A New MODIS C 6 Dark Target and Deep Blue Merged Aerosol Product on a 3 km Spatial Grid

In Moderate Resolution Imaging Spectroradiometer (MODIS) Collection (C6) aerosol products, the Dark Target (DT) and Deep Blue (DB) algorithms provide aerosol optical depth (AOD) observations at 3 km (DT3K) and 10 km (DT10K), and at 10 km resolution (DB10K), respectively. In this study, the DB10K is resampled to 3 km grid (DB3K) using the nearest neighbor interpolation technique and merged with DT3K to generate a new DT and DB merged aerosol product (DTB3K) on a 3 km grid using Simplified Merge Scheme (SMS). The goal is to supplement DB10K with high-resolution information over dense vegetation regions where DT3K is susceptible to error. SMS is defined as “an average of the DT3K and DB3K AOD retrievals or the available one with the highest quality flag”. The DT3K and DTB3K AOD retrievals are validated from 2008 to 2012 against cloud-screened and quality-assured AOD from 19 AERONET sites located in Europe. Results show that the percentage of DTB3K retrievals within the expected error (EE = ± (0.05 + 20%)) and data counts are increased by 40% and 11%, respectively, and the root mean square error and the mean bias are decreased by 26% and 54%, respectively, compared to the DT3K retrievals. These results suggest that the DTB3K product is a robust improvement over DT3K alone, and can be used operationally for air quality and climate-related studies as a high-resolution supplement to the current MODIS product suite.

The Moderate Resolution Imaging Spectroradiometer (MODIS) sensors onboard the Terra and Aqua satellites provide geophysical observations at 36 channels ranging from 0.4 to 14.4 µm with a temporal resolution of 1-2 days and spatial resolution of 250 m, 500 m, and 1000 m.In the MODIS Collection 5.1 (C5.1) level-2 operational aerosol product, daily AOD observations at 10 km resolution are available over dark surfaces from the Dark Target (DT 10K ) land algorithm [13,27,28], over ocean surfaces from the DT ocean algorithm [13,29], and over bright surfaces from the Deep Blue (DB 10K ) algorithm [16,30,31].These AOD observations are unable to resolve many local-level aerosol features due to their inherently coarse resolution.Therefore, the DT AOD product at 3 km resolution (DT 3K ) is introduced in the Collection 6 (C6) operational AOD product [32], as a supplement to the DT 10K [13] and DB 10K [16] AOD products.DT 3K is generated using the same inversion method as used in DT 10K , the only difference being in the selection of the dark target pixels [32].
For the development of the DT 3K algorithm over land [13,32], dark target pixels are selected using the top-of-atmosphere (TOA) reflectance between 0.01 and 0.25 in the 2.11 µm channel.Then, selected pixels are organized into retrieval windows of 6 pixels × 6 pixels (36 pixels) for subsequent aerosol retrievals.Pixels in the retrieval windows are masked for clouds, snow/ice, and other bright surfaces, and separated by land and water pixels.From the remaining pixels, the darkest 20% and brightest 50% in the retrieval window are discarded using the 0.66 µm channel with, at most, 11 pixels in the retrieval window being required to perform aerosol retrievals.In this process, pixels retained at 3 km resolution might be discarded at 10 km resolution.With fewer pixels contributing to the DT 3K retrieval, it yields a noisier product than the DT 10K retrieval [13,32].The DT 3K product has been validated over several regions and exhibits larger errors than the DT 10K product due to underestimation of the estimated surface reflectance and incorrect use of the available "look-up" aerosol models [13,[32][33][34][35][36].The expected error (EE) of the DT 3K over land is ±(0.05 + 20%) [13,32] which represents a one standard deviation confidence interval around the retrieved AOD (i.e., about 68% of points should fall within ±EE from the true AOD).
Initially, the MODIS DB algorithm was developed to retrieve AOD over bright surfaces [30,31].In C6, the Enhanced DB algorithm is used to retrieve AOD over both bright as well as dark surfaces [16,37,38].In developing the DB algorithm, pixels are masked for clouds and snow/ice surfaces, and surface reflectance is estimated for the remaining pixels at 0.412, 0.47, and 0.65 µm.Thus, AOD is retrieved at 1 km resolution by finding the best match between satellite TOA reflectance and pre-calculated TOA reflectance stored in a look-up-table (LUT), and then all available pixels are aggregated at 10 km resolution.The DB 10K AOD product has been validated in previous studies, which have reported better relative retrieval accuracy than the DT 10K AOD product [35,[38][39][40] with some exceptions [41].EE for Deep Blue is dependent on the viewing geometry, but is approximately 0.03 + 20% on average (i.e., the algorithms have different error characteristics).
Previous studies [33,34,36,40] have reported large uncertainty in the DT 3K AOD product at local-to-regional scales.For example, Nichol and Bilal [36] validated the DT 3K AOD retrievals over 16 AERONET sites in Asia corresponding with urban and vegetated land surfaces, and they found larger errors and overestimation in DT 3K .In addition, the DT and DB algorithms have different AOD spatial coverages over land due to differences in pixels selection criteria and their thresholds, the surface reflectance calculation method, and cloud mask.Therefore, a new product at higher resolution with low errors and more spatial coverage is preferable to understanding aerosol behavior at something approaching the level of an urban city center.
The main objective of this study is to describe and evaluate a new DT and DB-merged (DTB 3K ) AOD product on a 3 km grid to improve the quality of AOD retrievals and spatial coverage over vegetated and non-vegetated land surfaces (i.e., to retrieve AOD for those regions where the DT 3K does not retrieve AOD due to pixels selection criteria and cloud mask [13], and where DB 10K does not retrieve AOD due to errors in cloud mask that lead to removal of cloud free pixels [16,37]).This study validates DT 3K and DTB 3K AOD products over European AERONET sites located over vegetated surfaces, as the AOD product at 3 km resolution is only available for the DT algorithm which is supposed to retrieve AOD accurately over vegetated surfaces.However, the proposed product can be used over other global non-vegetated land surfaces since the product will weigh considerably more information from the DB algorithm which is designed to retrieve AOD accurately over non-vegetated surfaces.To support this hypothesis, one urban AERONET is also included in the validation experiment.Dataset and methods are described in Sections 2 and 3, respectively, and Sections 4 and 5 are about results and discussion, and conclusion, respectively.

Methods
DT 3K and merged DTB 3K AOD retrievals were validated from 2008 to 2012 against the 19 European AERONET sites.As the MODIS DT algorithm is designed to retrieve AOD over vegetated surfaces (NDVI > 0.3) [13], the AERONET sites selected for validation correspond with adjacent surfaces exhibiting NDVI values between 0.31 and 0.75, except one (Paris) with NDVI of 0.15 that is an urban site (Table 1).The methodology of this study is based on the following steps: (i) Only those DT 3K and DB 10K AOD retrievals at 0.55 µm passing recommended quality assurance (AQ) checks [13,16,37] were used (for DT, this corresponds to retrievals flagged QA = 3, and, for DB, retrievals flagged QA = 2 or QA = 3).Therefore, the DT 3K and DB 10K highest-quality retrievals were obtained from the Scientific Data Set (SDS) "Optical_Depth_Land_And_Ocean" and "Deep_Blue_Aerosol_Optical_Depth_550_Land_Best_Estimate", respectively.(ii) DB 10K AOD retrievals were resampled to 3 km spatial grid (DB 3k ) onto the DT 3K grid using the nearest neighbor interpolation algorithm [65,66] to match and overlap pixels of DB 3K with the pixels of DT 3K .As the DB algorithm first retrieves AOD at 1 km resolution, by finding the best match between satellite TOA reflectance and pre-calculated TOA reflectance stored in a LUT, all available pixels are then aggregated to 10 km resolution [16,37,38].It is expected that resampling from 10 to 3 km will not affect the accuracy and quality of the DB AOD retrievals.(iii) To reduce errors in DT 3K , the DTB 3k product is generated using the Simplified Merge Scheme (SMS) (DTB M1 in [39]).This technique is selected as it increases the number of collocations and decreases the errors, and is defined as "an average of the DT 3K and DB 3K AOD retrievals or the available one with highest quality assurance flag" independent of the NDVI values [39].This proposed technique differs from the operational DTB 10K technique [13], which uses "an average of the DT 10K and DB 10K AOD retrievals or available one for only 0.2 < NDVI < 0.3".Instead, the proposed technique uses "an average the DT 10K and DB 10K AOD retrievals or available one" for all available NDVI values.(iv) AERONET AOD is interpolated to 0.55 µm using a standard Ångström exponent (α) extrapolation [37], as the project does not provide AOD measurements directly at this common MODIS wavelength.(v) To increase the number of samples for validation, collocations are defined as the average of at least two AERONET AOD measurements between 10:00 and 12:00 local solar time and at least two pixels of MODIS AOD observations within a sampling window of 3 pixels × 3 pixels (average of 9 pixels) centered on the AERONET site.(i.e., an average within a 9 km × 9 km region).(vi) Retrieval errors are reported using the expected error (EE) of the DT algorithm at 3 km resolution over land [32], root mean square error (RMSE), and mean bias (MB).To compare DT 3K and DTB 3K statistically, the percent relative differences in N, EE Equation (1), RMSE Equation ( 2), MB Equation (3), and R Equation ( 4) are calculated using Equation (5).These relationships are defined as EE = ± 0.05 and

Results and Discussion
4.1.Validation of the DT 3K and DTB 3K AOD Products at Regional Scale DT 3K and DTB 3K AOD retrievals were validated from 2008 to 2012 (Figure 1 and Table 2) against AERONET.In Figure 1, red and black colors represent the coincident DT 3K and DTB 3K observations, respectively (dashed lines = EE envelopes, and the black solid line = 1:1 line).Figure 1 shows that the DT 3K AOD retrievals, in general, overestimate at all of the sites, although large variance was observed between them overall.This overestimation of AOD retrievals by DT 3K was observed at 13 out of 19 sites, while AOD retrievals at only six sites meet the requirement of the EE (>68% or 69% to 88% of the retrievals were within the EE).The greatest uncertainties were observed at Paris (NDVI = 0.15), Moscow_MU_MO (NDVI = 0.31), Leipzig (NDVI = 0.44), and Minsk (NDVI = 0.32), with only 8%, 14%, 26% and 27% of the retrievals, respectively, being within EE (Figure 1 and Table 2).This overestimation, occurring for both low and high aerosol loadings, probably implies an underestimation of the surface reflectance by the VIS vs. 2.11 µm relationship, and potentially an error in the aerosol schemes used in the LUT.Previous studies reported similar errors in the DT 3K AOD retrievals over different parts of the globe [32][33][34]36].This is also similar to the DT C6 algorithm at 10 km, which overestimates with positive offset [34][35][36]67,68].The aggregated results of all sites show a large and significant overestimation in the DT3K AOD retrievals, as 43% of the retrievals were above EE (Table 2).All these sites have different surface characteristics.For example, Paris is a pure urban site, whereas Leipzig is dominated by vegetated surfaces.For Paris and Leipzig, the slope between DT3K and AERONET was significantly greater than The aggregated results of all sites show a large and significant overestimation in the DT 3K AOD retrievals, as 43% of the retrievals were above EE (Table 2).All these sites have different surface characteristics.For example, Paris is a pure urban site, whereas Leipzig is dominated by vegetated surfaces.For Paris and Leipzig, the slope between DT 3K and AERONET was significantly greater than one (Paris = 1.99 and Leipzig = 1.47), which probably suggests too much absorption in the aerosol model used in the LUT [38,69].However, both generally experience a wide range of aerosol loading conditions.Thus, selection of an accurate aerosol model is important for accurate high AOD retrievals [13].Overall, the performance of DT 3K was relatively poor over the vegetated surfaces (NDVI > 0.30), as only 56% of the retrievals were within EE with RMSE of 0.131 and MB of 0.085.This is an important distinction, though, as the point of designing the retrieval was ultimately more accurate AOD over such surfaces.
Validation of the DTB 3K AOD retrievals show significant improvement in retrieval quality, as the percentage of retrievals within EE increased and RMSE and MB decreased at each site (Figure 1 and Table 2).For the Paris, Moscow_MU_MO, Leipzig, and Minsk sites, for instance, the percentage of retrievals within EE increased remarkably from 8% to 63%, 14% to 68%, 26% to 73%, and 27% to 65%, respectively; RMSE decreased from 0.362 to 0.188, 0.200 to 0.151, 0.164 to 0.120, and 0.163 to 0.122, respectively; and MB decreased from 0.311 to 0.083, 0.179 to 0.072, 0.137 to 0.063, and 0.135 to 0.066 (Table 2), respectively.These results suggest that the DB algorithm performs better at these sites compared with DT and the contribution of the DB AOD retrievals in the DTB 3K retrievals significantly improves the retrieval quality and reduces error.Again, the advantage of using the average of both DT and DB AOD retrievals is to minimize the error in the DT C6 algorithm [39].
For all sites, 77% of the DTB 3K AOD retrievals were within EE, which is 38% higher than the DT 3K AOD retrievals, RMSE and MB decreased from 0.131 to 0.097 and 0.087 to 0.039, which are 26% and 54%, respectively, lower than the DT 3K .These results suggest that a merged DTB 3K AOD product exhibits better retrieval quality than the DT 3K and can thus be applied with greater confidence for air quality studies at the relatively finer scales.

Validation of the DT 3K and DTB 3K AOD Products at Local Scales
The performance of the DT 3K and DTB 3K AOD products was further evaluated in terms of improvement in percentage of retrievals within EE, spatiotemporal data coverage, RMSE and MB and R at each AERONET site based on the following criteria [39]: if the relative difference using Equation ( 5) is (a) within 10% for the percentage of retrievals within EE; (b) within 20% for the data count (N); (c) within 5% for RMSE; (d) within 5% for MB; and (e) within 10% for R, then the DT 3K and DTB 3K are considered to perform equally well at that site, and these sites are denoted by a "plus" symbol in Figure 2. In Figure 2, DT 3K and DTB 3K are represented by "triangle" and "circle" symbols, respectively, when they performed better over the individual sites, and color variations represent the magnitude of the relative difference (%) between the DT 3K and DTB 3K AOD products.The point of this analysis is to highlight the robustness of the AOD product with respect to each statistical parameter for each individual site.
For the percentage of AOD retrievals within EE, the DTB 3K AOD product performed well, as 15 out of 19 sites showed improvement and the percentage of AOD retrievals within EE was increased by 11% to >100% compared with the DT 3K AOD product (Figure 2a).There were only four sites where DT 3K and DTB 3K performed equally, as the relative difference of the percentage of retrievals within EE is less than 10%.Overall, the DTB 3K method performed well and significantly improved retrieval quality, as the percentage of AOD retrievals within EE increased due to the contribution of the DB AOD retrievals.
For the data count, or number of collocations, the DT 3K and DTB 3K methods performed equally at 14 out of 19 sites, as the relative difference of data counts is within 20% (Figure 2b).For the remaining five sites, DTB 3K performed well compared with DT 3K as the data count increased by 21% to 60%.This indicates that the DTB 3K method is likely more skillful than the DT 3K method in terms of spatiotemporal data coverage.For correlation, the DT3K and DTB3K methods performed equally at 18 out of 19 sites, as the relative difference was within 10% (Figure 2e).There was only one site where the DT3K AOD retrievals have a better correlation with the AERONET AOD retrievals than the DTB3K AOD retrievals as the relative difference was between 11% and 20%.Overall, both methods performed equally in terms of correlation.
In full, these results suggest that the DTB3K method is robust, more efficient and performed better at relatively finer scales, with larger data count percentages within EE, greater data counts overall, and lower RMSE and MB than DT3K.

Summary and Conclusions
The Moderate Resolution Imaging Spectroradiometer (MODIS) Collection 6 aerosol product provides global aerosol optical depth (AOD) observations over land at 3 km and 10 km spatial resolutions based on the Dark Target (DT) algorithms, and at 10 km resolution based on the Deep Blue (DB) algorithm.The DT and DB algorithms have different spatial coverage of AOD observations over land due to differences in their retrieval approaches (i.e.pixel selection, cloud screening and surface reflectance estimation method).DT3K exhibits large errors over urban or non-vegetated surfaces, as the DT algorithm is designed to retrieve AOD over vegetated surfaces.Therefore, the objectives of this study included developing a new DT and DB merged aerosol product on a 3 km grid, which can reduce the errors and increase the spatiotemporal coverage by providing AOD observations for those surface types and regions where either of each (DT and DB) were unable to provide due to pixel selection criteria and cloud mask.
For this analysis: (i) only high quality-assured AOD observations were obtained from the Scientific Data Sets (SDS), including "Optical_Depth_Land_And_Ocean" and For RMSE and MB, the DTB 3K method significantly reduced the errors at 16 (Figure 2c) and 18 (Figure 2d) sites, respectively, compared with DT 3K .The RMSE and MB reduced by 6 to 60% and 21 to >100%, respectively.There were only three (one) sites where both methods exhibit the same RMSE (MB).These results suggest that DTB 3K is robust, with lower RMSE and MB errors than the DT 3K retrievals.
For correlation, the DT 3K and DTB 3K methods performed equally at 18 out of 19 sites, as the relative difference was within 10% (Figure 2e).There was only one site where the DT 3K AOD retrievals have a better correlation with the AERONET AOD retrievals than the DTB 3K AOD retrievals as the relative difference was between 11% and 20%.Overall, both methods performed equally in terms of correlation.
In full, these results suggest that the DTB 3K method is robust, more efficient and performed better at relatively finer scales, with larger data count percentages within EE, greater data counts overall, and lower RMSE and MB than DT 3K .

Summary and Conclusions
The Moderate Resolution Imaging Spectroradiometer (MODIS) Collection 6 aerosol product provides global aerosol optical depth (AOD) observations over land at 3 km and 10 km spatial resolutions based on the Dark Target (DT) algorithms, and at 10 km resolution based on the Deep Blue (DB) algorithm.The DT and DB algorithms have different spatial coverage of AOD observations over land due to differences in their retrieval approaches (i.e.pixel selection, cloud screening and surface reflectance estimation method).DT 3K exhibits large errors over urban or non-vegetated surfaces, as the DT algorithm is designed to retrieve AOD over vegetated surfaces.Therefore, the objectives of this study included developing a new DT and DB merged aerosol product on a 3 km grid, which can reduce the errors and increase the spatiotemporal coverage by providing AOD observations for those surface types and regions where either of each (DT and DB) were unable to provide due to pixel selection criteria and cloud mask.
For this analysis: (i) only high quality-assured AOD observations were obtained from the Scientific Data Sets (SDS), including "Optical_Depth_Land_And_Ocean" and "Deep_Blue_Aerosol_Optical_Depth_550_Land_Best_Estimate" for DT 3K and DB 10K , respectively; (ii) the DB 10K AOD retrievals were resampled to 3 km grid using nearest neighbor interpolation algorithm; and (iii) they were merged with DT 3K AOD retrievals using Simplified Merge Scheme (SMS) defined as "an average of the DT 3K and DB 3K AOD retrievals or the available one with highest quality assurance flag".DT 3K and DTB 3K AOD retrievals were validated from 2008 to 2012 against cloud-screened and quality-assured (Level 2.0 Version 2) AOD measurements obtained from the 19 AERONET sites in Europe located over the vegetated and non-vegetated surfaces.
Our primary conclusions are: (i) DT 3K AOD retrievals were overestimated over vegetated surfaces for both low and high aerosol loadings.(ii) The overestimation might be caused by the underestimation of the surface reflectance and inappropriate aerosol model.(iii) Only 56% retrievals of the DT 3K were within the EE which indicates that the DT 3K product does not meet the requirements of the EE.(iv) The DTB 3K method significantly improved the retrieval quality as the percentage of the retrievals and data counts were increased, and RMSE and MB were decreased.(v) The contribution of DB AOD retrievals in the DTB 3K helped to reduce the overestimation in the DT 3K AOD retrievals for both low and high aerosol loadings.(vi) The percentage within the EE for the DTB 3K retrievals increased up to 77% which indicates that the DTB 3K product meets the requirements of the EE, and this is a 38% relative increase over the DT 3K AOD retrievals.(vii) The DBT 3K method reduced the RMSE and MB errors by 26% and 54%, respectively, for all sites.
Overall, the DTB 3K merged method is robust and performed better over vegetated and non-vegetated land surfaces than the DT 3K algorithm, and is recommended for air quality and climate-related studies in such land-surface regions.

12 Figure 2 .
Figure 2. Maps showing the best performing retrieval at AERONET sites for the following evaluation statistics: (a) percentage within the EE; (b) data count (N); (c) root mean square error (RMSE); (d) mean bias (MB); and (e) correlation coefficient (R).

Figure 2 .
Figure 2. Maps showing the best performing retrieval at AERONET sites for the following evaluation statistics: (a) percentage within the EE; (b) data count (N); (c) root mean square error (RMSE); (d) mean bias (MB); and (e) correlation coefficient (R).

Table 1 .
Summary of the AERONET sites used in this study from 2008 to 2012.

Table 2 .
Validation summary of the DT 3K and DTB 3K AOD retrievals.