Retrieval and validation of long-term aerosol optical depth from AVHRR data over China

ABSTRACT Advanced Very High Resolution Radiometer (AVHRR) onboard National Oceanic and Atmospheric Administration (NOAA) satellites can provide over 40 years of global remote sensing observations, which can be used to retrieve long-term aerosol optical depth (AOD). This is of great significance to the study of global climate change. In this paper, we proposed an algorithm to jointly calculate AOD and land surface properties from AVHRR observations. With assumptions that AOD doesn’t vary in adjacent space and earth surface property doesn’t vary in two days, the algorithm considered non-Lambertian surface reflection based on the shape of bidirectional reflectance distribution function (BRDF shape) and obtained AOD and surface property by optimal estimation (OE) method. The algorithm has been applied to NOAA-7, 9, 11, 14, 16, 18, and 19 satellites and AVHRR-retrieved AOD with 5 × 10 km over China (15°–60° N, 70°–140°E) has been obtained from 1982 to 2016. Comparisons of AVHRR-retrieved AOD against AErosol RObotic NETwork (AERONET) (in and around China) and China Aerosol Remote Sensing Network (CARSNET) AOD show good consistency with 62.62% points within the uncertainty of Δτ =  ± (0.05 + 0.25τ) and root-mean-square error (RMSE) of 0.26. Further comparison of the monthly mean AOD of multiple AOD datasets in the ‘Beijing’, ‘Dalanzadgad’, ‘NCU_Taiwan’ and ‘Kanpur’ stations shows that the results of the algorithm are stable. The yearly averaged AOD data also has similar agreements with MERRA-2 (The Modern-Era Retrospective analysis for Research and Applications, Version 2) and AVHRRDB data (AVHRR ‘Deep Blue’ aerosol data set). The multi-year mean correlation coefficient is 0.70 and 0.61 and the percentages within the uncertainty are 80.01% and 67.29% compared with MERRA-2 AOD and AVHRRDB AOD respectively.


Introduction
Aerosols are one of the significant air pollutants that are produced from a variety of causes, including dust, flannel, smoke, haze, etc. Moreover, heavy aerosols mostly have the following two effects on human life: (1) haze or heavy dust aerosols can limit visibility, which interferes with transportation and people's everyday lives (Charlson 1969;Moses and Akinyemi 2017); (2) They are unhealthy for individuals, leading to respiratory and pulmonary disorders, especially fine particle matters (Tie, Wu, and Brasseur 2009;McGuinn et al. 2016). Additionally, aerosols have significant impacts on biogeochemical cycles and global climate change (Twomey 1977;Kaufman, Tanré, and Boucher 2002;Verma et al. 2017). Therefore, precise aerosol monitoring is crucial for quantitative assessment of the anthropogenic contribution to climate change and for guiding people to live healthy lives. It is essential to meet global grand challenges for sustainable development (Kulmala et al. 2021).
The capacity to quantify aerosol loading in the atmosphere makes aerosol optical depth one of the most important indicators for comprehending atmospheric physics and local air quality (Wei et al. 2019b). With the development of multi-spectral sensor, multi-angle sensor, high spectrum sensor, polarization sensor, radar, etc. and different platforms such as polar orbit satellite and earth synchronous satellite, researchers put forward many inversion algorithms which makes the retrieval of AOD from spaceborne sensor observations a major technique for monitoring aerosol loads on a large scale. As the first sensor that has been applied to retrieve AOD, AVHRR has the advantages of a wide swath, global observation and long-time observations and AOD from AVHRR over the ocean has been produced operationally (Zhao et al. 2002;Zhao 2004;Zhao et al. 2008). However, several widely used algorithms such as 'Dark Target' (DT) and 'Deep Blue' (DB) cannot be employed directly in AVHRR AOD retrieval over land since it only has a red channel in the visible spectrum and lacks the 2.1μm channel (Kaufman et al. 1997;Hsu et al. 2004Hsu et al. , 2013Levy, Remer, Mattoo et al. 2007).
Although there are many challenges such as a single visible band and problems with the AVHRR calibration (Li et al. 2014), many researchers try to develop algorithms that can be applied to obtain AOD over land from AVHRR observation. These algorithms can be categorized into one of the two categories listed below: (1) The surface reflectance is obtained first, and then AOD from AVHRR is retrieved from the atmospheric radiation transfer model. These methods obtain surface reflectance in different ways. Hauser et al. (2005) Takemata, Fukui, and Kawata (2006) achieved 0.64μm surface reflectance by fitting the linear relationship of surface reflectance between 0.64μm and 3.75μm and obtained the AOD on 24 and 25 April 1997. The study only got two days of AOD and only 31 points were used for fitting the surface reflectance. Wang et al. (2021) proposed a dual-channel aerosol optical depth retrieval algorithm that takes the BRDF effect into account and the algorithm was applied between 2008 and 2010. This method achieved a surface reflectance ratio of 0.63μm to 0.85μm from the corrected least top-of-atmospheric (TOA) reflectance ratio in 31 days. Li et al. (2021) proposed an improved dense dark vegetation algorithm, which applied MODIS (Moderate-resolution Imaging Spectroradiometer) vegetation index products to correct AVHRR NDVI (Normalized Difference Vegetation Index) to estimate AVHRR land surface reflectance, and retrieved AOD in mid-eastern America from June to October 2016. Moreover, Xue et al. (2017) presented an algorithm obtaining 0.64μm surface reflectance by fitting 3.75μm surface reflectance, NDVI, and scatter angle at different land types and different seasons and obtained AOD data set over North China and Central Europe for 35 years. Hsu et al. (2017) developed a physics-based algorithm to retrieve AOD from AVHRR both over land and ocean. The algorithm applied over land is an extension of the DB algorithm. The global AOD dataset (AVHRRDB) produced from NOAA-11, 14, and 18 satellite observations are available at the website https://portal.nccs.nasa.gov/datashare/AVHRRDeepBlue. (2) AVHRR AOD is obtained through experience relationships, such as regression methods or machine learning methods. Tian et al. (2022) generated AVHRR AOD products in 1998, 2006, and 2015 using machine learning training of AVHRR Level-2 data and MODIS AOD in 2014. Gao et al. (2016Gao et al. ( , 2019 got AOD over China from 1982 to 2011 using multiple regression relationships generated from Aqua/MODIS AOD and AVHRR Level_1b data. However, this data set was validated only on a monthly scale. All the above methods are favorable attempts for AVHRR AOD inversion, especially the studies of Xue et al. (2017), Gao et al. (2019), and Hsu et al. (2017) which provided AVHRR AOD results for long time series. However, the performance of AVHRR AOD results in a long-time sequence over China is still to be improved.
In this paper, AVHRR AOD results were obtained with the optimal estimation method with prior knowledge and reasonable assumptions. Different from the above two types of methods, in addition to taking into account the surface's BRDF features, our work concurrently estimated AOD and surface attributes from AVHRR measurements. We obtained the long-time range AOD data from 1982 to 2016 over China which have high precision in the research region compared to ground-based observations. The remaining part of the article is organized as follows. Section 2 gives the introduction of the data used in this paper and an overview of the study area. Section 3 describes the detailed implementation process of the proposed algorithm. Section 4 gives the evaluation of the AVHRR-retrieved AOD results against ground-based station measurements and comparisons with other algorithm results. The discussions and conclusion are given in the last section.

Data
In this study, many data are used for inversion including AVHRR level 1B Global Area Coverage (GAC) product, MCD43C1 Albedo/BRDF model parameters, MCD12C1 land cover type and the data used for validation and comparison include AERONET and CARSNET measurements, AVHRRDB AOD and MERRA-2 AOD. The data information is summarized as shown in Table 1.

AVHRR GAC
AVHRR is the sensor onboard the NOAA family of meteorological satellites. It has five spectral bands and a spatial resolution of 1.1 km at the nadir. AVHRR sensor scans with ± 55°from the nadir and has a large swath width of 2800 km which could scan the earth twice per day. The first AVHRR sensor carried by TIROS-N, was launched on 13 October 1978 and the latest AVHRR sensor, aboard Metop-C, was launched on 6 November 2018. During the period, AVHRR sensors were improved in four generations. At present, there are over 40 years of AVHRR sensor data records. It provides the following four level 1B data types: The Local Area Coverage dataset recording a portion of the orbit at the original resolution (1.1 km); the real-time downlink data -High Resolution Picture Transmission (1.1 km); the Global Area Coverage dataset which could provide 4 km reduced spatial resolution global observation; the daily data -Full Resolution Area Coverage (1.1 km) of the entire globe with the launch of Metop-A on 19 October 2006. In this paper, we have used AVHRR GAC data for AVHRR AOD inversion. The satellites with AVHRR sensors pass the equator twice per day. As some morning satellites cross the equator at 7:30, the zenith angle of solar is too large and there is less sun coverage, we have chosen afternoon satellites for research including 9,11,14,16,18,and 19. They have the observation at about 14:00 local time. The specific AVHRR information is as Table 2 shows. It must be pointed out that the calibration of AVHRR has problems due to changes in the instrument's response (Molling et al. 2010). In this study, we recalibrated AVHRR channel 1 by the calibration coefficients provided by Li et al. (2014).

MODIS products
MODIS sensors which are abroad on Terra and Aqua launched in 2000 and 2002 have 36 channels and have been used in many research fields such as meteorology, land, glaciers, and oceans (Myhre, Kvalevåg, and Schaaf 2005;Galford et al. 2008;Grobler et al. 2012;Natsagdorj et al. 2021;Jiang et al. 2021). In this study, MCD12C1 product was used to mask the water body over the study area. MCD43C1 provided BRDF model parameters which could produce BRDF shape.

Other products
AVHRRDB V001 product provides AOD parameter retrieved from AVHRR (NOAA-11:1989.07-1991.05, NOAA-14:1995.01-1999.12, NOAA-18:2006.01-2011). Level 2 product contains data from a 5-minute portion of a single AVHRR swath and has an 8.8 km spatial resolution. Level 3 product includes daily and monthly AOD with a spatial resolution of 1°× 1°. In this paper, monthly AOD was processed for cross-comparison with AVHRR-retrieved results.
MERRA-2 is a long time series of reanalysis data including various meteorological variables such as net radiation, temperature, relative humidity, wind speed, etc. It is released by NASA's Global Modeling and Assimilation Office in 2016. MERRA-2 AOD reanalysis product can provide AOD from 1980 to the present every three hours and have a certain accuracy (Aldabash, Balcik, and Glantz 2020;Zhang et al. 2020). MERRA-2 AOD was used for cross-comparison with AVHRR AOD in time series.
AERONET is a global ground-based atmospheric aerosol observation network. Its long-term observation data have been widely used in aerosol research and characterization, satellite data validation, etc. CARSNET is also the network created by China which has been proven to have the same accuracy as AERONET (Xie et al. 2011) and has more stations in China. They can provide accurate AOD at 1020, 870, 670, 440 nm, and Angstrom exponent which can calculate AOD at other wavelengths. Ground-based stations used in the paper are illustrated in Figure 1. Due to the dearth of long-term observation sites in China, some AERONET sites around China are also chosen for the verification of the results.

Study area
The study region with the extent of 15°-60°N, 70°-140°E over China and surrounding areas were selected to apply the algorithm as Figure 1 shows. This region has complex topography and various land cover types. The complex surface and atmospheric conditions and large altitude spans in the study area require high applicability of the algorithm.

Li-Qin model
Li-Qin model is a forward atmospheric radiative transfer model considering non-Lambert surface conditions. It can simulate atmospheric radiation transfer in the ultraviolet to near-infrared spectral range (Li et al. 1996;Qin, Herman, and Ahmad 2001). Lin-Qin model is expressed as follows: In the above, r TOA represents the spectral reflectance at the top of the atmosphere after atmospheric and surface reflection. r a is path reflectance reflected by atmospheric aerosols and molecules. t means AOD, m s and m v are angle-dependent variables denoting the cosine of solar zenith angle and satellite zenith angle respectively. t d refers to diffuse transmission and S is the spherical albedo of the atmosphere. Li-Qin model is based on the four flux theory and uses four albedo parameters r, r DHR , r HDR and r BHR to express surface BRDF property. r is bi-directional surface reflectance, that is, the direct reflection of the incoming direct sunlight. r DHR is directional-hemispherical reflectance (DHR), which describes the diffuse reflection of the incoming direct sunlight across the hemispherical. r HDR is hemispherical-directional reflectance (HDR) which describes the direct reflection of the incident sunlight from the entire hemispherical. r BHR is bi-hemispherical reflectance. Li-Qin Model is highly accurate in the ultraviolet to near-infrared spectrum and the average relative difference for different surface types is less than 0.7% (Qin, Herman, and Ahmad 2001). 3.1.2 RTLS model and BRDF shape For describing the surface BRDF effect, we adopted Ross-Thick-Li-Sparse (RTLS) model to simulate surface directional reflectance (Wanner, Li, and Strahler 1995). RTLS is a semi-empirical kernel-driven model and it has been applied in the MODIS MCD43 product (Schaaf et al. 2002).
It provides bi-directional reflectance R(u i , u v , w, l) (defined as r in Eq. (1)) as a function of three parameters and three geometric kernels: a 0 , a 1 , a 2 are coefficients and K iso , K vol , K geo are kernels depended on observation conditions and they can be calculated by (Strahler 1999). u i , u v , w separately refer to the zenith direction of light incoming and leaving and their relative azimuth. l represents wavelength but the kernels are wavelength-independent. At the same time, the other three reflectance parameters defined in Equation (1) can be calculated using the RTLS model, which can be found in (He et al. 2012;She et al. 2019).
To simplify the expression of BRDF parameters, (2) can be rewritten as: where and p 1 (l) = a 1 (l)/a 0 (l) p 2 (l) = a 2 (l)/a 0 (l) Eq. (4) is defined as BRDF shape which is stable and varies little in the same period of the year (Li et al. 2017). Now, the BRDF effect can be expressed with one parameter a 0 (l). BRDF shape was generated by the MCD43C1 product. The parameters where quality assessment is better than 2 were selected for averaging on a monthly scale. To obtain a more effective inversion, we use the results of 5-year data averages to fill in the missing values. To test the stability of BRDF shape over long time series, we have compared the effects of BRDF shape for surface bidirectional reflectance and found that the reflectance varies little under different observation conditions between 2001 and 2016. Here, we applied monthly averaged a 0 (l) to simulate the surface bidirectional reflectance of these two years. As illustrated in Figure 2, the simulated reflectance difference between 2001 and 2016 is minor when the zenith angle and azimuth angle of the sun and satellite are small, and big when these angles are large. At the same time, there is a large seasonal variance in the simulated reflectance difference, with a modest difference in July and a large difference in January. However, the simulated bidirectional reflectance is consistent each month and the root-meansquare errors of reflectance of the entire period are lower than 7% under different observation conditions. Therefore, 2001 year's BRDF shape parameters were applied for AVHRR inversion work before 2001.

Aerosol model
MODIS DT algorithm clustering AERONET relative parameters obtained four aerosol types: moderately absorbing, absorbing, non-absorbing, and spheroid aerosol models, together with the continental model (Levy, Remer, and Dubovik 2007), has a good performance in a long time series (Wei et al. 2019a;Che et al. 2019). Levy, Remer, and Dubovik (2007) determined seasonal 1°×1°maps of expected aerosol types and we directly use the five aerosol models and their seasonal distributions used by MODIS DT product for AVHRR AOD inversion in this study.

Retrieval strategy
In this study, surface reflectance is not directly estimated to retrieve AOD, but surface parameters and AOD are simultaneously estimated by an iteration process. Here, the 0.63μm channel observations were used to retrieve AOD. For each pixel, there are two parameters to be solved with only one observation. For more information, we selected adjacent two days and multi-pixels to jointly calculate based on two reasonable assumptions: (1) the surface property doesn't vary in two days, which means a 0 (l) in formula (3) doesn't change in the period.
(2) AODs are the same in a certain space with N pixels. So, we have 2×N observations and 2 + N unknowns. Obviously, when N≥2, we can achieve robust retrieval relatively. In this paper, we applied N = 2, that is, in the 2×1 strip range (10×5 km 2 ) for AVHRR AOD inversion. In order to solve the above process, we used the optimal estimation method to join the iteration Wagner, Govaerts, and Lattanzio 2010;She et al. 2019). It minimizes the cost function to obtain the optimal solution and the cost function is shown as follows: Where, J(x) is the cost function. x represents the parameters to be solved that are a 0 (l) and t in this research and x 0 denotes initial values for iteration. y 0 means TOA reflectance obtained from AVHRR channel 1 observation and F(x) represents the TOA reflectance simulated by Li-Qin model. Both S y and S a are error covariance matrices, represent the error matrix of AVHRR observation and initial guesses, respectively. For AVHRR channel 1, each observation and adjacent pixels are considered to be independent. Therefore, S y is defined as a diagonal matrix with 0.04 diagonal elements. S a is also defined as a diagonal considering that AOD and surface property a 0 (l) are independent. The initial AOD is set to a fixed value of 0.2 and a 0 (l) is set as the MODIS BRDF product's monthly average. The value of the diagonal S a is determined by the following three parts: (1) 10% uncertainty of MODIS BRDF product.
(2) the standard deviation of a 0 (l) each month (She et al. 2019) and the maximum standard deviation over the years applied to data before 2001.
(3) a large value to ensure that prior knowledge provides weak constraints.
To minimize the cost function, we used Levenberg-Marquardt method for iterative solutions. The whole process of the algorithm is illustrated in Figure 3.

Result
The algorithm was applied to AHVRR GAC data and the AOD results were obtained from 1982 to 2016 over the study area. To evaluate the accuracy of AVHRR-retrieved AOD, we used AERONET and CARSNET AOD for direct validation. In order to verify the validity of the AVHRR-retrieved AOD over a long period, we compared the algorithm results with AVHRRDB and MERRA-2 products.

Inversion results
The annual mean of AVHRR-retrieved AOD maps from 1982 to 2016 is shown in Figure 4. It can be seen that AOD increased significantly after 2000, which is closely related to China's economic development. After 2013, there was an obvious decline. The Chinese government implemented the air cleaning plan in 2013, which made a significant improvement (Chen et al. 2018). We can see that the AOD in North China, East China, and South China, Sichuan Basin, and Taklimakan Desert has an obvious high value. This is basically consistent with the research results of Filonchyk et al. (2019).

Validation by ground-based sites
The ground-based sites we used for validation are illustrated in Figure 1. The AERONET AOD products selected that are recorded during the study period in the research area are Level 1.5 (select Level 2.0 data if it exists. Level 1.5 product is cloud screened and Level 2.0 is cloud cleared and quality controls) and CARSNET AOD data are all from 2008. To evaluate our algorithm results, we used the correlation coefficient (R), root-mean-square error (RMSE), mean absolute error (MAE), and expect error (EE) index.
The time and space matching of AOD of the ground-based site and AVHRR inversion were carried out: AERONET and CARSNET AOD were averaged within ±30 min of the satellite observation time, and the AVHRR-derived AODs were averaged at a 25 × 25 km spatial box centered on the ground-based stations. Figure 5 shows the scatter diagram of AOD matching results of the satellite-retrieved and two ground-based observation networks. The two blue lines indicate EE of ± (0.05 + 0.25τ) following the comparison of AVHRRDB V001 dataset with AERONET observations over land (Hsu et al. 2017). The red one is the standard line y = x. As shown in Figure 5, the correlation between AVHRR AOD retrieved in this paper and ground measurement AOD is 0.78. With all matched 6713 points, 62.62% points are within EE envelope, 21.35% points are below EE and 16.03% points are above EE, indicating a trend of underestimation which is similar to the results of AVHRRDB in this study area (Che et al. 2018;Mei et al. 2019). We think that this may be the system error from the red channel measurements. Since the algorithm results are more overestimated when AOD is low, and more seriously, there are more underestimates when AOD is high, the algorithm has a little high RMSE of 0.26 and MAE of 0.16. However, we surprisingly found that there was no significant difference in the EE index from year to year, which indicated that the algorithm was independent and stable in time. In particular, the validation results of 55 ground observations in mainland China showed that 63.31% points are within EE, with RMSE = 0.26 and MAE = 0.17. This is similar to the results verified by all sites (63 sites in total) in the entire study area. Compared with the performance of AVHRRDB AOD over mainland China which has less than 55% of the points within EE lines and RMSE greater than 0.25 (Che et al. 2018;Mei et al. 2019), this algorithm result has more percentage points falling in EE and smaller RMSE.

Comparison with AVHRRDB and MERRA-2 products
Due to the paucity of ground site data for the results prior to 1997, we compared MERRA-2, AVHRRDB, and AVHRR-retrieved AOD for a wide time range in order to confirm their accuracy and rationality. Four AERONET station locations are selected, Beijing, Dalanzadgad, NCU_Taiwan, and Kanpur. Figure 6 shows monthly mean variations of AVHRR-retrieved AOD, AVHRRDB AOD, MERRA-2 AOD, and AERONET AOD at selected AERONET stations. In Beijing and Kanpur sites, the variation tendency of AVHRR-retrieved AODs before 2000 is basically the same as MERRA-2 AODs, and the results of AVHRRDB are relatively high. However, AVHRR-retrieved AOD after 2000 is relatively low compared with the other three products. In the Dalanzadgad site, the four datasets have good consistency in time, while AVHRRDB results have obvious overestimation in some months. In NCU_Taiwan site, 1982-1989, the spectrum of AVHRR-retrieved AOD is higher in some months compared to MERRA-2 AOD. Results after 2000 show that AVHRRDB results are mostly underestimated. Overall, AVHRR-retrieved AOD has a good performance in the time series.
To verify the robustness of the algorithm, we have made the following analysis on the annual mean of AVHRRDB AOD, MERRA-2 AOD, and AVHRR-retrieved AOD: (1) We have compared the correlation of annual mean AOD of the three products in the study region. As shown in Figure  7, the correlation coefficient of AOD in almost all years is greater than 0.5, and there is a good correlation between MERRA-2 and AVHRR AOD in the whole time, and MERRA-2 and AVHRRDB AOD also show a good relationship, especially from 2006 to 2010. However, AVHRRDB and AVHRR AOD show a relatively low correlation. (2) We have calculated the percentage of annual averaged AOD falling in EE lines for each of the two products. AVHRR-retrieved results and AVHRRDB AOD are consistent with MERRA-2 AOD with 80.01% and 75.36% percentage within EE as shown in Figure 8. We also found that the consistency between AVHRRDB AOD and AVHRR-retrieved AOD is low, and only 67.29% of points are within EE. Affected by factors such as the coverage of algorithm results and the existence of clouds, the above comparison is uncertain. However, it still shows the rationality and robustness of the algorithm in time on the whole.

Discussions and conclusions
Atmospheric aerosols not only cause environmental pollution, but also affect the earth's climate. To obtain long-term AOD from AVHRR GAC data over land, the manuscript presents a new algorithm based on assumptions that aerosol in the design window has the same AOD and the land surface property doesn't change in two days. The algorithm uses Li-Qin model to simulate the solar radiation received by the satellite, and the RTLS model and BRDF shape generated from MODIS BRDF/albedo product are used to simulate the surface BRDF effect. In addition, the stability of BRDF shape over a long period is also confirmed. The aerosol model defined in the MODIS DT algorithm is used in this paper. Finally, the AODs of AVHRR were obtained through an iteration process with the OE method.
The algorithm results were evaluated by comparison with the ground-based AERONET and CARSNET measurements, and crossly compared AVHRRDB and MERRA-2 products. The validation with AERONET and CARSNET shows that the algorithm has high accuracy. Moreover, there is a good correlation between AVHRR AOD and MERRA-2 AOD over a long time, and it also shows a certain correlation with AVHRRDB AOD, which illustrates the algorithm has good stability and certain accuracy over a long time. However, the algorithm presents a certain underestimation as a whole.
This long-term historic AOD records can be used to research the climate and environmental changes, especially before 2000. The data recording may also be used to reconstruct the historical  mass concentration of fine particles (PM2.5) and support environmental assessment and pollution prevention and control (Zhang et al. 2021). Moreover, the importance algorithm proposed in the manuscript and the AOD data set are reflected in the monitoring of atmospheric quality, helping to predict atmospheric changes, providing recommendations for atmospheric environment improvement, ensuring people's healthy lives, promoting the sustainability of urban development, which could contribute to the achievement of the United Nations Sustainable Development Goals (Varotsos and Cracknell 2020;Kulmala et al. 2021).
The future study still needs to address several issues. First, atmosphere composition has been changing for more than 30 years, and more adequate aerosol types should be considered. Second, although the BRDF shape generated by the MCD43C1 product is stable in the long-term, it is necessary to evaluate the error of using the BRDF shape as prior information for retrieving AOD before 2000. Third, since the early data lack real values to verify, a more convincing method is needed to evaluate the accuracy of data before 2000. In addition, the algorithm can be applied to AVHRR observation twice per day to improve coverage. In the future, we will focus on the above aspects, evaluate surface reflectance results and attempt to retrieve AOD on a global scale.