Validation of MODIS C6 Dark Target Aerosol Products at 3 km and 10 km Spatial Resolutions Over the China Seas and the Eastern Indian Ocean

: In this study, MODerate resolution Imaging Spectroradiometer (MODIS) Collection 6 (C6) level-2 Dark Target (DT) Aerosol Optical Depth (AOD) products at 3 km (DT3K) and 10 km (DT10K) spatial resolutions were validated over the China seas and the eastern Indian Ocean against Maritime Aerosol Network (MAN) Level 1.5 AOD measurements collected through 13 cruises from 2010 to 2014. For this, DT3K and DT10K AOD observations were obtained from four Scientiﬁc Data Sets (SDS), i.e., “Effective_Optical_Depth_Average_Ocean” (EODAO AOD ), “Effective_Optical_Depth_Best_Ocean” (EODBO AOD ), “Image_Optical_Depth_Land_And_Ocean” (IODLAO AOD ) and “Optical_Depth_Land_And_Ocean” (ODLAO AOD ). The MAN AOD measurements were ﬁltered within (i) ± 2 h, (ii) ± 4 h, (iii) ± 6 h, and (iv) ± 12 h of MODIS overpass time. Results showed that the DT10K and DT3K performed equally over the China seas and the eastern Indian Ocean in terms of retrievals quality and agreement with the MAN AOD measurements, whereas the DT3K has less coincident observations than the DT10K. For seasonal analysis, larger underestimation in the DT algorithm was observed in autumn followed by spring, whereas retrievals were well correlated with the MAN AOD data in summer. Overall, this study found that ODLAO AOD observations for the DT3K and DT10K were much better than EODAO AOD, EODBO AOD and IODLAO AOD in terms of high correlation and a large percentage of the AOD retrievals within the Expected Error (EE = +(0.04 + 10%), − (0.02 + 10%)).


Introduction
Aerosols are important components of the Earth's climate system and play an important role in the global energy budget [1], perturbing the hydrological cycle [2], and reducing atmospheric visibility [3], and in large concentrations are harmful to human health [4]. The particles travel from one place to another and accumulate in the atmosphere over periods of days and weeks [5]. Aerosol effects on weather and air quality are uncertain, due to lack of understanding of spatiotemporal variations in their optical properties. Therefore, characterizing aerosol spatial variations and distributions over time are necessary for understanding the present and possible future climate conditions. For this, ground-based Sun photometers such as Aerosol Robotic Network (AERONET) [6,7] over land and Maritime Aerosol Network (MAN) [8,9] over the ocean are used to provide multispectral aerosol properties including aerosol optical depth (AOD) at high temporal resolution but over limited locations.

Study Area
The combined waterbody of China seas and the eastern Indian Ocean, together, with a maximum depth greater than 5000 m, is the largest marginal sea in Southeast Asia. It extends from the equator to 23 • N and from 99 • E to 121 • E and joins the Pacific Ocean through the Luzon Strait between the Taiwan Island and the Luzon Island [55]. It is bounded in the north by the passive south China continental margin. To the south, small continental blocks that were drifted off mainland Asia, separate the basin from an extinct subduction zone along Palawan and northwest Borneo. The combined waterbody of China seas and the eastern Indian Ocean has an area of 3.3 million km 2 excluding the Gulfs of Thailand and Tonkin [56]. There are great differences in water type and quality over the region as turbidity of the coastal water is higher than deep ocean water [57].

MAN AOD
The Maritime Aerosol Network constitutes a component of the AERONET is affiliated with the AERONET calibration and data processing standards and procedures [7]. The proposed activity involved handheld Sun photometer measurements from various ships (scientific and non-scientific) and will complement island-based AERONET measurements, thus extending data collection to the vast regions where no islands exist [8]. The direct Sun measurements are acquired in five spectral channels within the spectral range between 340 nm and 1020 nm. The estimated uncertainty of the optical depth in each channel does not exceed ±0.02 [58], primarily due to inter-calibration against AERONET CIMEL instruments which are more accurate, i.e., uncertainty within 0.01 in the visible and near-infrared channels [59]. Thus, MAN provides high-quality AOD measurements with known uncertainty. In this study, MAN L1.5 AOD data from 13 different cruises (represented by different color symbols in Figure 1, and Table 1) are obtained for validation of the DT3K and DT10K AOD products.

MODIS Dataset
In this study, Aqua-MODIS C6 level-2 operational DT3K and DT10K aerosol products (MYD04) were downloaded from the "Level-1 and Atmosphere Archive and Distribution System (LAADS) Distributed Active Archive Center (DAAC)" (https://ladsweb.nascom.nasa.gov/) to obtain the DT AOD retrievals for validation. The MODIS daily level-3 Remote Sensing Reflectance (Rrs) product at 4 km resolution was downloaded from (https://oceandata.sci.gsfc.nasa.gov/MODIS-Aqua/Mapped/ Daily/4km/Rrs/) to match with MAN AOD and MODIS AOD. Table 2 gives a detailed summary of the dataset used in this study.

Methods
In this study, Aqua-MODIS C6 DT3K and DT10K aerosol products were obtained for the China seas and the eastern Indian Ocean according to the time period of available ground data collected by 13 different cruises from 2010 to 2014. Only those DT3K and DT10K AOD retrievals at 0.55 µm passing recommended quality flags (QF) checks [41] were used for validation. The DT3K and DT10K AOD retrievals were obtained from the SDS (i) "Effective_Optical_Depth_Average_Ocean", i.e., retrieved EODAO AOD over water for "average" solution at seven channels; (ii) "Effective_Optical_Depth_Best_Ocean", i.e., retrieved EODBO AOD over water for "best" solution at seven channels; (iii) "Image_Optical_Depth_Land_And_Ocean", i.e., retrieved IODLAO AOD over water for "average" solution (QF = 1, 2, 3) and over land for all quality flags (QF = 0, 1, 2, 3) at 0. 55 µm; and (iv) "Optical_Depth_Land_And_Ocean", i.e., retrieved ODLAO AOD over water for "average" solution (QF = 1, 2, 3) and over land for "best" (QF = 3). More information about each SDS can be found in [41]. The AOD observations from each SDS were used same as available except for the SDS "Optical_Depth_Land_And_Ocean", for which ODLAO AOD was filtered for the highest quality assurance flag (QF = 3) using the SDS "Land_Ocean_Quality_Flag". In other words, the DT3K and DT10K AOD observations were used from "average" to "best" solutions for validation. As MAN does not provide AOD measurements at 0.55 µm, the MAN AOD was interpolated to 0.55 µm using Ångström exponent (α) (Equations (1) and (2)) to match with the MODIS AOD: where, α is Ångström exponent, and λ is the wavelength.
To consider the atmospheric variability and increase the number of samples, the MAN AOD measurements were considered within (i) ±2 h; (ii) ±4 h; (iii) ±6 h; and (iv) ±12 h of MODIS over pass time for each sampling site, and an average of at least two pixels of MODIS AOD observations within a sampling window of 3 × 3 pixels (average of 9 pixels) centered on the MAN AOD location, i.e., average of 9 km × 9 km region for the DT3K, and 30 km × 30 km region for the DT10K, was considered. The coincident AOD observations for EODAO AOD were 397 for DT3K and 697 for DT10K, for EODBO AOD were 397 for DT3K and 697 for DT10K , for IODLAO AOD were 419 for DT3K and 730 for DT10K, and for ODLAO AOD were 272 for DT3K and 512 for DT10K. Unavailability of the satellite AOD observations corresponding to MAN AOD was due to the limitation of the aerosol algorithm, data collection gaps and cloud cover [43].
Four statistical indicators such as correlation coefficient (R), root mean square error (RMSE), mean absolute error (MAE), and expected error (EE) of AOD were used to evaluate MODIS AOD against MAN AOD. The correlation coefficient (R) is a dimensionless index, which is invariant to linear transformations of either variable [60]. It indicates agreement between the MODIS AOD and the MAN AOD, and the higher values the better agreement. In this study, R is obtained from Deming Regression (DR) instead of that from Linear Regression (LR) because DR estimates an unbiased slope by assuming the Gaussian distribution of errors in both x and y data points (which is typical of our data) [61][62][63]. While LR estimates a biased slope by assuming random measurement errors in the dependent variable (y) and an error-free independent variable (x) [61,[64][65][66], but is inappropriate for use when significant errors are expected in both variables.
The root mean square error (RMSE) was used to measure the differences between MODIS and MAN AOD which is sensitive to both systematic and random errors. The equation to calculate RMSE is as follow: The mean absolute error (MAE) is the most natural measure of mean error magnitude [67] and calculated as Equation (4): Expected error (EE) is estimated for the uncertainty of MODIS AOD over ocean [41].
The MODIS AOD retrievals were reported as a good quality if they fall within the following envelope (Equation (6)).  [36] which might be due to large errors in the estimated surface reflectance and aerosol schemes used in look-up-table (LUT) for aerosol retrievals [43]. The performance of the ODLA AOD retrievals was much better than the other SDS in terms of the agreement with the MAN AOD measurements, a large percentage of retrievals within the EE, and small RMSE. The good agreement indicates that the DT10K AOD retrievals followed the actual variations in aerosol concentrations as measured by MAN AOD [68,69]. Overall, the performance of the DT10K algorithm was not satisfactory in terms of data quality as maximum 61% of the retrievals for ±4 h were within the EE, which is less than one standard deviation confidence interval, i.e., about 66% of points should fall within ±EE from the true AOD, which indicates that the DT10K ocean algorithm does not meet the requirements of the EE over the China seas and the eastern Indian Ocean. MAN AOD measurements were not available for winter seasons, therefore seasonal validation of the DT10K AOD observations was conducted for spring, summer, and autumn seasons (Figure 3) using AOD observations obtained from four different SDS for ±12 h of the satellite overpass as for this time intervals, DT10K has large numbers of coincident observations. Larger and small numbers of coincident observations were available for spring and summer, respectively. DT AOD was underestimated for high aerosol loadings during spring and autumn, whereas overestimated during summer. This was might be due to the regional meteorology and dust aerosol generation and transport for each season [70]. The quality of the EODAO AOD (Figure 3a-c), EODBO AOD (Figure 3d-f), and IODLAO AOD (Figure 3g-i) retrievals were good in spring than summer and autumn, whereas agreement with MAN AOD was not good. All these of AOD retrievals do not meet the requirements of the EE for each season as the percentage within the EE was less than 66%. Good quality and agreement with MAN AOD were observed for ODLAO AOD observations (Figure 3j-l) in summer as the correlation was 0.856 and 75% of the retrievals were within the EE, although only 51 coincident observations were available.

Validation of DT AOD at 3 km Resolution
As with the DT10K, validation of the DT3K EODAO AOD (Figure 4a-d), EODBO AOD (Figure 4e-h), IODLAO AOD (Figure 3i-l), and ODLAO AOD observations (Figure 4m-p) was conducted against MAN AOD measurements over four different time intervals. Similar trends as DT10K was observed in several coincident observations, RMSE, correlation and the percentage of retrievals within the EE. However, the DT3K has a small number of coincident observations than the DT10K for each time interval. For different time intervals, validation showed that the retrieved AOD observations were well correlated with the MAN AOD measurements for ±2 h, shown small RMSE for ±12 h, and a large percentage of retrievals within the EE for ±4 h. The DT3K has large underestimation compared to the DT10K during both low to high pollution episodes which might be caused by an error in the surface reflectance [42,44,45]. The performance of the ODLA AOD retrievals was much better than the other SDS in terms of the agreement with the MAN AOD measurements, a large percentage of retrievals within the EE, and small RMSE. However, the DT3K did not perform well in terms of data quality as maximum 60% of the retrievals for ±4 h were within the EE, which is less than one standard deviation confidence interval, i.e., about 66% of points should fall within ±EE from the true AOD, which indicates that the DT3K ocean algorithm resolution does not meet the requirements of the EE over the South China Sea. As with the DT10K algorithm, a large number of coincident observations were available in spring, and large underestimation and RMSE were observed in autumn for each AOD dataset ( Figure 5). The DT3K algorithm has less number of coincident observations than the DT10K algorithm but performed well ( Figure 5). DT has good quality retrievals, but poorly correlated with MAN measurements during spring than autumn. However, ODLAO AOD observations were well correlated with MAN AOD with large percentage within the EE and small RMSE during the summer season. These are similar findings as the DT10K algorithm.

Comparison between DT10K and DT3K
DT3K EODAO AOD (Figure 6a), EODBO AOD (Figure 6b), IODLAO AOD (Figure 6c) and ODLAO AOD (Figure 6d) observations were compared with the DT10K observations for each cruise as shown in Figure 6. Many coincident observations were available for IODLAO AOD , and a small number of observations were available for ODLAO AOD which contains only highest quality flag AOD observations. All the parameters of DT3K and DT10K were well correlated with each other with R of 0.94 to 0.99, the slope of 0.92 to 1.02, and RMSE of 0.015 to 0.046. Overall, ODLAO AOD was robust and performed well with highest R of 0.99, slope equal to 1.0 and smallest RMSE 0.015 compared to the other parameters.

Error Analysis
To analyze the errors (MODIS-MAN) in the DT3K and DT10K with respect to seasons over water surfaces, MODIS remote sensing reflectance, which was converted to normalized water reflectance (Rs = Rrs × π), was obtained corresponding to each AOD values. Results show that errors in the DT10K (Figure 7) have a less negative correlation with normalized water reflectance than the errors in the DT3K (Figure 8), especially for EODAO AOD (Figures 7a and 8a), EODBO AOD (Figures 7b  and 8b), and IODLAO AOD (Figures 7c and 8c) observations. These results indicate that the DT3K mostly underestimates AOD over turbid waters where normalized water reflectances are high and overestimates AOD over clear water where normalized water reflectances are relatively low compared to the DT10K. For example, significant underestimation was observed in spring for high values of normalized water reflectance, and this was might be due to an error in the estimated surface reflectance used in the DT algorithm [43]. This underestimation in the DT3K AOD retrievals was more prominent because some pixels might be retained in the DT3K which are deselected by the DT10K during threshold criteria [44]. It is worth mentioning that ODLAO AOD (Figures 7d and 8d) has a weak correlation with the normalized water reflectance, and the DT algorithm for this parameter is less sensitive to the water quality and under-and over-estimations do not follow the increasing and decreasing trend of normalized water reflectance. These results showed that the ODLAO AOD observations for the highest quality flag are less sensitive to water types and suitable for accurate aerosol monitoring over the China seas and the eastern Indian Ocean compared to the other parameters.

Conclusions
This study evaluated the MODIS C6 aerosol products based on the DT algorithm at 3 km (DT3K) and 10 km (DT10K) resolutions over the China seas and the eastern Indian Ocean. For this, EODAO AOD , EODBO AOD , IODLAO AOD , and ODLAO AOD observations were obtained from different Scientific Data Sets, i.e., "Effective_Optical_Depth_Average_Ocean", "Effective_Optical_Depth_Best_Ocean", "Image_Optical_Depth_Land_And_Ocean" and "Optical_Depth_Land_And_Ocean", respectively. For validation, level 1.5 AOD measurements were obtained from the Maritime Aerosol Network (MAN) which were collected using 13 cruises from 2010 to 2014 over the China seas and the eastern Indian Ocean. The MAN AOD data were filtered within (i) ±2 h; (ii) ±4 h; (iii) ±6 h; and (iv) ±12 h of MODIS overpass. This study found that The DT3K and DT10K algorithm performed equally over the China seas and the eastern Indian Ocean in terms of the agreement with the MAN AOD and percentage of the retrievals within the Expected Error (EE = +(0.04 + 10%), −(0.02 + 10%)), i.
the DT3K coincident AOD observations were less than the DT10K, ii.
a large number of incident observations were available within ±12 h, small RMSE was observed within ±2 h, and a large percentage of the retrievals within the EE was observed within ±4 h, iii.
both the DT3K and DT10K extremely underestimates over water surfaces, and large underestimation being observed during autumn by summer, iv.
the algorithm performed well during summer, but it has fewer numbers of coincident observations for the both DT3K and DT10K, v.
ODLAO AOD observations from the DT3K and DT10K were found better and suitable for use over the China seas and the eastern Indian Ocean compared to the EODAO AOD , EODBO AOD , and IODLAO AOD in terms of high correlation with the MAN AOD data and large percentage within the EE, vi.
ODLAO AOD observations were less sensitive to the variations in normalized water reflectance, and vii.
overall, this study found that the DT10K and DT3K AOD retrievals do not meet the requirements of the EE as the percentage within the EE was less than 68%.