We consider CF calculated directly from the ARSCL (or KAZRARSCL) data products as a fraction of 10 s (or 4 s) cloudy returns over a given time (e.g., Xi et al., 2010). Thus, the CF defines the frequency of cloud occurrence from the combined narrow-FOV lidar–radar measurements. ShCu returns are defined using thresholds on 10 s (or 4 s) CBH and cloud top height information (Table 1) in a binary fashion. The three estimates of ShCu cover from lidar–radar observations are (1) CF from the merged lidar–radar products uses CBH information from the merged ceilometer–MPL and cloud top height from the merged MPL–radar. This method has the advantages of low missing data due to use of multiple instruments and incorporates information about cloud top height consistent with the definition of shallow convection. The MPL has the potential to attenuate in thick clouds, necessitating radar data for cloud top height retrieval. Insect contamination may contribute to significant uncertainty in the radar-based retrievals of cloud boundaries; however, we do not expect this to significantly impact our results for several reasons. First, when using cloud top height in cloud fraction, we still require cloud base to be identified by the ceilometer or MPL, so we will not misclassify insect-only layers as cloud. Second, our results are not very sensitive to the actual value of cloud top height as long as it is below 4 km, and as most insects will be found in the boundary layer or immediately above it (Kollias et al., 2016; Wainwright et al., 2017) they are not likely to cause the radar to misidentify a cloud height above 4 km if a cloud does not exist at that height. Finally, both the ARSCL and KAZRARSCL products rely primarily on lidar–based cloud detections in identifying insect-only events, so we do not expect any systematic differences in the cloud fraction determined by the two products as a result of insects. (2) CF from merged ceilometer–MPL uses CBH only and also has the advantage of low missing data as either ceilometer or MPL are used to determine CBH. (3) CF from the ceilometer alone is the most common method that only uses CBH information, but it has a disadvantage of missing data and a limited vertical range for detecting high-level clouds (Table 1). Additionally, total cloud fraction CF is calculated from merged ceilometer–MPL for any cloudy return detected within the narrow-FOV vertical column to screen multiple cloud layers (Sect. 3.4), this merged product has an extended range of 10 km from the MPL.

Narrow-FOV observations of CF, CBH and wind are computed for two averaging periods: (1) a 30 min period centered on the 15 min averaging time for FSC, meaning the CF time period begins 7.5 min before the FSC averaging time-bin and ends 7.5 afterwards; and (2) a 60 min period. The extended observation time (30 min instead of 15 min) for the CF measurement is aimed to compensate for the additional sky area observed by the wide-FOV TSI during a 15 min period. Hourly CF observations and FSC averages are computed over identical averaging periods for consistency with previous studies.

FSCs obtained from TSI observations within the 100 and 160∘ FOV are available from routine TSI processing. The 100∘ FOV is used in this analysis because it was previously established to best correspond to nadir observations due to a reduced impact of cloud sides on estimated cloud cover (Kassianov et al., 2005). Prior to 2005, the TSI used 20∘ FOV rather than 100∘; therefore, we calculate 100∘ FOV FSC for an additional period (2000–2005) directly from the TSI cloud mask as the fraction of thin and opaque pixels over total number of pixels. In our calculations, we remove a 25∘ FOV ellipse surrounding the sun location for consistency with the routine TSI processing. Our product compares well with the available TSI product; only 2 % of the calculated FSCs exceed the processed FSCs by more than 0.05. The total (opaque + thin) FSC is calculated for each cloud mask and then averaged in nonoverlapping 15 min and 60 min intervals. Note that 15 min is consistent with the expected decorrelation time of a ShCu field (Kassianov et al., 2005). The TSI does not contain CBH information; therefore selection of single-layer ShCu time periods requires the joint use of active sensors.

A narrow-FOV FSC is calculated from the TSI cloud mask with the sampling, sensitivity and processing of the wide-FOV FSC. These calculations are aimed to mimic the narrow-FOV lidar–radar measurements (“CF-like”). The narrow-FOV represents a 3 pixel×3 pixel region located at a zenith angle of 20∘ and azimuthal angle of 315∘ from each TSI cloud mask. There are three main reasons for selection of this region. This region (1) is close to zenith (the zenith region is obstructed by the camera box), (2) is located outside of the sun circle region during the study hours (local daylight 09:00–18:00) and (3) corresponds to the best agreement with the wide-FOV FSC observations (Wagner and Kleiss, 2016). The TSI image resolution has been increased from 352 pixels×288 pixels (original resolution) to 480 pixels×640 pixels (improved resolution) in August 2011. The narrow-FOV (3 pixel×3 pixel region) “CF-like” observation has 4.1 (original) and 2.3∘ (improved) angular resolution. For comparison, the narrow-FOV ARSCL cloud products, including the lidar–radar CF, have much finer angular resolution (about 0.2∘). The corresponding spatial resolution of the lidar–radar CF can be estimated by multiplying wind speed at cloud base height by lidar–radar dwell time (about 10 s). For example, its spatial resolution is about 100 m for 10 m s-1 wind speed.

Data selection and quality control procedures are applied to the averaged and merged dataset. The scatterplots for 30 min CF and 15 min FSC (Fig. B1) illustrate application of the quality control criteria, while resulting data completeness is provided in Table B1. These screening criteria are executed on the times with ShCu identified in Sect. 2 and includes the following steps:

Periods with clear (averaged FSC < 0.05) or overcast (averaged FSC > 0.95) conditions are removed from this analysis for all variables.

Instrument-dependent values of the estimated CF (Table 1, Sect. 3.1) and the narrow- and wide-FOV FSCs (Sect. 3.2, 3.3) are generated for the whole 18-year period (2000–2017) and two subperiods (2000–2010 and 2011–2017). The comparison of the cloud cover estimates from the subperiods illustrates the joint and individual impacts of instrumental upgrades (Sects. A2–A3) and data merging schemes (Sect. A4) on the long-term statistics of cloud cover. The comparison includes both 15 min (e.g., Lareau et al., 2018) and 60 min (e.g., Zhang et al., 2017) averaging windows to assess the impact of the shorter averaging time on the CF-FSC agreement.

We begin our investigation of differences in cloud cover estimates by comparing mean values (Tables 2 and 3). The 60 min average FSCs obtained for the two subperiods are comparable (0.34 vs. 0.33), and their distributions are very similar (comparison not shown), indicating weak changes in the mean ShCu cloud cover for the two subperiods. The ceilometer-based CF supports this interpretation as mean values show good agreement with FSC, particularly for the later subperiod (2011–2017) where the ceilometer data have good completeness. However, the mean CF values calculated from the merged instruments are larger than FSC values by 0.12 for the ceilometer–MPL and 0.08 for the merged lidar–radar in the early subperiod (2000–2010). The larger values of the CF from the merged data in comparison with the ceilometer and FSC in the early subperiod are consistent with previous studies which focus on all cloud types and sky conditions (Boers et al., 2010; Qian et al., 2012; Wu et al., 2014; Kennedy et al., 2014). However, the FSC–CF correlation obtained for the single-layer ShCu is greatly improved over those from the previous studies with all cloud types. For example, Wu et al. (2014) reported a correlation coefficient of 0.54 comparing hourly CF and FSC when ignoring clear and overcast conditions. Here, the correlation coefficients for hourly data are much higher (0.8–0.84) (Table B3). For the later subperiod (2011–2017), the mean CF from the merged instruments decreases by 0.05 for the combined ceilometer–MPL and 0.09 for the combined lidar–radar. As a result, the combined lidar–radar CF shows an improved agreement with the FSC, whereas the CF from ceilometer–MPL is still larger than the FSC by 0.08 (Table 2).

The changes in mean values of the merged instrument CFs between the two subperiods are examined in the 1D and 2D histograms in Fig. 1. The figure shows strong correlation between merged ceilometer–MPL and merged lidar–radar CFs in the early subperiod, indicating that the cloud top height threshold employed has little influence on cloud fraction. In contrast, in the later subperiod the RMSD between the CFs is twice as large, and the correlation is only moderate, indicating a higher influence of the cloud top height criteria. The 1D histograms (Fig. 1a, b) show that both variables shifted to have more observations of low CF in the second subperiod, although this is most pronounced for the lidar–radar. These results suggest that changes in merging the ceilometer–MPL cloud base height detection impacted CF moderately, whereas changes in cloud top height detection in the merged lidar–radar data created significant differences in CF between the two periods.

The 2D histogram comparisons of merged lidar–radar CF and FSC (Fig. 2) show the elimination of bias across the entire range of FSC in the later subperiod; however, this may be due to the introduction of compensating errors when the overprediction of clouds by the combined MPL–ceilometer cloud base height is compensated by the under prediction of clouds by the radar cloud top height in the later period when only radar is used to detect cloud top height below 3 km. Though the mean bias is reduced by 0.08 (Table 2) in the later period, the scatter about the 1:1 line is little affected as evidenced by only a 0.04 reduction in the RMSD in the later subperiod; additionally, 51 % of CF values differ from FSC by more than 0.1 in the early period compared to 41 % in the later subperiod. The modest improvement in agreement indicates the continued presence of cloud detection differences. The 2D histograms for the merged-lidar data CF vs. FSC (Fig. 3) show better agreement across the entire range in the later subperiod, consistent with an improvement in ShCu cloud detection. The compensating error hypothesis is supported further by the comparisons of ceilometer CF to FSC (Fig. 4), which return the highest correlation coefficients (Table B3) and the lowest RMSD (Tables 2 and 3).

The new cloud radar (deployed in 2010) is expected to have improved ShCu detection in the lower atmosphere over the previous instrument. Therefore, it is not expected that the changes in radar detection are responsible for the reduction of merged lidar–radar CF in the later subperiod. We should note that the small droplet size of continental cumuli has been reported to impede their detection by both the new (Lamer and Kollias, 2015) and retired radar (Chandra et al., 2013). For example, Lamer and Kollias (2015) report that 37 % of the hydrometeor detections by the ceilometer were missed by the radar at the SGP site. Most certainly, in either period, the radar would not contribute to an overestimation of CF. Instead it is expected that changes in how MPL and radar data are merged to retrieve cloud top height is likely the source of these significant differences seen between subperiods (Sect. A4). Though a number of differences exist, the incorporation of MPL data (below 3 km) in the original cloud top height retrieval would increase the number of detected cloud tops compared to those retrieved from the radar data alone for the initial period (2000–2010). Reliance only on the radar data for cloud top detection in the updated algorithm would result in fewer cloud top height detections and therefore a lower CF (see Sect. A4 for more details).

Since ceilometer CF does not demonstrate dependence on subperiods, it is unlikely that the decrease in merged ceilometer–MPL CF between the two periods is due to the ceilometer. The merged ceilometer–MPL CF is always the same or greater than CF from the ceilometer alone, and the bias decreases between the two subperiods from 0.19 to 0.11 (comparing times with complete ceilometer observations). This decrease in bias also corresponds to a decrease in the RMSD from 0.25 to 0.16 and an increase in Pearson's correlation from 0.86 to 0.91. The improvement in agreement might be attributable to improvements in the MPL mask that reduce false cloudy returns from boundary layer aerosols, resulting in a lower MPL CF (Sect. A3). It is also consistent with an increased reliance on the ceilometer for boundary layer retrievals in the cloud base best estimate retrieval algorithm (Sect. A4). Specifically, the updates of the merging algorithm would likely result in more frequent “clear” returns. The TSI and ceilometer data in comparison with the merged ceilometer–MPL and/or ceilometer–MPL–radar data may be preferable for consistent estimates of ShCu cloud cover at the SGP site from 2000 to present.

Recall that the CF obtained from lidar–radar observations with narrow FOV represents a transect of a cloudy sky along the wind direction, while FSC acquired from wide-FOV TSI data defines an area of cloudy sky. Both the CF and the FSC are widespread measurement-based estimates of cloud cover. Fractional sky cover (FSC) from a 100∘ image spans about a 3 km width of the sky for a 1.5 km cloud base, whereas the active remote sensing instruments (radar, MPL and ceilometer) observe only a “soda straw” approximately 10 m wide – essentially a point in the sky. Cloud fraction obtained from these narrow field of view instruments is thus essentially a 1D transect through clouds that pass overhead. Previous studies have reported sampling uncertainties in transect measurements for modeled random cloud fields (Astin et al., 2001; Berg and Stull, 2002; Kassianov et al., 2005). The term “random” refers to the random arrangement of clouds on a horizontal plane within a given domain. In particular, the previous model studies (Astin et al., 2001; Berg and Stull, 2002) have demonstrated that the cloud cover obtained from the transect measurements mimics the area-averaged cloud cover for non-organized (e.g., random) cloud fields well if the sample size is relatively large (or numerous individual clouds are sampled). Recently Oue et al. (2016) showed that 10 or more ceilometers equally spaced across a 25 km width in the cross-wind direction are required to estimate the simulated cloud cover in the small (30 km) domain. Certainly, the number of ceilometers, their locations and averaging time required for an accurate estimation of the cloud cover depend on the spatial arrangement of clouds and wind speed. Conversely, poor agreement is expected for organized cloud fields, such as so-called “cloud streets,” where individual clouds are arranged in rows along the mean wind direction within a given period of interest. To estimate the relative impact of the FOV configuration on the estimated cloud cover, we use the “CF-like” observations (Sect. 3.3) following an approach previously introduced by Wagner and Kleiss (2016). Recall that the narrow-FOV “CF-like” and the wide-FOV observations are both from the TSI and thus have identical sensitivities for ShCu detection. Thus, the differences between “CF-like” and FSC illustrate the FOV impact on the estimated cloud cover.

Comparison of “CF-like” and FSC shows symmetric variability around the 1:1 line (Fig. 5) and no indication of bias in average values (Tables 2 and 3). For the 1 h (30 min) observations, 23 % (41 %) of “CF-like” values differ from FSC by more than 0.1, and 5 % (14 %) differ by more than 0.2. Reducing the averaging time from 1 h to 30 min increases the RMSD by 0.05 (Table 2) and decreases the correlation from 0.92 to 0.85 (Table B3). Similar trends in the RMSD and correlation coefficients are seen for all the CF measurements as well. This result is consistent with a 15 min decorrelation time for the temporal FSC fluctuations (Kassianov et al., 2005). The “CF-like”–FSC comparison allows us to estimate that narrow-FOV observations are responsible for a ±0.1 uncertainty in hourly observations of CF (Table 2).

Using the passive “CF-like” observations, we also emphasize the relative impact of the instrument-dependent cloud detection and data merging for actively sensed CF measurements. The variability in the best fits in Figs. 2–4 (comparing FSC to CF) is greater than is expected from FOV impacts alone (Fig. 5) for both hourly and sub-hourly observations. For hourly measurements, 23 % of the “CF-like” values differ from FSCs by more than 0.1; this is the expected impact of FOV alone. For the earlier subperiod (2000–2010), the percentages of the CF values differing by more than 0.1 from the wide-FOV FSC are 62 %, 51 % and 34 % for the merged ceilometer–MPL, lidar–radar and ceilometer retrievals, respectively. For the later subperiod (2011–2017), which has the least bias, the corresponding percentages are 53 %, 41 % and 30 %. The large percentages of observations with significant uncertainty in both time periods highlight the importance of distinguishing between FOV effects from instrumental detection differences when using data for sub-hourly applications.

The effective spatial area sampled by either narrow or wide FOV instruments is a function of both sampling duration and wind speed. High wind speed in comparison with low wind speed (1) increases sample size for a given period and (2) tends to organize horizontal arrangement of clouds (e.g., Weckworth et al., 1999; Atkinson and Zhang 1996). These two factors associated with sample size and spatial arrangement of clouds should be considered when differences between cloud cover obtained from narrow- and wide-FOV observations as function of wind speed are considered (Fig. 6). In particular, Fig. 6 illustrates that both CF–FSC and “CF-like”–FSC differences are reduced noticeably as the wind speed increases from 1 to 3 m s-1 and continue to reduce slightly as the wind speed grows up to 11 m s-1. The CF–FSC and “CF-like”–FSC differences obtained at a higher wind speed (above 11 m s-1) should be considered with caution due to limited number of the corresponding cases with high wind speed (e.g., fewer than 100 cases for 60 min time average). The increased sampling area associated with increased wind speed does not necessarily result in an improved agreement between the narrow- and wide-FOV observations for both hourly and sub-hourly observations due to the impact of wind speed on cloud organization.

We developed a tool to help understand the impact of different sources on individual CF measurements. These sources include data quality control, detection differences and spatial cloud organization during the time period of interest (FOV-impact). Many factors contribute to cloud field organization including cloud cover, cloud size (Zhu et al., 1992), wind speed, cloud growth and decay rates, and the presence of cloud streets and clusters. The “quick-look” tool is used to distinguish the detection-based uncertainty from the FOV-impact during different sky conditions.

TSI cloud mask images are analyzed for FSC differences along the cross-wind direction to visually assess the spatial inhomogeneity of the cloud field and its potential impact on disagreement between narrow- and wide-FOV observations. Prior to our analysis, the 25∘ FOV sun circle is removed and the cloud mask is cropped to the 100∘ FOV and undistorted from hemispherical to rectilinear coordinates (Chow et al., 2011). Mean composite images are generated by summing all images within 15 min and dividing by the number of images (Fig. 7a). Each pixel in the averaged image can be interpreted as a 15 min CF measurement from a narrow-FOV sensor. The variability of CF in the cross-wind direction can indicate the possible influence of cloud field organization on cloud cover estimates by narrow-FOV observations. The composite image is then divided into 21 equally spaced “lanes” parallel to the wind direction at CBH (lane FOV is 5∘). The mean and interquartile CF from each lane is then computed for this 15 min average. Figure 7b shows that the cross-wind variability represented by the lane CF means exceeds substantially the within-lane variability (vertical bars). The low variability within each lane indicates small changes in FSC within a lane for this period.

Figure 8 shows an example of “quick-look” results generated for a single day (14 June 2017). The narrow-FOV (5∘) lane-averaged FSC (Fig. 7) is used to estimate the FOV impact for each 15 min interval. The variability in lane FSC for each 15 min interval is visually displayed as one row in the heatmap (Fig. 8a) and provides the spatial distribution of the narrow-FOV FSC perpendicular to the wind direction. The source of between-lane FSC variability can be inferred from the 15 min mean composite images themselves (Fig. 8b). For perfect data, the thumbnail images are a representation of the mean cloud field within the 100∘ FOV image. For typical wind speeds (4–15 ms-1) it is common to observe streaks, as is seen in this example day, generated by the motion of clouds (see also Fig. 9a, b). For low wind speeds (less than 4 ms-1), a mottled field of clouds emerges with pixel CF dependent upon the rate of cloud formation and decay (Fig. 9c). Less variability in cross-lane FSC is observed for large cloud cover (FSC > 0.7) especially for typical winds (Fig. 9d).

The lane-averaged FSC is compared to actual 30 min CF measurements in Fig. 8c. The intention of this plot is to compare the narrow-FOV CF to the range of possible narrow-FOV FSC observed over different sky regions. By quantifying the uncertainty in cloud cover estimates due to the spatial variations in the mean cloud field, we are better equipped to separate FOV impacts from detection differences. For example, a noticeable cloud street appears from about 12:30–13:15 local time, characterized by a wider spread in the lane-FSC min-max and interquartile range (IQR). The mean FSC is approximately 0.5 during this time, and the CF varies between 0.25 and 0.3 and is nearly equal to the lane-FSC minimum. The corresponding thumbnails in Fig. 8b appear nonuniform, suggesting that the CF measurement is impacted by the narrow FOV configuration. In contrast, from 13:15 to 14:15 local time a more uniform mean cloud field is depicted, and the lane-FSC IQR is within 0.1 of the FSC. During this period, ceilometer CF is also within the lane IQR. Surprisingly, CF from the merged lidar–radar diverges from the ceilometer value for the first time in the day and exceeds the lane-FSC maximum significantly. We interpret the larger merged lidar–radar CF to be an overestimation of cloud cover associated with detection differences or data merging issues.

Figure 8 is also useful for quality checks as the quick looks are not censored based on the criteria of Sect. 3.4. For example, on this day we can see little evidence of upper-level clouds by the close agreement of the merged ceilometer–MPL total CF (blue) and ShCu CF (red dashed). The thumbnails illustrate image artifacts such as incomplete removal of sun glare (17:00–19:00 LT), which would cause FSC to be overestimated. In this same period, the images otherwise show relatively uniform cloud cover; therefore CFceil might be preferred over FSC. Sun glare is common in the TSI, and a method has been proposed to identify and correct its effects on the FSC (Long, 2010). The wind-direction, indicated by the red arrow, is provided to verify cloud motion with observed wind direction. This is particularly helpful if missing wind data at CBH have been imputed from other sources, as it can be incorrect. Sun glare and incorrect wind direction are significant limitations of the extensibility of this method to a statistical analysis.

There are two main expected applications of the introduced “quick-look” tool. The first potential application is a classification of spatial organization of cloud fields using, for example, cross-wind cloud field variability (e.g., peaks and valleys in Fig. 7b) and within-lane variance in cloud amount (e.g., vertical bars in Fig. 7b). Numerous images generated by the “quick-look” tool (e.g., Fig. 8b) for the extended period (2000–2017) can be considered as a valuable training dataset for machine learning with focus on automated detection of desired features of the cloud fields (e.g., cloud streets) and unwanted contaminations of TSI images (e.g., Fig. 9). The second potential application is a visual inspection of the generated images for a given period of interest (e.g., a short-term field campaign) to check for the impact of instrumental detection differences and cloud field organization on the observed cloud amount. Visual inspection may be feasible given a limited number (about 40) of ShCu events annually during the warm season. For example, a spread of the lane CFs (gray region in Fig. 8c) gives an idea about the cross-wind cloud field variability within a given FOV and thus aids in understanding the difference between cloud amounts obtained from the narrow- and wide-FOV observations.