Evaluations of cirrus contamination and screening in ground aerosol observations using collocated lidar systems

Authors


Abstract

[1] Cirrus clouds, particularly subvisual high thin cirrus with low optical thickness, are difficult to screen in operational aerosol retrieval algorithms. Collocated aerosol and cirrus observations from ground measurements, such as the Aerosol Robotic Network (AERONET) and the Micro-Pulse Lidar Network (MPLNET), provide us with an unprecedented opportunity to systematically examine the susceptibility of operational aerosol products to cirrus contamination. Quality assured aerosol optical thickness (AOT) measurements were also tested against the Cloud-Aerosol Lidar and Infrared Pathfinder Satellite Observations (CALIPSO) vertical feature mask (VFM) and the Moderate Resolution Imaging Spectroradiometer (MODIS) thin cirrus screening parameters for the purpose of evaluating cirrus contamination. Key results of this study include: (1) quantitative evaluations of data uncertainties in AERONET AOT retrievals are conducted; although AERONET cirrus screening schemes are successful in removing most cirrus contamination, strong residuals displaying strong spatial and seasonal variability still exist, particularly over thin cirrus prevalent regions during cirrus peak seasons; (2) challenges in matching up different data for analysis are highlighted and corresponding solutions proposed; and (3) estimates of the relative contributions from cirrus contamination to aerosol retrievals are discussed. The results are valuable for better understanding and further improving ground aerosol measurements that are critical for aerosol-related climate research.

1. Introduction

[2] Satellite data play irreplaceable roles in large-scale aerosol observations and relevant global climate change studies [e.g.,Andreae, 1991; Bréon et al., 2002; Menon et al., 2002; Huang et al., 2009]. However, the accuracy of satellite aerosol retrievals heavily relies on ground measurements because ground-based aerosol observations play an important role in calibrating and validating their spaceborne counterparts [Holben et al., 1998]. Uncertainties associated with satellite data retrieval algorithms are still largely not well quantified [e.g., Myhre et al., 2005]. Meanwhile cloud screening and quality control in ground data retrievals are also challenging [Smirnov et al., 2000; Schaap et al., 2009]. The existence of cirrus clouds with low optical thickness is still sometimes observed in the satellite and ground aerosol data products [e.g., Gao et al., 2002a; Kaufman et al., 2005; Huang et al., 2011; Chew et al., 2011]. Therefore, it is imperative to perform rigorous and systematic global evaluations on the severity of cirrus contamination in ground aerosol products and to investigate better alternatives for cirrus screening schemes.

[3] With concurrent cirrus observations from ground-based or spaceborne lidars, quantitative evaluation of cirrus contamination in the operational aerosol products becomes possible [e.g.,Huang et al., 2011]. For ground observations, the Aerosol Robotic Network (AERONET) [Holben et al., 1998] and the Micro-Pulse Lidar Network (MPLNET) [Welton et al., 2001] provide simultaneous measurements of aerosol and atmospheric profiling at their collocated sites. For satellite observations, with the advent of the A-Train satellite constellation [L'Ecuyer and Jiang, 2010], global cirrus cloud coverage and its temporal and spatial variability can be comprehensively observed for the first time [Sassen et al., 2008; Massie et al., 2010; Jiang et al., 2010]. The collocated MODIS-derived cirrus parameters and the Cloud-Aerosol Lidar and Infrared Pathfinder Satellite Observations (CALIPSO) measurements provide us with an unprecedented opportunity to examine the susceptibility of the ground aerosol products to cirrus contamination and to evaluate the robustness of current cirrus screening techniques. Such evaluation and examination are valuable for improving operational ground aerosol retrieval algorithms in related to cirrus screening and potential cirrus contamination correction.

[4] For the cloud screening in the current AERONET aerosol optical thickness (AOT) retrievals, a series of procedures are adopted by examining the temporal variability of measured AOT [Smirnov et al., 2000]. This temporal variability based technique is effective for eliminating most cloud signals [e.g., Smirnov et al., 2000; Kaufman et al., 2006]; however, residual cirrus contamination in the operational aerosol products are still observed [e.g., Gao et al., 2002a; Kaufman et al., 2005; Schaap et al., 2009; Huang et al., 2011], that warrant in-depth investigations as done in this study by taking advantage of ground and spaceborne lidar observations for detecting cirrus.

[5] For the sites where AERONET and MPLNET are collocated, lidar measurements from MPLNET can provide observational evidence of cirrus to help verify the susceptibility of aerosol data to cirrus contamination. Similarly, spaceborne observations from the CALIPSO lidar can provide an alternative cirrus observation reference for the AERONET sites that are close to one of the CALIPSO tracks. Additionally, because cirrus clouds usually occur at higher altitudes (>10 km in the tropical region) and mostly consist of ice particles, detecting cirrus from passive satellites, such as the Moderate Resolution Imaging Spectroradiometer (MODIS), can be achieved based on apparent reflectance at 1.38 μm, 0.66 μm, and 1.24 μm, and brightness temperature differences in the thermal bands [e.g., Gao and Kaufman, 1995; Gao et al., 2002a, 2002b; Roskovensky and Liou, 2003; Roskovensky et al., 2004]. In order to scale the effect of water vapor absorption, reflectance at a second channel is usually required in the practical algorithms [Gao et al., 2002b]. A ratio between the MODIS apparent reflectance at bands 1.38 μm and 0.66 μm is preferred over other satellite-derived cirrus screening parameters for detecting cirrus over Southeast Asia during the cirrus prevailing season due to its overall cirrus screening effectiveness by contrasting 1.38μm and 0.66 μm bands that have different sensitivities to atmospheric water vapor [Huang et al., 2011].

[6] Therefore, as an extension of a detailed regional study in the Biomass-burning Aerosols in South East-Asia: Smoke Impact Assessment (BASE-ASIA) campaign [Huang et al., 2011], this study aims to: (1) investigate the consistency and comparability of detecting cirrus using MPLNET and CALIPSO; (2) investigate the susceptibility of ground aerosol measurements to cirrus contamination and to quantify its influence at selected AERONET sites; this goal is achieved by exploring the susceptibility of valid and quality assured aerosol retrievals to identifying cirrus in the following pairs of matched up data: AERONET versus MPLNET, AERONET versus CALIPSO, and AERONET versus MODIS; (3) evaluate the relative optical depth contributions from cirrus contamination to ‘cirrus-contaminated’ AOT observations for those cirrus contaminated cases and examine the corresponding changes in the Ångström exponent; and (4) discuss various factors that impact the data matchup schemes used in this study and recommend solutions for future studies.

[7] This paper is arranged as follows. Section 2 describes the main data sets used in this study, followed by a detailed demonstration of results given in section 3. Last, section 4 presents our main findings and conclusions.

2. Data and Data Processing

[8] Because the main focus of the study is on ground measurements, the primary data sets for this study are concurrent ground aerosol and cirrus observations, complemented by cirrus observations from satellites. For aerosol retrievals, we used aerosol products from AERONET; for cirrus identification, we employed data from MPLNET, the CALIPSO vertical feature mask (VFM) and the MODIS-derived cirrus parameter.

2.1. AERONET

[9] The AERONET provides a long-term, continuous and readily accessible public domain database of aerosol optical, microphysical and radiative properties for aerosol research and characterization, validation of satellite retrievals, and synergism with other databases [Holben et al., 1998]. In the current AERONET AOT retrieval, the cloud screening is performed by a series of procedures that examine the temporal variability of measured AOT [Smirnov et al., 2000], including the standard deviation of three consecutive AOT measurements (triplet) over a one-minute time interval, the standard deviation of the remaining AOT (500 nm) data points acquired during a day, and observations of AOT (500 nm) and Ångström exponent with variability higher than three standard deviations within the daily intervals.

[10] For this study, only cloud-screened and quality-assured Level 2.0 AERONET AOT data at 440 nm with the highest operational quality was selected to validate against concurrent cirrus observations for computing susceptibility statistics.

2.2. MPLNET

[11] The collocated MPLNET and AERONET super sites provide both column and vertically resolved measurements of aerosols and clouds. The parameters provided include optical depth, single scatter albedo, size distribution, aerosol and cloud heights, planetary boundary layer (PBL) structure and evolution, and profiles of extinction and backscatter [Welton et al., 2001; Berkoff et al., 2008] (http://mplnet.gsfc.nasa.gov). The standard designs of Micro-Pulse Lidar (MPL) instruments, including on-site maintenance and calibration techniques, are described in details byCampbell et al. [2002]. The lidar instrument related data uncertainties were further elaborated by Welton and Campbell [2002]. Out of 16 collocated MPLNET and AERONET sites, 13 sites with overlapping temporal data coverage were selected. We primarily use MPLNET Level 1.0 normalized relative backscatter (NRB) data for cirrus visualization and cirrus flag derivation. The statistical NRB-derived cirrus flag is independent from the standard MPLNET products and it is developed in this study for automated cirrus detection. This flag is generated based on the statistical characterization of the NRB data in each time-space window (300-m in range and 10-min in time). To be discriminated from a more theoretical based cirrus flag, this cirrus flag is named as ‘Statistical Cirrus Flag’ (SCF) in this paper. Although MPLNET has both day and night observations and noise level generally increases in daytime, we had to use daytime data because AERONET data are daytime measurements. The following criteria were applied in each time-space window of the NRB data to identify the existence of cirrus cloud and to minimize the influence from noise: (1) the total number of samples has to exceed 30; (2) the averaged NRB value has to exceed 0.35; and (3) cloud base height has to be higher than 8 km.

[12] The selection of the threshold values were based on visual inspections of many cases by comparing the cirrus flag to the NRB profiles to ensure the cirrus features were separated from surrounding noise and from the aerosol and low cloud layers underneath. It is noteworthy, however, that for the Monterey and Trinidad-Head sites, the trans-pacific aerosol layers can be as high as cirrus base heights [e.g.,Eguchi et al., 2009]. In such circumstances, we only count cirrus occurrence when cloud base is greater than 10 km to avoid misidentifying aerosol layers at high altitude as cirrus. Although this conservative solution may underestimate the occurring frequency of cirrus clouds, it gives us more confidence on cirrus detection.

[13] Moreover, once SCF identifies cirrus during a 10-min window, a cirrus persistence flag (CPF) is designed to count the continuity of NRB samples that have the 300-m range averaged NRB values exceeding 0.35 at each 1-min lidar sampling step within the 10-min time window. This threshold value of NRB was determined based on its effectiveness to distinguishing cirrus features from ambient noise. CPF will be used to test the persistence of cirrus during each 10-min window. The effectiveness of SCF and CPF in cirrus detection will be elaborated insection 3.

[14] Cirrus identification highly depends on selection criteria. We will test four sets of cirrus selection criteria of SCF and CPF within different time windows (TW): ‘TW10 existence’, ‘TW30 existence’, ‘TW30 overall persistence’ and ‘TW30 strong persistence’, from less strict to most strict, respectively.

[15] 1. ‘TW10 existence’ requires SCF at each 10-min time window without any additional cirrus persistence testing.

[16] 2. ‘TW30 existence’ requires SCF at three consecutive 10-min time windows, without any additional cirrus persistence testing.

[17] 3. ‘TW30 overall persistence’ tests on both SCF and CPF at three consecutive 10-min time windows. Within each 30-min time window, it requires CPF values higher than 20 out of 30 samples collected from each one-minute MPL sampling resolution.

[18] 4. ‘TW30 strong persistence’ is the strictest test that looks at both SCF and CPF at three consecutive 10-min time windows and requires CPF values higher than 9 out of the 10 samples within each 10-min time window, and such requirements have to be met for all three consecutive windows. The difference in the results of these four settings will be discussed when they are used for the AERONET-MPLNET matchup insection 3.3.

2.3. CALIPSO

[19] The CALIPSO payload includes an active lidar instrument (CALIOP) and two passive imagers in the infrared and visible regimes to probe the vertical structure and properties of clouds and aerosols over the globe [Vaughan et al., 2005, 2009]. CALIOP provides a unique capability to closely examine the vertical profiles of aerosol and clouds from space [Young and Vaughan, 2009; Winker et al., 2010; Yorks et al., 2011]. For this study, the CALIPSO lidar version 3.01 Level 2 vertical feature mask (VFM) products, that includes a ‘transparent cirrus’ cloud subtype [Vaughan et al., 2005, 2009; Liu et al., 2009], were used as baseline for cirrus cloud detection. Only daytime CALIPSO data were used when comparisons are made with the coincident AERONET aerosol measurements. For comparison with the MPLNET cirrus detection that is performed during both day and night, both daytime and nighttime CALIPSO data were used.

2.4. MODIS

[20] For this study, considering its better sensor stability in recent years and its better comparison to CALIPSO results within the A-Train constellation, only Aqua MODIS data were used for cirrus identification. The primary data sets for assessing the cirrus screening are the MYD021KM level 1B Collection 5 data. The apparent reflectance at 1.38μm (R1.38) and the reflectance ratio (RR) between bands 1.38 μm and 0.66 μm (RR1.38/0.66) were used as indicators for the presence of optically thin cirrus at relatively large scale [Huang et al., 2011].

3. Results

3.1. Thin Cirrus Climatology From CALIPSO

[21] We study the cirrus climatology and its seasonal and regional variability to better understand their possible links to the data uncertainties in aerosol products associated with the cirrus screening. In this study, thin cirrus occurrence frequency is calculated solely based on the CALIPSO VFM products. The following three criteria are set accordingly to ensure the classification of cirrus clouds is appropriate:

[22] 1. The confidence level for a feature reported in VFM has to be high with the cloud-aerosol discrimination (CAD) score greater than 70.

[23] 2. The feature type is ‘cloud’, and the feature subtype is ‘transparent cirrus clouds.’

[24] 3. Surface return signal is detected. This ensures that the cirrus cloud detected is optically thin [Sassen et al., 2008].

[25] Based on these criteria, we calculated distributions of thin cirrus occurrence frequency. Only daytime statistics are shown in Figure 1 because the AERONET aerosol retrievals are only available at daytime. A global average of 18% in Figure 1a is comparable to 15% in Sassen et al. [2008] where they also constrained cloud top temperature to be less than −40°C from the CloudSat data in order to distinguish pure ice clouds from mixed phase clouds [Sassen and Campbell, 2001]. Average altitude of thin cirrus and its latitudinal dependence in Figure 1b are also similar to Sassen et al. [2008]. While the global distribution of cirrus occurrence is highly consistent with Sassen et al. [2008], it is noteworthy that the Tibet Plateau features much higher thin cirrus occurrence frequency than that in Sassen et al. [2008], which might be attributable to their additional control of cloud top temperature that can exclude most mixed phase clouds [Sassen and Campbell, 2001]. Moreover, Sassen et al. [2008]used version 2 of the VFM while this work employed the latest version 3.01 of the VFM, therefore the significant changes made to the cloud-aerosol discrimination algorithms between the two versions may also partly attributable for the discrepancies betweenFigure 1 and Sassen et al. [2008]. The seasonal migrations of thin cirrus prevailing regimes are also clearly seen in the thin cirrus occurrence frequencies (plots for each individual season are not shown).

Figure 1.

(a) Daytime thin cirrus occurrence frequency (%) and (b) daytime thin cirrus daytime average height (km) in each 5° × 5° grid as calculated from CALIPSO VFM (December 2006–November 2007).

3.2. MPLNET Versus CALIPSO

[26] Before comparing concurrent aerosol and cirrus observations, it is intriguing to compare the cirrus detection capability of ground lidar and its spaceborne counterpart, by crosschecking the effectiveness of MPLNET NRB-derived SCF and the CALIPSO VFM. While a quantitative direct comparison between the MPLNET and CALIPSO cloud measurements may be challenging, an indicative qualitative comparison in terms of cirrus existence is feasible [Berkoff et al., 2008; Dupont et al., 2010; Thorsen et al., 2011]. The most challenging remaining issue is the distance between the MPLNET sites and CALIPSO overpass tracks. The CALIPSO satellite does not always pass right over the MPLNET sites. Another challenge is that, in some cases, the CALIPSO overpass time is close to the time when the MPL is turned off around solar noon to prevent damage to the instrument's optics by avoid receiving strong sunlight, which is more critical for tropical sites that have very small solar zenith angle around high noon. Additionally the CALIPSO 16-day repeat cycle also significantly reduces the sample size of MPLNET-CALIPSO collocation.

[27] A first-step crosscheck between MPL and CALIPSO is the cirrus occurrence seasonality. We selected four AERONET-MPLNET sites (GSFC, COVE, Trinidad_Head and NCU_Taiwan) that exhibit the longest multiple year data coverage to give equal sampling weight to different seasons.Table 1tabulates the cirrus occurrence seasonality derived from both MPLNET and CALIPSO. For MPLNET data, we defined the MPLNET cirrus occurrence frequency as a ratio (percentage) of the number of 10-min time windows that observed cirrus over the total number of 10-min time windows during an observation time period of one hour (30 min before and after the CALIPSO overpass time). The CALIPSO satellite usually passes over the selected MPLNET sites between 13:00 and 14:00 local time in daytime. For each site, the values for CALIPSO are thin cirrus occurrence frequency observed from all CALIPSO track overpasses with nearest distance less than 100 km to the site. The statistics were calculated independently for each data source without point-to-point data matching. Listed inTable 1 are the corresponding seasonal and annual mean values from MPLNET and CALIPSO.

Table 1. Comparison Between MPLNET and CALIPSO on the Seasonality of Daytime Thin Cirrus Occurrence Frequency (%) Over Four MPLNET Sites With Longest Data Recordsa
SiteUTCMPLNET (%)CALIPSO (%)
MAMJJASONDJF4-SeasonMAMJJASONDJF4-Season
  • a

    For each site, the values for MPLNET are the percentage of cirrus cases over the total MPLNET measurements during +/−30 min around the CALIPSO daytime overpass time. For each site, the values for CALIPSO are thin cirrus occurrence frequency observed from all CALIPSO track overpasses with nearest distance less than 100 km to the site. The highest and lowest seasons are boldfaced for each site.

GSFC (39N, 76W)18.0–19.011.37 (360/3165)15.65 (525/3355)15.24 (439/2880)14.14 (398/2815)14.10 (1722/12215)20.21 (230/1138)21.50 (263/1223)12.51 (142/1135)16.20 (175/1080)17.70 (810/4576)
COVE (37N, 76W)17.9–18.912.42 (281/2262)13.95 (327/2337)10.35 (187/1806)13.24 (278/2100)12.49 (1073/8505)15.73 (135/858)24.57 (243/989)14.40 (126/875)18.81 (149/792)12.35 (434/3514)
Trinidad_Head (41N, 124W)20.7–21.712.27 (276/2250)3.38 (105/3102)7.27 (178/2448)11.34 (264/2328)8.13 (823/10128)28.10 (222/790)10.85 (112/1032)16.42 (123/749)15.95 (123/771)14.33 (479/3342)
NCU_Taiwan (25N, 121E)5.0–6.05.33 (116/2175)9.59 (211/2200)5.31 (103/1940)0.76 (15/1985)5.36 (445/8300)13.10 (87/664)22.98 (191/831)6.64 (42/632)4.75 (36/758)12.34 (356/2885)

[28] Overall, the annual means of cirrus frequency from MPLNET and CALIPSO are comparable in their orders of magnitudes: 14.10 versus 17.70, 12.62 49 versus 12.35, 8.13 versus 14.33, 5.36 versus 12.34 percent for GSFC, COVE, Trinidad_Head and NCU_Taiwan, respectively. There are systematic viewing geometry differences between the ground-based and spaceborne lidar systems. Moreover, the annual and seasonal means listed inTable 1were not from one-to-one cirrus case matching up and they were calculated for each sensor separately. Therefore caution should be advised for quantitative comparisons between the two instruments. For all four sites, the MPLNET and CALIPSO agreed on the thin cirrus peak seasons: GSFC, COVE and NCU for June–July–August (JJA), and Trinidad_Head for March–April–May (MAM). Three out of the four sites agreed on the least cirrus occurrence frequency (COVE for September–October–November (SON), Trinidad_Head for JJA, and NCU for December–January–February (DJF)) except GSFC where MPLNET exhibited a low cirrus season for MAM but for CALIPSO the low cirrus season was SON. Although they both agree on the cirrus peaks seasons, the discrepancy is also significant: the CALIPSO detected cirrus frequencies for the peak seasons were generally higher than those from the MPLNET: 21.50% versus 15.65%, 24.57% versus 13.95%, 28.10% versus 12.27%, and 22.98% versus 9.59% for GSFC, COVE, Trinidad_Head and NCU sites, respectively. There are three possible reasons for such discrepancies: First, the CALIPSO's ‘top-down’ viewing geometry allows better detection of high clouds before the lidar signal become attenuated; However, in the MPL's ‘bottom-up’ viewing geometry, lidar signal could be attenuated by aerosol layers, fog and low clouds significantly before it reaches cirrus altitude. For example, the ground lidar systems located at NCU_Taiwan and Trinidad_Head are frequently impacted by extremely heavy fogs during much of the year, while CALIPSO still reports higher cirrus presence from those sites. With the further developments of both MPLNET and CALIPSO products, systematic screening of those aerosol, fog and low clouds cases should be carefully conducted to reach a more reasonable comparison between MPLNET and CALIPSO thin cirrus occurrence frequency for future efforts. Second, noontime measurements are always difficult for ground lidar, because the noise levels are usually much higher when the solar zenith angle is low which makes automated cirrus detection more challenging. Moreover the MPL lidar noontime shut-down protective measure also prevents continuous observations of thin cirrus around local noontime. This second factor is expected to have a bigger impact on tropical sites during boreal summer time, such as NCU_Taiwan with a 22.98% versus 9.59% difference. Last, cirrus detection limits from the SCF method and the CALIPSO cirrus scheme are not necessarily equal, which causes differences in the cirrus un-detectable rates. Supporting the effectiveness of the SCF approach, previous studies [e.g.,Dupont et al., 2010; Thorsen et al., 2011] also observed large discrepancies between ground- and space-based data sets in terms of cirrus optical properties and reported relatively larger cumulative cirrus occurrence of subvisible thin cirrus from CALIOP than from ground-based lidars.

[29] To gain more insight on the comparability between MPLNET and CALIPSO, we further matched up 9 MPLNET-AERONET collocated sites (seeTable 2). To ensure a one-to-one matchup of the data, we only chose those data pairs with the closest distance of CALIPSO track to the site and the closest MPLNET data collection time (within ±5 min) to the CALIPSO overpass. Because CALIPSO overpass tracks shift slightly within a range of ∼15–20 km between tracks during the 16-day repeat cycle at each site, the distance between the sites and CALIPSO tracks also varies. Seen fromTable 2, among the 9 sites, some sites (i.e., Gosan_SNU) have a distance less than 10 km, but other sites (i.e., GSFC) can have larger ranges up to 90 km. Despite all the challenges and the limited sample number of collocated cases, close examination of all cirrus cases from June 2006 to December 2010 indicated that, in terms of cirrus detection, for the 8 sites (except Singapore) that have more than 20 matchups (∼one year of day or night data coverage considering 16-day CALIPSO data cycle), MPLNET and CALIPSO reached a percentage agreement of 71–88% when both daytime and nighttime cases were counted. The agreement results are not much different between daytime and nighttime. This not only establishes the general comparability of the MPLNET L1.0 SCF and the CALIPSO VFM in terms of cirrus detection, but it also demonstrates the effectiveness of MPL L1.0 SCF for detecting cirrus without significant impacts from large noise during the daytime. A very noteworthy point is that when MPLNET cirrus criteria were set much tighter, for example, from “TW10 existence” to “TW30 strong persistence,” the number of cirrus cases decreased significantly. Such sensitivity to cirrus detection criteria impacts the AERONET-MPLNET matchup significantly, which contributes to the discrepancy between the results from the AERONET-MPLNET matchup and the results from the AERONET-CALIPSO matchup, in addition to the already existing temporal and spatial differences of matched up samples. Similarly for CALIPSO cirrus detection, the multiresolution averaging scheme also occasionally results in spatial smearing, both horizontally and vertically, of broken cirrus clouds [Yorks et al., 2011]. This sensitivity will be further discussed in the following sections.

Table 2. Statistics on the MPLNET-CALIPSO Matchup Over the 9 AERONET-MPLNET Collocated Sites During Daytime and Nighttimea
SiteCasesDaytimeNighttimeAgreement % (Day and Night)
DayMPLVFMBothAgreement % (Day)NightMPLVFMBothAgreement % (Night)
  • a

    In the ‘MPL’ column, the numbers outside parentheses are from the ‘TW10 existence’ tests, and the numbers inside parentheses are from the ‘TW30 strong persistence’ tests. Similarly, in the ‘Both’ row, the numbers are the corresponding MPL cases that agreed with the CALIPSO VFM cirrus testing. In the last column, ‘Agreement %’ is the percentage of MPL and CALIPSO agreed cases over the total matchup cases. The distance (km) in the ‘Day’ and ‘Night’ columns are allowance thresholds of the distance between the site and the CALIPSO overpass tracks.

GSFC (39N, 76W)Cirrus90 km5 (1)54 (1) 40 km10 (3)145 (3)  
Non-Cirrus 222221  514742  
Total Cases 27272592.59% 61614777.05%81.82%
COVE (37N, 76W)Cirrus50 km24 (0)138 (0) 20 km15 (3)239 (3)  
Non-Cirrus 485943  544640  
Total Cases 72725170.83% 69694971.01%70.92%
NCU (25N, 121E)Cirrus30 km2 (0)71 (0) 70 km5 (1)72 (1)  
Non-Cirrus 605554  545249  
Total Cases 62625588.71% 59595186.44%87.60%
Trinidad_head (41N, 124W)Cirrus30 km2 (0)21 (0) 40 km8 (3)74 (2)  
Non-Cirrus 202019  252622  
Total Cases 22222090.91% 33332678.79%83.63%
Gosan_SNU (33N, 128E)Cirrus10 km5 (0)92 (0) No samples     
Non-Cirrus 433936       
Total Cases 48483879.17% 79.17%    
Monterey (37N, 122W)Cirrus30 km4 (0)71 (0) 20 km4 (2)84 (2)  
Non-Cirrus 363330  383434  
Total Cases 40403177.50% 42423890.48%84.14%
Barbados (13N, 59W)Cirrus80 km0 (0)00 (0) 50 km12 (3)2110 (3)  
Non-Cirrus 111  352624  
Total Cases 111100% 47473472.34%72.92%
Singapore (1N, 104E)Cirrus80 km2 (0)41 (0) No samples     
Non-Cirrus 1087       
Total Cases 1212866.67%     66.67%
Kanpur (26N, 80E)Cirrus35 km4 (2)31 (1) 70 km2 (0)32 (0)  
Non-Cirrus 212219  544  
Total Cases 25252080% 77685.71%81.25%

3.3. AERONET Versus MPLNET

3.3.1. AERONET-MPLNET Matchup

[30] The AERONET AOT retrievals were paired up with the MPLNET NRB-derived SCF and CPF to calculate susceptibility percentage (%, SP), an indicator of how many percentages of best quality assured L2.0 AOT retrievals are potentially contaminated by cirrus. Results about SP will be discussed insection 3.3.2. The MPLNET SCF and CPF calculations and the four MPLNET cirrus detection criteria settings were discussed in section 2.2. Additional requirements for the one-to-one AERONET-MPLNET matchup are as follows.

[31] 1. For each of the four MPL cirrus criteria settings, AERONET has to have valid quality assured L2.0 AOT retrievals at 440 nm within the central MPL SCF 10-min time window, to be counted as being potentially susceptible to cirrus contamination.

[32] 2. At each matchup, the solar zenith angle (SZA) has to be less than 20°. This is because the MPLNET lidar's viewing angle is fixed to vertical direction while the AERONET sunphotometer is always tracking and looking at the sun and measures an optical thickness along a slant path. The less the solar zenith angle, the atmospheric paths as observed by both instruments are better matched up. They never exactly overlap however, because the micro-pulse lidar cannot look into the sun.

[33] To further elaborate on the AERONET-MPLNET matchup,Figure 2shows the MPL NRB, SCF, and CPF in their respective (a)-(c) panels for the cirrus case over the COVE site on June 7, 2007. The persistent cirrus layer around 11–12 km altitude was clearly seen from the NRB profile (Figure 2a). After statistical analysis, both SCF and CPF showed consistent results with the NRB observations. SCF in Figure 2bshows the corresponding NRB values when cirrus existence was identified at each 10-min time window. In comparison toFigure 2a, Figure 2b shows that the SCF filtering process removed most of ambient noises effectively, demonstrating that SCF is capable of distinguishing cirrus layers from noise very effectively. CPF in Figure 2c on the other hand described the continuity of the cirrus layers that had persistent strong lidar scattering signals (NRB > 0.35). Therefore when low cloud attenuated the lidar signal significantly, for example the case around 10:40AM, the CPF number decreased correspondingly because of the weaker NRB strength. The corresponding AERONET measurements, including AOT (440 nm), AOT (500 nm), Ångström exponent (440–675 nm) susceptible to cirrus contamination, and solar zenith angle, are also shown in Figure 2d. It is noteworthy however that the aerosol measurements around 9AM local hour were not counted as cirrus contaminated cases because SZA was 41°, which did not pass the SZA < 20° test.

Figure 2.

An example cirrus occurrence case over COVE AERONET and MPLNET site on June 7, 2007: (a) MPL L1.0 normalized relative backscatter (NRB) higher than 0.35; the square and triangle markers are for cloud top and base heights as in the MPL standard cloud height products, respectively; (b) MPL statistical cirrus flag (SCF); (c) MPL statistical cirrus persistence flag (CPF); and (d) AERONET AOT and Ångström exponent measurements, and solar zenith angle (SZA) (note the SZA for the data measurement around 9am was 41°, which did not pass the SZA < 20° test and is therefore off the chart).

3.3.2. Susceptibility Percentage (SP)

[34] Susceptibility percentage (SP) is defined as the percentage of aerosol retrievals that are susceptible to cirrus contamination to the total numbers of aerosol retrievals that contains both ‘cirrus-contaminated’ and ‘cirrus-free’ cases. Because matchup criteria can be less strict or very strict, SP values change with different settings of matchup criteria.Table 3 summarizes the sensitivity of SP to cirrus existence and persistence criteria settings, time window selections, and SZA, for all 13 sites with their temporal coverage sorted in order. As seen in Table 1, changes in SP can be an order of magnitude simply because of different cirrus selection criteria. For example, the SP values at GSFC were 7.74%, 3.61%, 3.44% and 1.55% for ‘TW10 existence’, ‘TW30 existence’, ‘TW30 overall persistence’ and ‘TW30 strong persistence’ respectively. The reasons are twofold: one is the actual spatial and temporal variability of cirrus clouds, the other is the way that lidar looks upright for high cloud detection and gets attenuated along the atmospheric path. Although cirrus usually occur at synoptic scales, low clouds, aerosol and the atmosphere can significantly attenuate the MPL lidar signal, before it reaches more than the 10 km height to detect cirrus. Therefore any occurrence of heavy low or middle cloud or heavy aerosol could prevent continuous observation of cirrus. Note that this impact gets particularly stronger around noontime when noise levels usually increase significantly (see Figure 2a), which makes cirrus detection even more challenging as it requires relatively stronger lidar signals in order to discriminate cirrus from ambient noises. Moreover, the MPL lidar shutdown around high noon at low SZA hours also prevented continuous observations of cirrus persistence, particularly for tropical sites. Thus additional strong persistence testing (e.g., ‘TW20 strong persistence’) resulted in much lower SP values than relatively weaker persistence testing (e.g., ‘TW10 existence’). SP values for the top 10 AERONET-MPLNET sites from the ‘TW30 overall persistence’ testing are plotted on top of the CALIPSO thin cirrus occurrence frequency map inFigure 3. With the ‘TW30 overall persistence’ testing and the SZA filtering (SZA < 20°), all 10 sites have SP values less than 5% and 4 of them (40%) are actually less than 1% (Figure 3); but for the ‘TW10 existence’ testing, 6 out of 10 sites (60%) have SP values within 4–10%, and the other 4 (40%) within 1–3%. Similarly in Table 3, when the time window becomes larger, for example, changing cirrus detection from 15-min time window to 30-min or 60-min time windows, the requirements for cirrus strong persistence also become higher, thus less cirrus cases were detected, and SP values become lower correspondingly. For example, at GSFC, the SP values for TW15, TW30 and TW60 were 3.10%, 1.55% and 1.20% respectively.

Table 3. Susceptibility Percentage (SP, %) of AERONET Level 2.0 AOT Retrievals to Cirrus Contamination, and Its Sensitivity to Cirrus Existence and Persistence Criteria Settings, Time Window (TW) and Solar Zenith Angle (SZA)a
Site NameMPLNET CoverageSensitivity of SP (%) to Cirrus Existence and Persistence Criteria Settings (Sample Size N: Cirrus/Total)Sensitivity of SP (%) to Time Window (Sample Size N: Cirrus/Total)Sensitivity of SP (%) to SZA (Sample Size N: Cirrus/Total)
SZA < 20º TW10 ExistenceSZA < 20º TW30 ExistenceSZA < 20º TW30 Overall PersistenceSZA < 20º TW30 Strong PersistenceSZA < 20º TW15 Strong PersistenceSZA < 20º TW30 Strong PersistenceSZA < 20º TW60 Strong PersistenceSZA < 20º TW30 Strong PersistenceAll SZA TW30 Strong Persistence
  • a

    Samples are from all seasons. The numbers inside parentheses are the sample size of ‘cirrus cases’ over the total sample size of ‘cirrus and non-cirrus cases’, as the calculations of SP values. The SP values with the ‘TW30 overall persistence’ tests were plotted inFigure 3).

GSFC (38N, 76W)2001.11–2010.067.74 (45/581)3.61 (21/581)3.44 (20/581)1.55 (9/581)3.10 (18/581)1.55 (9/581)1.20 (7/581)1.55 (9/581)3.29 (1265/38437)
COVE (37N, 76W)2004.05–2010.094.44 (11/248)2.82 (7/248)2.42 (7/248)2.82 (7/248)4.03 (10/248)2.82 (7/248)2.82 (7/248)2.82 (7/248)2.21 (186/8409)
Trinidad_Head (41N, 124W)2005.05–2010.091.14 (2/175)0 (0/175)0 (0/175)0 (0/175)0 (0/175)0 (0/175)0 (0/175)0 (0/175)2.10 (223/10612)
NCU_Taiwan (25N, 121E)2005.01–2009.104.00 (6/150)0.67 (1/150)0.67 (1/150)0.67 (1/150)1.33 (2/150)0.67 (1/150)0.67 (1/150)0.67 (1/150)4.10 (139/3394)
Gosan_SNU (33N, 128E)2007.04–2010.024.76 (5/105)0.95 (1/105)0.95 (1/105)0.95 (1/105)1.90 (2/105)0.95 (1/105)0 (0/105)0.95 (1/105)0.68 (11/1620)
Monterey (37N, 122W)2003.04–2003.10 2004.01–2004.04 2007.03–2009.042.44 (19/780)1.41 (11/780)1.41 (11/780)0.26 (2/780)1.41 (11/780)0.26 (2/780)0 (0/780)0.26 (2/780)2.80 (537/19152)
Ragged_Point (13N, 59W)2008.06–2011.091.95 (16/819)0.49 (4/819)0.49 (4/819)0.49 (4/819)0.98 (8/819)0.49 (4/819)0.24 (2/819)0.49 (4/819)4.17 (344/8242)
Singapore (1N, 104E)2009.09–2011.096.35 (24/378)2.65 (10/378)2.65 (10/378)0.79 (3/378)2.91 (11/378)0.79 (3/378)0 (0/378)0.79 (3/378)4.85 (239/4923)
Kanpur (26N, 80E)2009.05–2010.092.36 (9/381)1.31 (5/381)1.31 (5/381)1.31 (5/381)1.60 (6/375)1.31 (5/381)1.60 (6/375)1.31 (5/381)2.74 (127/4634)
Pimai (15N, 102E)2006.02–2006.058.33 (14/168)3.57 (6/168)3.57 (6/168)2.38 (4/168)5.52 (9/163)2.38 (4/168)0 (0/163)2.38 (4/168)8.65 (148/1711)
Skukuza (25S, 32E)1999.08–1999.09 2000.08–2000.09----------------0.26 (3/1176)
Mongu (15S, 23E)2000.08–2000.090 (0/16)0 (0/16)0 (0/16)0 (0/16)0 (0/16)0 (0/16)0 (0/16)0 (0/16)0.72 (7/978)
XiangHe (40N, 117E)2005.02–2005.05----------------1.74 (21/1207)
Figure 3.

Susceptibility percentage (SP, %) of AERONET L2.0 AOT retrievals to cirrus contamination as tested against the MPLNET statistical cirrus flag.

[35] Viewing geometry differences between the sunphotometer and micro-pulse lidar can affect the SP assessment dramatically. For example for GSFC, the SP value increases significantly from 1.55% to 3.29% when the SZA constraint changes from SZA < 20° to all SZA applying the ‘TW30 strong persistence’ test (seeTable 3). The ‘SZA < 20°’ control is conducted to account for the viewing geometry differences between sunphotometers and lidar instruments. A ‘SZA < 20°’ criterion ensures a better matchup. On the downside however, a ‘SZA < 20°’ screening significantly reduced the sample sizes. For example, the number of cirrus cases for ‘TW30, all SZA’ at GSFC (Table 3) was found to be 1265 versus 9 for ‘TW30, SZA < 20°’. However, it is worthwhile to emphasis that the AERONET-MPLNET matchups that sample at higher SZA (i.e., SZA > 20) are less indicative of cirrus contamination in the AERONET measurements because the two instruments were more likely looking at different atmospheric paths when their viewing angles were widely separated.

[36] Seasonal variability was also found in the SP statistics. The derived SP values shown in Figure 3 and tabulated in Table 3 features strong seasonal signals. Table 4 compares cirrus statistics of SP values and samples for their seasonality over the 13 sites. For example, cirrus cases occurred more frequently in boreal spring for Pimai and in boreal summer for GSFC and COVE (also see Table 1). All the 10 cirrus cases in the ‘TW30 strong persistence’ testing over GSFC were from boreal summer. Both ‘TW10 existence’ and ‘TW30 overall persistence’ tests indicate similar seasonality of cirrus occurrence at each site.

Table 4. Seasonality of Susceptibility Percentage (SP, %) of AERONET Level 2.0 AOT Retrievals to Cirrus Contamination, and Its Sensitivity to Cirrus Existence and Persistence Criteria Settingsa
Site NameTW10 ExistenceTW30 Overall Persistence
MAM SP%, SZA < 20°JJA SP%, SZA < 20°SON SP%, SZA < 20°DJF SP%, SZA < 20°All Seasons, SP (%), SZA < 20°MAM SP%, SZA < 20°JJA SP%, SZA < 20°SON SP%, SZA < 20°DJF SP%, SZA < 20°All Seasons, SP (%), SZA < 20°
  • a

    Two types of cirrus persistence criteria settings (TW10 existence and TW30 overall persistence) are shown.

GSFC (39N, 76W)7.09 (9/127)7.93 (36/454)7.74 (45/581)0.79 (1/127)4.19 (19/454)3.44 (20/581)
COVE (37N, 76W)2.54 (3/118)6.15 (8/130)4.44 (11/248)0.85 (1/117)4.10 (5/122)2.42 (7/248)
Trinidad_Head (41N, 124W)0 (0/2)1.16 (2/173)1.14 (2/175)0 (0/2)0 (0/173)0 (0/175)
NCU_Taiwan (25N, 121E)11.1 (4/36)2.04 (2/98)(0/16) 4.00 (6/150)2.78 (1/36)0 (0/98)0 (0/16)0.67 (1/150)
Gosan_SNU (33N, 128E)0 (0/59)10.87 (5/46)4.76 (5/105)0 (0/59)2.17 (1/46)0.95 (1/105)
Monterey (37N, 122W)5.88 (11/187)1.35 (8/593)  2.44 (19/780)4.28 (8/187)0.51 (3/593)1.41 (11/780)
Ragged_Point (13N, 59W)1.32 (4/303)1.11 (4/360)5.13 (8/156)1.95 (16/819)0.33 (1/303)0.28 (1/360)1.28 (2/156)0.49 (4/819)
Singapore (1N, 104E)5.56 (3/54)9.33 (7/75)6.81 (13/191)1.72 (1/58)6.35 (24/378)1.85 (1/54)2.67 (2/75)3.66 (7/191)0 (0/58)2.65 (10/378)
Kanpur (26N, 80E)1.43 (2/140)2.90 (7/241)2.36 (9/381)0.71 (1/140)1.66 (4/241)1.31 (5/381)
Pimai (15N, 102E)8.33 (14/168)8.33 (14/168)3.57 (6/168)3.57 (6/168)
Skukuza (25S, 32E)
Mongu (15S, 23E)0 (0/16)0 (0/16)0 (0/16)0 (0/16)
XiangHe (40N, 117E)

3.3.3. Cirrus Optical Depth Calculation for Selected Cases

[37] We further investigated each individual cirrus case identified in the AERONET-MPLNET matchup for more details. With given NRB and molecular backscatter profiles, molecular optical depth can be calculated from molecular extinction profiles based on the National Centers for Environment Prediction (NCEP) vertical temperature and pressure profiles, thus theoretically cirrus optical depth can also be calculated [e.g.,Young, 1995; Del Guasta, 1998]:

display math
display math

where subscripts b and t denote cirrus base and top, respectively. β and τ are molecular backscatter and optical depth respectively, while m and c stand for molecular and cirrus. C is a coefficient that counts for lidar performance and lidar signal attenuation due to other aerosol or cloud layers beneath cirrus. All these parameters are retrieved at cirrus base and cirrus top heights. From (1) and (2), cirrus optical depth can be calculated as [e.g., Young and Vaughan, 2009]:

display math

The challenge however comes from the following two influential factors that prevent precise measuring of NRB values at high altitude in daytime: (1) MPL signals become extremely weak when it reaches an altitude higher than 10 km where cirrus layers reside, particularly after being further attenuated by cirrus; and (2) during daytime, particularly around local noon time when the AERONET-MPLNET matchup requires the closeness of viewing geometries from both instruments (SZA < 20°), noise level also increases significantly (seeFigure 2a). Therefore operational cirrus optical depth estimation based on the MPL data set faces extreme difficulties. In this work, we selected a very limited number of quality cirrus cases for testing an empirical approach for calculating cirrus optical depth, in the scope of evaluating relative contribution of cirrus optical depth to total optical depth observed by the sunphotometer. We assessed all cases for lidar operational stability, lidar signal strengths before and after cirrus layers, and persistence of cirrus layers. Results from two test cases over the GSFC site on June 7th 2007 are shown in Figure 4.

Figure 4.

Cirrus optical depth estimation for cirrus cases over GSFC on June 7, 2007: (a) NRB profile from 12 to 20 UTC (local 7am to 5pm). The two matchup cases are highlighted by vertical lines. (b) Cirrus optical depth calculation results for the case of 16:12UTC. (c) Cirrus optical depth calculation results for the case of 16:22UTC. The green dashed lines in Figures 4b and 4c mark the NRB threshold value of 0.35.

[38] Figure 4a shows a very persistent cirrus layer lasting for more than 8 h over GSFC on June 7th 2007. To overcome the influence from noise particularly above cirrus layers, we used the data distribution pattern from the concurrent molecular backscatter profile (calculated from NCEP data source) to proxy the NRB data distributions beneath and above cirrus layers (Figures 4b and 4c). The assumption is that in the clear portions of the atmosphere (i.e., above aerosol and low clouds but below cirrus clouds), the data distribution pattern of the NRB profile is similar to the data distribution pattern of the molecular backscatter profile. Such data similarity of NRB profile to molecular scattering profiles that are not interfered by cloud, aerosol and noises, has been broadly discussed in previous literatures [e.g., Sassen et al., 1989; Vaughan et al., 2005, 2009; Young, 1995; Del Guasta, 1998; Young and Vaughan, 2009]. This assumption was further verified by the observations at MPLNET night scenes when noise levels were significantly low. For these two particular cases, the measured NRB profile data from 4 km to 10 km and the collocated molecular backscatter profile data from the same vertical range were trained to find a best linear fit function between the two data sets. This ‘trained’ best fit function was then applied to the molecular backscatter data to approximately calculate the NRB data “right beneath” and “just after” cirrus layers, the required inputs to the equation (3). The best linear fit training and utilization can be formulated as the following equations:

display math

where F is the best fitting function derived from:

display math

[39] Then, cirrus optical depth can be calculated in equation (3)by using the approximated NRB values, the molecular backscatter and molecular optical depth data as inputs. The molecular backscatter and optical depth were calculated from a Rayleigh radiative model based on NCEP reanalysis temperature and pressure profiles. Results show roughly 30–50% relative contributions from cirrus to the possibly ‘cirrus-contaminated’ AOT retrievals at 527 nm, 0.0926 versus 0.270 for 16:12UTC case, and 0.123 versus 0.253 for the 16:22UTC case. However, despite the residual profile-fitting uncertainties, this level of cirrus optical depth did not seem to decrease Ångström exponent significantly to a very low value. The Ångström Exponents were 1.0 for both cases under cirrus contaminations.

3.4. AERONET Versus CALIPSO

[40] Another approach for assessing cirrus contamination in the AERONET AOT retrievals is to compare them with CALIPSO cirrus observations. The complication, however, comes from the limited CALIPSO temporal coverage at each site because of the 16-day repeating cycle and the distant distance between the CALIPSO overpass tracks and most AERONET sites. We selected the AERONET and CALIPSO data by the following criteria: (1) the distance between the AERONET site and a nearest CALIPSO track is smaller than 10 km; (2) the CALIPSO overpass time is within +/− 10 min of the AERONET data collection time, and (3) the total sample size of matched-up data has to exceed 20 for each AERONET site, roughly corresponding to about one-year of CALIPSO and AERONET paired data because CALIPSO overpass the same site every 16 days.

[41] After matching up the data, the resulting SP values for the 18 AERONET sites that passed time and space filtering are presented in Figure 5, superimposed on an annual mean thin cirrus occurrence frequency map. Over half (10 out of 18) sites have SP values less than 10%, which means there is a relative low level of susceptibility of AOT retrieval to cirrus contamination (Figure 5). This level of SP values is relatively comparable in the order of magnitude to the AERONET-MPLNET ‘TW10 Existence’ testing (seeTable 3). However, some sites showed much larger SP values, for example, 33% for CARTEL, 23% for CEILAP-BA, and 21% for Xianghe that are outside of the cirrus prevailing regions, and 25% for Ilorin which is within the tropical cirrus region. Because the cirrus occurrence frequencies (Figure 5) for those sites outside of the cirrus prevailing regions are not high, more strict cloud screenings in the AERONET observations at these sites are recommended. Statistics were also calculated for four boreal seasons separately but sample sizes are rather limited. Similar to the AERONET-MPLNET comparison, strong seasonal and regional variability were also found for the distributions of SP values over these sites, which tend to be higher during the local cirrus prevailing seasons. Statistics also indicate that sample size issues can affect SP values significantly. For example, if we increase the sample size requirement to 40 (equivalent to about two years of CALIPSO and AERONET matched-up data) instead of 20, only 6 sites would have passed the threshold and all of them would have SP values less than 15%, which is closer to the AERONET-MPLNET evaluation results from the ‘TW10 existence’ testing.

Figure 5.

Susceptibility percentage (SP, %) tests of AERONET L2.0 AOT retrievals against the CALIPSO vertical feature mask. Refer to section 3.4for more details of the one-to-one matchup criteria.

[42] The SP values from the majority of sites in the AERONET-MPLNET and the AERONET-CALIPSO matchups are comparable in the order of magnitude. For example, 60% of the sites have SP values of 4–10% in the ‘TW10 existence’ testing shown inTable 3, and about half the sites with less than 10% in Figure 5(note that all sites have SP values less than 15% if the sample size requirement is set to 40). However, the discrepancy between AERONET-MPLNET (Tables 3 and 4 and Figure 3) and AERONET-CALIPSO (Figure 5) was also observed. Possible explanations are the following: (1) The AERONET-MPLNET and AERONET-CALISPO matchups are based on different spatial-temporal domains. The former and latter are related more to time/distance constraints, respectively. (2) MPL and CALIPSO observe cirrus occurrence frequency differently, while the MPL usually has lower values than CALIPSO during cirrus peak seasons, as explained insection 3.2 (Table 1). For the cases when MPLNET were shut down or attenuated by aerosol, fog or low clouds and couldn't reach cirrus altitude, the AERONET will mostly likely not have valid aerosol retrievals. Therefore the MPLNET and AERONET observation conditions are more self-consistent by nature than the top-down CALIPSO observation. (3) The SP values are highly sensitive to the selection of cirrus detection criteria (seeTable 24). The tighter the cirrus detection requirements are the less cirrus cases were identified.

3.5. AERONET-MPLNET-CALIPSO 3-Way Matchup

[43] To extend investigations in susceptibility percentage discrepancies between AERONET-MPLNET and AERONET-CALIPSO beyond the matchups of MPLNET-CALIPSO (section 3.2), AERONET-MPLNET (section 3.3), and AERONET-CALIPSO (section 3.4), it is intriguing to see whether we can identify sufficient samples for a 3-way AERONET-MPLNET-CALIPSO matchup. Such data matching is only valid for daytime because AERONET aerosol data are only measured during daytime. A two-step matchup procedure was adopted: 1) match up MPLNET-CALIPSO as described insection 3.2, and identify the MPLNET data collection times that are closest to the CALIPSO overpass; 2) match up AERONET aerosol data around the MPLNET data collection time identified in Step 1. Two different temporal limitations were tested for comparison: (1) any AERONET aerosol AOT 440 nm measurements within 0.5 h of MPLNET cirrus cases matched up with CALIPSO overpass were considered ‘cirrus susceptible’; and (2) any AERONET aerosol measurements within 1 h of MPLNET cirrus cases matched up with CALIPSO overpass were considered ‘cirrus susceptible’. Unfortunately very few ‘cirrus susceptible’ cases were found from the 3-way comparison for the 9 sites inTable 2. For the GSFC site, 27 AERONET-MPLNET-CALIPSO matchup cases were identified, where both MPL and CALIPSO agreed on four cirrus cases. Of the four cases, one AERONET matchup was identified as ‘cirrus susceptible’ using the ‘TW30 overall persistence’ testing for the 1-h time allowance, and none were identified for the 0.5-h time allowance. It is noted that the numbers are not statistically significant due to the insufficient sample sizes. However, the study successfully demonstrates the 3-way matchup approach, which will prove to be more valuable as longer CALIPSO data sets and more MPLNET-AERONET collocated sites become available. Collective information resulting from a 3-way data yields improved constraints for cirrus susceptibility testing because it provides two independent verification channels for concurrent cirrus detection.

3.6. AERONET Versus MODIS

[44] One of the important objectives of this study is to investigate the feasibility of using satellite derived cloud screening parameters for cloud screening of AERONET aerosol retrievals around the satellite overpass time. Therefore, it is essential to explore the susceptibility of AERONET retrievals to cirrus contamination at AERONET sites during the MODIS overpass times. Because RR1.38/0.66 is indicative of cirrus [Roskovensky and Liou, 2003; Huang et al., 2011], AERONET AOT and Fine Mode Fraction (FMF) measurements were collocated with the MODIS-derived RR1.38/0.66 over select AERONET sites. The 15 AERONET sites were chosen according to their L2.0 AOT data availability and their representativeness on a global map: 4 of them have 5+ year data records and the other 11 have 7+ year data records. Further spatial and temporal constraints for the collocations are: (1) spatially, considering the 1 km resolution of MODIS L1B data, the closest RR1.38/0.66 value are retrieved within 1 km distance from each AERONET site; and (2) temporally, the closest AERONET data points are collected within a ±30 min time window centered at the MODIS overpass time.

[45] Figure 6 shows overall susceptibility levels of AERONET AOT and FMF data at the 15 sites. For both AOT and FMF, there are 13 (93%) sites having the SP value less than 10%, a comparable SP level to the previous comparisons in AERONET versus CALIPSO, indicating the effectiveness of current AERONET cloud screening schemes on successfully identifying cirrus existence.

Figure 6.

Susceptibility percentage map of AERONET aerosol retrievals against MODIS derived RR1.38/0.66 over 15 AERONET sites. The four eastern most sites were selected with 5+ years of L2.0 AOT data record; and all the remaining sites were selected with 7+ years of L2.0 AOT data records available. SP values (%) in red are for AOT and yellow for FMF.

[46] Because cirrus cloud particle sizes are larger than aerosols, potential cirrus contamination can be reflected in the changes of the aerosol's particle size distribution; and this phenomenon should become more significant over aerosol emission regions where fine aerosol particles (such as smoke) usually prevail. In order to see the changes of Ångström exponent (AE) and FMF transitioning from cirrus-free to cirrus-contaminated cases, we selected three representative AERONET sites having the longest L2.0 AOT data record over three smoke predominant regions during their peak smoke seasons respectively: Alta_Floresta in Amazon during SON, 2004–2009; Mukdahan in Southeast Asia during MAM, 2004–2009; Mongu in Southern Africa during JJA, 2003–2010. The changes in the Probability Distribution Function (PDF) of AE and FMF in response to high RR1.38/0.66 at these three sites are shown inFigure 7. Because there were no MPLNET data available at these sites, the collocated MODIS reflectance ratio RR1.38/0.66 was used to distinguish cirrus-contaminated cases from cirrus-free cases. A threshold value of RR1.38/0.66 = 0.1 was used for cirrus cloud identification. Systematic PDF shifting in AE and FMF were observed for all three sites. In comparison to cirrus-free cases, AE and FMF in cirrus-contaminated cases tend to have smaller values, indicating more frequent presence of large particles as a result of possible cirrus contamination. Kolmogorov-Smirnov tests, which are usually used for testing the significance level of differences between two data distributions, indicate that the data distributions of AE and FMF in cirrus-free and cirrus-contaminated cases, as shown inFigure 7, are significantly different at a confidence level of 95%. These evidences are consistent with the theoretical prediction that cirrus contamination in the aerosol retrieval would lead to larger retrieved particle sizes, more evidence of potential cirrus contamination in AERONET aerosol retrievals.

Figure 7.

PDF of AE and FMF for cirrus and non-cirrus cases over three representative AERONET sites for smoke prevailing regions during peak smoke seasons: (a and d) Alta_Floresta in Amazon during SON, 2004–2009; (b and e) Mukdahan in Southeast Asia during MAM, 2004–2009; and (c and f) Mongu in Southern Africa during JJA, 2003–2010. Figures 7a–7c are for AE and Figures 7d–7f are for FMF. RR1.38/0.66 > 0.1 was used for thin cirrus case identification.

[47] Such tests of collocating AERONET AOT (or FMF) with the MODIS RR1.38/0.66 cirrus detection parameterization suggest feasible operational routines that can be used to crosscheck aerosol and cirrus retrievals from AERONET and operational satellites. This becomes more important for satellite product calibration/validation field campaigns where in situ measurements are closely examined along with collocated satellite observations in near real-time to verify the atmospheric environment and to validate satellite retrievals.

4. Summary and Discussions

[48] Concurrent aerosol and cirrus observations from ground measurements and satellites were used to evaluate the susceptibility of ground aerosol retrievals to cirrus contamination. We first compared MPLNET and CALIPSO in terms of their cirrus detection capabilities. Their agreement rate is about 71–88% for both day and night matchup cases. For the cirrus occurrence frequency, both agreed on the cirrus peak seasons at four selective sites; however, MPLNET detected relatively lower cirrus frequency than CALIPSO during the cirrus peak seasons.

[49] To quantify the susceptibility of the AERONET aerosol products to cirrus contamination, the following pairs of data sets were matched up: (1) AERONET versus MPLNET, (2) AERONET versus CALIPSO, and (3) AERONET versus MODIS. In the AERONET-MPLNET matchup, challenges come from the different viewing geometries of the two instruments and difficult cirrus observations at high altitude when the lower atmosphere significantly attenuates lidar signals. For a ‘SZA < 20° and TW30 overall cirrus persistence’ testing, all susceptibility percentages at 10 collocated AERONET and MPLNET sites are less than 5%, and 40% of the sites are less than 1%; for the ‘SZA < 20° and TW10 existence’ testing, 6 out of 10 sites (60%) have SP values within 4–10%, and the other 4 (40%) within 1–3%. The SP values are sensitive to different cirrus detection criteria, such as cirrus persistence test settings, time window selections, and solar zenith angle constraints. Selection of cirrus detection criteria depends on the cirrus detection limit requirements, such as a more strict cirrus screening or a relatively less strict screening. The “TW30 with overall persistence and SZA < 20°” criteria represent the middle choice between the two requirement ends. An empirical approach for cirrus optical depth calculation based on MPLNET NRB profiles was established and successfully implemented for selective cases to roughly estimate the relative contribution of cirrus contamination to AOT retrievals.

[50] Despite various challenges in collocating AERONET with CALIPSO, such as insufficient sampling and distance between CALIPSO daytime tracks and AERONET sites, over half of the 18 AERONET-CALIPSO collocated sites also have a susceptibility percentage less than 10%, a similar order of magnitude to the AERONET-MPLNET assessment results. A promising 3-Way AERONET-MPLNET-CALIPSO matchup scheme was established and tested in this study. As the CALIPSO lidar acquires more data and the number of the AERONET-MPLNET supersites increases, the 3-Way comparison will become more valuable when sufficient matchup samples are available. AERONET aerosol retrievals were also paired up with MODIS cirrus parameters, such as RR1.38/0.66, to test the ground-satellite matchup techniques in terms of using satellite derived cirrus detection to evaluate cirrus contamination in ground aerosol retrievals. The AERONET-MODIS comparison showed that 93% of all sites have the SP value less than 10%, a comparable SP level to the AERONET-CALIPSO matchup. For three smoke dominant regions during their biomass burning seasons, cirrus-free cases and cirrus-contaminated cases were discriminated from each other using the MODIS cirrus parameter. Smaller AERONET Ångström exponents and Fine Mode Fractions were also found in their probability data distributions for ‘cirrus-contaminated’ cases than in the ‘cirrus-free’ cases, another indication that cirrus potentially contaminates the AERONET aerosol retrievals.

[51] Statistical results from this study demonstrated the effectiveness of the current cloud screening schemes in the AERONET retrieval on successfully detecting cirrus existence although residual cirrus contaminated cases may still exist. It is also noteworthy that the susceptibility evaluation is highly dependent on both season and region. Moreover, influential factors, such as viewing geometry differences between sunphotometers and micro-pulse lidars when AERONET and MPLNET are compared, and the sample size threshold values when AERONET and CALIPSO data are compared, can significantly impact the susceptibility percentage. From a cirrus contamination perspective, this study improves our understanding of data uncertainties of ground aerosol products. Similar evaluations on satellite aerosol retrievals are underway. Further improvement of ground aerosol product quality is valuable for calibration and validation of satellite aerosol retrievals, and also very important for any consequential aerosol-related climate research.

Acknowledgments

[52] This work is supported by grant from the NASA EOS Program, managed by Hal Maring. Authors thank David Giles, Bo-Cai Gao, Steve Ou, Larry R. Belcher and Zhien Wang for their constructive comments on the use of in situ and satellite data, analysis methodology and cirrus climatology. Aqua MODIS L1B data were obtained from NASA L1 and Atmosphere Archive and Distribution System (LAADS). CALIPSO data were obtained from the NASA Langley Research Center Atmospheric Science Data Center. The NASA Micro-Pulse Lidar Network is funded by the NASA Earth Observing System and Radiation Sciences Program. The data at the COVE site are funded by the NASA Earth Observing System project. Authors acknowledge site PIs Brent N. Holben, Ellsworth J. Welton, Si-Chee Tsay, Jeffrey S. Reid, James R. Campbell, Soon-Chang Yoon, Gregory L. Schuster, Neng-Huei Lin, Joseph M. Prospero, John E. Barnes, Soo Chin Liew and all their colleagues for the AERONET and MPLNET sites data that are significantly used in this paper, and all other site PIs and their colleagues for all the site data used in this paper. Their tremendous efforts to collect the quality data sets and put them publicly available are sincerely appreciated. Authors thank the three reviewers for their valuable and insightful comments that helped to improve the paper significantly.