A new multiangle implementation of the atmospheric correction (MAIAC) algorithm has been applied in the Moderate Resolution Imaging Spectroradiometer (MODIS) sensor and has recently provided globally high-spatial-resolution aerosol optical depth (AOD) products at 1 km. Moreover, several improvements have been modified in the classical Dark Target (DT) and Deep Blue (DB) aerosol retrieval algorithms in MODIS Collection 6.1 products. Thus, validation and comparison of the MAIAC, DT, and DB algorithms are urgent in China. In this paper, we present a comprehensive assessment and comparison of AOD products at a 550 nm wavelength based on three aerosol retrieval algorithms in the MODIS sensor using ground-truth measurements from AErosol RObotic NETwork (AERONET) sites over China from 2000 to 2017. In general, MAIAC products achieved better accuracy than DT and DB products in the overall validation and accuracy improvement of DB products after the QA filter, demonstrating the highest values among the three products. In addition, the DT algorithms had higher aerosol retrievals in cropland, forest, and ocean land types than the other two products, and the MAIAC algorithms were more accurate in grassland, built-up, unoccupied, and mixed land types among the three products. In the geometry dependency analysis, the solar zenith angle, scattering angle, and relative azimuth angle, excluding the view zenith angle, significantly affected the performance of the three aerosol retrieval algorithms. The three products showed different accuracies with varying regions and seasons. Similar spatial patterns were found for the three products, but the MAIAC retrievals were smaller in the North China Plain and higher in Yunnan Province compared with the DT and DB retrievals before the QA filter. After the QA filter, the DB retrievals were significantly lower than the MAIAC retrievals in south China. Moreover, the spatiotemporal completeness of the MAIAC product was also better than the DT and DB products.
Aerosols are a multi-compartment system consisting of suspended solid and liquid particles in the atmosphere, which play an important role in radiative forcing (Rajeev et al., 2001), regional climate (Qian and Giorgi, 1999; Feng et al., 2019), and urban air pollution (Dominici et al., 2014). The aerosol optical depth (AOD) is the key aerosol optical parameter, defined as the vertical integration of the aerosol extinction coefficient from the ground to the top of the atmosphere (TOA). Ground measurements from the AErosol RObotic NETwork (AERONET) provide high-quality multiband aerosol optical and microphysical properties at 15 min sampling frequencies on a global scale (Holben et al., 1998). High-quality ground measurements are often employed to validate satellite aerosol products (Chu et al., 2002) and to provide a regional aerosol model for the satellite aerosol retrieval algorithm (Levy et al., 2013). However, they cannot grasp the high aerosol spatial variability due to the sparse ground sites where spatial variability information is still necessary. Though some active remote-sensing methods, e.g., spaceborne lidar, can monitor vertical distribution of aerosol, they still cannot observe high aerosol spatial variability (Huang et al., 2007; Jia et al., 2015; Liu et al., 2015). Although model-simulated AOD can obtain spatially continuous data, its very coarse resolution and large uncertainties limit its application (Sun et al., 2019; Cesnulyte et al., 2014). In contrast, the satellite aerosol retrieval algorithm has the ability to achieve continuous spatial measurements with high spatial resolution (She et al., 2017).
The Moderate Resolution Imaging Spectroradiometer (MODIS) sensor with its multiband detection ability from the visible band to thermal infrared spectrum band (Salomonson et al., 1989) can readily detect aerosol properties. With the Terra satellite and Aqua satellite carrying the MODIS sensor successfully launched in 2000 and 2002, respectively, MODIS has stored over 17 years of historical globally monitored data. Recently, a new multiangle implementation of the atmospheric correction (MAIAC) algorithm has been applied in the MODIS sensor, which provides high-spatial-resolution aerosol data at 1 km (Lyapustin et al., 2018). Moreover, some important improvements in classical Dark Target (DT; Mattoo, 2017) and Deep Blue (DB; Hsu, 2017) aerosol retrieval algorithms have been revised in MODIS Collection 6.1 products. However, all satellite aerosol retrieval algorithms are under some hypothesis and approximation assessments, and the accuracy should be validated before applying a satellite aerosol product in related studies.
China is experiencing severe aerosol pollution, and numerous studies on
aerosol pollution have utilized MODIS Collection 6.0 aerosol retrievals to
map aerosol pollution and to analyze its spatiotemporal trends (Fang et al.,
2016; Ma et al., 2014; He and Huang, 2018a, b; Zou et al., 2016, 2019; Zhai et al.,
2018). Few studies have applied 1 km MAIAC aerosol retrievals to map finer
aerosol concentrations in regional China, e.g., the Yangtze River Delta (Xiao
et al., 2017) and Shandong Province (Li et al., 2018). Before widely applying
MAIAC and C6.1 products in China, the accuracy differences and applicable
conditions of the three aerosol retrievals should first be recognized to
guide the utilization of these products. Recently, the global validation
(Lyapustin et al., 2018) and regional validation in South America (Martins et
al., 2017), North America (Superczynski et al., 2017), and South Asia (Mhawish
et al., 2019) for MAIAC products has shown that more than 66 % of
retrievals fall within the expected error (EE
In this context, we provide the first comprehensive understanding and comparison of the aerosol retrieval uncertainties for MAIAC, DT, and DB products in China based on spatiotemporal accuracy differentiation patterns, spatiotemporal completeness, land type dependence characteristics, view geometry dependence characteristic aspects, and other features. The following paper is organized as follows: Sect. 2 briefly introduces three satellite products with their retrieval algorithm and ground AERONET data, the validation approach is clarified in Sect. 3, and Sect. 4 provides the detailed validation results and discussion. The conclusions are presented in Sect. 5.
Three aerosol products, e.g., MAIAC, DT, and DB, are stored in Hierarchical
Data Format (*.hdf) files, and we obtain corresponding *.hdf files in the
China region from 2000 to 2017 from the NASA Earthdata Search website
(
The DT algorithm retrieves AOD parameters based on the assumption that the surface reflectance in two visible bands, e.g., 470 and 644 nm, presents a good linear relationship with the surface reflectance in the shortwave infrared (SWIR) band, e.g., 2119 nm, in dark, dense vegetated area, and the measurement in the SWIR band is transparent with the aerosol particle (Kaufman et al., 1997; Levy et al., 2013). The surface and aerosol information can then be decoupled from the TOA spectral reflectance. Compared with the DT algorithm in Collection 6.0, the DT algorithm in Collection 6.1 mainly revises the surface characterization over the land surface when the urban percentage is larger than 20 % (Gupta et al., 2016).
The DT algorithm produces two aerosol resolution products in Collection 6.0
and 6.1, e.g., 3 km
The DB algorithm retrieves the AOD parameter under the hypothesis that the surface reflectance in the Deep Blue band, e.g., 412 nm, is much smaller than in longer bands over bright surfaces, such as urban and desert regions (Hsu et al., 2004). First, the DB algorithm retrieves 1 km aerosol properties using the global surface reflectance database in visible bands, e.g., 412, 470, and 650 nm, and then aggregates 1 km pixels into a 10 km scale. In Collection 6.0, the surface reflectance database is improved using knowledge of the normalized difference vegetation index, scattering angle, and season (Hsu et al., 2013). The ability to retrieve aerosol data over a bright surface for the DB algorithm greatly expands the coverage of aerosol retrieval. The general principles for collection of the 6.1 DB products are still the same as those in the Collection 6.0 version. The major improvements for Collection 6.1 DB products are in the radiometric calibration, heavy smoke detection, artifact reduction over heterogeneous terrain, surface model in elevated terrain and regional/seasonal aerosol optical models (Hsu, 2017).
The same as the DT products, SDSs named
“Deep_Blue_Aerosol_Optical_Depth_550_Land” without the QA filter and
“Deep_ Blue_Aerosol_Optical_Depth_550_Land_Best_Estimate” with the
QA filter (QA
The MAIAC algorithm relies on the assumption that the surface reflectance changes slowly over time and shows high variability over space, whereas the aerosol loading changes very fast over time and varies only on a limited space scale. The main procedure of MAIAC is as follows: first, MAIAC resamples MODIS L1B measurements into a fixed 1 km grid, and then it adopts 4–16 d time series of resampled MODIS measurements to retrieve the surface Ross–Thick Li–Sparse (RTLS) bidirectional reflectance distribution function (Lucht et al., 2000) using the measurements in the SWIR band. Subsequently, the linear spectral regression coefficient (SRC) between 470 and 2119 nm for each 1 km grid is retrieved instead of using the empirical regression coefficient in the DT algorithm. Finally, the AOD parameter at 470 nm can be retrieved by searching the minimum spectral residual between the theoretical TOA reflectance of the lookup table and the measurements in the red and SWIR bands. The AOD is originally retrieved at 470 nm, and the AOD parameter at 550 nm is computed using the AOD parameter at 470 nm based on spectral properties, expressed by the regional aerosol model from the MAIAC lookup table. The detailed MAIAC algorithm has been described by Lyapustin et al. (2011).
Data used in this study were from the “Optical_Depth_055” and “AOD_QA” SDSs, and data were collected from the Terra satellite. The data type of the “AOD_QA” SDSs is a 16-bit unsigned integer, and the best retrieved quality can be selected if 8–11 bytes of “AOD_QA” SDS bits are “0000”, which indicates the retrieval pixel and its adjacent pixel is clear (Lyapustin et al., 2018). The solar zenith angle in the “cosSZA” SDSs, view zenith angle in the “cosVZA” SDSs, relative azimuth angle in the “RelAZ” SDSs, and scattering angle in the “Scattering_Angle” SDSs were also selected to analyze the view geometry dependence for MAIAC products.
Selected AERONET sites used in this study. The number of the match column statistics matches the number between the satellite observations before the QA filter and the ground AERONET observations in the selected spatiotemporal window presented in Sect. 3.1.
Locations of the selected AERONET sites around China displayed on the land cover map from 2013. BTH: Beijing–Tianjin–Heibei; YRD: Yangtze River Delta; PRD: Pearl River Delta; NW: northwestern China.
AERONET is a global ground-based aerosol monitoring network that provides
continuous optical and microphysical properties of aerosols at a 15 min
sampling rate. The total uncertainty for the AERONET AOD parameter under
cloud-free conditions is lower than
One key difficulty in the aerosol retrieval algorithm is to decouple surface and atmosphere information in the satellite apparent reflectance. Land cover information greatly affects atmosphere properties (Xu et al., 2018; Feng and Zou, 2019). Understanding the uncertainties in a satellite aerosol retrieval algorithm for different land cover types is necessary (W. Wang et al., 2019). GIMCP land cover data with 30 m resolution in the years 2000, 2005, 2008, 2010, and 2013 were used in this study. The first level of GIMCP land cover data includes cropland, forest, grassland, water, and built-up and unoccupied land. Among them, unoccupied land includes desert, saline–alkaline soil, swampland, bare land, and bare rock gravel, which mainly includes bright surfaces. The high spatial resolution and abundant land cover types support our studies. Figure 1 shows the first level land cover type across the China mainland in 2013.
Land cover type for each AERONET site in 2013.
There are only a small number of matchup data between the satellite data and ground data when using the direct matching method, e.g., use only 1 pixel where the AERONET sites are located and ground measurement at the exact satellite overpass time, due to large numbers of missing data in AERONET or satellite data and the time delay between the satellite overpass time and AERONET sampling time. Therefore, under the assumption that aerosol information is homogeneous in a limited spatial and temporal lag, a suitable spatiotemporal window is often adopted to increase the matchup data number. Thus, satellite measurements in the spatial window around the AERONET sites are averaged, and ground measurements in the temporal window centered on the satellite overpass time are averaged.
For 10 km DT and DB products, the selected spatial window is often
50 km
The first level of GIMCP land cover data was used to label the AERONET site
group. Due to the selected 30 km
Table 2 shows the land cover types for each AERONET site in 2013. There were no land cover type changes for most sites except Hangzhou_City, Muztagh_Ata, and NAM_CO. For the Hangzhou_City site, the land cover type changed from cropland to mixed group from 2005 to 2008, potentially due to the process of urbanization. For the Muztagh_Ata site, the land cover type changed from unoccupied land to grassland from 2008 to 2010, and the land cover type for the NAM_CO site varied from grassland to the mixed group between 2008 and 2010. We labeled each matchup dataset for the three sites using the land cover type in the nearest year to the AERONET sampling time.
The expected error (EE) envelope is often used to validate satellite
retrieval uncertainties. More than 66 % of retrievals falling within the
expected error lines indicate good accuracy. For the DT algorithm, the EE
envelope is generally defined as
In order to compare the spatiotemporal completeness of three products, daily
spatial completeness and the temporal completeness are defined by Eqs. (7)
and (8).
Figure 2 shows the overall evaluation for MAIAC, DT, and DB products before
and after the QA filter. In total, MAIAC products have more matchup data than
DT and DB products, which indicates the completeness of the MAIAC product may
be higher than the DT and DB products. Before the QA filter, the statistic
showed that 69.84 % of retrievals fall within the EE envelope, indicating
a good accuracy for MAIAC products in China. Compared with DT and DB
products, only 53.64 % and 55.66 % of retrievals were determined for
DT and DB products. Based on the
Overall accuracy evaluation of MAIAC, DT, and DB AOD versus AERONET
AOD at 550 nm before and after the QA filter. The black line, red line, and
dashed line in the scatterplot are the 1 : 1 reference line, regression
line, and expected error (EE
To analyze and compare the retrieval accuracy at different AOD levels for
three products, four bins with different levels, low level (
Evaluation of the MAIAC accuracy for different land cover types
before and after the QA filter. The black line, red line, and dashed line in
the scatterplot are the 1 : 1 reference line, regression line, and expected
error (EE
Accuracy evaluation of MAIAC, DT, and DB at the low level (
Figure 3 shows a scatterplot figure of the MAIAC products in different land cover types before and after the QA filter. In total, MAIAC retrievals in cropland, built-up, grassland, and ocean types were more accurate than forest, unoccupied land, and mixed types according to the Within_EE results. After the QA filter, except for grassland, the accuracies all improved, and the improvement effect in ocean type was more obvious.
Evaluation of the MAIAC accuracy in the forest area for each AERONET
site before and after the QA filter. The black line, red line, and dashed line
in the scatterplot are the 1 : 1 reference line, regression line, and
expected error (EE
The high aerosol loading, e.g., AODs
In evergreen forest areas (Fig. 3b-i and b-ii), the retrievals showed a good
correlation with ground measurements, with
Evaluation of the MAIAC accuracy in the grassland area for each
AERONET site before and after the QA filter. The black line, red line, and
dashed line in the scatterplot are the 1 : 1 reference line, regression
line, and expected error (EE
For the grassland type (Fig. 3c-i and c-ii), over 83.68 % of MAIAC
retrievals fell into the EE lines before the QA filter, and the
MAIAC had good accuracy in the unoccupied land cover type (Fig. 3e-i and e-ii),
with Within_EE results of 67.44 % and 71.43 % before and after the
QA filter, and
Comparison of the retrieval accuracy of the MAIAC, DT, and DB products for different land cover types before and after the QA filter. “–” means no matchup pairs or that the matchup pairs number fewer than 10. The bold number is the highest peformance among three algorithms by each indicator.
In comparison to DT and DB products, Table 4 shows the validation of the
statistical results for the MAIAC, DT, and DB products with different land
type covers. In cropland area, the accuracy of the DT product was evidently
better than that of the MAIAC and DB products according to the
Table 5 shows the validation accuracy for three products after the QA filter in four seasons. In cropland, the retrieval accuracies in autumn for the three products were better than in other seasons. For forest land types, three products showed a higher correlation in autumn than the other seasons, but the Within_EE values demonstrated the best results in winter, and the corresponding results for DB products were clearly higher than for the other two products. In terms of grassland type, MAIAC and DB products were more accurate in summer and spring, respectively. In the built-up region, all products showed a high correlation in all seasons, but DT products were seriously overestimated. In unoccupied land, matchup pairs for MAIAC and DB products were more focused in spring, and MAIAC products performed better than DB products. A high correlation was also found for the three products in mixed and ocean regions in all seasons, but more MAIAC retrievals met the EE envelope line.
Comparison of the retrieval accuracy of the MAIAC, DT, and DB products for different land cover types in four seasons after the QA filter. “–” means no matchup pairs or that the matchup pairs number fewer than 10. The bold number is the highest peformance among four seasons by each indicator.
The Ångström exponent (AE) is a key parameter to describe aerosol
particle size, and in general, local aerosol sources play a dominant role in
aerosol regimes (Mhawish et al., 2019). To discover aerosol particle sizes in
different land covers, Fig. 6 shows a scatterplot of the AE (440–675 nm)
parameter versus AOD for different land cover types. Our results were similar
to those of Martins et al. (2017). The aerosol types in China are mainly
fine-mode aerosol particles (AE
Scatterplot of AOD at 550 nm against the Ångström exponent for different land cover types. We selected AERONET sites with maximum observations for each land cover type: XiangHe (cropland); Taipei_CWB (forest); QOMS_CAS (grassland); Beijing (built-up); Dunhuang (unoccupied land); Hong_Kong_PolyU (ocean).
Scatterplot of the AOD bias from matchup data versus the AERONET Ångström exponent (440–675 nm) before and after the QA filter.
To determine how the view geometry influences the accuracy for three retrieval algorithms, we analyzed view geometry dependency using the following four angles: solar zenith angle (SZA), view zenith angle (VZA), scattering angle (SA), and relative azimuth angle (RAA) (Superczynski et al., 2017; W. Wang et al., 2019). We separated each kind of angle into 10 bins and statistically analyzed the AOD bias distribution in each bin. The results are displayed in Fig. 8.
In terms of the solar zenith angle, the three retrieval algorithms all showed
a strong dependency with different characteristics. A slight downtrend along
with SZA was found in the MAIAC algorithm, and the MAIAC retrievals seemed
slightly overestimated when SZA was less than 40
The MAIAC and DB algorithms showed no dependency on the view zenith angle, and the corresponding mean bias lines did not fluctuate much along with VZA. Compared with the results obtained before and after the QA filter, the mean bias line for the MAIAC algorithm slightly increased, and the mean bias line for the DB algorithm moves down to a relatively large degree. VZA slightly affected the DT performance with a little downtrend. After the QA filter, the mean bias line slightly declined.
The scattering angle also greatly impacted the performance of the three
retrieval algorithms. MAIAC retrievals seemed to be underestimated when the
SAs were less than 100
For the MAIAC algorithm, positive biases occurred as RAA approached the
extremes of 0
Dependency of the AOD bias on the solar zenith angle, view zenith
angle, scattering angle, and relative azimuth angle for the
To investigate retrieval accuracy of the three algorithms at different
regions and different times, Fig. 9 shows the
Three products presented different retrieval accuracies in different regions.
In the BTH region (marked by the black box in Fig. 1), three products showed
a good correlation with the ground measurements, e.g.,
Figure 10 presents the monthly validation results for the three products. We
overlooked the specific QOMS_CAS site for this purpose due to its poor
performance after the QA filter, which would affect the overall accuracy.
Three products showed a good correlation with the ground measurements for all
months with
Evaluation results for MAIAC, DT, and DB after and before the QA filter in each AERONET site. The subscript QA denotes the corresponding results after the QA filter.
Validation of MAIAC, DT, and DB in different months before and after the QA filter.
We investigate the annual change in retrieval accuracy for three products to
ascertain whether the MODIS instrument maintains its performance due to it
exceeding its designed lifetime. However, according to Table 1, the time
durations of each AERONET site were significantly different. Thus, the
matchup observation pair during each year was from different sites. This
phenomenon may result from incomparable validation results for each year.
However, if only considering the sites with the same monitoring time, most
sites will be discarded, and fewer matchup numbers will cause unreliable
corresponding statistical results. Thus, we still adopted all site
measurements. We ignored the results for the years 2000, 2001, 2002, and 2003
due to fewer matchup numbers in these years. According to Fig. 11, three
products showed a high correlation with ground measurements according to the
Validation of MAIAC, DT, and DB in different years before and after the QA filter from 2004 to 2017.
Bias plot for the three products before the QA filter at five selected AERONET sites with a monitoring period containing most of the study years from 2004 to 2017.
To compare the difference in spatial variations for the three products, we upscaled the MAIAC product to match the grid of the DT and DB products; thus, 1 km pixels falling within the 10 km grid were averaged. Such a protocol can aid in investigating differences in different regions between the three products.
Figure 13 presents multiyear averaged and difference results between MAIAC,
DT, and DB products, with aerosol loading presenting a noteworthy assembly
characteristic. Higher AOD values were concentrated in the North China Plain
and Sichuan Basin where the land cover types were mainly cropland-oriented,
as shown in Fig. 1. Before the QA filter, compared with the DT and DB
observations, the MAIAC AODs were smaller in the North China Plain and larger
in Yunnan Province and east Taiwan. After the QA filter, the DB AODs became
smaller in the North China Plain and southeast region. Compared with the DB
AODs, the MAIAC AODs became slightly higher in the North China Plain
(difference over 0.1) and obviously higher in southeast China (difference
over 0.3). Recall the statistical result presented in Fig. 9, in which the DT
and DB products were overestimated in the BTH region, the DB product was
underestimated in the YRD region, and the MAIAC product seemed to be
overestimated in east Taiwan. These findings indicate that MAIAC retrievals
are more accurate than DT and DB in the North China Plain and southeast
region, and DB retrievals are more accurate than MAIAC in east Taiwan.
However, due to the lack of the AERONET site in Yunnan Province, we could not
evaluate the accuracy of the three products in Yunnan Province. The
difference before and after the QA filter for the MAIAC product was very
small, except for some individual pixels in the Tibet region. In addition,
there was an obvious boundary in the 30
Averaged AOD distributions throughout the year for MAIAC, DT, and DB before the QA filter and their differences after the QA filter from 2000 to 2017. The subscript QA denotes the corresponding results after the QA filter.
Figure 14 shows the seasonal comparison results among three products before and after the QA filter. The AOD spatial variation for the three products showed apparent seasonal characteristics. The AODs in the North China Plain in summer were higher than in other seasons, and the AODs in the Tarim Basin in spring were higher than in other seasons. Based on the AOD spatial variation difference map, the difference between MAIAC and DT in the North China Plain evolved gradually from negative in spring to positive in winter. The negative difference between MAIAC and DB in the North China Plain was higher in summer and winter than in spring and autumn. The positive difference in Yunnan Province between MAIAC and DT was slightly lower than that between MAIAC and DB. After the QA filter, AODs in south China for the DB product were extremely low compared with those for the MAIAC product.
Seasonal averaged AOD distributions for MAIAC, DT, and DB and their differences before and after the QA filter from 2000 to 2017. The subscript QA denotes the corresponding results after the QA filter.
Based the upscale MAIAC 10 km data in Sect. 4.4, the spatial completeness in Eq. (7) and temporal completeness in Eq. (8) for three products are shown in Figs. 15 and 16. According to Fig. 15, the spatial completeness of the MAIAC product was higher than the DT and DB products before and after the QA filter. The spatial completeness of the DT product was smallest due to its retrieval failure on a bright surface. The spatial completeness for all the products showed an obvious periodical trend change. Table 6 shows the statistics for the spatial completeness of the three products in different seasons. Before the QA filter, the averaged spatial completeness of MAIAC (46.87 %) was higher than DT (16.66 %) and DB (34.80 %). After the QA filter, the reduced proportion of MAIAC (17.18 %) exceeded DB (15.30 %) and DT (8.66 %) because many climatology values in the Tibet Plateau were discarded. Comparison of the spatial completeness in four seasons revealed a higher spatial completeness for the three products in autumn than the other three seasons due to the reduced cloudiness in the dry autumn season. The spatial completeness in winter was smallest due to the influence of the surface snow cover and large deciduous trees. Compared with MAIAC and DB products, the spatial completeness of the DT product in winter was minimal due to the bright surface in winter.
Daily spatial completeness for MAIAC, DT, and DB from 2000 to 2017 before and after the QA filter.
Seasonal averaged spatial completeness for MAIAC, DT, and DB before and after the QA filter and their declining proportions after the QA filter.
Figure 16 presents the temporal completeness in China for the three products. Due to the climatology values in the Tibet Plateau, the temporal completeness of the MAIAC product in this region was very high (over 80 %). After the QA filter, the temporal completeness rapidly decreased in this region. In the other region, the declining proportions of temporal completeness for MAIAC were mostly lower than 10 %, except for Yunnan Province (nearly 15 %), Hainan Province (nearly 20 %), and east Taiwan (nearly 20 %). Compared with the MAIAC and DB products, DT retrievals were very scarce in the Tarim Basin due to failure on the bright desert surface. DT retrievals were more concentrated on the North China Plain and in Yunnan Province. After the QA filter, a dramatically reduced proportional area of temporal completeness (nearly 30 %) for DT products was observed in the cropland region in northeast China. The severely reduced proportional area (nearly 40 %) for the DB product after the QA filter was mainly focused on unoccupied land, e.g., gobi, saline–alkaline soil, at the top of the Tibet Plateau. Compared with the MAIAC product, before the QA filter, the DB product showed more retrievals in the Tarim Basin, North China Plain, and southeast China and fewer retrievals in Yunnan Province and northeast China. After the QA filter, the temporal completeness of the MAIAC product was better than the DB product in all regions.
Spatial distributions of temporal completeness for MAIAC, DT, and DB before and after the QA filter and their differences from 2000 to 2017. The subscript QA denotes the corresponding results after the QA filter.
In this study, we present the first comprehensive validation and comparison of three MODIS aerosol retrieval algorithms (i.e., MAIAC, DT, and DB) across China in terms of overall accuracy, land cover dependency, viewed geometry dependency, spatiotemporal retrieval accuracy, spatial distribution difference, and spatiotemporal completeness. These validation results may guide users to utilize the three products appropriately. The main results and conclusions are presented below.
In terms of overall accuracy, the MAIAC product is more accurate than the DT and DB products. The DT and DB products are positively biased before the QA filter, and the positive bias for the DB product is alleviated by the QA filter.
DT retrievals in cropland, forest, and ocean seem to be more accurate but with a positive bias than retrievals by the MAIAC and DB algorithms. The MAIAC algorithm performs better in grassland, built-up, and mixed areas than the DT and DB algorithms.
Three algorithms show a strong dependency on SZA, SA, and RAA. VZA only marginally affects the retrieval accuracy of the three algorithms.
The MAIAC product performs better in the BTH, YRD, PRD, and NW regions than the DT and DB algorithms, and the DB product performs better than the DT and MAIAC products after the QA filter in east Taiwan. The MAIAC algorithm performs better than the DT and DB algorithms in most months except June, July, August, and September. In these four months, MAIAC retrievals appear to be overestimated, and DB retrievals after the QA filter are more accurate than MAIAC retrievals.
Three AOD products present a similar spatial pattern with high aerosol loading in the North China Plain and Sichuan Basin. In comparison, MAIAC retrievals are lower in the North China Plain and Sichuan Basin than DT and DB retrievals and are higher in Yunnan Province and east Taiwan than DT and DB retrievals. After the QA filter, the DB AOD values are significantly reduced and obviously lower than the MAIAC product in southeast China.
Based on spatiotemporal completeness analysis, the MAIAC product has more retrievals in the spatiotemporal domain than the DT and DB products. The spatial completeness exhibits a strong periodical change, and the temporal completeness is highest in autumn compared to other seasons due to the decreasing cloud cover in this dry season, which is lowest in winter due to the snow cover and deciduous vegetation. In terms of temporal completeness, MAIAC has more retrievals in the Tarim Basin and the cropland in northeast China compared with the DT algorithm. Compared with the DB algorithm, MAIAC has fewer retrievals in the Tarim Basin and southeast China and more retrievals in northeast China. After the QA filter, the temporal completeness of MAIAC in all regions of China is better than that of the DB product.
MAIAC, DT, and DB data are publicly
available from NASA Earthdata Search at
NL and BZ designed the whole experiment. NL and YL developed the experiment code and performed it. The paper was initially written by NL and fully revised by BZ. HF, WW, and YT provided a lot of constructive comments on the experiment.
The authors declare that they have no conflict of interest.
We thank NASA for providing MAIAC, DT, DB products, and AERONET data. We would like to thank Yujie Wang and Alexei Lyapustin in NASA for answering our question concerning the MAIAC product and the help it provided.
This research has been supported by the National Key Research and Development Program of China (grant no. 2016YFC0206205), the National Natural Science Foundation of China (grant no. 41871317), and the Innovation Driven Program of Central South University (grant no. 2018CX016).
This paper was edited by Jianping Huang and reviewed by three anonymous referees.