A long-term climatology of classified cloud types has been generated for 13 years (1997–2009) over the Atmospheric Radiation Measurement (ARM) Southern Great Plains (SGP) site for seven cloud categories: low clouds, congestus, deep convection, altocumulus, altostratus, cirrostratus/anvil, and cirrus. The classification was based on the cloud macrophysical quantities of cloud top, cloud base, and physical thickness of cloud layers, as measured by active sensors such as the millimeter-wavelength cloud radar (MMCR) and micropulse lidar (MPL). Climate variability of cloud characteristics has been examined using the 13-yr cloud-type retrieval. Low clouds and cirrus showed distinct diurnal and seasonal cycles. Total cloud occurrence followed the variation of low clouds, with a diurnal peak in early afternoon and a seasonal maximum in late winter. Additionally, further work has been done to identify fair-weather shallow cumulus (FWSC) events for 9 years (2000–08). Periods containing FWSC, a subcategory of clouds classified as low clouds, were produced using cloud fraction information from a total-sky imager and ceilometer. The identified FWSC periods in our study show good agreement with manually identified FWSC, missing only 6 cases out of 70 possible events during the spring to summer seasons (May–August).
Various types of clouds have different radiative forcing (Chen et al. 2000); thus, an accurate cloud-type classification is necessary to understand the role of clouds on the energy budget and the regional/global hydrological cycle. Mace et al. (2006) and McFarlane et al. (2013) categorized cloud types based on typical values of cloud top, cloud base, and physical thickness of cloud layers, over the U.S. Department of Energy’s Atmospheric Radiation Measurement (ARM) Southern Great Plains (SGP) and tropical western Pacific Ocean atmospheric observatory sites. An advantage of using a simple definition of cloud types relying on cloud macrophysical quantities, such as cloud height and thickness, is that it can be easily duplicated in large-eddy simulation (LES) models. However, classifying cloud types using this method will be sensitive to predefined threshold values. Another classification method seen in the literature utilizes a trained network based on expertly categorized samples according to different characteristics of cloud types (Penaloza and Welch 1996; Wang and Sassen 2004). The trained network is then applied to unknown cloud samples to categorize them into desired cloud types. A trained network method can better handle ambiguous situations, but it does not guarantee improved performance versus a simple method using threshold values (Tovinkere et al. 1993).
The ARM SGP site, established in 1993, is suitable for studying a continental climate in midlatitudes. Since 1997, this site has provided continuous measurements of cloud vertical distribution using active sensors such as the millimeter-wavelength cloud radar (MMCR) (Moran et al. 1998) and the micropulse lidar (MPL) (Spinhirne 1993) as well as radiation, aerosols, and vertically integrated cloud properties. Previous studies of cloud characteristics over the ARM SGP have shown a distinct seasonal cycle of cloud fraction (CF) with a maximum during late winter and a minimum during summer (Dong et al. 2006; Kennedy et al. 2014; Wang and Sassen 2001; Xie et al. 2010). The diurnal cycle, another fundamental mode of climate variability, commonly showed an increase of low-level clouds in early afternoon over the SGP site (Dong et al. 2006; Mace et al. 2006; Wang and Sassen 2001; Xie et al. 2010). The diurnal cycle of high-level clouds differed between studies, which could be due to different analysis periods, seasons, definitions of high-level clouds, or different measurements used to detect clouds.
Recently, the Department of Energy (DOE) ARM facility has expanded its activities to include routine LES modeling through the LES ARM Symbiotic Simulation and Observation (LASSO) data stream to complement high-density observations at the SGP site (Gustafson et al. 2017b, 2018). Shallow cumulus clouds at the SGP have been the initial focus of the LES modeling efforts because they are an important part of the radiation budget, having an average shortwave radiative forcing of −45.5 W m−2 (Berg et al. 2011), and are challenging to simulate accurately using climate models. This is partly due to the small spatial scale of these clouds compared to model grid spacing and due to complicated interactions between microphysical and boundary layer processes (Gustafson et al. 2017a). To study climatological changes of cloud types and to give guidance in choosing shallow cumulus events for the routine LES modeling, we developed a cloud-type classification algorithm based on predefined values of cloud base, top, and thickness over the SGP site. We used this algorithm to create a database of classified cloud types as an initial step for further categorization of low clouds into fair-weather shallow cumulus (FWSC).
In this study, a 13-yr (1997–2009) climatology of the classified cloud types was produced. Further, a 9-yr (2000–08) dataset of automatically identified FWSC periods was generated and compared with manually determined FWSC (Berg and Kassianov 2008, hereafter BK08; Zhang and Klein 2013, hereafter ZK13) during the spring to summer seasons from May to August. Details of the algorithms developed to classify cloud types and select FWSC events are explained in section 2. Sections 3 and 4 present results and a summary with discussion, respectively.
a. Classified cloud types
Cloud top, cloud base, and thickness of cloud layers were calculated from the active remote sensing of clouds (ARSCL) (Johnson and Jensen 2009) data product at the SGP Central Facility. ARSCL data include the top and bottom heights of each cloud layer for up to 10 layers, detected by either the MMCR or MPL (Clothiaux et al. 1998). Cloud boundaries and thickness derived from the combination of these two instruments provides more reliable cloud layer identification because of the complementary capabilities of the two active sensors (i.e., MMCR and MPL) (Clothiaux et al. 2000; Uttal et al. 1995). Lidar can detect most mid- and/or high-level clouds, but strong optical signal attenuation prevents penetration of thick low and midlevel clouds with high hydrometeor concentrations. In contrast, radar often fails to detect clouds containing small particles, yet can effectively detect mid- and/or high-level clouds above lower cloud layers because it penetrates low clouds that do not contain significant precipitation.
The ARSCL data are given at a high vertical resolution (30 m). When selecting cloud layers from the ARSCL data, we eliminated thin layers and merged layers that are separated by a small vertical distance to simplify the cloud classification and reduce false cloud layers that result from lidar or radar artifacts. Figure 1a illustrates the screening method used to remove thin cloud layers using a hypothetical column of clear and cloudy retrievals. Since the vertical resolution of ARSCL data is 30 m, the depth of the first cloud layer in this scenario is 150 m. The top and bottom of the first cloud layer were denoted cB1 and cT1, respectively. We required cloud layers to be contiguous over 120 m to be retained for further analysis. Therefore, the first cloud layer was retained. The depth of the second layer was 60 m, so this section was not retained. In addition to removing small cloud layers (Fig. 1a), a second screen was applied to merge two cloud layers separated by less than 120 m (Fig. 1b).
Each cloud layer was assigned to one of seven cloud types based on the top height, base height, and physical thickness of each layer. Table 1 shows the cloud base, top, and thickness criteria used to define cloud types. A threshold value of 3.5 km was used to differentiate low clouds from other cloud types, as previous studies had shown a minimum in cloud amount at SGP between 3 and 4 km (e.g., Naud et al. 2005; Mace and Benson 2008). Ka-band cloud radars, like the MMCR, attenuate during heavy rainfall periods; thus, high-level clouds can be missed, and the detected cloud top can be underestimated (Wang and Sassen 2001). Therefore, we did not retrieve cloud types during times when the precipitation rate was larger than 1 mm h−1, which might reduce the frequency of deep convection and congestus in our retrieval. Surface precipitation from the ARM surface meteorology system (MET) (Cialella et al. 1990) was used for this procedure. Cloud-type retrievals were generated for 13 years from 1997 to 2009, with a temporal resolution of 1 min (Riihimaki and Shi 2018).
b. Detection of fair-weather shallow cumulus
Single-layer low-cloud layers detected by the cloud-type algorithm were further processed to select FWSC events (Sivaraman et al. 2018). This was done by incorporating additional CF information from the total-sky imager (TSI; Morris 2005, 1994) and the ceilometer (Morris 2016; Ermold and Morris 1996) located at the SGP site to complement and check the low cloud layers detected by the ARSCL-based cloud types from the MPL and MMCR. Partially cloudy conditions were the main criteria used to distinguish FWSC from other low cloud types. The TSI gives CF with a hemispheric field of view, providing a broader contextual view of the sky than the narrow, zenith-pointing field of view of the ARSCL-derived cloud types. The TSI processing software analyzes charge-coupled device (CCD) images to determine the fraction of opaque and optically thin clouds over a 180° field of view centered on zenith. The ceilometer provides a narrow, zenith-pointing field of view like the ARSCL cloud boundaries; however, the infrared wavelength and proprietary processing software is optimized to accurately identify low cloud base. We used the ceilometer cloud fraction to check that the ARSCL-based cloud-type product was accurately detecting the presence of low clouds, since the instruments used in the ARSCL-based cloud-type retrieval can sometimes misclassify aerosol layers as low cloud (MPL) or miss the small cloud droplets of shallow convection (MMCR).
The detailed procedure for automated identification of FWSC is illustrated in Fig. 2. In addition to single-layer FWSC, FWSC cases with overlying cirrus were also identified using the same procedure. Because FWSC only partially covers the sky, CF is required to reside within a certain range to be classified as FWSC. Hourly averaged opaque CF from the TSI (CFTSI) and retrieved hourly CF from the ceilometer lowest cloud base (CFceilometer) were used.
Figure 2b shows an example of the procedure used to identify FWSC events. First, single-layer low clouds, noted in groups B, C, and D, or low clouds with overlying cirrus, noted in group A, in Fig. 2b are determined from the cloud-type data product. The algorithm then requires hourly CFTSI to be between 0.5% and 80%, CFceilometer to be greater than 0%, and at least 2 min of the hour to be identified as low clouds by the cloud-type algorithm. If CFTSI or CFceilometer do not satisfy the criteria during a given low cloud event, the event is rejected as an FWSC event (e.g., CFTSI for the fourth low cloud in B is 90% and is now marked in a gray color in second panel of Fig. 2b). Note that FWSC is characterized by smaller CF and composed of a smaller number of larger-size individual cloud cells, relative to altocumulus (Ac). To be defined as an FWSC event, the low cloud occurrence during 1 h should be greater than 2 min and the duration of the FWSC must be longer than 1.5 h. The low cloud event in period D does not meet this length criteria; thus, clouds in D were rejected and noted in a gray color (see third panel of Fig. 2b). If there is a time gap longer than 2.5 h between each cloud among the identified events of FWSC, the algorithm separates the events into two different FWSC events. Clouds in A and B represent one FWSC event because the time between the two cloud events was shorter than 2.5 h. By contrast, clouds in A and C represent separate FWSC events.
FWSC events were further subclassified into isolated FWSC and transition cases where FWSC was proceeded or followed by other mid- and upper-level cloud types. Specifically, FWSC transitions to and from cirrus/cirrostratus (Ci/Cs), low-level stratus (St), and Ac/altostratus (As) were identified separately from FWSC in isolation from other cloud types. Layers classified as low clouds were defined as St when CFTSI was greater than 80%. Ci, Cs, Ac, and As were defined using the cloud-type classification (Table 1) as cloud types 7, 6, 4, and 5, respectively. If any of these cloud types (Ci/Cs, St, or Ac/As) existed with a duration exceeding 2 h during the 3-h period preceding the start time of FWSC, we identified the corresponding FWSC events as transition events from this cloud type to FWSC. The opposite transition cases from FWSC to Ci/Cs, St, or Ac/As were identified in an analogous way but using the 3-h period after the FWSC ending time.
a. Classified cloud types
Figure 3b shows an example of the classified-cloud types for 24 May 2008 and Fig. 3a shows the corresponding radar reflectivity and the best estimate of lowest cloud base from the lidar. During this day, well-organized convection developed to the northwest of the SGP site, moved toward the site, and produced an abundant amount of rainfall, with a maximum rate of 1.3 mm min−1 at 0900 UTC. After 1800 UTC on the same day, a lightly precipitating convective cloud with precipitation rates of 0.3 mm min−1, formed by locally driven conditions around the SGP site. The algorithm categorized the cloud types on this day, such as deep convection during the period between 0800 and 1130 UTC and the subsequent low-cloud period. This simple algorithm can be duplicated for other ARM sites by adjusting the threshold values in accordance with different cloud characteristics in the corresponding regions.
Diurnal and seasonal frequencies of cloud types at SGP were examined over 13 years (1997–2009) to provide a consistent and long-term assessment of the variability of cloud characteristics over these time scales (Fig. 4). Seasonally, the low cloud-type maximum frequency was found in February and minimum in July (Fig. 4a). Cirrus cloud occurrence, the most frequent cloud type over the SGP site, peaked in May followed by a decrease until September. Seasonal variation of total cloud occurrence followed the variation of low clouds (black and blue solid lines in Fig. 4a). A maximum in total cloud occurrence was seen during late winter and a minimum in summer. Other cloud types (congestus, deep convection, alto cumulus/stratus, cirrostratus) did not substantially contribute to the total cloud occurrence. Each of these categories occurred less than 10% of the time. In addition to cirrus and low cloud types, altocumulus and cirrostratus also had distinct seasonal cycles (see green and orange lines in Fig. 4a). Altocumulus frequency was at a maximum during summer and minimum during winter. The opposite trend is seen in the seasonal cycle of cirrostratus.
Cirrus and low clouds had a clear diurnal cycle (Fig. 4b). Low cloud occurrence increased until the early afternoon. The total cloud occurrence was also slightly higher in early afternoon, mainly due to the increase in low cloud occurrence (black line in Fig. 4b). By contrast, the minimum frequency of cirrus was in the afternoon and the maximum at night. The diurnal cycle of deep convention/congestus and alto cumulus/stratus, showed little variability compared to that of low cloud and cirrus. Our findings for the seasonal and diurnal cycle (Figs. 4a,b) are consistent with previous studies (Dong et al. 2006; Kennedy et al. 2014; Wang and Sassen 2001; Xie et al. 2010).
Segele et al. (2013) analyzed warm boundary layer clouds during 4 years (1997–2000) over the SGP site and showed no significant interannual differences in cloud-base heights of boundary layer clouds. However, our climatology showed distinct diurnal and seasonal variations of cloud height with higher cloud-base height during daytime and summer (Fig. 5). Del Genio and Wolf (2000) also found similar seasonal patterns as in our study. To examine how the seasonal and diurnal variation of classified cloud types were sensitive to the predefined threshold values, the threshold value of 3.5 km (Table 1) was increased to 4.5 km. As expected, low cloud amount increased because of the increased depth over which low clouds can reside when using the 4.5-km threshold (cf. Figs. 4a,c and4b,d). However, even though the amount of Ac, As, and Cs decreased and that of deep convection increased, the overall effect of changing the threshold did not change the features of diurnal and seasonal variations of the cloud types (Figs. 4c,d).
b. Detection of fair-weather shallow cumulus
The identified FWSC periods in our study were compared with manually selected periods from previous studies by BK08 and ZK13. Using ARSCL data (Clothiaux et al. 1998) and TSI (Morris 2005) movies, BK08 manually selected FWSC periods during the spring to summer seasons from May to August for 5 years (2000–04) to study the climatology of cloud macroscale properties over the SGP site. Their study focused on identifying cases with single-layer shallow cumuli, so it excluded cases that appeared to have multilayer clouds. That 5 years of data from BK08 was extended to 9 years (2000–08) by Berg et al. (2011). ZK13 also manually identified FWSC events over the SGP during 13 years (1997–2009) and examined the factors controlling the vertical extent of FWSC. Besides ARSCL and TSI data, precipitation from the Arkansas Red Basin River Forecast Center and data from Geostationary Operational Environmental Satellites (GOES) were incorporated in the study by ZK13 to identify and exclude both precipitation days and FWSC impacted by large-scale phenomena. ZK13 also required that the observed lowest cloud base had to rise during the daytime, hinting at the link to boundary layer development and including only locally generated FWSC. The coincident 9-yr dataset of FWSC periods (2000–08) from both ZK13 and BK08 are compared with the FWSC climatology generated using our automated algorithm.
Table 2 shows the results of comparing our automated dataset, generated following the schematic diagram in Fig. 2, with manually determined FWSC periods from BK08 and ZK13. Because of inconsistencies between the datasets from BK08 and ZK13, we evaluated our dataset with the FWSC periods identified by both BK08 and ZK13 studies. One of the main differences between the two datasets is the exclusion of FWSC affected by large-scale phenomena in ZK13. A total of 81 FWSC cases were identified by both BK08 and ZK13 during nine years, though 11 of these cases were missing some of the input data used in our automated algorithm, so we used the 70 cases with all available data as listed in the “total man. w/all data” column as the reference dataset. Our algorithm, labeled LR18 in Table 2, identified 40 of the remaining 70 FWSC validation cases as single layer, isolated FWSC, as listed in the “hit” column. An additional 24 of those cases were identified as FWSC by our algorithm but labeled as cases with overlying cirrus, transition cases of FWSC from or to other cloud types, or both (Table 2, overlap only, transition only, and overlap and transition, respectively). Six cases were rejected by our algorithm because the duration of low clouds was too short to be classified as a shallow cumulus event (Table 2, miss).
Figure 6 shows examples of hit, miss, and overlap cases. In Fig. 6a, LR18 identifies a FWSC period from 1700 to 2300 UTC, similar to BK08 and ZK13. The TSI image at 2200 UTC confirms the occurrence of FWSC during the identified period. It is interesting to note that the time periods identified as containing FWSC are slightly different between the three datasets (BK08, ZK13, LR18) for this hit case (1600–2500 UTC in BK08 vs 1600–2400 UTC in ZK13 vs 1700–2300 UTC in LR18). The TSI image at 1900 UTC in Fig. 6b assures us that FWSC existed on 27 August 2000. The reason our algorithm missed this FWSC event was due to the restriction that low cloud occurrence must be greater than or equal to 2 min h−1 (Fig. 2). During the time between 1930 and 2030 UTC, the frequency of low cloud occurrence does not meet this criterion. Relaxing this criteria about the frequency of low clouds, however, increased the number of false positive cases.
A total of 14 cases identified as FWSC in both BK08 and ZK13 were categorized as FWSC overlapping with cirrus clouds in LR18. In Fig. 6c, classified cloud types using ARSCL and the TSI image at 2000 UTC verified the existence of high-level cirrus clouds above FWSC. For some purposes, we wish to be able to separate single-layered FWSC from those with overlying cloud layers; thus, these cases were labeled separately. Another situation that we classified into categories other than single-layered FWSC was the transition between other cloud types (St, Ci/Cs, and Ac/As) and FWSC. We note that BK08 and ZK13 identified many of these transition cases as FWSC. Table 2 shows that 20 cases from BK08 and ZK13 were confirmed as transition cases by TSI movie inspection. Among the 20 cases, change from Ci/Cs to FWSC happened the most frequently (9 cases). Figure 7 shows six examples of transition cases that our algorithm distinguished from FWSC. All these cases were confirmed as transition cases through inspection of TSI movies.
The biggest challenge for improving our algorithm is to reduce the number of false positive cases. To give modelers maximum flexibility to choose cases of interest, we designed our algorithm to prioritize capturing all possibly relevant cases over excluding false positives. As a result, LR18 produced 43 more cases of single-layer FWSC than identified by both BK08 and ZK13 (Table 2). Through inspection of TSI movies and GOES data, we classified 39 of the 43 events into three categories. The other four cases could not be judged because of an absence of TSI movies on the corresponding dates. False positive cases in LR18 could be categorized into three groups according to their causes: smoke, Ac, and FWSC impacted by large-scale weather phenomena. It would be advantageous to be able to separate locally driven FWSC from FWSC created by large-scale phenomena, consistent with the ZK13 method. While FWSC events influenced by large-scale weather patterns are still shallow cumulus clouds and are not precisely false positives, they are more challenging to simulate well in the limited domain of an LES model. For the LR18 false positive cases impacted by large-scale phenomena, the classified cloud types and opaque CF from TSI (or detected cloud base from ceilometer) of these false positive cases did not show any differences from the true FWSC case. TSI images alone could not distinguish cases impacted by large-sale phenomena from locally driven FWSC cases (cf. Figs. 8a and 6a).
To investigate how to distinguish this type of FWSC from locally generated FWSC, we utilized cloud data from visible infrared solar-infrared split-window technique (VISST)-retrieved satellite products, which are based on 4-km resolution data from the 0.65-, 3.9-, 11-, and 12-μm channels of the GOES-8 imager since March 2000 (Minnis et al. 2008). Eleven false positive cases (Table 2) were confirmed as FWSC impacted by large-scale weather phenomena from inspection of visible images of GOES data (http://www.aviationweather.gov/adds/satellite) and surface weather charts from the Weather Prediction Center of the National Weather Service (http://www.wpc.ncep.noaa.gov). Time-averaged CF (>25%) and cloud-top height (>7.5 km), calculated from three snapshots at 0830, 1130, and 1430 LST of VISST data, over a region (33°–41°N, 102°–93°W) centered in the SGP site were used to identify cloud systems impacted by large-scale phenomena. Four out of the 11 cases were classified as impacted by large-scale weather using the above criteria. A cloud system on 11 August 2006 is an example (Fig. 8a). Another six cases could not be detected using these criteria because there were squall-like narrow clouds with insignificant CF. A cloud system on 4 July 2008 is an example of these squall-like clouds. VISST data problems on 7 August 2011, the final case, made it difficult to examine whether the FWSC event was impacted by large-scale weather on that date.
For the smoke and Ac false positive cases, TSI images showed the features of Ac and smoke (see right column in Figs. 8b,c). We conducted additional sensitivity tests to distinguish characteristics of false positive cases from true FWSC cases. From the sensitivity test results (LR18_TSI50 in Table 2), in which we reduced the maximum threshold value of CFTSI from 80% to 50% (see Fig. 2), we did see that we could eliminate some Ac false positives cases with larger opaque TSI CF. However, by reducing the maximum threshold value of CFTSI, several FWSC cases were also missed. Because we did not want to exclude true FWSC cases, we kept the maximum threshold value of CFTSI at 80%.
Only one false positive case was found due to the misclassification of smoke as a cloud and is shown in Fig. 8c. The ceilometer can detect the backscatter signal from clouds but does not detect smoke as cloud; thus, a ceilometer is a critical instrument to distinguish clouds from smoke. However, the presented false positive case could not be eliminated even with the incorporation of ceilometer measurements because the smoke plume existed with FWSC (note the FWSC on the left-top side of the TSI image in Fig. 8c).
The seasonal variation of detected FWSC events derived using the LR18 method over the 9-yr (2000–08) period is shown in Fig. 9. FWSC events occurred most frequently during the summer season. The number of FWSC events decreased dramatically in October and remained low through March. Even though changing the threshold value from 3.5 to 4.5 km in the definition of the low cloud type in Table 1 increased the number of FWSC events except in January (cf. solid and dotted lines in Fig. 9), it did not affect the pattern of the FWSC seasonal cycle, with the majority of FWSC found during the warm season.
4. Summary and discussion
An algorithm that can classify cloud type based on predefined threshold values of cloud-top height, cloud-base height, and cloud thickness has been developed for the SGP site. Cloud layers were detected from surface-based active remote sensors, specifically millimeter-wavelength cloud radar (MMCR) and micropulse lidar (MPL). This classification was based on the method by Mace et al. (2006). However, differently from Mace et al. (2006), in which cloud optical depth from the multifilter rotating shadow-band radiometer (MFRSR) was used to give information on cloud thickness to define the cloud types, the cloud thickness in our study was directly calculated from cloud-top and cloud-base heights following Burleyson et al. (2015) and McFarlane et al. (2013). Even though the cloud classification algorithm was sensitive to threshold values, this simple definition of cloud types had the advantage of easy duplication using a large-eddy simulation model. Cloud-type classification using simple cloud boundary threshold values can be duplicated for other ARM sites by adjusting the threshold values according to different cloud characteristics in the corresponding regions. In addition to the general cloud classification, an automated method has been devised to select fair-weather shallow cumulus (FWSC) periods. FWSC is a subcategory of the cloud layers identified as low clouds by the cloud classification algorithm using opaque cloud fraction from a total-sky imager (TSI) and detected cloud-base information from a ceilometer. A 13-yr (1997–2009) climatology of the classified cloud types and a 9-yr (2000–09) dataset of FWSC periods were produced.
The variability of cloud characteristics, including diurnal and seasonal variations of cloud types, was examined over 13 years. Low-level and cirrus cloud types had distinct diurnal and seasonal cycles, and the variation of total cloud occurrence followed the variation of the low cloud type. Similar to low clouds, the diurnal cycle of total cloud occurrence peaked in the early afternoon and the seasonal cycle peaked during late winter.
Periods of FWSC identified by the automated identification algorithm were compared with manually selected FWSC periods from previous studies (BK08; ZK13). Our algorithm subset FWSC events into isolated FWSC, those that transition between FWSC and other cloud types, and FWSC overlapping with cirrus clouds. Of the 70 cases selected as FWSC in the studies of both BK08 and ZK13 that included all data needed for our algorithm, 24 cases were judged as overlap and transition cases using our algorithm, 40 were identified as isolated FWSC events, and only 6 FWSC cases were missed.
Our automated algorithm found 43 additional FWSC events that were not identified in the manually selected datasets. Three main causes were found for these false positive cases including smoke, large-scale weather phenomena, and altocumulus. Sensitivity tests showed that some altocumulus false positive cases had larger opaque TSI cloud fractions than true FWSC cases and could be eliminated by reducing the maximum threshold value of CFTSI. However, reducing the maximum threshold value of CFTSI below 80% also caused the algorithm to miss true FWSC cases, so this change was not made. Altocumulus showed distinct features with a greater number of more closely spaced individual cloud cells compared to features shown from FWSC (cf. differences in Figs. 6a and 8b). Thus, we hope that future work incorporating advanced techniques to identify the visual patterns of clouds from TSI images will improve our automated algorithm to identify FWSC events. More efforts should be pursued to eliminate the contamination of ARSCL by insect clutter and to identify FWSC related to large-scale phenomena as well. From tests incorporating VISST satellite data, we saw the possibility to detect some cases when FWSC was created by large-scale phenomena. Being able to separate locally forced FWSC events from those influenced by large-scale weather phenomena could help better separate cases that we expect an LES model to be able to simulate well from those that require a larger domain, and is a subject for future work.
We greatly express our thanks to the ARM value-added products (VAP) science sponsors and scientists, Andrew M. Vogelmann, Jennifer M. Comstock, Chitra Sivaraman, Michael Jensen, and Justin W. Monroe, for their helpful discussions and contributions to VAP development. This research was supported by the Office of Biological and Environmental Research (BER) of the U.S. Department of Energy (DOE) as part of the Atmospheric Radiation Measurement (ARM) facility, an Office of Science user facility and by the National Research Foundation of Korea (NRF) grant funded by the South Korean government (MSIT) (2019R1C1C1008482). Data were obtained from the ARM facility, a U.S. DOE Office of Science user facility sponsored by the Office of BER. All data, including active remote sensing and TSI used in this study, are freely downloadable online (https://www.arm.gov/). Larry Berg and Yunyan Zhang were supported by Atmospheric System Research (ASR) program in the Office of Biological and Environmental Research, Office of Science, DOE. Lawrence Livermore National Laboratory is operated for the DOE by Lawrence Livermore National Security, LLC, under Contract DE-AC52-07NA27344. The Pacific Northwest National Laboratory is operated for DOE by Battelle Memorial Institute under Contract DE-AC05-76RLO 1830.
Current affiliation: Cooperative Institute of Research in Environmental Sciences, NOAA/Earth System Research Laboratory, Boulder, Colorado.