HTML
--> --> -->AOD from satellite measurements can provide global AOD trends. A number of studies have assessed tendencies in AOD from several satellite products, such as those from the Advanced Very High Resolution Radiometer (Mishchenko et al., 2007; Zhao et al., 2008), the Sea-viewing Wide Field-of-view Sensor (SeaWiFS; Yoon et al., 2011; Hsu et al., 2012), the Multi-angle Imaging SpectroRadiometer (MISR; Kahn et al., 2007), the Moderate Resolution Imaging Spectroradiometer (MODIS; Yu et al., 2009; Zhang and Reid, 2010; Zhang et al., 2017), and the Along Track Scanning Radiometer (Thomas et al., 2010). The AOD retrieval algorithms, underlying assumptions and uncertainties involved in creating the products differ between the satellite sensors (Petrenko et al., 2012). These AOD products have substantial discrepancies (Kokhanovsky et al., 2007; Mishchenko et al., 2007; Thomas et al., 2010). The discrepancies remain so large that they can hardly be utilized to obtain a consistent and unified AOD trend (Li et al., 2009). Especially on regional scales, the differences can be much larger and more complex. Note that most of the studies mentioned above focused on the AOD trends from satellites over the ocean. Few studies have been carried out on AOD trends using satellite aerosol products over land (Yoon et al., 2011; Hsu et al., 2012), where the aerosol and surface characteristics are more complicated than those over the ocean. (Yoon et al., 2011) and (Hsu et al., 2012) evaluated the linear trends of SeaWiFS AOD using AERONET measurements over a limited number of land and coastal sites. Significant positive trends derived from SeaWiFS AOD were found over the Arabian Peninsula, India, southern China and eastern China, with linear slopes of 0.0092 0.0013, 0.0063 0.0020, 0.0049 0.0018 and 0.0032 0.0018 yr-1, respectively. However, (Li et al., 2014) reported declines of -0.10 and -0.18 (10 yr)-1 in AERONET AOD at Beijing (2002-13) and Xianghe (2005-12) on the North China Plain. Negative AOD trends over the east coastal region of China since 2008 were also revealed by MODIS and MISR satellite measurements (Zhang et al., 2017).
Several recent studies have analyzed aerosol variations and trends using different models and data from multiple platforms. (Chin et al., 2014) investigated aerosol tendencies and variations over a period of 30 years (1980-2009) using the Goddard Chemistry Aerosol Radiation and Transport model and multiple satellite sensors and ground-based networks. They found that the model-calculated global annually averaged AOD values showed little trend——a result that highlights the necessity of studying regional-scale AOD trends. (Pozzer et al., 2015) compared the AOD trends from MODIS, SeaWiFS and MISR satellite measurements with simulated values from the ECHAM/Modular Earth Submodel System Atmospheric Chemistry (EMAC) general circulation model covering the decade (2001-10) on a global scale. The simulation results showed that the model could qualitatively reproduce the MODIS AOD trends over North America, Eastern Europe, North Africa and the Middle East, while some discrepancies were found for other regions. The application of a newly developed satellite AOD dataset, such as the MODIS collection 6, will be helpful to improve the reliability of these datasets and reduce artificial trends (Lyapustin et al., 2014; Pozzer et al., 2015).
The MODIS sensor onboard NASA's Aqua satellite has been collecting valuable data about the Earth system for more than 14 years, which has made it possible to detect the long-term variation in aerosol loading across the globe. However, the long-term AOD trends derived from MODIS need careful validation and assessment, especially over land. It is necessary to compare the AOD trends from the MODIS sensor with those from ground-based observations, which was the primary objective of the present study. Furthermore, most satellite aerosol products are derived from polar-orbiting satellites that pass overhead at a fixed time of day; for example, Terra passes overhead at 1010-1130 LST and Aqua passes overhead at 1230-1400 LST. Therefore, only 1-2 overpasses with aerosol retrievals collected at a fixed time of the day are available from these polar-orbiting satellites. When we use these satellite aerosol products to study long-term trends, one question that arises is whether the long-term AOD trend derived from satellite measurements of 1-2 overpasses can be representative of that determined from daily measurements. This was the second objective of the present work, which is somewhat similar to the analysis performed by (Kaufman et al., 2000), who investigated whether daily aerosol abundances and properties could be represented by the aerosol products from the Terra and Aqua polar orbiting satellites. They showed that the annual average AOD could be calculated from Terra and Aqua aerosol products within a 2% error level, and this conclusion was independent of the particle size or range of the AOD.
2.1. Multi-sensor Aerosol Products Sampling System AOD
The Multi-sensor Aerosol Products Sampling System (MAPSS) is a framework that provides statistics on spatial and temporal subsets of level 2.0 aerosol scientific datasets from spaceborne sensors as MODIS and MISR. The data system is described in detail in (Petrenko et al., 2012). The AERONET sites were identified as focal points for spatial statistics. The process of generating the statistics for each spatial spaceborne aerosol product involved extracting values of the pixels that fell within a diameter of approximately 50 km centered on the chosen AERONET site. Statistics for ground-based temporal observations collected at a particular station were derived from measurements taken within 30 min of each satellite overpass of this location. The MODIS dataset obtained from MAPSS used in this paper is a merged dataset comprising the dark target (DT) and deep blue (DB) AOD products filtered by quality assurance (QA) processes. Unless otherwise specified, all the analyses in this paper are based on AOD at 550 nm. The highest quality (QA = 3) AODs retrieved by the DB algorithm have been shown to have an absolute uncertainty of approximately 0.03+0.20τ M for a typical Atmosphere Mass Factor of 2.8 (Sayer et al., 2013), where τ M is MODIS retrieved AOD. The uncertainty of MODIS DT AOD has been estimated to be 0.05+0.15τ A over land (Levy et al., 2010, 2013), where τ A is collocated AERONET AOD. Data at 53 AERONET stations were chosen for this analysis because their measurements are available for a relatively longer period compared with other stations. Figure 1 shows the spatial distribution of these stations, with their data lengths represented by the colors.
Figure1. Spatial distribution of stations with at least 70 months of monthly median AOD during 2002-15. The data length is represented by the colors.2
2.2. Data analysis
We used quality-assured and cloud-screened level 2.0 AERONET AOD data, with a low uncertainty of 0.01-0.02 (Holben et al., 1998). Three daily series were calculated from instantaneous AOD products as follows: AERONET daily mean AOD was calculated from all instantaneous AOD measurements each day; AERONET morning and afternoon mean AOD was calculated from instantaneous AOD measurements within 30 min of the Terra and Aqua overpass times; and monthly median AOD was calculated from the daily mean AERONET AOD if there were more than five daily measurements each month [AOD reference value (AOD RE)], which was taken as a benchmark for comparison with other estimates. The monthly median AOD in the morning (AOD AM) and in the afternoon (AOD PM) were also calculated if at least five daily measurements were available each month. The monthly median was used in the trend test instead of the monthly mean because AOD does not follow a normal distribution. The AOD trends derived from AOD RE, AOD AM and AOD PM were compared to determine whether the trends derived from AERONET AOD measured at the Terra and Aqua overpass times were representative of those derived from daily mean AERONET AOD. The monthly medians of AOD MOD were calculated from all collocated AOD from MODIS/Aqua provided by the MAPSS system if there were more than five daily measurements each month. The AOD trends from AOD MOD were compared with those of AOD RE to determine whether AERONET AOD trends could be reproduced by MODIS. The trends from AERONET and MODIS/Aqua AOD at each site were calculated based on the measurements during the same measurement period. Although the comparison of the AOD RE and AOD MOD trends mentioned above was based on measurements during the same study period, AOD measurements were not necessarily available simultaneously from AERONET and MODIS each day. Therefore, the monthly median AOD from AERONET and MODIS were probably calculated from different daily measurements. In other words, it was possible that daily AOD was available from AERONET, but MODIS/Aqua had no AOD product, or vice versa. This situation arose because of insufficient temporal coverage by MODIS or insufficient spatial coverage by AERONET. Since the difference in temporal sampling was not considered in calculating the monthly median AOD values, the trends in the MODIS and AERONET data may not have been consistent. The AOD measurements from AERONET and MODIS/Aqua were matched day-by-day, and the monthly medians were then recalculated for all the sites to assess the potential effects of sampling issues on the comparison of AOD trends.2
2.3. Trend detection
The MK test was used to detect monotonic trends. The MK test is a rank-based non-parametric test for assessing the significance of a trend, which requires that the data be independent. The null hypothesis (H0) is that a sample of data, {X1,X2…,Xn}, is independent and identically distributed. The alternative hypothesis (H1) is that a monotonic trend exists in X. The slope b of the trend is computed using the method proposed by (Sen, 1968) as follows: \begin{equation} b={\rm Median}\left(\dfrac{X_i-X_j}{i-j}\right),\quad \forall i>j , \ \ (1)\end{equation} where b is the slope of the trend, Xi and Xj are the ith and the jth observations respectively. The slope is robust for estimating the magnitude of a trend (Yue and Wang, 2002) and much less sensitive to outliers compared with linear regression coefficient (Li et al., 2014).The non-parametric test is more suitable for non-normally distributed, censored, and missing data (Yue and Wang, 2002), such as the AERONET AOD data (Li et al., 2014). However, the AOD time series, which have distinct seasonal variability, frequently display statistically significant serial correlation. The existence of positive serial correlation in a time series increases the probability of the MK test detecting a significant trend (Hirsch and Slack, 1984; Yue and Wang, 2002). Therefore, a pre-whitening scheme proposed by (Yue et al., 2002) was used to eliminate the influence of serial correlation on the MK trend detections. First, the slope (b) of the trend in an AOD time series was estimated using Eq. (1). If b was almost equal to zero, then it was not necessary to continue performing the trend analysis. If it differed from zero, the AOD time series were detrended using \begin{equation} X'_t=X_t-T_t=X_t-bt ,\ \ (2) \end{equation} where X't is the `trend-removed' residual series, Xt denotes the raw AOD time series, and Tt is the identified trend with the slope b of the trend at time t. Second, the lag-1 autocorrelation component was removed from the detrended time series (X't) using \begin{equation} Y'_t=X'_t-r_1X'_{t-1} , \ \ (3)\end{equation} where r1 is the lag-1 autocorrelation coefficient, X't-1 is the detrended series at time t-1. Third, the identified trend was added back to the residual Y't according to the following equation: \begin{equation} Y_t=Y'_t+T_t . \ \ (4)\end{equation} It was evident that the blended time series (Yt) preserved the true trend and was no longer influenced by the effects of autocorrelation. The MK test and Sen's slope estimator were applied to the blended series to estimate the trend in AOD time series. (Hirsch and Slack, 1984) developed a seasonal MK test and estimated the annual trend as the median of the seasonal trends——an approach that is resistant to data seasonality and serial dependence. Thus, the pre-whitening scheme described by (Yue et al., 2002) was adopted first. Then, the seasonal MK test of (Hirsch and Slack, 1984) and Sen's slope estimator were applied to the pre-whitened time series of AOD.
Table 2 shows that the majority of sites exhibited negative trends of AOD RE, including all stations in Europe, North America, South America and most sites in Africa (eight of nine) and Asia (seven of nine), although the trends were not statistically significant at some sites. A significant increasing trend in AOD RE [0.12 (10 yr)-1] was found only at Solar_Village in Saudi Arabia. The strong positive trend there has also been noted in AOD measurements collected by SeaWiFS (Hsu et al., 2012), MISR (de Meij et al., 2012a) and MODIS/Aqua (this study), as well as EMAC model simulations (de Meij et al., 2012b; Pozzer et al., 2015). AOD RE at Skukuza in South Africa, Kanpur in India, and Birdsville in Australia showed trends that were increasing but were not statistically significant.
Comparisons of the trends between AOD RE and AOD AM, between AOD RE and AOD PM, and between AOD RE and AOD MOD are illustrated in Figs. 2-4, respectively. The tendencies in AOD RE were reproduced by AOD AM and AOD PM at nearly all stations. The trend magnitudes of AOD AM and AOD PM were also close to those from AOD RE at most stations. The results suggest that AOD measured at Terra or Aqua overpass times could represent the trends derived from daily measurements.
Figure2. Comparison of decadal trends between AOD RE and AOD AM in different regions. The trends estimated from AOD RE and AOD AM are shown on the x-axes and the y-axes, respectively, in the scatterplots. Blue solid diamonds indicate that the trends from both AOD RE and AOD AM were statistically significant. Red solid squares indicate that the AOD RE trend was statistically significant but the AOD AM trend was statistically non-significant. Green asterisks indicate that the AOD RE trend was statistically non-significant but the AOD AM trend was statistically significant. Cyan crosses indicate that the trends from AOD RE and AOD AM were statistically non-significant. The black dots in the map represented the locations of sites, and the numbers of sites in the different regions are given in parentheses in the scatter-plots.
Figure3. As Fig. 2, except that the trend comparisons are between AOD RE and AOD PM.
Figure4. As Fig. 2, except that the trend comparisons are between AOD RE and AOD MOD.
Figure5. Comparison of AERONET AOD (AOD RE) and MODIS/Aqua AOD (AOD MOD) at (a) Birdsville in Australia, (b) Trelew in Argentina, and (c) SEDE_BOKER in Israel. The AOD anomalies are shown as blue solid lines, which are matched to the right-hand ordinate axis, and the blue dashed lines indicate that the anomaly is 0.The trends in AOD RE and AOD MOD were consistent over 44 of the sites. All sites in Europe and North America showed the same tendencies. Opposite MK trends between AOD RE and AOD MOD were found at one of nine sites in Africa (Banizoumbou), four of nine sites in Asia (Beijing, Xianghe, Nes_Ziona and SEDE_BOKER), three of four sites in Oceania (Birdsville, Canberra and Jabiru), and three of four sites in South America (CEILAP-BA, La_Parguera, and Trelew).
The tendencies of the strictly matched monthly medians of AOD RE and AOD MOD (i.e., monthly medians calculated from simultaneous measurements from both products) remained unchanged except at Beijing in China, which suggests that the temporal differences in sampling between AOD RE and AOD MOD has little effect on the detection of AOD trends in most cases. If the temporal sampling difference was not accounted for, the AOD RE at Beijing showed a significant decreasing trend [-0.08 (10 yr)-1] but the AOD MOD exhibited a weak increasing trend [0.01 (10 yr)-1]. The trend in AOD MOD changed to -0.03 (10 yr)-1 if the daily measurements from AERONET and MODIS/Aqua were strictly matched. Additionally, the seasonal MK tests of MODIS and AERONET AOD exhibited the same trend signs at Xianghe as those reported by (Zhang et al., 2017). Increasing trends of 0.001 yr-1 (MODIS) and 0.004 yr-1 (AERONET) were found from 2002 to 2007, but there were downward trends of -0.006 yr-1 (MODIS) and -0.004 yr-1 (AERONET) from 2008 to 2015. However, consistent declines in AERONET AOD and MODIS/Aqua AOD were found at Beijing during the two periods (2002-07 and 2008-15).
AOD over Australia is typically low, with occasional high AOD events mostly resulting from locally derived or transported smoke from wildfires and mineral dust (Qin and Mitchell, 2009). Another interesting feature seen in this region is that the surface reflectance is relatively high, because of the abundance of bare soil and desert. Both factors indicate that it is very hard to retrieve AOD with high quality. Validation has shown that the correlation between AERONET and MODIS AOD is lower over Australia than in other regions (Sayer et al., 2013). Figure 5a shows the time series of the monthly medians of AOD RE and AOD MOD (shown as a dependent variable on the left-hand y-axis) and the AOD anomaly (shown as a dependent variable on the right-hand y-axis) defined as the difference between the AOD MOD and the monthly averaged AOD RE over all years at Birdsville, Australia. An interesting feature is that all the AOD anomalies were positive before September 2008, i.e., MODIS/Aqua overestimated AOD, but AOD was underestimated by MODIS afterwards. The shift in the AOD anomalies resulted in the artificially negative trend in AOD MOD seen at Birdsville. This same scenario also occurred at Canberra. Trelew is in the middle of the Argentinean Patagonia, where the monthly median AOD is less than 0.10. A shift in AOD anomalies after March 2008 (Fig. 5b) resulted in an artificially positive tendency in AOD MOD. The reason for these abrupt changes in MODIS performance compared with AERONET at these stations needs further study.
Banizoumbou, in the semi-arid Sahel region and influenced by dust emissions and/or biomass burning (Levy et al., 2010), showed high AOD. Studies have shown that, from the mid-1980s to the mid-2000s in the Sahara and Sahel regions, near-surface wind speed decreased, precipitation increased, and therefore dust emissions decreased (Chin et al., 2014). This finding of a reduction in AOD is also supported by in-situ aerosol measurements (Li et al., 2014) and model simulations (Chin et al., 2014). The AOD RE and AOD MOD at other stations in the Sahel showed a declining trend. Therefore, the increasing trend at Banizoumbou (although non-significant) seen in the MODIS data seems likely to be spurious.
The sites of SEDE_BOKER and Nes_Ziona are in Israel, where aerosols are influenced both by mineral dust transported from the Saharan and Arabian deserts and fine-mode pollution emitted by the petroleum industry in this region (Yoon et al., 2012). The AOD at these two sites showed a decreasing tendency based on AERONET measurements, which is consistent with the results reported by (Yoon et al., 2012) and (Li et al., 2014). (Yoon et al., 2012) noted that the decreased AERONET AOD at SEDE_BOKER was due to a decrease in coarse-mode particles [approximately -0.29 (10 yr)-1 for coarse-mode AOD]. MODIS/Aqua AOD generally produced overestimates at SEDE_BOKER, where 98% of the monthly median anomalies were positive (Fig. 5c). The retrieval uncertainty of MODIS AOD under conditions with heavy aerosol loading is mainly related to the aerosol mode assumptions. The MODIS retrieval algorithm assumes a constant single-scattering albedo value; however, the AERONET single-scattering albedo at SEDE_BOKER shows an increasing trend [0.004 (10 yr)-1] (Li et al., 2014). Therefore, the temporal change in single-scattering albedo may produce a spurious tendency in the AOD retrieved by MODIS (Mishchenko et al., 2007, 2012; Li et al., 2014).
In addition, we found a few high-altitude sites (>1.2 km) in North America (BSRN_BAO_Boulder, Railroad_Valley, Sevilleta, TABLE_MOUNTAIN_CA) and Asia (Dalanzadgad) where AERONET AOD was not suitable for representing the regional AOD trend due to a lack of spatial representation. The retrieval accuracy of MODIS AOD at such sites is impacted by the weak aerosol signal due to the very low aerosol loading (Levy et al., 2010; Sayer et al., 2013). Thus, high-altitude sites (>1.2 km) were excluded in the AOD trend analyses.
AOD trends calculated from daily measurements can be accurately reproduced by AOD measurements collected at the overpass time of MODIS, which indicates that MODIS has the potential to capture AOD trends.
Consistent AOD tendencies calculated from the monthly medians of the AERONET and MODIS/Aqua AOD products were derived at 44 of 53 sites. However, this was not true for sites in Australia, where the retrieval accuracy of MODIS/Aqua AOD data was found to be poor due to the very low aerosol loading (monthly median AOD <0.15) and the relatively higher surface reflectance of bare soil and/or desert. Disagreement in the AOD trends between MODIS and AERONET was also found at a few sites in South America. This was due to the low retrieval quality in MODIS/Aqua AOD and likely resulted from improper surface reflectance parameterization and uncertainties in the aerosol model used in the algorithm (Levy et al., 2010). Furthermore, sampling issues should be evaluated carefully when deriving long-term trends in AOD in regions where AOD shows substantial day-to-day variation.
Merged data based on the MODIS AOD products derived from both the DT and DB algorithms are used in many studies to investigate the long-term AOD trends over land areas in various locations throughout the world. In some regions, the typical AOD and its variability are low, close to the magnitude of the retrieval uncertainty, which affects the estimation of the AOD trends. However, the current MODIS AOD products do provide a continuous aerosol data record for monitoring long-term climate change.
