AERONET, a network of globally distributed ground-based sun photometers, is widely used to understand global AOPs and to validate satellite-based aerosol products. The AERONET sun photometer measurements of direct solar radiation provide accurate measurements of AOD (∼ 0.01 in the visible and near-infrared and ∼ 0.02 in the ultraviolet) under cloud-free conditions (Eck et al., 1999; Holben et al., 1998, 2001), and sky radiance measurements in an almucantar scenario can be inverted to calculate AOPs such as size distribution, single scattering albedo, phase functions, and the complex index of refraction (Dubovik and King, 2000; Dubovik et al., 2000, 2002).

During the DRAGON-NE Asia campaign in 2012, 20 Cimel sun–sky radiometer instruments were deployed in Seoul, as well as in eastern and western parts of South Korea. In Japan, about 20 instruments were deployed in Osaka, west Japan, and Fukushima valley. The distribution of DRAGON Korea and Japan sites is shown in Fig. 1, along with the number of AOD data provided in level 2.0 (cloud-screened and quality-assured; Smirnov et al., 2000) direct products during the campaign. These deployed sun photometers provided the high spatial-resolution information to address characteristics of mega city aerosol. Figure 2 shows the average and standard deviation for each AOD (500 nm) and Ångström exponent (AE, 440–870 nm) measured during the campaign. In Fig. 2a, the average AOD ranged between 0.23 and 0.52, and showed a decreasing behavior towards the southeast. The maximum value of 0.52 was found at two sites in Fukue (128.68∘ E, 32.75∘ N) and Sanggye (127.07∘ E, 37.66∘ N), while a minimum value of 0.23 was found at Kohriyama site (140.38∘ E, 37.36∘ N). In terms of local average, the mean AOD of 0.43 in Seoul was higher than the value of 0.30 in Osaka. Similarly, the standard deviation of AOD in Fig. 2b was low in the eastern part of Korea. While the standard deviation varied between 0.22 and 0.31 in Seoul, the values in Japan were between 0.11 and 0.16. The regional difference is also calculated in terms of AE in Fig. 2c. The respective average AE of 1.20 and 1.27 in Seoul and Osaka represents that the particle size in Seoul is larger than that in Osaka, in general. The spatial distributions of AOD and AE can be related closely with transport of aerosol in East Asia during winter and spring (Park et al., 2014).

In this study, the extensive AERONET inversion data (level 2.0 daily products) over East Asia (20–50∘ N, 95–145∘ E) were used to analyze optimized AOPs; the retrieved volume size distribution and complex refractive indices are utilized to compute the spectral SSA. Duvobik et al. (2000) recommended that the quality of refractive index and SSA becomes reliable when the AOD (440 nm) is higher than 0.4 and the solar zenith angle is higher than 45∘. To avoid insufficient data points for a low AOD case, the daily averaged product was applied. Level 2.0 AOD data sets measured for the DRAGON-NE Asia 2012 campaign with more than 50 data points were used to validate the retrieval results. The AERONET sites used, including the campaign sites, are listed in Table 1, along with the period of the inversion products. The campaign sites are numbered, and sites indicated in bold represent the validation sites selected randomly to test the consistency of the retrieval accuracy. The inversion products obtained at those validation sites were not applied to analyze the aerosol model, but direct AOD products were used to validate the algorithm. While a total of 12 126 inversion data sets from 1999 to 2012 were compiled, 84 091 AOD data sets were applied from the campaign at 39 sites in the spring of 2012.

A multi-purpose geostationary satellite, COMS, designed to orbit at a longitude of 128.2∘ E, was launched on 27 June 2010 by the Korean government. The satellite performs meteorological and ocean monitoring by using the MI and Geostationary Ocean Color Imager (GOCI) instruments. The MI measures the single visible reflectance (0.55–0.80 µm) at a 1 km spatial resolution, and the brightness temperature (BT) at four infrared (IR) wavelengths at a 4 km spatial and 30 min temporal resolution. The four IR channels cover spectral ranges of 10.3–11.3 (IR1), 11.5–12.5 (IR2), 6.5–7.0 (IR3), and 3.5–4.0 µm (IR4). The MI can cover a full disk from its equatorial position at 128.2∘ E, though this study focuses mainly on images from East Asia. The MI measurements from the single visible and four IR channels are applied to retrieve land and ocean surface temperature, incoming and outgoing radiance, and atmospheric variables including aerosol, cloud properties, precipitable water, and upper tropospheric humidity. The level 2.0 products can be obtained from the National Meteorological Satellite Center (http://nmsc.kma.go.kr/html/homepage/ko/main.do) of Korea.

To estimate the BAOD distribution over East Asia over a long period, an AOD product from the Moderate Resolution Imaging Spectroradiometer (MODIS) (Collection 5.1; MYD04_Lv2.0) was used at 10 × 10 km2 resolution. The AOD at 550 nm from a dark target algorithm (Levy et al., 2007b, 2010; Remer et al., 2005) was interpolated onto a grid of 0.25∘ × 0.25∘ to find the minimum value for each area. Considering spatial variation of BAOD, the MODIS product was applied to cover a wider area over the long term, although satellite measurements have larger uncertainty than the ground-based measurements. The expected error in the AOD product is ±(0.05 + 15 %), and over 66 % of the retrieved AODs from the MODIS algorithm lie within the error range, with a correlation coefficient of 0.9 (Levy et al., 2010). Despite the seasonal variation of atmospheric conditions over northeast Asia, the seasonal variation of the BAOD was not considered because of insufficient data points for winter and summer depending on snow surface and summer monsoon. The uncertainty related with the BAOD assumption will be discussed in Sect. 3.5.

The AOD was retrieved only for cloud-free pixels which satisfy threshold tests of TOA reflectance and BT. The threshold of 0.35 for the TOA reflectance at the visible channel separated bright cloud pixels, and the threshold of 5 K for the BT difference between the maximum BT for the previous 30 days and the BT of the current pixel separated cold cloud pixels. The pixels which have a BT lower than 265 K were also masked out. Additionally, thresholds for BT differences between IR1 and IR2, and IR1 and IR4 were taken from Frey et al. (2008). The thresholds to distinguish cloud and aerosol pixels (IR1 - IR2 brightness temperature difference (BTD)), and to detect low-level clouds (IR1 - IR4 BTD) were adjusted as follows by trial and error. The positive BTD between IR1 and IR2, and the largely negative BTD (< -6 K) were found in cloud pixels. Thus, the cloud masking procedure includes the following tests: visible reflectance>0.35IR1-IR2>0.5K&IR1<268KIR1-IR2>0.5K&IR1max-IR1>5KIR1-IR2>1.5K&IR1-IR4<-6K for oceanIR1-IR2>0.5K&IR1-IR4<-10K for oceanIR1-IR2>1.5K&IR1-IR4<-14K for land.

The BAOD represents a residual AOD value even in the clearest conditions, i.e., the minimum AOD for each location. According to analyses of global AERONET direct measurements, the minimum AOD over urban areas or near an aerosol source region is non-zero due to the steady emission of aerosol (Kim et al., 2016). An underestimation of BAOD results in an underestimation of retrieved AOD. In an environment of continuous development, population growth, and desertification, the BAOD is not negligible, particularly over East Asia. Accordingly, Kim et al. (2014) used the monthly BAOD obtained from AERONET direct measurements in Hong Kong for AOD retrieval in the region. Subsequently, the BAOD was estimated from the MODIS AOD product for 7 years from 2006 to 2012, and used here in order to take advantage of the fine spatial resolution of the satellite measurements. The BAOD ranged from 0.00 to 0.56, with an average value of 0.03 (Fig. 4). Over the ocean, spatial variation of BAOD was not significant because the background aerosol is most likely sea salt with the median value of 0.022. Over land, however, the spatial distribution of BAOD was related to surface type. While the median of BAOD over land was 0.017, the values near metropolitan areas such as Beijing, Seoul, Tokyo, and Hong Kong were generally higher than 0.1. Over the industrialized region located in the lower reaches of the Yangtze River and near Hong Kong, the values even reached over 0.30. Conversely, the region located far from the aerosol source showed low BAODs. Overall, the BAOD map clearly reveals the most heavily polluted region as a hotspot.

The surface reflectance was estimated from the minimum TOA reflectance, after correcting for atmospheric and BAOD effects. For details of the atmospheric correction, see Kim et al. (2014).

The calculated TOA reflectance from RTM simulations is affected by the concentration, particle size/shape, and scattering properties of aerosol. Consequently, an increase in the SSA of the particle correlates positively with TOA reflectance for the same AOD. The use of a well-defined aerosol model to generate the LUT is therefore crucial to obtain accurate AOD values from the inversion method. Although spatial variation of the aerosol characteristics shown in Fig. 2 was not taken into account, a regionally integrated aerosol model over the area of interest suggests typical properties from these areas, since the geostationary MI steadily observes the same field of view from a fixed location. In this study, the aerosol models were obtained from a seasonal average of AERONET inversion data sets over East Asia. There are two groups of inversion data sets applied to examine the effect of the DRAGON-NE Asia campaign on the retrieval accuracy of aerosol. The first data sets were compiled from 18 AERONET sites from 1999 to 2010, with a total of 4898 data points as used by Kim et al. (2014). This group was named as the “original” data set, where the name and location of these sites are represented by italic type. The full list shown by normal characters in Table 1 summarizes the sites used to construct the new data set as described in Sect. 2.1.

The new group includes 40 additional AERONET sites and extends the measurement period by up to 2 years (2011–2012) including the campaign. The greater quantity of data, from the increased number of sites for the extended measurement periods, allows us to optimize the aerosol model for the region of interest.

To compare the effects of the temporal extension and spatially more dense measurements, the integrated AOPs for each case are presented in Table 2. In the table, AOPs considered to calculate LUT for the MAM (March, April, and May) season were listed for each AOD bin in order of SSA, refractive index, effective radius and standard deviation of volume size distribution, and the number of integrated data. To consider the change in AOP with respect to AOD suggested by Levy et al. (2007a), the AOPs were categorized into six AOD bins. The bins are categorized by 0.0–0.3, 0.3–0.6, 0.6–1.0, 1.0–1.4, 1.4–1.8, and 1.8–3.0, and the median values of each AOD bin are shown in Table 2. Though AERONET inversion data provide four spectral SSAs at 440, 675, 870, and 1020 nm, the values at 675 nm were analyzed considering the spectral range of the MI visible channel. For the LUT calculation, however, wavelength dependence of the refractive index was obtained from the AERONET retrieval and applied. Based on the wavelength dependence, the AOD was retrieved at 550 nm. The total average and standard deviation of the SSA for the original group (Table 2a) was 0.92 and 0.035, respectively. The SSA ranged between 0.911 and 0.925 in order of AOD. Accordingly, the real part of the refractive index showed positive correlation with the AOD. The increase of AOD caused the increase of effective radius and standard deviation of fine-mode size distribution, too. With the quality criteria of the inversion products, the number of data points was significantly low for the low AOD bin. The number of data was also decreased with the increasing AOD. In Table 2b, the AOPs obtained from the temporally extended data sets from the same sites were listed. A slight increase of the effective radius for coarse-mode particles was found for the low AOD cases in accordance with the increase of the number of data. When the data set from the DRAGON-NE Asia campaign, and a few additional sites in China not included in the original study, were applied, all AOD bins showed an increase in SSA by more than 0.005, and the average value was 0.93 ± 0.035. The larger data set resulted in an increase in SSA by about 1 %, though the variation is lower than the standard deviation of SSA. The increase in SSA may also be due to a temporal change in SSA which was suggested in Lyapustin et al. (2011a). The previous study showed increases in SSA in eastern China from 2000 to 2010 by about 0.02 at 470 nm. The imaginary part of the refractive index was generally decreased, and the decrease was more significant for low AOD condition than high AOD conditions. Meanwhile, the increasing effective radius of coarse particles was also found. Figure 5 shows the volume size distribution analyzed from the original (Fig. 5a) and the new data (Fig. 5b) group for each AOD bin. In general, the coarse-mode particles of a bimodal log-normal size distribution tend to dominate due to sporadic dust events (e.g., Lee et al., 2010). With the increase in AOD, the mode radius of fine particles is increased, while that of coarse particles is decreased (Levy et al., 2007a). The effective radius and standard deviation values for fine and coarse mode are listed in Table 2a and c.

Using aerosol models derived from both the original and new data sets, LUTs were calculated by using the 6SV (Second Simulation of a Satellite Signal in the Solar Spectrum–Vector) RTM (Vermote et al., 1997; Kotchenova et al., 2006; Kotchenova and Vermote, 2007). In addition to measurement geometry (i.e., solar zenith angle, viewing zenith angle, and relative azimuth angle), the surface reflectance, aerosol model, and AOD were provided as input variables to calculate the LUTs. Surface elevation was also included to increase the accuracy of Rayleigh scattering correction.

As mentioned above, the AOD is retrieved by comparing measured and calculated TOA reflectance for a given set of measurement conditions. The values in the LUTs were linearly interpolated with the values in the neighboring bins because the calculation of TOA reflectance is performed as a function of several input variables. To test the effects of the changes in aerosol models, the AODs were respectively derived by using the original and the new LUTs.

To estimate the accuracy of retrievals from the inversion of the single-channel algorithm, and to understand its sensitivity to uncertainty in the assumed SSA, a reference test was performed. In this test, the TOA reflectance was analyzed within a ±4% variation in SSA relative to the reference condition, from simulations using the RTM for four different reference conditions of both AOD and SSA with assumed geometries. The 4 % variation covers the standard deviation of 0.035 for the integrated SSA of 0.92 mentioned in Sect. 3.3. In the simulation, the surface reflectance was assumed to be 0.05 and 0.10, and the scattering angle was varied from 135.7 to 173.2∘ with respect to the geostationary measurement conditions. The surface elevation was at sea level, and cloud-free conditions were assumed. The retrieved AOD from the simulated reflectance was then compared with the assumed reference AOD value. Because the AOD was retrieved from the simulated TOA reflectance by assuming the reference SSA, the ±4 % variation in SSA causes an error in AOD. The results for the comparison between the reference value and retrieved AODs for each simulated reflectance are shown in Fig. 6. The case with zero SSA error indicates that the assumed SSA for the retrieval was the same as the reference SSA. In other cases, the positive error in SSA indicates that the SSA used to calculate the LUT was overestimated when compared with the reference value. The errors in AOD and SSA were calculated as follows:

AOD error[%]=[(retrieved AOD-reference AOD)/referenceAOD]⋅100SSA error[%]=[(assumed SSA-reference SSA)/reference SSA]⋅100. Strong negative correlation was found between the errors in SSA and AOD. The error in SSA was negatively correlated with the error in AOD, and thus the overestimation of SSA leads to an underestimation of AOD. In terms of the absolute value of AOD error, the effects of the positive and negative errors in SSA are symmetric in general, though the effect of the negative error in SSA is slightly greater. The effect of assumed errors in SSA is more significant in scenarios with higher AOD. The SSA error of ±3 % results in an AOD error of -18.70 % (-0.03, an absolute difference) and +20.34 % (+0.03), respectively, when the reference AOD is 0.15 and the surface reflectance is 0.05. The range of error is increased when the reference AOD is higher, with retrieval errors of -20.03 % (-0.24) and +23.31 % (+0.28) caused by a ±3 % SSA error when the reference AOD is 1.20.

The error in AOD also increases with the increase of assumed surface reflectance relative to true reflectance. When the surface reflectance is increased from 0.05 to 0.10, the errors in the reference AOD of 0.15 ranged between -35 % (-0.05) and 36 % (+0.05). The increase of effect of the SSA assumption was related with the one-to-one correlation between the critical reflectance and SSA reflectance (Castanho et al., 2008; Fraser and Kaufman, 1985). Whereas the increase in aerosol contributes to the increase of TOA reflectance over a dark surface, the increase in AOD reduces the TOA reflectance by shielding the upwelling reflectance from a bright surface. There exists, therefore, a surface reflectance at which the positive and negative contributions of aerosol to the TOA reflectance are canceled out; then the surface reflectance is known as the critical reflectance. In consideration of the positive relationship between the critical reflectance and SSA, the sensitivity to the SSA assumption of the AOD retrieval can be increased near the critical reflectance.

Various uncertainties result in an error in retrieved AOD as the algorithm is based on a single channel, where most dominant uncertainties come from estimating the surface reflectance and assumed aerosol model. To investigate the retrieval error, several sensitivity tests were conducted. The effects of linear inversion error, assumptions of BAOD, the aerosol model, and surface elevation were estimated in a quantitative manner in addition to the aerosol model error shown in Fig. 6.

The LUT approach has been widely used to obtain aerosol information from satellite measurements by reducing the operation time. In the LUT approach, the calculated value is interpolated linearly from the neighboring bins for geometry, AOD, surface reflectance, and elevation. Thus, the number of entries for LUT calculation must be selected carefully to save operation time and maintain retrieval accuracy at the same time. LUT applied in this study presents TOA reflectance calculated as a function of geometrical angles of sun and satellite with 10∘ interval, and surface reflectance with 0.1 intervals. As long as the LUT approach is applied to the retrieval algorithm, the linear interpolation of TOA reflectance between each bin leads to the inversion error. Figure 7a and b show the percentage difference between retrieved and reference AODs in terms of scattering angle, surface reflectance, and AOD condition. Two different AODs of 0.15 and 1.20 were applied to calculate the reference reflectance with two surface reflectances of 0.05 and 0.10, and solar zenith angles ranging from 0 to 57∘ by intervals of 3∘. The satellite zenith and azimuth angle were assumed to be 10 and 40∘, respectively. In Fig. 7a and b, the percentage errors increase by increasing difference between the reference condition and LUT bin in terms of both scattering angle and surface reflectance. The inversion error varied from 0 to 8 %, which mainly increased with the increase of scattering angle, and decreased with the increase of AOD. In the figure, the solid lines represent the inversion error which has arisen solely by the angle interpolation in intervals of 0.1 for the surface reflectance in LUT. The dashed lines representing the inversion error for the surface reflectance of 0.05 show that the assumption about linearity between bins of surface reflectance increased the error negatively.

In the estimation of surface reflectance, the BAOD correction was applied to consider continuous emission of air pollutant over East Asia. However, the BAOD estimated from MODIS products contains retrieval uncertainty of the dark target algorithm. As mentioned above, the expected error range of MODIS AOD is ±(0.05 + 15 %). The BAOD is very low in general, and thus the expected error range can be over ±100 % when the BAOD is lower than 0.05. According to a sensitivity test, the ±100 % error in the BAOD of 0.05 led to a 7 % error in surface reflectance of 0.05 and an 11 % error in AOD of 0.45. The effects of BAOD error in surface reflectance and AOD are shown in Fig. 7c and d, respectively, under the conditions of BAOD of 0.15, three surface reflectances of 0.05, 0.10, and 0.15, and three AODs of 0.45, 0.80, and 1.20. In general, the underestimation of the BAOD leads to the overestimation of the surface reflectance. The -100 % error in the BAOD assumption caused 5.6 % overestimation of surface reflectance when the surface reflectance was 0.1. Meanwhile, the 5 % error in surface reflectance led to 25.56 % underestimation of the AOD when the reference AOD and surface reflectance were 0.45 and 0.10, respectively. The uncertainty decreased with the increase of surface reflectance, and the sensitivity to the error in surface reflectance was more significant for the low AOD condition than the high AOD. In this test, the inversion error was avoided by using reference reflectance calculated under the condition of LUT bins.

Lastly, the effect of assumption in surface elevation was analyzed, as shown in Fig. 7e. The assumption of surface elevation is linked with the Rayleigh scattering correction. The underestimation of surface elevation leads to the overestimation of atmospheric pressure, and thus the over-correction of the Rayleigh scattering, which eventually results in the overestimation of surface reflectance, and thus the underestimation of the AOD. The sensitivity was tested for an elevation of 1 km, two AODs of 0.15 and 0.80, and a surface reflectance of 0.10. The ±0.5 km errors in surface elevation resulted in +9.63 % and -10.56 % errors in AOD when the reference condition was assumed to be an AOD of 0.15. The increasing AOD significantly reduced sensitivity to the uncertainty, and a ±0.5 km error led to +1.30 and -1.43 % when the AOD was 0.80. The dependence on surface reflectance and elevation was not significant.

From the uncertainty tests, the largest uncertainty was found in the aerosol model assumption by about 30 % although the effect of each uncertainty was changed by conditions of AOD, surface reflectance, and sun-satellite geometry.

The greatest advantage of geostationary measurements is the availability of more cloud-free observations by continuous measurements at high temporal resolution. Besides, the AOD derived from geostationary satellite measurements can minimize the uncertainty caused by the different and limited sampling of polar-orbiting satellites in the trend estimation (Yoon et al., 2014). Figure 8 shows examples of retrieved AOD from the geostationary measurements from MI, using the single-channel algorithm. The red, green, and blue images, obtained from GOCI onboard the same platform measured at 01:16, 02:16, 03:16, 04:16, 05:16, and 06:16 UTC on 27 April 2012, show dust flow from the Shandong peninsula to the northern Korean peninsula. Similarly, the images of retrieved AOD show values greater than 1.0 in the dust plume, in contrast to the values lower than 0.4 over other regions. Compared with the MODIS AOD, the retrieved AODs over dusty regions are generally higher, though the distribution of MI AOD is spatially well matched over non-dusty regions. The spatially averaged value of the MI AOD in dusty regions [110–125∘ E, 35–40∘ N] decreased steadily from 2.67 at 00:00 UTC to 1.69 at 07:00 UTC, and the minimum value of 1.43 was found at 03:30 UTC. Meanwhile, the spatial mean values of AOD obtained from the MODIS TERRA and AQUA measurements were 1.11 at 03:55 UTC and 1.18 at 05:15 UTC respectively. In Fig. 8, the AOD images of TERRA and AQUA represent the measurements between 00:00 and 05:00 UTC, and between 02:00 and 06:00 UTC, respectively.

The results from MI also show the transport and concentration of aerosol in 30 min intervals, while the MODIS product can provide only two images per day. The map of MI AOD in hourly intervals shows that the high concentration of aerosol was mostly observed over northern China and the Yellow Sea before 03:00 UTC, with the dust plume extending to the East Sea across the northern Korean peninsula. We can deduce from the change in the dust plume that the wind field changed straight flow from southwest to northeast in the morning to a wave pattern, following a low pressure system located in Manchuria. Neither the dark target algorithm of MODIS nor the single-channel algorithm of MI could retrieve AOD over regions of brighter surfaces, due to the low sensitivity of the aerosol compared with the surface. However, unlike the MI retrieval, part of the dust scene over the ocean was missed in the MODIS retrieval due to sun glint masking.

For quantitative validation, the retrieved AODs were compared with the measured values from the 39 AERONET sun photometer sites in Korea and Japan. To investigate the effect of the new aerosol model as an input parameter to calculate the LUTs, the results of the original and new AOD retrievals were compared, and the comparisons are shown in Fig. 9. The measured AODs from all of the numbered DRAGON-NE Asia sites listed in Table 1 were used in the comparison shown in the top panel. In the lower panel, part of the AERONET AOD used as a validation group to test the consistency of the algorithm and to validate the retrieval accuracy is shown. The data from the validation group were not included in the AOP analysis due to a lack of inversion data sets. The comparison results are shown in the bottom panel of Fig. 9. The left and right panels show evaluations of the original and new AOD, respectively.

Using the original aerosol model, the retrieved AODs agree very well with the linear regression as follows: τMI[original LUT]=1.08τDRAGON-Asia-0.08,RMSE=0.18,r=0.87. Although the Pearson coefficient of 0.87 indicates a significant correlation, the regression slope indicates that the retrieved AOD is overestimated by 8 % compared with the AERONET value. Comparison with the validation group, however, shows a tendency to systematic underestimation, with a slope of 1.01 and y-offset of -0.05.

By applying the new aerosol model, the regression slope was improved to 1.00, although other measures remained similar: τMI[new LUT]=1.00τDRAGON-Asia-0.07,RMSE=0.17,r=0.85. In Sect. 3.4, the analysis of the retrieval sensitivity to the SSA assumption showed that the underestimation of the SSA in the aerosol model results in the overestimation of AOD. Thus, the overestimation of the original AOD suggests that the radiative absorptivity of the aerosol during MAM was slightly underestimated prior to the campaign. According to Fig. 6, a 1 % underestimation of SSA can result in an overestimation of AOD by up to 7 %. The uncertainty can vary with measurement geometry, AOD, or surface reflectance. Therefore, to a large degree, the 8 % decrease in AOD can be explained by a 1.1 % increase in SSA in the new aerosol model during MAM. The large RMSE and the underestimation for the validation group, however, are attributed to the spatial and temporal variation in AOPs, which cannot be standardized by the single aerosol model. Moreover, the change of aerosol model results in a decrease of percentage of the comparison data within 30 % difference range from 79.15 to 77.30 %. In terms of the comparison of the validation group, the regression slope decreased from 1.01 to 0.93, though the comparison still shows strong correlation between the retrieved and measured AOD. As long as a single aerosol model is applied, the spatial and temporal variations of aerosol properties are the largest uncertainty of the AOD retrieval algorithm. When the difference between assumed and actual SSA becomes higher than 3 %, the retrieval error exceed 30 %. The degradation of the comparison statistics shows the limitation of the single-channel algorithm. The uncertainties in estimation of surface reflectance and assumption of linearity between LUT bins have effects on the accuracy of low AOD as described in Sect. 3.5. The sensitivity tests showed that the effects of each retrieval uncertainty depend on the condition of AOD. For the condition of low AOD, the effect of the aerosol model assumption on the retrieval uncertainty in AOD is significantly lower than the effects of surface reflectance estimation. However, an insufficient number of inversion data for an AOD bin between 0.0 and 0.3, where the AOD is lower than the criteria of quality assurance of 0.4 (440 nm), increases the uncertainty in the assumption of the aerosol model for the condition of low AOD. Consequently, it was found that the validation statistics for low AOD (< 0.4 at 550 nm) were significantly lower than that for high AOD. While the correlation coefficient and regression slope of the low AOD comparison were 0.49 and 0.35, those for the high AOD condition were 0.78 and 0.86. The ratio of the low AOD to the total comparison data set was 41.72 %. To show the retrieval accuracy for each campaign site, the Taylor diagram (Taylor, 2001) is shown in Fig. 10. This diagram summarizes how closely a set of retrievals matches observations in terms of r, RMSE, and standard deviation. The polar angle of the point from the x axis indicates the correlation coefficient, and the radial distance represents the normalized standard deviation, which in this case describes the ratio of the standard deviation of the retrieved MI AOD to that of the AERONET (Yoon et al., 2014) values. The distance between the symbol and the dashed arc, which represents the standard deviation of the AERONET value, shows the similarity of the amplitude of their variations; a radial distance of > 1 indicates that the standard deviation of the MI AOD is greater than that of AERONET. On the other hand, the RMSE between the MI and AERONET AODs is proportional to the distance to the point on the x axis identified as “AERONET”, marked with a dotted arc. Consequently, the decrease in distance between the “AERONET” point and the position of the symbol indicates an increase in similarity between the retrieved and measured AODs. The normalized standard deviations of retrieved AOD generally range from 1 to 1.5, except for the Kohriyama (site number 12) and Matsue (site number 19) in Japan. In spite of the high correlation coefficients of 0.85 and 0.78 at the sites, the high regression slopes of 1.58 and 1.35 suggest that the radiative absorptivity was underestimated in this region, and thus the AOD was significantly overestimated in the case of high-AOD conditions. The large negative y-intercepts of -0.12 and -0.25 could be caused by the underestimation of AOD following an overestimation of BAOD in the case of low-AOD conditions.

The comparison statistics of the original and new AOD, plotted in the Taylor diagram, are also listed in Tables 3 and 4, respectively. The correlation coefficients obtained from the 39 DRAGON sites range from 0.66 to 0.95 and the average was 0.84 when the original aerosol model was applied. The maximum value was found at Anmyeon (site number 3) and Kunsan_NU (National University) in Korea, and the minimum value of 0.66 was found at Nishi-Harima (site number 25) in Japan. The Anmyeon site was located in a rural area near the ocean to monitor background conditions of the atmosphere (e.g., Kim et al., 2007), and thus the dark surface contributes to a reduction in the uncertainty in AOD retrieval. The Kunsan-NU site, as with the Anmyeon site, was surrounded by mountains, reservoirs, and an area that is rural. Meanwhile, the Nishi-Harima site was located on the top of Mount Onade (435.9 m altitude, Nishi-Harima Astronomical Observatory) among trees, and thus the uncertainty caused during surface correction can also be reduced. However, the comparison statistics showed systematic underestimation of the AODs by a regression slope of 0.86 and a y-intercept of -0.06. To compare the difference between the AOD correlations for each sites, temporal variation of the AODs obtained from MI and AERONET measurements were represented in Fig. 11. In Fig. 11, the AOD variations for the four aforementioned sites are shown in order of the (a) Anmyeon, (b) Kunsan_NU, (c) Kohriyama, and (d) Nishi-Harima sites. The red boxes and black circles, which indicate the MI AOD and the AERONET values, were well matched at (a) Anmyeon and (b) Kunsan-NU with good correlation statistics. The vertical distribution of symbols for each day represents diurnal variation of AOD, and the variations were also highly correlated regardless of time. The temporal variations showed an increase in AOD during the period from 1 to 15 May at both sites. In two other sites in Japan, Kohriyama and Nishi-Harima, no temporal pattern can be found because of the low number of comparison data, though the variation of MI AOD was closely related with the AERONET value. A notable feature in the comparison was the low number of data. Table 3 showed that most Japanese sites (except Fukue) have a lower number of comparison data than the Korean sites, and the low number trend was related with the frequency of the direct measurements by the sun photometer in Japanese sites. While the total number of direct AOD products in the level 2.0 data set ranged between 99 and 3630 in Japan, the number ranged from 1296 to 5191 in Korea. The difference in data counts indicates that there were frequent rain and cloud events over Japan, resulting in uncertainty in the AOD retrieval in Japan, including the Koriyama and Nishi-Harima sites. However, a reason for the significant underestimation trend of the MI AOD at Nishi-Harima is not clear yet.

Excluding the Fukue_2 site, which has a low number of comparison data of only 4, the regression slopes at 32 AERONET sites were higher than 1.0, and the values at 9 sites exceeded 1.2. As well as the Kohriyama and the Matsue sites, the comparison results for all but three sites (2, 30, and 32) show negative y-intercepts between -0.003 and -0.25. As with the improved correlation seen in the scatter plot, the Taylor diagram and regression statistics listed in Table 4 also show improvements in retrieval accuracy at each site. The distances between the data point and the AERONET value at each site were generally reduced, especially at Tsukuba (site number 32). At this site, the systematic overestimation was significantly reduced by applying the new aerosol model, also leading to an improved correlation coefficient. The regression slope over all sites decreased by about 0.08, while the y-intercept changed within a range from -0.03 to 0.06, in accordance with the increased SSA in the new aerosol model. Whereas most of the comparisons were improved by the decrease in the slope, some sites (11, 21, 25, 26, 28, and 36) show a better result using the original aerosol model in terms of the regression slope. The change in correlation coefficient and RMSE was not significant.