The study area was South Korea (33–39∘ N, 124–131.5∘ E), located in northeast Asia, a region known to have relatively poor air quality. Our study area is located in the midlatitude region where the prevailing westerlies carry particulates from the two most rapidly developing countries in Asia (i.e., China and India). The annual mean temperature of South Korea ranges from 10 to 15 ∘C, and the annual precipitation ranges from 1000 to 1900 mm. More than half of the precipitation occurs in summer during the Asian monsoon. Wind direction is seasonal, with northwesterly winds prevailing in winter and southwesterly winds in summer.

Data used in this study are ground observations as the target variable and remote-sensing data, model-based data, and other ancillary spatial data as explanatory variables. We selected the explanatory variables considering the recent literature that estimated ground PM concentrations (He and Huang, 2018; Chen et al., 2018; Brokamp et al., 2018), which are explained in the following sections.

Many of the in situ observation data used in this study showed low concentrations, while there were a relatively small number of observations of high concentrations. This imbalance in samples could result in biased estimation with a significant underestimation of high-concentration data. Thus, over-/subsampling approaches were conducted for the training datasets to overcome the problem caused by the unbalanced samples (Table 3).

The oversampling approach is based on the assumption that the PM concentration of a training sample (i.e., at a pixel) is not significantly different from those of its neighboring pixels. The pixels within a circular window with a radius of three pixels (i.e., 37 pixels including the focus cell) were considered as potential neighboring pixels (see Supplement Fig. S1). Those 37 neighboring pixels were numbered based on the proximity to the center (i.e., the closer the pixel is to the center, the lower the number considering the direction from the focus). In order to perform oversampling, the intervals of 30 and 20 µg m-3 were first applied to the PM10 and PM2.5 samples, respectively (i.e., 0–30, 30–60, …, 360–390, and > 390 µg m-3 for PM10 and 0–20, 20–40, …, 100–120, > 120 for PM2.5). The second groups (i.e., 30–60 µg m-3 for PM10 and 20–40 µg m-3 for PM2.5) had the largest sample sizes, and thus the subsampling approach based on simple random sampling (i.e., 50 %) was applied to the second groups. For the other groups, we multiplied an integer value ranging from 1 to 37 by the sample size of each group to produce a more balanced sample distribution (i.e., the smaller the sample size, the larger the integer). Oversampling was then performed based on the order of the neighboring pixels. Input variables in the adjacent pixels of high-concentration samples were extracted with the corresponding target variables (i.e., PM2.5 and PM10) that were randomly perturbed within 5 % of the focus pixel concentrations. This oversampling approach can effectively reduce the underestimation of high PM concentrations that results from the small training sample size of high-concentration data.

RF is an ensemble model based on classification and regression trees (CART) with randomized node optimization and bootstrap aggregating (a.k.a. bagging; Breiman, 2001). RF generates numerous independent trees to overcome the limitations of a single-decision (or regression) tree method, such as the dependency on a single tree and the problem of overfitting the training data, resulting in better performance than single CARTs (Kim et al., 2015; Lee et al., 2016; Liu et al., 2018). A multitude of independent trees are ensembled to reach a solution by majority voting for classification or averaging for regression (e.g., Amani et al., 2017; Im et al., 2016; Latifi et al., 2018). RF provides information on how a variable contributes to model development using out-of-bag (OOB) data that are not used in training a model (Sonobe et al., 2017; Park et al., 2017). When a variable from OOB data is randomly permuted, the change in mean square error in percentage is calculated (Breiman, 2001). The larger the increase in the error for a variable, the more contributing the variable is. RF was applied to the training data to develop the models for estimating ground-level PM concentrations. The models were evaluated using the test and hindcast validation data.

Accuracy assessment of the developed models were conducted using the test and hindcast validation datasets based on the following five metrics: coefficient of determination (R2), RMSE, relative RMSE (rRMSE), mean bias (MB), and mean error (ME). rRMSE, MB, and ME are calculated as rRMSE=RMSEy‾×100%,MB=1N∑i=1Nfi-yi,ME=1N∑i=1Nfi-yi, where yi is the observed data, y‾ is the mean of the observed data, fi is an estimated value, and N is the number of observations. The rRMSE is the RMSE normalized by the mean value of observed data, which is useful for comparing results with different scales. The MB and ME are the averages of variation between the model-derived and observed values, with the exception that ME uses only absolute difference. The MB presents a tendency of overestimation or underestimation by a given model. The ME is the difference between observation and estimation (Boylan and Russell, 2006).

MODIS AOD is one of the widely used satellite-based aerosol products and has often been used to estimate PM concentrations. The developed RF models were compared with those using MODIS AOD instead of GOCI aerosol products. Unlike GOCI, MODIS only provides AOD with 3 km resolution (i.e., MYD04_3K) over land. AOD was used for developing MODIS-based models without incorporating other aerosol-related variables (i.e., AE, FMF, and SSA). In order to compare the performance between MODIS- and GOCI-based RF models, 50 % of the samples that were commonly included in both MODIS and GOCI datasets were used to develop the models, while the remaining samples were used to validate the models.

In addition, the ground-level PM concentrations predicted using the GOCI-based RF models were compared to the simulated and predicted results by GEOS-Chem and CMAQ models. The GEOS-Chem v10-01 was utilized with the Global Forecast System (GFS; produced by the National Centers for Environmental Prediction (NCEP)) as meteorological fields and the MIX Asian emission inventory was used as emissions. The nested domain for the GEOS-Chem simulation is 70–150∘ E and 15–55∘ N, which covers East Asia. The horizontal resolution of the nested model is 0.25∘ × 0.3125∘. The boundary conditions for the nested model are from the GEOS-Chem global simulation at 2∘ × 2.5∘ horizontal resolution. The CMAQ model version 4.7.1 was used to simulate the ground-level PM10 and PM2.5 concentrations. Meteorological fields simulated by the Weather Research and Forecasting (WRF) model and emission data from the SMOKE model were utilized to run the CMAQ model. The comparison among the GOCI-based model, GEOS-Chem, and CMAQ to in situ measurements was conducted using the hindcast validation dataset. For comparison to in situ measurements, the results from the GOCI-based models were resampled to the GEOS-Chem grid with 0.25∘ × 0.3125∘ from January to September 2016 and to the CMAQ grids with 9 km × 9 km for 2015–2016. The approach by van Donkelaar et al. (2010) that uses the ratio between the ground-level data and total column of AOD to satellite-based AOD (i.e., here GOCI AOD) using the vertical profile of AOD from GEOS-Chem was adopted to predict ground-level PM concentrations (i.e., GOCI-GEOS-Chem fused PM estimation).

The evaluation results of the developed models for estimating PM10 and PM2.5 concentrations using the test datasets over South Korea are presented in Table 4. The models (the improved models hereafter) based on the balanced training samples through over-/subsampling, resulted in R2 values of 0.78 and 0.73 and RMSEs of 17.08 and 8.25 µg m-3 for PM10 and PM2.5, respectively. There was a significant improvement in using the balanced training samples instead of the original samples (decrease in RMSE ∼30 % and rRMSE ∼10 %). MB and ME also confirmed that the balanced samples improved the models estimating ground-level PM concentrations (Table 3; Fig. 3). In particular, high-concentration data (over 150 µg m-3 for PM10 and 50 µg m-3 for PM2.5) were well estimated by the improved models. The slopes of the trends were also improved from 0.46–0.48 to 0.77–0.78. The slopes were still lower than 1, which is due to the slight overestimation of low PM concentration data (Fig. 3). This significant improvement in the estimation performance was mainly due to the proposed sampling strategies in order to use more balanced training data. The use of the balanced training data resulted in the huge increase in the estimation accuracy of ground-level PM concentrations especially for high-concentration samples at the expense of a slight accuracy decrease for low concentrations.

Although it is not possible to directly compare the present results with those from other studies, the results from this study agreed well with those from recent literature that used machine learning approaches for estimating PM concentrations (Gupta and Christopher, 2009b; Wu et al., 2012; Li et al., 2017a; Yeganeh et al., 2017; Hu et al., 2017b; Chen et al., 2018). Hu et al. (2017b) estimated surface PM2.5 concentrations using RF, resulting in the cross-validation R2 of 0.8 and RMSE of 2.83 µg m-3. Similarly, Chen et al. (2018) compared three different methods (i.e., RF, the generalized additive model (GAM), and the nonlinear exposure–lag–response model (NEM)) to estimate surface PM2.5 concentrations over China during 2014–2016. Their daily estimation results show cross-validation R2 of 0.83, 0.55, and 0.51 for RF, GAM, and NEM, respectively, implying the robustness of machine learning compared to traditional statistical models. A geographically adjusted deep belief network (Geoi-DBN) was used to estimate PM2.5 over China and showed a good correlation with observation data (R2=0.88 and RMSE = 13.68 µg m-3; Li et al., 2017a). The literature shows that empirical models using statistical and machine learning approaches often underestimate high PM concentrations (Wu et al., 2012; Li et al., 2017a). However, the RF-based models developed in our study has proved to be effective for modeling high ground-level PM concentrations.

In addition, the seasonal variation in model performance for 2015 and 2016 is shown in Table 5. The R2 values for PM10 estimations are the highest (0.87) in winter with an RMSE of 12.78 µg m-3 and the lowest (0.50) in summer with an RMSE of 12.62 µg m-3, as compared to R2 values of 0.77 and 0.74 with RMSEs of 16.61 and 13.07 µg m-3 in fall and spring, respectively. The summer season resulted in relatively high rRMSE for estimating ground-level PM concentrations compared to the other seasons. This is mainly because ground-level PM concentrations are typically low in summer in South Korea. The cloud contamination and the relatively small sample size in summer might lead to estimation errors (Shi et al., 2014; Sogacheva et al., 2017).

Figure 4 depicts the top 10 input variables that were identified as the most contributing variables by the improved RF models for estimating PM10 and PM2.5 concentrations. The results indicate that AOD, DOY, MaxWS (i.e., maximum wind speed), RSDN (i.e., solar radiation), and Dew (i.e., dew-point temperature) were commonly identified as contributing variables by the RF models to estimate both ground-level PM10 and PM2.5 concentrations. The AOD was identified as the most significant factor, which agreed well with the existing literature (Yu et al., 2017; Zang et al., 2017; Chen et al., 2018). Although most high PM concentration samples had high AOD values, some high-PM samples had low AOD values. Careful examination of the samples shows that there were Asian dust events at low altitudes in those cases, which were not effectively included in the AOD derived from satellite sensor systems. In other words, the satellite-derived AOD has a weak sensitivity in capturing aerosols at low altitudes (Choi et al., 2018). This could be an error source, implying that altitude information of such dust events can be used to further improve the models for estimating ground-level PM concentrations.

Some meteorological variables indicating the atmospheric conditions also contributed to the estimation of ground-level PM concentrations in the improved models. There is a relationship between solar radiation and aerosols in which solar radiation reaching the surface increases with decreasing aerosol concentration (Préndez et al., 1995; Hu et al., 2017a; Borlina and Rennó, 2017). Prior studies noted that there is an inverse relationship between wind speed and both PM10 and PM2.5 (Gupta et al., 2006; Maraziotis et al., 2008; Krynicka and Drzeniecka-Osiadacz, 2013). This relationship causes an increase in PM concentrations under low wind speed conditions but a decrease under high wind speed conditions, which is also confirmed in the present study. This means that atmospheric conditions such as air stagnation have significant impacts on surface PM concentrations. The results correspond to previous studies (e.g., You et al., 2015; Yeganeh et al., 2017; Hu et al., 2017b; Yu et al., 2017) showing that meteorological factors are strongly effective in improving PM estimation models. Interestingly, the anthropogenic factors such as LC_ratio (urban ratio), PopDens (population density), NH3, and SO2 were more important for PM2.5 estimation than PM10. This implies that the sources of PM2.5 are mainly anthropogenic in South Korea (Moon et al., 2011; Ryou et al., 2018).

Figure 5 illustrates the spatial distribution of 2-year (2015–2016) averaged surface PM10 and PM2.5 concentrations at 1 km resolution with station-based in situ PM10 and PM2.5 concentrations over South Korea. The pixels that have concentration values for more than 5 % of the period (> 36 days for the 2 years) were used to produce the spatial distribution maps to secure the reliability of the distribution. The predicted PM10 and PM2.5 have similar spatial patterns with relatively high concentrations for urban areas especially around metropolitan areas and agree well with observed concentrations (Fig. 5).

The seasonal maps of PM10 and PM2.5 concentrations are also shown in Fig. 6. South Korea usually has the rainy season in June and July. For this reason, cloud contaminants are much more significant in the summer than in the other seasons, which resulted in many no-data pixels for the summer maps (Fig. 6). The ground-level PM concentrations in the spring and winter are much higher than in summer and fall for PM10. The results agree well with the general seasonal patterns of PM10 concentrations of South Korea, where PM concentrations are much higher in spring due to Asian dust inflow carried by westerly winds (Park and Shin, 2017). In addition, anthropogenic emissions generally increase PM concentrations in winter (Lu et al., 2011b; Li et al., 2016). The seasonal distribution of PM2.5 concentrations is similar to that of PM10. However, high concentrations were predominantly found in fall for PM2.5. The cold Siberian high pressure might explain this. When warm air from the south flows into the study area, and while the force of the Siberian anticyclone stops, an inversion layer is formed. Then, PM is trapped because the atmospheric circulation becomes stagnant. Another reason may be the relative overestimation of PM2.5 by the RF model in the fall season (Table 5). MB was greatest for the fall season among the four seasons, indicating an overestimation of PM2.5. A more careful data configuration between training and test samples with larger sample size may mitigate such an overestimation.

The existing studies have generally used MODIS-derived AOD to estimate surface PM concentrations for various countries because of its global coverage and high quality (Remer et al., 2006; Gupta and Christopher, 2009a, b; Van Donkelaar et al., 2010; Wang et al., 2010; Chudnovsky et al., 2014; You et al., 2015; Hu et al., 2017b; Yu et al., 2017; He and Huang, 2018). In this section, the estimated ground-level PM10 and PM2.5 concentrations are compared based on GOCI AOD and MODIS AOD. Figure 7 displays the scatterplots showing the cross-validation results of the RF-based models using GOCI-derived and MODIS-derived AODs. Although there was no statistically significant difference between the two types of models through ANOVA tests, the GOCI-based RF models produced slightly better accuracy metrics (i.e., R2, RMSE, and rRMSE) than MODIS-based RF models for estimating ground-level PM concentrations. When comparing ground PM concentrations to AODs derived from the two sensor data (i.e., MODIS and GOCI), GOCI-derived AOD showed slightly higher correlation with the ground PM concentrations than MODIS-derived ones (Supplement Fig. S2). Considering the advantages of GOCI as a geostationary satellite sensor (i.e., moderate spatial and temporal resolutions; eight times a day with a 6 km grid size of the aerosol product), it is very promising to use GOCI-derived products as input to PM estimation models. It should also be noted that GOCI-2, which has enhanced sensor specifications (i.e., 10 data collections per day at 3 km spatial resolution of the aerosol product) is planned to be launched in 2019.

The RF-based models for estimating ground-level PM10 and PM2.5 concentrations were further compared with process-based models, i.e., GEOS-Chem and CMAQ. Figure 8 shows the comparison of the accuracy metrics of the three models: the GEOS-Chem simulated, GOCI-GEOS-Chem fused, and the RF-predicted PM concentrations using the hindcast validation datasets (Table 3). The GOCI-GEOS-Chem fused PM10 concentrations have less errors than the GEOS-Chem simulated PM10 concentration, which agrees well with the existing literature. However, both tend to significantly underestimate the ground-level PM10 concentrations when compared to the proposed RF models. Consequently, the proposed RF models have the lowest RMSE, MB, and ME among those models. Although the results of GOCI-GEOS-Chem fused PM2.5 showed that R2 (GEOS-Chem PM2.5: 0.00; GOCI-GEOS-Chem fused PM2.5: 0.14) and slope (GEOS-Chem PM2.5: -0.02; GOCI-GEOS-Chem fused PM2.5: 1.41) improved more than those of GEOS-Chem PM2.5, the RMSE and ME of the fused model were higher than the GEOS-Chem model because the fused model overestimated PM concentrations. The RF models also produced a better performance than CMAQ for estimating both PM10 and PM2.5 concentrations (Fig. 9). Similar to the GEOS-Chem models, CMAQ tends to underestimate PM concentrations showing a large negative MB value.