Corresponding author: T. X.-P. Zhao, National Climatic Data Center, NOAA/NESDIS, 151 Patton Ave., RM-445, Asheville, NC 28801, USA. (Xuepeng.Zhao@noaa.gov)
 Subpixel cloud contamination is one of the major issues plaguing passive satellite aerosol remote sensing. Its impact on the aerosol optical thickness (AOT) retrieval has been analyzed/evaluated by many studies. However, the question of how it influences the AOT trend remains to be answered. In this paper, four long-term advanced very high resolution radiometer (AVHRR) AOT data sets from 1981 to 2009 over global oceans for four different definitions of clear sky, respectively, are produced by applying a two-channel aerosol retrieval algorithm to the AVHRR clear-sky reflectances derived by combining NOAA Pathfinder Atmosphere's Extended AVHRR climate data record level-2b all-sky reflectances with the cloud probability parameter determined from the Bayesian probabilistic cloud detection technique. A global analysis of the effect of cloud contamination on the AVHRR AOT retrieval as well as on its long-term trend is then performed by comparing the results from the four data sets. It was found that cloud contamination imposes not only a positive bias on AOT values but also a positive bias on its long-term trend such that negative trends become less negative and positive trends become more positive. A cloud probability value of ≤1% has been identified as an optimal criterion for clear-sky definition to minimize the cloud contamination in the AVHRR aerosol retrieval while still retaining strong aerosol signals. In order for a satellite aerosol product to be useful and reliable in aerosol trend detection, the cloud contamination effect on aerosol trends needs to be studied/evaluated carefully along with the effects of calibration error, surface disturbance, and aerosol model assumptions.
If you can't find a tool you're looking for, please click the link at the top of the page to "Go to old article view". Alternatively, view our Knowledge Base articles for additional help. Your feedback is important to us, so please let us know if you have comments or ideas for improvement.
 Aerosols perturb the Earth's energy budget by scattering and absorbing radiation [e.g., Ohmura, 2009; Wild et al., 2009] and by altering cloud properties and lifetimes [e.g., Hobbs, 1993; Lohmann and Feichter, 2005]. They also exert large influences on weather, air quality, hydrological cycles, and ecosystems. Global long-term aerosol observation and data product with climate quality are needed to monitor the global aerosol distributions and study how they are changing annually and seasonally. They are also needed to find out if there is any trend in global aerosol distributions over the years and what effect the long-term change will ultimately have on the global environment and climate.
 Satellite remote sensing of aerosols is the only way to perform global aerosol measurement for multidecadal time periods [Kaufman et al., 2002; King et al., 1999; Mishchenko et al., 2007b]. There are several factors in passive satellite aerosol remote sensing that are critical in determining the quality of retrieved aerosol optical thickness (AOT) as well as the truthfulness of the AOT trend derived from the corresponding long-term AOT products [Li et al., 2009]. The first factor is instrument calibration. Calibration errors may corrupt the accuracy of retrieved AOT and cause spurious AOT long-term trend [e.g., Mishchenko and Geogdzhayev, 2007a; Zhang and Reid, 2010; Zhao et al., 2008]. The second factor is the aerosol model assumption, especially the absorption properties and size distribution, which may impact the accuracy of retrieved AOT as well as change the sign of the AOT trend [e.g., Jeong et al., 2005; Mishchenko et al., 2012; Zhao et al., 2008]. The third factor is surface disturbance caused by rough ocean surface. Errors in the treatment of the surface roughness associated with surface wind may infect the AOT quality and result in false AOT trend [e.g., Jeong and Li, 2005; Jeong et al., 2005; Mishchenko and Geogdzhayev, 2007a]. The last factor is the subpixel cloud contamination due to incomplete cloud screening. Actually, cloud contamination is also a very disturbing issue for the surface aerosol remote sensing [e.g., Smirnov et al., 2000; Yoon et al., 2012]. This paper will focus specifically on the analysis of cloud contamination on the aerosol retrieval from the advanced very high resolution radiometer (AVHRR).
 Subpixel cloud contamination may corrupt the quality of passive satellite aerosol remote sensing. This is because all satellite aerosol retrieval algorithms perform retrieval for cloud-free pixels identified by the cloud screening schemes. It is hard to completely avoid misclassifying pixels containing some clouds as cloud free since there are always some clouds (called subpixel clouds) smaller than the pixel size of satellite measurements. As a result, both global analyses [e.g., de Vries and Wagner, 2011; Jeong and Li, 2005; Jeong et al., 2005; Kaufman et al., 2005; Liu et al., 2009; Zhao et al., 2005a] and regional case studies [e.g., Gao et al., 2002; Heidinger et al., 2002b; Huang et al., 2011; Zhang et al., 2005; Zhao et al., 2009, 2005b] have been performed to investigate the cloud contamination and its effect on the aerosol products derived from passive remote sensing measurements of various satellite instruments. These studies found out that cloud contamination generally produces a positive bias on aerosol optical thickness (a column amount parameter) and a negative bias on aerosol Ångström exponent (a size parameter) since clouds generally have a higher reflectance and lower spectral dependence than aerosols.
 However, we are not aware of any efforts that have been performed to study the effect of cloud contamination on global aerosol long-term trend (such as aerosol optical thickness long-term trend), which is an effective indicator of the global changes due to natural or anthropogenic emissions. This is mainly due to the following two reasons: (1) almost all the cloud screening schemes used in the satellite aerosol retrievals are threshold based, which poses difficulty to perform sensitivity studies for determining optimal screening thresholds for a whole long-term satellite aerosol data record; and (2) several long-term data sets that differ only in cloud screening are normally needed in an analysis to separate cloud contamination effect from other error sources, such as instrument calibration error, aerosol model assumptions (size distribution and absorption properties), and surface disturbance, of satellite aerosol remote sensing. It is too expensive (computing resources and storage spaces) to produce these long-term data sets through reprocessing massive satellite observations.
 A new cloud detection algorithm based on the naïve Bayesian probabilistic methodology has been developed recently and applied to NOAA's operational AVHRR observation by Heidinger et al. . Analysis of collocated AVHRR and Cloud-Aerosol Lidar and Infrared Pathfinder Satellite Observations (CALIPSO)/Cloud-Aerosol Lidar with Orthogonal Polarization (CALIOP) observations was used to automatically and globally derive the Bayesian classifiers. Thus, cloud detection is valid for not only opaque clouds but also semitransparent/high-level cirrus clouds since the cloudiness in the collocated training data base is based on CALIPSO/CALIOP cloud-layer product. For example, for an AVHRR Global Area Coverage (GAC) pixel (~5 km in size), its cloud fraction (or probability) was computed from 5 CALIOP pixels that were closest to the center of the AVHRR GAC pixel. Detailed description of the detection algorithm can be found in Heidinger et al. , which is the paper dedicated to this detection algorithm. One of the key strengths of this Bayesian approach is that the probability of cloud varies smoothly over the range of the classifier rather than jump from 0 to 1 when passing over the chosen threshold in a threshold-based approach. This makes the sensitivity study of cloud contamination on aerosol long-term trend possible by using different clear-sky conditions (criteria) defined by the cloud probability parameter. The Bayesian technique is applied within the NOAA Pathfinder Atmosphere's Extended (PATMOS-x) climate data set. The resultant AVHRR all-sky reflectances and cloud probability parameter are mapped onto a 0.1° × 0.1° descending and ascending orbital grid (rather than pixel) and named level-2b product, which can be used to generate long-term AOT product efficiently with an aerosol retrieval algorithm (faster retrieval and less data volume on the orbital grid level compared to the orbital pixel level). It makes possible to produce several long-term AVHRR AOT products with different clear-sky definitions based on the Bayesian cloud probability parameter through reprocessing. Thus, the products can be used for the study of cloud contamination on the AOT and its long-term trend.
 In this paper, we specifically present a global survey of the effect of cloud contamination on both AOT and its long-term trend by using the long-term AVHRR AOT products produced from the PATMOS-x level-2b clear-sky reflectances derived from the Bayesian cloud screening approach. PATMOS-x level-2b data (including aerosol retrieval algorithm) used in the current study are described in section 2. Analysis of cloud contamination effect on AOT and its long-term trend is performed in section 3 and 4, respectively. Some discussions are provided in section 5. Summary and conclusions are given in the closing section.
2 PAMOS-x Data
 A new level-2b data set [http://cimss.ssec.wisc.edu/patmosx/overview.html] of the PATMOS-x AVHRR climate data product is used in the current study, which was generated by Heidinger et al.  for the NOAA climate data record (CDR) program through reprocessing five-channel AVHRR observations from the NOAA-6, NOAA-7, NOAA-9, NOAA-10, NOAA-11, NOAA-12, NOAA-14, NOAA-15, NOAA-16, NOAA-17, NOAA-18, and NOAA-19 polar-orbiting satellites. The product provides a continuous atmospheric record (including top-of-atmosphere reflectances/radiances and cloud properties) spanning over 31 years from 1978 to 2009. The product variables are mapped on a 0.1° × 0.1° daily global ascending and descending orbital grid. AVHRR instrument calibration used in the reprocessing has been improved through intersatellite calibration with more accurate MODIS measurement [Heidinger et al., 2002a, 2010; Molling et al., 2010] and becomes comparable to the MODIS calibration. A new cloud detection algorithm based on Bayesian probabilistic methodology [Heidinger et al., 2012] has been used to produce the cloudy probability (from a range of 0% to 100%) on every orbital grid point so that users can form their own clear-sky criteria through sensitivity studies by selecting different cloudy probability values. An analysis of collocated AVHRR and CALIPSO lidar observations was used to automatically and globally derive the six Bayesian cloud mask classifiers. The uncertainties of the cloud mask are also provided in the PATMOS-x level-2b data set.
 The consistency and homogeneity of observations from multiple AVHRR sensors used for producing three decades of PATMOS-x level-2b data has been improved greatly through intercalibration with MODIS [Heidinger et al., 2002a] and retrospective intercalibration among AVHRR sensors [Heidinger et al., 2010] to mitigate the degradation of individual AVHRR sensors with time. An examination of cloud amount and cloud optical depth time series had demonstrated that the resultant AVHRR intersatellite biases are less than 1%. A paper dedicated to the climate quality (including consistency and homogeneity) of PATMOS-x AVHRR level-2b climate data record (CDR) is being prepared by Heidinger et al. , which is out the scope of the current paper.
 A two-channel AVHRR AOT retrieval algorithm developed by Zhao et al.  has been applied to the AVHRR clear-sky reflectances over global oceans at 0.63 and 0.86 µm channels from the PATMOS-x level-2b product. The algorithm has been validated by comparing with the AERONET ground AOT measurement [Zhao et al., 2003, 2004] and the MODIS satellite AOT observation [Zhao et al., 2005a, 2005b]. It had been applied to the previous version of the PATMOS-x reflectance product to generate long-term AOT data for the study of the AOT long-term trend [Zhao et al., 2011, 2008]. Only the data for the time period from 1981 to 2009 are used in the current study considering too many missing observations in the first 3 years (1978–1980) of the AVHRR data record.
 For the current study of cloud contamination effect on the AOT long-term trend, four long-term AOT data sets have been produced by applying the two-channel AVHRR AOT retrieval algorithm to the PATMOS-x level-2b reflectances of AVHRR channels 1 (0.63 µm) and 2 (0.86 µm). The clear-sky condition is defined for the four data sets (or cases) by selecting the cloudy probability values of 0.5%, 1%, 5%, and 15%, respectively. The data sets are summarized in Table 1. Although both AOT at AVHRR channels 1 (τ1) and 2 (τ2) are produced, only τ1 is considered as an AOT climate data record and used for the subsequent analysis of cloud contamination effect on the AOT long-term trend since the aerosol retrieval in the AVHRR broad spectral channel 2 is contaminated by water vapor absorption [Zhao et al., 2004].
Table 1. Four Long-Term AOT Data Sets From AVHRR Observations for Four Different Clear-Sky Conditions (Cases) Defined by Using the Cloud Probability (cpb) Parameter of the PATMOS-x AVHRR Level-2b Producta
Data Sets (Case #)
Spatial Resolution (Orbital Descending and Ascending Grid)
Temporal Coverage (Resolution)
aτ1 and τ2 are aerosol optical thicknesses of AVHRR channels 1 (0.63 µm) and 2 (0.86 µm), respectively.
cpb ≤ 0.5%
τ1 and τ2
0.1° × 0.1°
1981.8–2009.12 (daily and monthly)
cpb ≤ 1%
τ1 and τ2
0.1° × 0.1°
1981.8–2009.12 (daily and monthly)
cpb ≤ 5%
τ1 and τ2
0.1° × 0.1°
1981.8–2009.12 (daily and monthly)
cpb ≤ 15%
τ1 and τ2
0.1° × 0.1°
1981.8–2009.12 (daily and monthly)
 Most imager cloud masks, such as AVHRR and MODIS, typically provide four classifications of pixel cloudiness: clear, probably clear, probably cloudy, and cloudy. PATMOS-x uses a threshold cloud probably value of 90% to separate cloudy from probably cloudy, 10% to separate clear from probably clear, and 50% to separate probably clear from probably cloudy. Thus, data set 4 (or case 004) is expected to contain cloud contamination more easily than the other three data sets, while data set 1 (or case 001) is the clearest among the four data sets.
3 Cloud Contamination Effect on AOT
 Figure 1a shows the long-term averaged τ1 over the global ocean from data set 1. Figures 1b–1d display the difference of long-term averaged τ1 of data sets 2, 3, and 4 from that of data set 1, respectively. AOT of data set 2 is slightly higher than that of data set 1 almost uniformly over the global ocean (see Figure 1b). More AOT increase is observed in data set 3 (see Figure 1c), especially at the high latitudes of both hemispheres and in the offshore oceanic regions with relatively high AOT, such as the west coastal oceans of central Africa, south and east Asian coastal waters, Arabian Sea, Bay of Bengal, and tropical east coastal oceans of Central America. AOT increases further in data set 4 (see Figure 1d), and the largest increases are still in the offshore regions with relatively high AOT. This result corroborates the finding of previous studies [e.g., Zhang et al., 2005, 2009] that subpixel cloud contamination tends to bias the retrieved AOT high. The positive biases in the AVHRR AOT due to the cloud contamination prevail nearly uniformly on all the latitudes (see Figure 2a) and in all months/seasons (see Figure 2b). The magnitude of long-term and globally averaged biases observed in Figure 2 for data set 4 is more than 50% of the AOT values of data set 1.
 To compare individual months for the four data sets, Figure 3 shows the scatterplot of monthly and globally averaged τ1 from August 1981 to December 2009 (total of 341 months) for data set 1 versus data sets 2, 3, and 4. Data points of data set 1 versus 2 fall almost perfectly on the 1:1 line, while data points of data set 1 versus 3 (or 4) bias systematically above 1:1 line and the dispersion of the data points also gradually increases from data set 2 to 4. The effect of cloud contamination on AOT displays both systematic and random attributes.
 The sampling effect due to different cloud screening criteria was also examined. The global distributions of total samples of clear-sky grids from long-term monthly mean PATMOS-x AVHRR level-2b data are compared in Figure 4 for cases 001 and 004. The regions with evident difference in sampling number are over the tropical convergence zone, where high-level cirrus clouds are prevailing. This suggests that the stringent cloud screening used in case 001 is also effective to filter out semitransparent cirrus clouds. Even though the absolute difference in clear-sky sampling number is bigger in tropical regions than at high latitudes, the relative difference is reversed with a bigger difference at high latitudes since the sampling number of clear sky at high latitudes is much lower than its tropical counterparts due to the difficulty of cloud screening for the observation near the terminator at high latitudes (with higher solar zenith angle and longer observing path length). Bright surface contamination due to sea ice is also a culprit for defective cloud screening and the dilemma of aerosol retrieval at high latitudes. Thus, the impact of sampling issue caused by different cloud screening criteria in the four data sets can be important at high latitudes due to sparse clear-sky grids. This explains the patches of suspicious high AOT (see Figure 1a) and the largest positive biases in AOT (see Figures 1c and 1d) at the high latitudes of both hemispheres. Meanwhile, the sampling effect due to cloud screening can also exert some impact in the regions with higher AOT values since the grids added with less stringent cloud screening in data set 4 naturally contain relatively high AOT values. This explains the larger AOT differences (see Figures 1c and 1d) in coastal regions at low latitudes, such as the west coastal oceans of central Africa influenced by dust emission and biomass burnings and south and east Asian coastal waters, Arabian Sea, and Bay of Bengal affected from industrial emissions.
 Due to the dilemma in the cloud screening and aerosol retrieval at high latitudes mentioned above, our analysis of cloud contamination effect on AOT and its long-term trend in this paper will focus only within ±60° latitudes. The above analysis suggests that separating clear from probably clear with a 10% cloud probability in the PATMOS-x cloud mask is not sufficient for aerosol optical thickness retrieval. The criterion should be at least a value of cpb ≤ 1%, which is satisfied by both data sets 1 (case 001) and 2 (case 002). Since the AOT difference between data sets 1 and 2 are small over the global ocean (see Figures 1b and 3) and fewer data samples are averaged to obtain the AOT value for each orbital grid (0.1° × 0.1°) of data set 1, case 002 (data set 2) will be treated as the standard clear-sky case.
4 Cloud Contamination Effect on the AOT Trend
 As we mentioned above, using the PATMOS-x AVHRR level-2b product makes it possible to produce several AOT long-term data products different in the clear-sky definition and study the effect of cloud contamination on the AOT long-term trend, which has not been performed before for any long-term satellite aerosol products based on our knowledge. The linear trend model is commonly used by scientists in environment and climate change studies and is familiar to policy makers and the public. In this paper, we will focus our study only on the linear trend of the AOT since it allows a simple approximation of the direction and magnitude of the changes in the AOT data and is adequate for many practical purposes. The linear change of a variable, such as AOT, in a climate study is often measured with the linear long-term trend (LLT) in the unit of absolute (or percentage) changes per decade. The AOT LLT is the slope of the linear regression line for the time series of monthly or annually averaged AOT. We shall adopt the commonly used decision rule that a real trend is indicated at the 5% significance level or the 95% confidence level when |LLT/σ| > 2. Here σ is the standard deviation of the AOT LLT. We further name LLT/σ as the significance of the AOT LLT. A significance >+2 (or <−2) indicates that the increasing (or decreasing) tendency detected is above the 95% confidence level. Readers can refer to Weatherhead et al.  for a detailed description of this linear trend model and its application to environmental data.
 Figure 5 shows the distribution of the AOT LLT over global oceans for data sets 1–4. Negative trends prevail over global oceans, and the largest negative trend appears over the North Atlantic Ocean. The negative trend over broad open oceans was also obtained by Mishchenko and Geogdzhayev [2007a] from the analysis of the Global Aerosol Climatology Project (GACP) data based on independent aerosol retrieval from AVHRR observations with different calibration schemes. Recently, Mishchenko et al.  have performed a sensitivity study by using the aerosol model with different absorption in their GACP aerosol retrieval algorithm for the earlier and recent AVHRR observations. Their result indicates that the general negative trend over global ocean may be partially due to using the fixed aerosol model (especially the fixed absorption property) over the duration of AVHRR observations, which is an interesting topic for our future study.
 The positive trends appear only in the offshore oceanic regions under the influence of biomass burning and industrial pollutions, such as the west coastal oceans of central and south Africa, south and east Asian coastal waters, Arabian Sea, and Bay of Bengal. These patterns of the AOT trend had been noticed in other studies using different satellite observations. [e.g., Massie et al., 2004; Mishchenko et al., 2007a; Zhang and Reid, 2010; Zhao et al., 2008]. Cloud contamination generally imposes a positive bias on the AOT trend (see Figures 5 and 6), which generally makes negative trends become less negative (over most oceanic regions) and positive trends become more positive (such as the west coastal oceans of central and south Africa, south and east Asian coastal waters, Arabian Sea, and Bay of Bengal) except at northern tropical Atlantic Ocean, where dust particles originated from Sahara desert prevail year round and the negative AOT trend is magnified rather than compressed by cloud contamination (more discussion will be provided in section 5).
 For the case with relatively strong cloud contamination (e.g., data set 4), we have noticed that there is an evident increase in aerosol grids contaminated by cloud for some orbital grids poleward of ±40° north/south, which explains the spurious positive trends in AOT observed in Figure 5d over several oceanic regions at high latitudes of South Hemisphere and over northern Pacific Ocean and the east coastal waters of Canada. Examining the difference of clear-sky reflectance for cases 1 and 4 (see Figure 7) confirms evident disturbance from residual cloud in case 4 in these regions, which has noticeable impact on the corresponding AOT values (see Figure 1d) and, in consequence, on their trend (see Figure 6) due to sparse clear-sky grids and less samples at high latitudes (compare to low latitudes) as revealed in Figure 4.
 The significance of the zonal mean AOT trend for the four cases is shown in Figure 8a. For the data sets 1 and 2, negative trends above the 95% confidence level prevail nearly at all latitudes, except at southern tropical latitudes where the negative trend is below the 95% confidence level. For data set 4 (with the highest potential of cloud contamination among the four data sets), latitudes with a negative trend above the 95% confidence level are significantly reduced, and only those at the northern middle latitudes are still above the 95% confidence level. AOT trends in each month (or season) (see Figure 8b) are also negative for the past three decades but below the 95% confidence level for all four data sets, which is due to the smoothing effect by including more data samples in averaging to produce long-term averaged monthly (or seasonally) AOT data.
 The effect of cloud contamination on the AOT trend identified in above analysis may have two possible origins. The first origin is that the trend of cloud cover itself may propagate into the AOT trend through cloud screening. The second origin is due to the trend of the contributions of residual cloud contamination to the clear-sky reflectance used for aerosol retrieval. To find out which one is the real cause for the changes of the AOT trend due to cloud contamination, the trend of cloudy probability (see Figure 9a) from the PATMOS-x AVHRR level-2b product and the trend of the contributions of residual cloud contamination to the clear-sky reflectance (see Figure 9b) used in our AOT retrieval have been examined. The distribution patterns in Figure 9b are very similar to those observed in Figure 6 for the difference of the AOT trend between cases 001 and 004, which suggests that the trend of the contribution of residual cloud contamination to the clear-sky reflectance is the major cause for the changes of the AOT trend due to cloud contamination observed in Figure 6, especially over several oceanic regions at high latitudes of South Hemisphere and over northern Pacific Ocean and the east coastal waters of Canada. However, over the northern tropical Atlantic Ocean where dust particles originated from Sahara desert prevail year round, the trend of cloudy probability (even though its magnitude is generally small and less significant) may also propagate into the AOT trend through incomplete cloud screening as indicated by the similarity in the distribution pattern over the region in Figures 6, 9a, and 9b. This may be also the explanation to the negative difference in the AOT trend caused by subpixel cloud contamination (as observed in Figure 6) over the northern tropical Atlantic Ocean.
 In the past decade, significant efforts have been made to minimize the errors associated with instrument calibration, surface boundary conditions (e.g., surface wind disturbance), atmospheric state (e.g., water vapor amount), aerosol microphysical assumptions (e.g., size distribution and absorption property), and cloud contamination in AVHRR aerosol observations. The effect of the uncertainties in the first four components on the AOT long-term trend has been investigated, and spurious AOT trends may result from these error sources [see Jeong and Li, 2005; Li et al., 2009; Mishchenko and Geogdzhayev, 2007b; Mishchenko et al., 2012; Zhao et al., 2008]. However, the effect of subpixel cloud contamination on the AOT trend is missing in these previous studies due to the two difficulties mentioned in the introduction. Application of our efficient two-channel aerosol retrieval algorithm to the PATMOS-x AVHRR level-2b clear-sky reflectance based on Bayesian cloud probability makes the study of cloud contamination effect on the AOT long-term trend become possible. Optimal clear-sky criteria (cpb ≤ 1%) has been specifically identified through the sensitivity studies for the 29 year AVHRR data record and applied to the production of a new version-2 AVHRR AOT CDR product (i.e., data set 2), which is the major enhancement to this version of the data product compared to the previous versions. Current analysis is from a global perspective and detailed regional analysis needs to be added for a complete assessment of the cloud contamination effect on the AVHRR AOT long-term trend. This regional analysis is underway and the results will be reported in a separate paper in the near future.
6 Summary and Conclusions
 Four AVHRR AOT long-term (1981–2009) data sets for four different clear-sky definitions are produced by applying our two-channel aerosol retrieval algorithm to the clear-sky reflectances derived by combining PATMOS-x AVHRR CDR level-2b all-sky reflectances with the cloud probability parameter determined from the Bayesian probabilistic technique. Since the only difference of the four data sets is clear-sky definitions, comparing their global AOT distributions and long-term changes provides a good opportunity for studying the global effect of subpixel cloud contamination on the AVHRR AOT CDR as well as on its long-term trend.
 The results indicate that cloud contamination produces not only positive bias on AOT values but also positive bias on its long-term trend—negative trends become less negative (over most oceanic regions) and positive trends become more positive (over the coastal oceans influenced by enhanced offshore industrial pollutions and biomass burnings). Evident spurious positive trends in AOT can be caused by residual subpixel clouds of cloud screening over several oceanic regions at high latitudes of South Hemisphere or over northern Pacific Ocean and the east coastal waters of Canada. For zonal mean AOT, a negative trend above the 95% confidence level prevails nearly at all the latitudes (except at southern tropical latitudes), and cloud contamination tends to alleviate these negative trends below the 95% confidence level except at northern middle latitudes. A cloud probability value of ≤1% has been identified as an optimal criterion for clear sky to minimize cloud contamination while retaining strong aerosol signals in the AVHRR aerosol retrieval, and the corresponding data set 2 is selected as the new version of the AVHRR AOT CDR product.
 The current study is focused on global analysis, and detailed regional analysis needs to be added as future work for a complete assessment of the cloud contamination effect on the AVHRR AOT and its long-term trend. Based on this and other previous studies [e.g., Jeong and Li, 2005; Li et al., 2009; Mishchenko et al., 2012; Zhao et al., 2008], we conclude that, in order for a satellite aerosol product to be useful and reliable in aerosol trend detection, the effect of the four major error sources plaguing passive satellite aerosol remote sensing (calibration error, surface disturbance, aerosol model assumptions, and cloud contamination) on the aerosol trend needs to be studied/evaluated carefully.
 The authors would like to acknowledge the support from the CDR program at the National Climatic Data Center (NCDC) of NOAA/NESDIS. The constructive suggestions from three anonymous reviewers are greatly appreciated. The views, opinions, and findings contained in this paper are those of the author (s) and should not be construed as an official National Oceanic and Atmospheric Administration or U.S. Government position, policy, or decision.