Comprehensive Validation and Comparison of Three VIIRS Aerosol Products over the Ocean on a Global Scale

: Three parallel Visible/Infrared Imager Radiometer Suite (VIIRS) aerosol products (SOAR, NOAA, and AERDT) provided data since 2012. It is necessary to study the performances and advantages of different products. This study aims to analyze the accuracy and error of these products over the ocean and compare them with each other. The results show that the three VIIRS ocean aerosol retrievals (including total aerosol optical depth (AOD), ﬁne mode fraction, Ångström exponent (AE), and ﬁne AOD (AODF)) correlate well with AErosol RObotic NETwork (AERONET) retrievals (e.g., correlation >0.895 for AOD and >0.825 for AE), which are comparable to the newest moderate-resolution imaging spectro-radiometer (MODIS) retrievals. Overall, the SOAR retrievals with quality ﬁltering have the best validation accuracy of all parameters. Therefore, it is more recommended to use. The differences in the annual AOD spatial patterns of different products are small (bias < 0.016), but their AE spatial patterns are evidently different (bias > 0.315), indicating the large uncertainty of VIIRS AE. Error analysis shows that the scattering angle and wind speed affect aerosol retrieval. Application of the non-spherical dust model may reduce the dependence of retrieval bias on the scattering angle. Overall, this study provides validation support for VIIRS products usage and possible algorithm improvements.


Introduction
Aerosols, as an important component of the atmosphere, influence the cloud property, Earth's radiation balance, and global climate change [1][2][3][4]. Excessive aerosols affect human health and life as pollutants (such as PM 2.5 ) [5]. With the development of remote sensing technology, satellite aerosol observation from space has been widely used in related scientific research [6].
The ocean surface reflectance is stable and low, which is suitable for aerosol measurements using satellites [7,8]. This is because the aerosol signal accounts for a higher proportion of satellite-observed reflectance at a low surface reflectance. For example, Geogdzhayev et al. [9] developed a two-channel aerosol retrieval algorithm over the ocean and applied it to global Advanced Very High-Resolution Radiometer (AVHRR) observations for retrieving aerosol optical depth (AOD) and Ångström exponent (AE). Preliminary validation results of the AOD retrieved by the algorithm show that it correlates well with the AOD measured by the sun photometer [10]. Owing to the good and stable radiation performance of sea-viewing wide field-of-view sensor (SeaWiFS) and moderate-resolution imaging spectro-radiometer (MODIS) sensors, it is reliable to use them to retrieve aerosol Remote Sens. 2022, 14,2544 3 of 17 from that of MODIS. However, because of the difference between the spectral response functions of VIIRS and MODIS sensors, parameters such as atmospheric water vapor, ozone correction coefficients, and Rayleigh optical depth of the VIIRS spectral bands need to be re-derived [25]. The AERDT algorithm takes as input the 750 m resolution TOA reflectance of 6-minute granule data processed by NASA. Then unsuitable pixels (such as clouds, ice, ocean glints, and land) are removed. A 6 × 6 spatial window is used to aggregate pixels to a coarser resolution for a higher signal-to-noise ratio and better retrieval efficiency. In the 6 × 6 box, the algorithm removes the darkest and brightest 25% of pixels in the 0.87 µm band. Additionally, it takes the average of the remaining pixels as the reflectance of the box, which is an approximately 6-km resolution. The box reflectance will be trace gas corrected to obtain the final TOA reflectance used for the aerosol retrieval. The appropriate TOA spectral reflectance (0.55, 0.67, 0.87, 1.24, 1.61, and 2.25 µm) is compared with the spectral reflectance calculated by the look-up table, and the average of the acceptable retrieval solutions is taken to obtain the AOD and FMF, and then the AE is calculated. The spatial resolution of all parameters of the AERDT product is 6 km ( Table 1).

NOAA Aerosol Algorithm and Product
NOAA's ocean aerosol algorithm, described in Jackson et al. [16], is very similar to that of the AERDT. They share similar retrieval strategies and aerosol model parameters, but there are some differences. The NOAA algorithm does not retrieve aerosols over large inland lakes. It replaces the 0.55 µm band used by the AERDT algorithm with the 0.746 µm band. Additionally, it reports the retrieval of the optimal solution instead of the average of the acceptable solutions. The upper limit of AOD retrieval of the NOAA algorithm is 2.0, while that of the AERDT algorithm is 5.0. Their pixel aggregation methods are different. NOAA algorithm first retrieves 750 m resolution aerosol properties and then aggregates them to a 6-km coarse resolution (Table 1).

SOAR Aerosol Algorithm and Product
The SOAR algorithm used by the AERDB product is described in Sayer et al. [18], which differs from the NOAA and AERDT algorithms. It has two retrieval schemes, a full retrieval, and a backup retrieval, depending on whether the ocean surface is turbid. In the open ocean, SOAR uses the full retrieval based on seven bands (0.49, 0.55, 0.67, 0.87, 1.24, 1.61, and 2.25 µm), while on turbid water surfaces it uses the backup retrieval based on the four near-infrared bands (0.87, 1.24, 1.61, and 2.25 µm). SOAR's aerosol model is based on internal limited mixing of the coarse and fine modes, i.e., limiting FMF to a reasonable range ( Table 2 in [18]), rather than the free mixing approach used by AERDT and NOAA algorithms. SOAR uses a non-spherical dust optical model instead of the spherical one used by AERDT and NOAA [26]. The improvement of the dust model has a large impact on the retrieval results, which will be discussed later. In terms of retrieval strategy, SOAR uses the optimal estimation theory, with AOD and FMF as free parameters, and its initial value is determined by the AOD and FMF nodes in the look-up table with the smallest residual error. When the spectral AOD and FMF are determined, then AE is calculated using 550 and 870 nm AOD. SOAR first retrieves aerosol properties at a 750 m resolution and then aggregates them to a 6-km resolution (Table 1).

MODIS Aerosol Data
MODIS aerosol products (MOD04 and MYD04) over the ocean have been widely validated and used (e.g., [12,27,28]). Therefore, this study used MYD04 data as a comparison to characterize the relative accuracy of VIIRS aerosol products. MYD04 is chosen instead of MOD04 because MYD04 has better validation accuracy and is close to the VIIRS imaging time [29]. MYD04 also records AOD, AE, and FMF properties, but at a 10-km resolution.

AERONET Data
The uncertainty of direct-solar AOD observations from AERONET is small (~0.01 in visible), so it serves as a ground benchmark to validate satellite retrievals [30][31][32]. The newly released version 3 data applies automated cloud screening and instrument anomaly quality control technique, which greatly speeds up the production of level 2 data [33]. In this study, level 2 AERONET AOD and AE data of version 3 are used. AERONET Spectral Deconvolution Algorithm (SDA) and standard inversion products provide FMF parameters [34,35]. Because of the better temporal sampling of SDA FMF, it is used to compare satellite FMF and AODF retrieval [36]. Note that the uncertainty of the SDA FMF is approximately 0.1 [34].

Validation Method
All AERONET sun photometers provide AOD at 440, 675, 870, and 1020 nm, some instruments provide other additional wavelengths (e.g., 340, 380, and 500 nm). However, none of them provide 550 nm AOD, instead, 550 nm AOD is the most commonly used by satellite retrieval. Therefore, the quadratic log interpolation method was used to calculate the AERONET 550 nm AOD from the other wavelengths of AOD using Equation (1) [20,30,36]. Then, the 550 and 870 nm AODs are used to calculate the AE, which is for direct comparison with the satellite-retrieved AE. AERONET SDA product only provides 500 nm FMF. For direct comparison with FMF and AODF retrievals from VIIRS, the SDA total and fine mode AOD is interpolated to 550 nm using corresponding AE, and then the 550 nm FMF is calculated using Equation (2) [20]. Since the spatiotemporal sampling of AERONET retrieval and satellite retrieval is inconsistent, they require spatiotemporal matching. In this study, a temporal window of ±30 min is used to calculate the median AERONET retrieval, and a spatial window of ±25 km is used to calculate the median satellite retrieval. Note that only the satellite retrievals with quality filtering are used for comparison. The minimum number of AERONET values allowed within the window is one, and the minimum fraction of valid satellite values is 20% [17,20]. log(τ λ ) = a 0 + a 1 log(λ) + a 2 log(λ) 2 (1) where τ λ is the AOD at wavelength λ; a 0 , a 1 , and a 2 are the fitting coefficients, which are commonly calculated by AOD at 440, 500, 675, and 870 nm using each AERONET record.
f 550 = τ f 550 /τ t 550 (2) where τ f 550 , τ t 550 , and f 550 are fine mode AOD, total AODT, and FMF at 550 nm, respectively. The metrics used to evaluate VIIRS ocean aerosol products include: (1) the number (N) is the volume of matching pairs. (2) the correlation coefficient (R) is used to characterize the variation consistency between satellite retrieval and sun photometer retrieval. (3) Mean bias (MB) is used to characterize the overall overestimation or underestimation of satellite retrieval. (4) Mean absolute error (MAE) and root mean square error (RMSE) are used to characterize the overall error of satellite retrieval. Note that MAE and RMSE are related to the level of comparison values and RMSE is easily affected by outliers. (5) The expected error is used to characterize the fraction of satellite retrieval that meets the expected accuracy, and it is generally believed that the fraction should be greater than 68% (i.e., one standard deviation). Equations (3)-(5) represent the expected errors for the AOD (EE), AE (EE_AE), and FMF (EE_FMF), respectively. The Global Climate Observing System (GCOS) provides the expected accuracy of AOD for long-term climate change research as Equation (6).

Aerosol Optical Depth
The validation results of the total AOD are shown in Figure 1. SOAR AOD metrics (EE:0.681 and GCOS:0.545) are similar to those of Sayer et al. [20] (EE:0.66 and GCOS:0.49-0.52). The NOAA AOD RMSE (0.075) is slightly larger than the result (0.065) of Huang et al. [19]. The EE of MYD04 AOD in the study of Levy et al. [12] is higher than that of our result (Figure 1d), which is owing to their use of a more relaxed EE formula. The ocean AOD of all products has a good correlation with the AERONET AOD with the R of 0.895-0.917. SOAR AOD has the best performance with smaller errors (MAE and RMSE) and higher EE and GCOS fractions. The fraction of SOAR AOD falling on EE is 0.681, which meets the goal of EE (~0.68). The fraction within EE of NOAA AOD is 0.634. Both of them have better accuracy than MYD04 AOD (EE of 0.607). In contrast, AERDT using the same algorithm as MYD04 has lower accuracy than MYD04 AOD. Overall, except for AERDT AOD, SOAR and NOAA AOD are equal to or better than the MYD04 AOD level.

Ångström Exponent
AE utilizes the different responses of two wavelengths to aerosol to characterize the size of aerosol particles. This parameter is useful for distinguishing between fine and coarse dominant aerosol types. Therefore, the AE parameter is also the focus of this study. When the AOD loading is low, there is a non-negligible uncertainty in the calculated AE; therefore, the AE parameters are compared only when the AERONET 550 nm AOD is >0.2.
The validation results of AE are shown in Figure 2. Since only the AE with AOD > 0.2 is compared, validations for SOAR, NOAA, and MYD04 AE are better than previous studies [12,18,19]. Although each product captures the variation in AE, i.e., R > 0.825, they all overestimate low AE and underestimate high AE. The same problem was also found in the previous validation of MODIS AE (e.g., [37]). Zhou et al. [38] proved that using a non-spherical dust model can reduce the overestimation of low AE and FMF. Fortunately, most of their retrievals are within the EE_AE (>0.760), and their RMSEs are all smaller than EE_AE (defined as ±0.4), indicating that their AE retrievals are quantitatively meaningful. From the two metrics of EE_AE and RMSE, the AEs of SOAR and AERDT are better than that of MYD04, while the AE of NOAA is slightly worse than that of MYD04. Remote Sens. 2022, 14, x FOR PEER REVIEW 6 of 18

Ångström Exponent
AE utilizes the different responses of two wavelengths to aerosol to characterize the size of aerosol particles. This parameter is useful for distinguishing between fine and coarse dominant aerosol types. Therefore, the AE parameter is also the focus of this study. When the AOD loading is low, there is a non-negligible uncertainty in the calculated AE; therefore, the AE parameters are compared only when the AERONET 550 nm AOD is >0.2.
The validation results of AE are shown in Figure 2. Since only the AE with AOD > 0.2 is compared, validations for SOAR, NOAA, and MYD04 AE are better than previous studies [12,18,19]. Although each product captures the variation in AE, i.e., R > 0.825, they all overestimate low AE and underestimate high AE. The same problem was also found in the previous validation of MODIS AE (e.g., [37]). Zhou et al. [38] proved that using a non-spherical dust model can reduce the overestimation of low AE and FMF. Fortunately, most of their retrievals are within the EE_AE (>0.760), and their RMSEs are all smaller than EE_AE (defined as ±0.4), indicating that their AE retrievals are quantitatively meaningful. From the two metrics of EE_AE and RMSE, the AEs of SOAR and AERDT are better than that of MYD04, while the AE of NOAA is slightly worse than that of MYD04. The definitions of VIIRS-retrieved FMF and AERONET-retrieved FMF differ slightly, but they are related and can be compared with each other [13] (Figure 3). The comparative metrics (Figure 3a) of SOAR FMF are similar to previous studies (e.g., [18,20]). Except for

Fine Mode Fraction and AOD
The definitions of VIIRS-retrieved FMF and AERONET-retrieved FMF differ slightly, but they are related and can be compared with each other [13] (Figure 3). The comparative metrics (Figure 3a) of SOAR FMF are similar to previous studies (e.g., [18,20]). Except for NOAA FMF, the performance of SOAR and AERDT FMF is close to or better than that of MYD04 FMF. The SOAR algorithm is better at retrieving low FMF cases, while other algorithms overestimate low FMF. This suggests that NOAA, AERDT, and MYD04 using similar algorithms (spherical dust aerosol model) tend to increase the fraction of fine modes to fit satellite-observed reflectance. Using non-spherical dust models may improve their FMF retrieval [38]. Overall, the FMFs of all algorithms are quantitatively meaningful, i.e., they capture variations in FMF (R > 0.789), and the EE_FMF is >0.713. A related parameter to FMF is the fine mode AOD (Equation (2)), which can be used to estimate anthropogenic radiative forcing [14,15]. The accuracy of SOAR AODF is very high (e.g., R:0.872, EE:0.803, GCOS:0.727). This indicates that it can be applied to scientific research. Conversely, the scatter plot (Figure 4b-d) and EE metrics show that the NOAA, AERDT, and MYD04 algorithms significantly overestimate AODF. The overestimated total AOD ( Figure 1) and FMF ( Figure 3) together lead to the overestimation of AODF. Therefore, these algorithms need to be improved in the future. A related parameter to FMF is the fine mode AOD (Equation (2)), which can be used to estimate anthropogenic radiative forcing [14,15]. The accuracy of SOAR AODF is very high (e.g., R:0.872, EE:0.803, GCOS:0.727). This indicates that it can be applied to scientific research. Conversely, the scatter plot (Figure 4b-d) and EE metrics show that the NOAA, AERDT, and MYD04 algorithms significantly overestimate AODF. The overestimated total AOD ( Figure 1) and FMF ( Figure 3) together lead to the overestimation of AODF. Therefore, these algorithms need to be improved in the future.

Quantitative Comparison
To reduce the impact of sample differences, the same matchups (same site and same time) of all VIIRS retrievals with AERONET retrievals are used to discuss the relative accuracy of different VIIRS products ( Figure 5). The AOD accuracy of SOAR and NOAA is similar and they are close to the 1:1 line. While the AERDT systematically overestimates the AOD, it remains within the expected error envelope. AE and FMF comparison results (Figure 5b,c) show similar product differences, i.e., NOAA and AERDT overestimated low parameter values, which is consistent with the overall validation results (Figures 2 and 3). In contrast, the AE and FMF of SOAR are close to the 1:1 line. For the AODF parameter, SOAR underestimates high AODF. Interestingly, the AODFs of AERDT and NOAA almost overlap, they both overestimate the low AODF and underestimate the high AODF. Overall, the comparison results show that SOAR has the best performance on all parameters.

Quantitative Comparison
To reduce the impact of sample differences, the same matchups (same site and same time) of all VIIRS retrievals with AERONET retrievals are used to discuss the relative accuracy of different VIIRS products ( Figure 5). The AOD accuracy of SOAR and NOAA is similar and they are close to the 1:1 line. While the AERDT systematically overestimates the AOD, it remains within the expected error envelope. AE and FMF comparison results (Figure 5b,c) show similar product differences, i.e., NOAA and AERDT overestimated low parameter values, which is consistent with the overall validation results (Figures 2 and 3). In contrast, the AE and FMF of SOAR are close to the 1:1 line. For the AODF parameter, SOAR underestimates high AODF. Interestingly, the AODFs of AERDT and NOAA almost overlap, they both overestimate the low AODF and underestimate the high AODF. Overall, the comparison results show that SOAR has the best performance on all parameters.

Aerosol Optical Depth
To understand the differences in spatial patterns of different VIIRS total AOD, a pixel-by-pixel comparison is analyzed (Figures 6-8). The three VIIRS products were resampled to a daily mean at 1 • spatial resolution for pixel-by-pixel matching. As can be seen in Figure 6a, the aerosols around the coast are heavily influenced by the land aerosols, and the AOD is higher than 0.15, while in the open ocean, the AOD is less than 0.15. In general, the AOD spatial patterns of SOAR and NOAA have little difference (MB = 0.001), while the AOD spatial patterns of AERDT and the other two products have larger differences. Their biases (−0.014 for SOAR and AERDT, 0.016 for AERDT and NOAA) are approximately half of the SOAR annual mean (0.028). Figure 7 shows that correlations between different AOD retrievals are high (R > 0.85) in most marine areas, especially in high AOD loading areas (R > 0.9). In regions under low AOD loading, their Remote Sens. 2022, 14, 2544 9 of 17 correlations are slightly reduced, which is understandable because the AOD variations in these regions are small. Figure 8 shows that when the latitude is <±40 degrees, the retrieval frequencies of the three products are similar, and their retrieval frequency ratio is approximately 0.9-1.1. In coastal zones, SOAR had better retrieval frequencies, possibly due to its backup algorithm retrieving a portion of turbid water [18]. With increasing latitude, AERDT has the highest retrieval frequency, followed by SOAR, and NOAA has the lowest retrieval frequency. When all quality AODs instead of high-quality AODs are used to calculate the retrieval frequency (not shown), SOAR outperformed that of AERDT except above 65 degrees north latitude and except above 60 degrees south latitude, NOAA's retrieval frequency exceeds that of the SOAR, indicating that SOAR and NOAA adopts a stricter quality control approach.

Aerosol Optical Depth
To understand the differences in spatial patterns of different VIIRS total AOD, a pixel-by-pixel comparison is analyzed (Figures 6-8). The three VIIRS products were resampled to a daily mean at 1° spatial resolution for pixel-by-pixel matching. As can be seen in Figure 6a, the aerosols around the coast are heavily influenced by the land aerosols, and the AOD is higher than 0.15, while in the open ocean, the AOD is less than 0.15. In general, the AOD spatial patterns of SOAR and NOAA have little difference (MB = 0.001), while the AOD spatial patterns of AERDT and the other two products have larger differences. Their biases (−0.014 for SOAR and AERDT, 0.016 for AERDT and NOAA) are approximately half of the SOAR annual mean (0.028). Figure 7 shows that correlations between different AOD retrievals are high (R > 0.85) in most marine areas, especially in high AOD loading areas (R > 0.9). In regions under low AOD loading, their correlations are slightly reduced, which is understandable because the AOD variations in these regions are small. Figure 8 shows that when the latitude is <±40 degrees, the retrieval frequencies of the three products are similar, and their retrieval frequency ratio is approximately 0.9-

Ångström Exponent
The AE spatial pattern of SOAR is shown in Figure 9a. In the oceans close to the mainland, high AE indicates fine-mode-dominated aerosols in these regions. In most oceans, the MB of SOAR AE and AERDT, and NOAA AE are negative, which is consistent with previous studies [20]. The AE pattern MB (0.062) of AERDT and NOAA is small, which is related to their use of similar algorithms (Section 2.3). This does not mean that they produce more accurate AE retrievals. Because the overall validation result (Figure 5b) shows that compared with AERDT and NOAA algorithms, SOAR AE has a better validation performance, which implies that the spatial pattern of SOAR AE may be more accurate. The correlation coefficients (Figure 10) of SOAR AE and other algorithm AEs are related to the annual mean of AE, that is, in regions with high AE, their correlation coefficients are high, and in regions with low AE, their correlation coefficients are low. This is because the open ocean is rarely polluted by aerosols from the land, and therefore, AE variations are slight. For the AE parameter, the serious differences between different algorithms indicate that the uncertainty of AE retrieved by VIIRS is large, and more validations are needed to improve the algorithm in the future.

Å ngström Exponent
The AE spatial pattern of SOAR is shown in Figure 9a. In the oceans close to the mainland, high AE indicates fine-mode-dominated aerosols in these regions. In most oceans, the MB of SOAR AE and AERDT, and NOAA AE are negative, which is consistent with previous studies [20]. The AE pattern MB (0.062) of AERDT and NOAA is small, which is related to their use of similar algorithms (Section 2.3). This does not mean that they produce more accurate AE retrievals. Because the overall validation result ( Figure  5b) shows that compared with AERDT and NOAA algorithms, SOAR AE has a better validation performance, which implies that the spatial pattern of SOAR AE may be more

Ångström Exponent
The AE spatial pattern of SOAR is shown in Figure 9a. In the oceans close to the mainland, high AE indicates fine-mode-dominated aerosols in these regions. In most oceans, the MB of SOAR AE and AERDT, and NOAA AE are negative, which is consistent with previous studies [20]. The AE pattern MB (0.062) of AERDT and NOAA is small, which is related to their use of similar algorithms (Section 2.3). This does not mean that they produce more accurate AE retrievals. Because the overall validation result ( Figure  5b) shows that compared with AERDT and NOAA algorithms, SOAR AE has a better validation performance, which implies that the spatial pattern of SOAR AE may be more

Ångström Exponent
The AE spatial pattern of SOAR is shown in Figure 9a. In the oceans close to the mainland, high AE indicates fine-mode-dominated aerosols in these regions. In most oceans, the MB of SOAR AE and AERDT, and NOAA AE are negative, which is consistent with previous studies [20]. The AE pattern MB (0.062) of AERDT and NOAA is small, which is related to their use of similar algorithms (Section 2.3). This does not mean that they produce more accurate AE retrievals. Because the overall validation result ( Figure  5b) shows that compared with AERDT and NOAA algorithms, SOAR AE has a better coefficients are high, and in regions with low AE, their correlation coefficients are low. This is because the open ocean is rarely polluted by aerosols from the land, and therefore, AE variations are slight. For the AE parameter, the serious differences between different algorithms indicate that the uncertainty of AE retrieved by VIIRS is large, and more validations are needed to improve the algorithm in the future.   Figure 11 shows the dependence of AOD bias (AODVIIRS-AODAERONET) on imaging angles and sea surface wind speed (Table 1). Almost all AOD biases are positive, indicating a systematic overestimation of AOD. As the solar zenith angle increases, the AOD bias decreases. One potential reason is that low latitudes (usually small solar zenith angles) have high AOD loadings and complex sources (Figures 6a and 9a), while high latitudes (usually large solar zenith angles) have low AOD loadings. This leads to a larger AOD bias for small solar zenith angles. Another potential reason is that the solar zenith angle (3D radiative effect) affects the selection of the non-spherical model and AOD retrieval in the algorithm [39,40]. When the scattering angle is <135°, all VIIRS products show similar AOD bias; however, when the scattering angle is >135°, their AOD bias is very different, suggesting that the scattering angle (>135°) may have a large influence on the ocean AOD retrieval. According to previous studies (e.g., [26,38]), this AOD bias can be attributed to the scattering phase function of the dust aerosol model. The dust model coefficients are high, and in regions with low AE, their correlation coefficients are low. This is because the open ocean is rarely polluted by aerosols from the land, and therefore, AE variations are slight. For the AE parameter, the serious differences between different algorithms indicate that the uncertainty of AE retrieved by VIIRS is large, and more validations are needed to improve the algorithm in the future.   Figure 11 shows the dependence of AOD bias (AODVIIRS-AODAERONET) on imaging angles and sea surface wind speed (Table 1). Almost all AOD biases are positive, indicating a systematic overestimation of AOD. As the solar zenith angle increases, the AOD bias decreases. One potential reason is that low latitudes (usually small solar zenith angles) have high AOD loadings and complex sources (Figures 6a and 9a), while high latitudes (usually large solar zenith angles) have low AOD loadings. This leads to a larger AOD bias for small solar zenith angles. Another potential reason is that the solar zenith angle (3D radiative effect) affects the selection of the non-spherical model and AOD retrieval in the algorithm [39,40]. When the scattering angle is <135°, all VIIRS products show similar AOD bias; however, when the scattering angle is >135°, their AOD bias is very different, suggesting that the scattering angle (>135°) may have a large influence on the ocean AOD retrieval. According to previous studies (e.g., [26,38]), this AOD bias can be attributed to the scattering phase function of the dust aerosol model. The dust model  Figure 11 shows the dependence of AOD bias (AOD VIIRS -AOD AERONET ) on imaging angles and sea surface wind speed (Table 1). Almost all AOD biases are positive, indicating a systematic overestimation of AOD. As the solar zenith angle increases, the AOD bias decreases. One potential reason is that low latitudes (usually small solar zenith angles) have high AOD loadings and complex sources (Figures 6a and 9a), while high latitudes (usually large solar zenith angles) have low AOD loadings. This leads to a larger AOD bias for small solar zenith angles. Another potential reason is that the solar zenith angle (3D radiative effect) affects the selection of the non-spherical model and AOD retrieval in the algorithm [39,40]. When the scattering angle is <135 • , all VIIRS products show similar AOD bias; however, when the scattering angle is >135 • , their AOD bias is very different, suggesting that the scattering angle (>135 • ) may have a large influence on the ocean AOD retrieval. According to previous studies (e.g., [26,38]), this AOD bias can be attributed to the scattering phase function of the dust aerosol model. The dust model of spherical approximation used by AERDT and NOAA produces scattering phase functions that deviate from the actual shape, while the SOAR using the non-spherical dust model maintains a stable AOD bias when the scattering angle is >135 (Figure 11c). The study by Lee et al. [26] illustrates that improving the dust aerosol model can reduce the dependence of the AOD bias on the scattering angle. However, it is difficult to determine the right non-spherical model at the pixel scale, therefore, most studies are based on several fixed non-spherical models [41]. Figure 11d shows that the sea surface wind speed affects the AOD bias, that is, the greater the wind speed, the greater the bias. Although all VIIRS algorithms account for variations in wind speed, suggesting that none of the algorithms can eliminate the effect of wind speed on retrieval. of spherical approximation used by AERDT and NOAA produces scattering phase functions that deviate from the actual shape, while the SOAR using the non-spherical dust model maintains a stable AOD bias when the scattering angle is >135 (Figure 11c). The study by Lee et al. [26] illustrates that improving the dust aerosol model can reduce the dependence of the AOD bias on the scattering angle. However, it is difficult to determine the right non-spherical model at the pixel scale, therefore, most studies are based on several fixed non-spherical models [41]. Figure 11d shows that the sea surface wind speed affects the AOD bias, that is, the greater the wind speed, the greater the bias. Although all VIIRS algorithms account for variations in wind speed, suggesting that none of the algorithms can eliminate the effect of wind speed on retrieval.  Figure 12 shows the dependence of AE bias. Unlike AOD, the solar zenith angle has little effect on AE bias. However, the AE bias of NOAA and AERDT algorithms shows a dependence on the scattering angle, which is also caused by a scattering phase function that does not match the actual shape of the dust aerosol [38]. The SOAR algorithm uses a non-spherical dust model, so its AE bias does not depend on the scattering angle [26]. Therefore, it is recommended that current algorithms update the non-spherical dust model and aerosol model selection strategy to improve aerosol retrieval. Figure 11d shows that the AE bias of all VIIRS products is positively correlated with sea surface wind speed.  Figure 12 shows the dependence of AE bias. Unlike AOD, the solar zenith angle has little effect on AE bias. However, the AE bias of NOAA and AERDT algorithms shows a dependence on the scattering angle, which is also caused by a scattering phase function that does not match the actual shape of the dust aerosol [38]. The SOAR algorithm uses a non-spherical dust model, so its AE bias does not depend on the scattering angle [26]. Therefore, it is recommended that current algorithms update the non-spherical dust model and aerosol model selection strategy to improve aerosol retrieval. Figure 11d shows that the AE bias of all VIIRS products is positively correlated with sea surface wind speed. Remote Sens. 2022, 14, x FOR PEER REVIEW 14 of 18

Discussion
VIIRS has provided data since 2012 and has three different aerosol algorithms applied to it simultaneously. An issue is whether these products can achieve the accuracy of existing products (such as MODIS) and can be applied scientifically. Another issue is the different performances and advantages of the three VIIRS products. This study uses AERONET ground-based observations to simultaneously validate and compare the total AOD, AE, FMF, and AODF of the three VIIRS aerosol products over the ocean using a consistent approach and data.
The validation results on the global scale show that the VIIRS ocean aerosol retrieval performance is good, reaching and possibly better than that of the MODIS aerosol retrieval in the same period. Liu et al. [42] compared the NOAA ocean AOD and the MODIS Collection 5.1 ocean AOD and also found that their accuracy was comparable. Note that users need to be wary of AERDT's AOD and NOAA's FMF and AE, which may not reach the accuracy of MODIS. Sawyer et al. [17] compared AERDT and MODIS AOD over the ocean and also found that the validation accuracy of the AERDT ocean AOD was slightly worse than that of the MODIS. Overall, SOAR shows the best validation accuracy of the aerosol retrievals (including total AOD, AE, FMF, and AODF) in the three products. Especially, the fraction of SOAR AODF better than the GCOS goal (EE goal) is 0.727 (0.803), which suggests that it meets the requirements of scientific research. Notably, we found that all quality retrievals for SOAR do not filter land pixels fully (not shown), so quality filtering for SOAR is recommended. Conversely, NOAA, AERDT, and MODIS significantly overestimate AODF with above EE > 0.418 (Figure 4). This is unavoidable because they overestimate both total AOD and FMF (Figures 1, 3 and 5). Therefore, the AODF of these algorithms should be used with caution for scientific research. There is a small difference in the spatial pattern of AOD between SOAR and NOAA, which is

Discussion
VIIRS has provided data since 2012 and has three different aerosol algorithms applied to it simultaneously. An issue is whether these products can achieve the accuracy of existing products (such as MODIS) and can be applied scientifically. Another issue is the different performances and advantages of the three VIIRS products. This study uses AERONET ground-based observations to simultaneously validate and compare the total AOD, AE, FMF, and AODF of the three VIIRS aerosol products over the ocean using a consistent approach and data.
The validation results on the global scale show that the VIIRS ocean aerosol retrieval performance is good, reaching and possibly better than that of the MODIS aerosol retrieval in the same period. Liu et al. [42] compared the NOAA ocean AOD and the MODIS Collection 5.1 ocean AOD and also found that their accuracy was comparable. Note that users need to be wary of AERDT's AOD and NOAA's FMF and AE, which may not reach the accuracy of MODIS. Sawyer et al. [17] compared AERDT and MODIS AOD over the ocean and also found that the validation accuracy of the AERDT ocean AOD was slightly worse than that of the MODIS. Overall, SOAR shows the best validation accuracy of the aerosol retrievals (including total AOD, AE, FMF, and AODF) in the three products. Especially, the fraction of SOAR AODF better than the GCOS goal (EE goal) is 0.727 (0.803), which suggests that it meets the requirements of scientific research. Notably, we found that all quality retrievals for SOAR do not filter land pixels fully (not shown), so quality filtering for SOAR is recommended. Conversely, NOAA, AERDT, and MODIS significantly overestimate AODF with above EE > 0.418 (Figure 4). This is unavoidable because they overestimate both total AOD and FMF (Figures 1, 3 and 5). Therefore, the AODF of these algorithms should be used with caution for scientific research. There is a small difference in the spatial pattern of AOD between SOAR and NOAA, which is consistent with a previous study [18]. However, the AOD spatial patterns of AERDT versus NOAA and SOAR show larger differences. Since NOAA and SOAR AODs show better validation accuracies, their AOD spatial patterns are more credible. The AE of SOAR is systematically smaller than that of NOAA and AERDT. Although the difference in AE between AERDT and NOAA is small (MB = 0.062), it does not mean that they retrieve more accurate AEs because SOAR shows a better validation of AE (Figures 2 and 5). Therefore, VIIRS AEs require further validation and analysis. The FMFs and AEs retrieved by VIIRS are correlated [13]; therefore, they generally exhibit the same validation characteristics. For example, AERDT and NOAA overestimated low AE and low FMF ( Figure 5). Conversely, SOAR performs well under low AE and FMF conditions. Using the non-spherical dust model may improve the retrievals of FMF and AE [38]. Proper selection of dust pixels and limited mixing ratios of coarse and fine modes are also possible directions for improvement [18,43].
Error analysis results show that scattering angle and sea surface wind speed have an impact on AOD and AE biases, suggesting potential points for algorithm improvement. For example, the non-spherical dust model is used in the SOAR algorithm; therefore, its retrieval has a smaller dependence on the scattering angle [26]. Previous studies have shown that simply adding non-spherical models to existing algorithms, while tempting, may not solve the problem. This is because the simple modification does not guarantee that the algorithm can stably select a suitable non-spherical model for the retrieval of dust aerosols, or a spherical model for the retrieval of non-dust aerosols [38,41]. Although all VIIRS algorithms consider the sea surface wind speed parameter, retrieval biases are still dependent on it. Previous studies have also found a similar problem [44,45]. The possible reason is that high wind speeds increase the reflectance of glint and foam, further enlarging the reflectance of the ocean surface [44]. Therefore, the aerosol signal accounts for a smaller proportion of satellite-observed reflectance, leading to greater AOD and AE bias. AOD and AE both have a relatively strong relationship with wind speed, which may be another possible reason why the retrieval bias is dependent on wind speed [46].

Conclusions
As the successor to sensors such as MODIS, three algorithm teams have developed different aerosol products to continue long-term MODIS aerosol observations. Over the ocean, all three algorithms can retrieve AOD, AE, FMF, and AODF properties. However, the validation of FMF and AODF is ignored by most studies. They are important and can better determine anthropogenic radiative forcing [1]. The main conclusions of the validation and comparison results in 2018 are: (1) SOAR total AOD retrievals (quality filtering) have a better validation accuracy. After quality and AOD loading filtering, the AE and FMF validation accuracies of the three algorithms are similar, but SOAR may be slightly better. (2) SOAR AODF meets the goal of GCOS (its retrieval stability is also necessary to be studied (e.g., [31])), while the AODF of AERDT and NOAA should be used with caution in the scientific study due to their lower validation accuracy. (3) The differences in AOD spatial patterns of the three VIIRS products are small, but the differences in AE spatial patterns are large. (4) Scattering angle and sea surface wind speed affect the retrieval bias. The suitable non-spherical dust model may reduce the dependence of retrieval bias on the scattering angle. Higher wind speeds result in greater ocean surface reflectance, which results in a reduced sensitivity of satellite-observed reflectivity to aerosol signals. This may explain the dependence of retrieval bias on wind speed. Since this study only used AERONET observation data in 2018 and on coasts and islands, the result may not represent the full view of the performance of these three ocean products. More robust studies will be carried out in the future.
Author Contributions: X.S. and L.F. designed the research; W.L. and X.S. performed the experiments and analyzed the data; W.L. and X.S. wrote the manuscript; L.F., J.W., Y.Z. and M.C. revised the manuscript. All authors have read and agreed to the published version of the manuscript.