with each PM 2.5 component for consistency with measurement protocols. PM 2.5 concentration is at 35 % relative humidity. IT1 refers to the WHO air quality interim target-1 of 35 μg m -3 .