The WRF model has been developed for providing mesoscale numerical weather prediction (NWP). It has also been used to provide meteorological input fields for CTM simulations (Appel et al., 2010; Chemel et al., 2010; Foley et al., 2010; Lee et al., 2016; Park et al., 2014). In this study, WRF v3.8.1 with the Advanced Research WRF (ARW) dynamical core was applied to prepare the meteorological inputs for the CMAQ model simulations. Dynamical and physical configurations for the WRF model simulations were selected as follows: the Yonsei University (YSU) scheme for planetary boundary layer (Hong et al., 2006); the WRF single-moment 6-class (WSM6) scheme for the microphysics (Hong and Lim, 2006); the Grell–Freitas ensemble scheme for cumulus parameterization (Grell and Freitas, 2014); the Noah-MP land surface model (Niu et al., 2011; Yang et al., 2011); the Rapid Radiative Transfer Model for Global Circulation Models (RRTMG) for shortwave/longwave options (Iacono et al., 2008); and the revised MM5 scheme for surface layer options (Jimenez et al., 2012). The National Centers for Environmental Prediction (NCEP) Final (FNL) Operational Global Analysis data on 1∘×1∘ grids were chosen for the ICs and boundary conditions (BCs) for the WRF simulations. In order to minimize meteorological field errors for the applications of ICs and BCs to the WRF simulations, the objective analysis (OBSGRID) nudging was conducted using the NCEP Automated Data Processing (ADP) Global Upper Air Observations/Surface Observational weather data via the Cressman's (1959) successive correction method. The adjusted meteorological variables were temperature, geopotential height, relative humidity and zonal/meridional winds.

The model domain for the WRF simulations covers northeast Asia with a horizontal resolution of 15km×15km, having a total of 223 latitudinal and 292 longitudinal grid cells. The size of the WRF domain is slightly larger than that of the CMAQ domain, as shown in Fig. 1. The meteorological data have 27 vertical layers from the surface (1000 hPa) to 50 hPa. The WRF meteorological fields (e.g., temperature, pressure, wind, humidity, and clouds) were then transformed into the CMAQ-ready format via the Meteorology-Chemistry Interface Processor (MCIP; Otte and Pleim, 2010) v4.3, which is a software to serve for transforming horizontal and vertical coordinates while trying to maintain dynamic consistency between WRF and CMAQ model simulations.

The CMAQ v5.1 model was used to estimate the concentrations of the atmospheric chemical species over the domain, as shown in Fig. 1. The CMAQ domain has 204 latitudinal and 273 longitudinal grid cells in total and also has a 15km×15km horizontal resolution and 27 vertical layers. The CMAQv5.1 model was configured to use. Chemical and physical configurations for the CMAQ model simulations were selected as follows: SAPRC07tc for the gas-phase chemical mechanism (Hutzell et al., 2012); AERO6 for aerosol thermodynamics (Appel et al., 2013); Euler Backward Iterative (EBI) chemistry solver (Hertel et al., 1993), which is a numerically optimized photochemistry mechanism solver; M3DRY for dry deposition velocity (Pleim and Xiu, 2003; Xiu and Pleim, 2001); global mass-conserving scheme (Yamartino and WRF) for horizontal and vertical advection (Colella and Woodward, 1984); MULTISCALE (Louis, 1979), which is a simple first-order eddy diffusion scheme for horizontal diffusion; and the Asymmetric Convective Model, version 2 (ACM2; Pleim, 2007a, b), for vertical diffusion.

For anthropogenic emissions, KORUS v1.0 emissions (Woo et al., 2012) were used. The emissions cover almost all of Asia and are based on three emission inventories: the Comprehensive Regional Emissions inventory for Atmospheric Transport Experiment (CREATE) for East Asia excluding Japan; the Model Inter-Comparison Study for Asia (MICS-Asia) for Japan; and the Studies of Emissions and Atmospheric Composition, Clouds and Climate Coupling by Regional Surveys (SEAC4RS) for South and Southeast Asia.

Biogenic emissions were prepared by running the Model of Emissions of Gases and Aerosols from Nature (MEGAN v2.1; Guenther et al., 2006, 2012) with a grid size identical to that of the CMAQ model simulations. For the MEGAN simulations, the MODIS land cover data (Friedl et al., 2010) and improved leaf area index (LAI) based on MODIS datasets (Yuan et al., 2011) were utilized. Pyrogenic emissions were obtained from the Fire Inventory from NCAR (FINN; Wiedinmyer et al., 2006, 2011). The lateral BCs for the CMAQ model simulations were prepared using the global model results of the Model for Ozone and Related chemical Tracers, version 4 (MOZART-4; Emmons et al., 2010) at every 6 h. The mapping and regridding of the MOZART-4 data were conducted by matching the CMAQ grid information.

A Korean geostationary satellite of Communication, Ocean and Meteorological Satellite (COMS) was launched on 26 June in 2010 over the Korean Peninsula. The COMS is a geostationary orbit satellite, and it is stationed at an altitude of approximately 36 000 km at a latitude of 36∘ N and a longitude of 128.2∘ E, with a horizontal coverage of 2500km×2500km (refer to Fig. 1). Among the three payloads of the COMS, Geostationary Ocean Color Image (GOCI) is the first multichannel ocean color sensor with visible and near-infrared channels. The GOCI instrument provides hourly spectral images with a spatial resolution of 500m×500m from 00:30 to 07:30 coordinated universal time (UTC) for eight spectral (six visible and two near-infrared) channels at 412, 443, 490, 555, 660, 680, 745 and 865 nm.

The Yonsei aerosol retrieval (YAER) algorithm for the GOCI sensor was initially developed by Lee et al. (2010) to retrieve the aerosol optical properties (AOPs) over ocean areas and was then improved by expanding to consider nonspherical aerosol optical properties (Lee et al., 2012). Choi et al. (2016) further extended the algorithm for application to land surfaces, and the algorithm was referred to as the GOCI YAER version 1 algorithm. With the GOCI YAER algorithm, hourly aerosol optical depths (AODs) at 550 nm were produced over East Asia. Choi et al. (2016) compared the retrieved GOCI AODs with other satellite-retrieved and ground-based observations and found several errors in the cloud masking and surface reflectances. These errors were corrected in the recently updated second version of the GOCI YAER algorithm (Choi et al., 2018), which used the updated cloud masking and more accurate surface reflectances. In this study, the most recent GOCI AOD products from the GOCI YAER version 2 algorithm were used.

In addition to the satellite data, ground-based observations in South Korea and China were also collected for use in the air quality prediction system for PM and gas-phase pollutants. The orange, red and blue dots in Fig. 1 represent the ground-based observation sites in China, Air Korea and supersite stations in South Korea, respectively. These observations provide real-time concentrations of air quality criteria species such as PM10, PM2.5, CO, O3, SO2 and NO2.

Throughout the period of the KORUS-AQ campaign, ground-based observation data were obtained from 1514 stations in China, 264 Air Korea stations and seven supersite stations in South Korea. In this study, 80 % of the ground-based observations in China and Air Korea stations in South Korea were randomly selected for use in the prediction system. The other 20 % of the data and supersite observations were used to evaluate the performances of the developed air quality prediction system.

In addition, AErosol RObotic NETwork (AERONET) AODs were used to conduct an independent evaluation of the air quality prediction system. AERONET is a federated global ground-based sun photometer network (Holben et al., 1998). Cloud-screened and quality-assured level 2.0 AODs for the AERONET were used in this study.

CMAQ AODs are calculated by integrating the aerosol extinction coefficient (σext) using the following equation: 1AODλ=∫0zσextλdz, where z represents the vertical height; σext is defined as the sum of the absorption coefficient (σabs) and the scattering coefficient (σsca); and σabs and σsca can be estimated by Eqs. (3) and (4), respectively, as shown below. 2σextλ=σabsλ+σsca(λ)3σabsλMm-1=∑in∑jm{(1-ωijλ)⋅βijλ⋅fijRH⋅[C]ij}4σscaλMm-1=∑in∑jm{ωijλ⋅βijλ⋅fijRH⋅[C]ij}, where i and j denote the particulate species and size bin (or particle mode), respectively; ωijλ is the single scattering albedo; βijλ is the mass extinction efficiency (MEE) of particulate species i for the size bin or particle mode j; [C]ij is the concentration of particulate species including NH42SO4,NH4NO3, black carbon, organic aerosols (OA), mineral dust and sea-salt aerosols; RH is the relative humidity; fijRH is the hygroscopic factor; and the single scattering albedo (ω) implies to the fraction (portion) of scattering in the total extinction.

Using Eqs. (2)–(4), AODs were calculated from the aerosol composition and RH. There have been intensive tests using different β and fRH values in the following three previous studies: (1) Chin et al.'s (2002) study with the Goddard Chemistry Aerosol Radiation and Transport (GOCART) model; (2) Martin et al.'s (2003) study with the GEOS-Chem model; and (3) Malm and Hand's (2007) study with the CMAQ model. Lee et al. (2016) tested these methods and then found that Chin et al.'s (2002) method reproduced the best results in estimating AODs at 550 nm over East Asia. On the basis of Lee et al.'s (2016) work, σext was estimated with the β and fRH values suggested by Chin et al. (2002). After that, σext was integrated with respect to altitude, in order to calculate the AODs. The calculated AODs were used in the air quality prediction system in order to prepare the ICs for the PM predictions.

The ground-based observations, together with GOCI-derived AODs, were used to prepare the ICs for the air quality predictions with the CMAQ model simulations. In order to achieve this, the following steps were taken: (i) the CMAQ-calculated concentrations of CO, O3 and SO2 were combined with the concentrations of CO, O3 and SO2 obtained from ground-based observations in South Korea (Air Korea) and China; (ii) the CMAQ-calculated AODs were assimilated with the GOCI AODs; (iii) the assimilated AODs were converted into PM10; (iv) the converted PM10 was again assimilated at the surface in South Korea and China; and (v) after the DA at the surface, the ratios of the assimilated species concentrations to the original CMAQ-simulated concentrations were applied so as to the adjust vertical profiles of the chemical species above the surface. In the air quality prediction system, the DA cycle is 24 h, and the assimilation takes place every day at 00:00 UTC (refer to Fig. 3).

The optimal interpolation (OI) method with the Kalman filter was chosen in the air quality prediction system. The OI method was originally used for meteorological applications (Lorenc, 1986) and has also been used in the assimilations for trace gases (Khattatov et al., 1999, 2000; Lamarque et al., 1999; Levelt et al., 1998). Recently, the OI technique has also been applied to aerosol fields (Collins et al., 2001; Yu et al., 2003; Generoso et al., 2007; Adhikary et al., 2008; Carmichael et al., 2009; Chung et al., 2010; Park et al., 2011, 2014; Tang et al., 2015, 2017).

Aerosol assimilation using the OI method was first applied by Collins et al. (2001) as follows: 5τ′m=τm+K(τo-Hτm)6K=BHT(HBHT+O)-17O=[foτo2+εo2]I8Bdx,dz=[fmτm2+εm2]exp⁡-dx22lmx2exp⁡-dz22lmz2, where τ′m, τm and τo represent the assimilated products by the OI method, the modeled values and the observed values, respectively; K is the Kalman gain matrix; H is the observation operator (or forward operator), which is an interpolator from model to observation space; B and O are the background and observation error covariance matrices, respectively; (⋅)T denotes the transpose of a matrix; fo is the fractional error in the observation-retrieved value; εo is the minimum root-mean-square error in the observation-retrieved values; I denotes the unit matrix; fm is the fractional error in the model estimates; εm is the minimum root-mean-square error in the model estimates; dx is the horizontal distance between two model grid points; lmx is the horizontal correlation length scale for the errors in the model; dz is the vertical distance between two model grid points; and lmz is the vertical correlation length scale for the errors in the model. In this work, the OI technique was applied for the DA of atmospheric gaseous species as well as particulate species.

Six free parameters (fm, fo, εm, εo, lmx and lmz) were used to calculate the error covariance matrices of the observations and model, the mathematical formalisms of which are described in Eqs. (7) and (8), respectively. Several previous studies have used fixed values for free parameters (Collins et al., 2001; Yu et al., 2003; Adhikary et al., 2008; Chung et al., 2010). These runs are called “static” runs. In contrast to those previous studies, “non-static” free parameters were applied in this study by minimizing the differences between the assimilated values and observations via an iterative process at each assimilation time step. This non-static free-parameter scheme is possible due to the fact that the OI technique with the Kalman filter is much less costly in terms of computation time than other DA techniques, such as the 3D or 4D variational methods. This is another advantage of using the OI technique in this system. It typically takes less than 20 min with a workstation environment (dual Intel Xeon 2.40 GHz processor).

In the procedure of DA, PM10 was assimilated in this study, because the PM10 data were more plentiful than PM2.5. The assimilated PM10 then needs to be allocated to the PM composition for the CMAQ-model prediction runs. In order to achieve this, the differences between the assimilated PM10 and background PM10 (ΔPM10) were first calculated. Then, ΔPM2.5 was estimated using the ratios of PM2.5 to PM10 from the background CMAQ model runs (i.e., ΔPM2.5=ΔPM10×PM2.5/PM10). ΔPM2.5 was then allocated to the PM2.5 composition according to the comparison between two PM2.5 compositions observed at the seven supersites and simulated from the CMAQ model runs over South Korea. Both of the compositions are shown in Fig. 4. In Fig. 4, the PM “OTHERS” indicates the remaining particulate matter species after excluding sulfate, nitrate, ammonium, organic aerosol (OA) and elementary carbon (EC). The PM OTHERS occupies 25 % of the total PM2.5 observed at supersites. The other fraction, ΔPM10×(1-PM2.5/PM10), was also distributed to the coarse-mode particles (PM2.5–10) as crustal elements.

Figure 5 shows the time-series plots of PM10, PM2.5, CO, O3 and SO2 concentrations from the BASE RUN and the DA RUN. Here, the observation data (OBS) obtained from the Air Korea network were compared with the results of the two sets of the CMAQ model simulations, i.e., (1) BASE RUN and (2) DA RUN. As mentioned previously, 20 % of the Air Korea observations used in the evaluation were randomly selected during the period of the KORUS-AQ campaign. The other 80 % of the Air Korea data were used in the DA at 00:00 UTC. For the forecast hours from 01:00 to 23:00 UTC, all of the ground observations (254 Air Korea and seven supersite stations) were used to evaluate the performances of the developed air quality prediction system. As shown in Fig. 5, we achieved some improvements in the prediction performances by applying the ICs to the CMAQ model simulations. The BASE RUN significantly underpredicted PM10, PM2.5 and CO, while the DA RUN produced concentrations that were more consistent with the observations than those of the BASE RUN.

In the case of CO, the observed CO mixing ratios were about 3 times higher than those from the BASE RUN. These large differences are well known and have been attributed to the underestimated emissions of CO (Heald et al., 2004). However, when the DA was applied, the predictions of the CO mixing ratios improved. Similarly, the performances of the PM10 and PM2.5 predictions improved with the application of the DA. Unlike PM10, PM2.5 and CO, the O3 mixing ratios and its diurnal trends from both the BASE RUN and DA RUN tend to be well matched with the observations. By contrast, the poorest performances of the BASE RUN and the DA RUN were shown for SO2.

In addition, a dust event took place between 6 and 8 May. This event is captured by the DA RUN (check red peaks in Fig. 5a and b), while the BASE RUN cannot capture this dust event. This demonstrates the capability of the current system to possibly predict dust events in South Korea. In the DA RUN, dust information is provided to the CMAQ model runs through both/either GOCI AOD and/or ground PM observations measured along the dust plume tracks.

The effectiveness of the DA with respect to prediction time was also analyzed by calculating the aggregated average concentrations of atmospheric species (see Figs. 6, 7 and 9). Figure 6 depicts the CMAQ-calculated average concentrations of PM10, PM2.5, CO and SO2 against the Air Korea observations. Our air quality prediction system regenerated relatively well-matched concentrations for PM10, PM2.5 and CO from the DA RUN.

Figure 7 shows the case of ozone from the DA RUN by assimilating CMAQ outputs with Air Korea-observed (a) O3 mixing ratios and (b) both O3 and NO2 mixing ratios for a preliminary test run. The ozone mixing ratios from the DA RUN in Fig. 7a were reasonably consistent with the observations at 00:00 UTC but disagreed with those between 04:00 and 09:00 UTC (13:00 and 18:00 KST), when solar insolation is the most intense. This may be attributed to the chemical imbalances between ozone production and ozone destruction (or titration). However, if CMAQ NO2 was assimilated with ground-based observations in South Korea (Air Korea) and China, the predicted ozone mixing ratios became substantially closer to the observations, as shown in Fig. 7b. This is clearly due to the fact that NOx is an important precursor of ozone. In the prediction of the ozone mixing ratios, both 1 h peak ozone (around 15:00 KST) and 8 h averaged ozone mixing ratios (between 09:00 and 17:00 KST) are important. Figure 7 clearly shows that the prediction accuracies of both the ozone mixing ratios were improved after the DA of NO2 mixing ratios.

Although the DA for NO2 provided better ozone predictions, one should take caution in using the NO2 observations. The NO2 mixing ratios measured at Air Korea sites are known to be contaminated by other nitrogen gases such as nitric acid (HNO3), peroxyacetyl nitrates (PANs) and alkyl nitrates (ANs), since the Air Korea NO2 mixing ratios are measured through a chemiluminescent method with catalysts of gold or molybdenum oxide at high temperatures. These are known to be “NO2 measurement artifacts” (Jung et al., 2017), which is one of the reasons that the DA of NO2 was not shown in Fig. 6. The NO2 mixing ratios are corrected from the Air Korea NO2 data and are then used to prepare the ICs via the DA for more accurate ozone and NO2 predictions. Currently, such corrections of the observed NO2 mixing ratios are being standardized with more sophisticated year-long NO2 measurements. After the corrections of the NO2 measurement artifacts, more evolved schemes of ozone and NO2 predictions will be possible in the future. As shown in Fig. 7, about a 20 % reduction (average fraction of non-NO2 mixing ratios in the observed NO2 mixing ratios) was made for these demonstration runs (Jung et al., 2017).

Another practical issue is now discussed. Although the assimilation with the observed NO2 mixing ratios can enhance the accuracy of the predictions of the daytime ozone mixing ratios, the nighttime ozone mixing ratios tend to be consistently overpredicted in the aggregated plot of the ozone mixing ratios at the observation sites (see Fig. 7). This can be caused by underestimated NO2 mixing ratios and thus not enough nighttime ozone titration. As mentioned before, reliable NO2 prediction via the correction of the NO2 measurement artifacts will be made in the future. Another possible reason of the overpredicted ozone mixing ratios during the nighttime can be underestimation of the mixing layer height (MLH). Figure 8 shows a comparison between lidar-measured MLH (black dashed line) and WRF-calculated MLH (with the option of the Yonsei University planetary boundary layer scheme by Hong et al., 2006; see red line). As shown in Fig. 8, the nocturnal lidar-measured MLH is about 2 times higher than the nocturnal WRF-calculated MLH as measured at a lidar site inside the campus of Seoul National University (SNU) in Seoul. Such underestimated MLH in the model tends to compress the ozone molecules within the mixing layer during the nighttime, which leads to consistently overpredicted nocturnal ozone mixing ratios. Based on this discrepancy shown in Fig. 8, more intensive comparison study is being carried out by comparing lidar-measured MLH with model-calculated MLH at multiple sites in South Korea.

In this work, the aerosol composition (including EC, OA, sulfate, nitrate and ammonium) was further compared with the composition observed at the supersites shown in Fig. 9. As shown in Fig. 9, agreement was observed between the DA RUN and observations for all of the major PM constituents. Again, a strong capability of our DA system is to improve the predictions of the aerosol composition.

Figure 10 shows the spatial distributions and bias of PM and chemical species throughout the entire period of the KORUS-AQ campaign over the Seoul metropolitan area (SMA). Noticeable improvements are observed to have been achieved in the spatial distributions by applying the ICs to the CMAQ model simulations, particularly for PM10 (Fig. 10a), PM2.5 (Fig. 10b) and CO (Fig. 10c). As shown in Fig. 10, the underpredicted concentrations of PM10, PM2.5 and CO were adjusted to concentrations closer to the observations. In the case of SO2 (see Fig. 10d), the DA RUN produced better agreement with the observations than the BASE RUN, but there were still underpredicted SO2 concentrations over the northeastern part of the SMA.

By contrast, relatively lower ozone mixing ratios from the DA RUN against the BASE RUN were found in the southwestern part of the SMA (see Fig. 10e). Due to the nonlinear relationship between NOx and O3, high mixing ratios (or emissions) of NOx in the SMA can lead to depletion of ozone. In these runs, the precursors of ozone such as NOx and volatile organic compounds (VOCs) were excluded in the preparation of the ICs for CMAQ model simulations. Again, this is because the Air Korea NO2 mixing ratios are contaminated by several reactive nitrogen species, so the data cannot be directly used in the assimilation procedures. In the case of VOCs, a limited number of datasets are available in South Korea for the DA. Improvements in the prediction of ozone mixing ratios can be achieved when the NO2 mixing ratios are corrected and a sufficient amount of VOC data (possibly from satellite data in the future) are available.

In order to achieve a better understanding of the performances of the DA RUN, analyses of statistical variables such as index of agreement (IOA), Pearson's correlation coefficient (R), root-mean-square error (RMSE) and mean bias (MB) were conducted using observations from the Air Korea stations for PM10, PM2.5, CO, SO2 and O3 (see Fig. 11). Definitions of the statistical variables are given in Appendix A.

After the applications of the ICs, both RMSE and MB became lower, while the correlation coefficient became higher for the entire set of predictions. In addition, it was found that the differences between the BASE RUN and the DA RUN tended to diminish as the prediction time progressed. The results of the statistical analysis are listed in Table 1. The results of the DA RUN were reasonably consistent with the observations for PM10 (IOA =0.60; R=0.40; RMSE =34.87; MB =-13.54) and PM2.5 (IOA =0.71; R=0.53; RMSE =17.83; MB =-2.43), as compared to the BASE RUN for PM10 (IOA =0.51; R=0.34; RMSE =40.84; MB =-27.18) and PM2.5 (IOA =0.67; R=0.51; RMSE =19.24; MB =-9.9). In terms of bias, an improvement was found for C, with MB = -0.036 for the DA RUN and MB =-0.27 for the BASE RUN. Regarding O3 and SO2, the DA RUN showed slightly better performances than the BASE RUN.

Table 2 presents the results of the statistical analysis at 00:00 UTC when the DA was conducted, with the results clearly showing how much closer the DA makes the CMAQ-calculated chemical concentrations to the observed concentrations. Collectively, the DA improved model accuracy by a large degree in terms of R, particularly for PM10 (R: 0.3→0.75; slope: 0.17→0.66) and O3 (R: 0.09→0.61; slope: 0.07→0.42). In addition, for all species, MB and RMSE decreased significantly with the DA RUN as compared with the BASE RUN.

In this section, a sensitivity analysis was conducted with different implementation time intervals of the DA (i.e., 24, 6 and 3 h) for AOD (refer to Fig. 12). As shown in Fig. 12, more frequent implementation of the DA is expected to make the predicted results closer to the observations. Although the DA RUN with a shorter assimilation time interval tends to produce a better prediction, it is not always the most appropriate choice, since the shorter assimilation time interval results in increased computational cost. Therefore, an optimized assimilation time interval should be found to achieve the best performances from the given DA system with the consideration of its own computational ability.

In addition, sensitivity analyses of the developed air quality prediction system applied to multiple implementations of the DA with different time intervals were also investigated for (a) PM10, (b) PM2.5, (c) CO, (d) SO2 and (e) O3, shown in Fig. 13. Figure 13 shows a soccer plot analysis for BASE RUN (blue crosses) and DA RUNs with different DA time intervals of 24 h (OI; red circles), 2 h (2 h OI; black diamonds) and 1 h (1 h OI; dark-green triangles). This set of testing was designed based on the fact that the performances are expected to improve if the DAs are implemented multiple times prior to the actual predictions at 00:00 UTC. Here, for the 2 h OI run, the DA was implemented three times a day at 20:00, 22:00 and 00:00 UTC, while for the 1 h OI run, the DA was implemented at 22:00, 23:00 and 00:00 UTC. The performances of all of the chemical species excluding ozone improved, as expected, with DA RUNs with more frequent and longer DA time intervals (i.e., three-time implementation with a 2 h time interval in our cases). In the case of ozone, the best performance was found for the air quality prediction system with the DA time interval of 24 h.

Unsurprisingly, more frequent DAs prior to the actual prediction mode (i.e., before 00:00 UTC in our system) with a longer time interval (such as 2 h) will be computationally costly. There will certainly be a tradeoff between the precision of air quality prediction and the computational cost. The system should be designed under the consideration of these two factors.