Statistical Characteristics of Cloud Occurrence and Vertical Structure Observed by a Ground-Based Ka-Band Cloud Radar in South Korea

The cloud measurements for two years from the vertical pointing Ka-band cloud radar at Boseong in Korea are used to analyze detailed cloud properties. The reflectivity of the cloud radar is calibrated with other vertical pointing radars compared with the two disdrometers. A simple threshold-based quality control method is applied to eliminate non-meteorological echoes (insects and noise) in conjunction with despeckling along the radial direction. Clouds are classified into five types: high (HC), middle (MC), low (LC) for non-precipitating clouds, and deep (RainDP) and shallow (RainSH) for precipitating clouds. The average cloud frequency was about 35.9% with the maximum frequency of 50% in June for the total two-year sampling period. The RainDP occurred most frequently (11.8%), followed by HC (9.3%), MC (7.4%), RainSH (4.4%), and LC (2.9%) out of the average occurrence of the total 35.9%. HC and RainDP were frequently observed in summer and autumn, while RainSH, LC, and MC were dominant in the winter due to the dominant cloud development by the air-sea interaction during the cold air outbreak. The HC showed a significant seasonal variation of the maximum height and the rapid growth in the layer above 7 km (about −15 °C) in summer and autumn. This rapid growth appears in HC, MC, LC, and RainDP and is linked with rapid increases in Doppler velocity and mass flux. Thus, this growth is originated from the dominant riming processes in addition to depositional growth and is supported by an updraft in the layer between 6 and 8 km. MC showed a single frequency peak around 6 km with rapid growth above and strong evaporation below. The Doppler velocity of MC rapidly increases above 8 km and is nearly constant below this height due to strong evaporation except in the summer. LC had a similar trend of reflectivity (rapid growth in the HC region and strong evaporation in the lower region) lacking high frequency in the MC region. Unlike LC, the RainDP had continuous growth toward the ground in the entire layer with rapid growth in the HC and MC regions. In addition, two modes (cloud and precipitation) appear on the ground in spring and fall with the vertical continuity of the high frequency in the precipitation mode. The precipitation growth was most efficient in RainSH in summer with a reflectivity gradient of about 20 dBZ km−1 and frequent updrafts larger than 1 m s−1 and was smaller in the MC and HC regions.

. Location of the Boseong Standard Weather Observatory (BSWO) site (red cross), Deungryang-myeon, Boseong-gun in which the cloud radar is installed.
The cloud radar with a frequency of 33.44 GHz radiates horizontal polarized electromagnetic waves and receives the returned signal in the horizontal and vertical polarization. Thus, it observes only a dual-polarization variable, (LDR, dB), and other single-polarization variables. Detailed specifications are shown in Table 1. The sensitivity of radar reflectivity is about −30 dBZ at 5 km and the measurement uncertainty of the reflectivity is as low as 0.28 dB [25]. The calibration of the cloud radar is critical for the quantitative use of the data but difficult due to its sensitivity and attenuation by precipitation. As typically performed for the weather radar, the cloud radar measured reflectivity can be directly compared with the reflectivity calculated from the disdrometers [10,28]. However, the disdrometer battles to measure smaller sized drops, leading to a high uncertainty in low rainfall intensity and reflectivity [29][30][31][32]. Although the disdrometers had a better performance in moderate and heavy rain, the cloud radar had strong attenuation by rain as The cloud radar with a frequency of 33.44 GHz radiates horizontal polarized electromagnetic waves and receives the returned signal in the horizontal and vertical polarization. Thus, it observes only a dual-polarization variable, (LDR, dB), and other single-polarization variables. Detailed specifications are shown in Table 1. The sensitivity of radar reflectivity is about −30 dBZ at 5 km and the measurement uncertainty of the reflectivity is as low as 0.28 dB [25]. The calibration of the cloud radar is critical for the quantitative use of the data but difficult due to its sensitivity and attenuation by precipitation. As typically performed for the weather radar, the cloud radar measured reflectivity can be directly compared with the reflectivity calculated from the disdrometers [10,28]. However, the disdrometer battles to measure smaller sized drops, leading to a high uncertainty in low rainfall intensity and reflectivity [29][30][31][32]. Although the disdrometers had a better performance in moderate and heavy rain, the cloud radar had strong attenuation by rain as well as the Mie scattering effect. Thus, the direct comparison of the cloud radar and disdrometer did not provide proper unbiased results in different ranges of reflectivity, and we applied a two-step calibration into the cloud radar. First, the vertically pointing radars (VertiX and MRR) were calibrated with the disdrometers during the moderate rain period. The weaker echoes from the vertically pointing radar and cloud radar were compared to avoid possible attenuation in the cloud radar. Furthermore, the sensitivity of the VertiX is much higher than the MRR and strong cumulus is often observed in the VertiX. Thus, the comparison of the VertiX and cloud radar is possible in much lower values of reflectivity than that of the MRR and cloud radar. We derived the calibration bias of the cloud radar with the limited data set of the VertiX and compared it with the one derived from the comparison of the MRR and cloud radar to demonstrate the possibility of using different calibration pairs.
The reflectivity from the cloud radar was directly compared with that from the VertiX and MRR that were calibrated with the two ground-based disdrometers (two-dimensional video disdrometer, 2DVD, and particle size and velocity, PARSIVEL). The 2DVD has a two-dimensional optical system with a sampling area of 100 × 100 mm, a horizontal size resolution of 0.18 mm, and a vertical size resolution of 0.2 mm [33]. The size, shape, and velocity of the particles passing through the sampling area are observed by using two cameras that are aligned with two light sources at different heights. The PARSIVEL uses laser beams to measure the intensity of light and its occlusion time. These measurements are then converted into the size and fall speed of the hydrometeor. The PARSIVEL has a measuring area of 54 cm 2 , a diameter range of 0.2-25 mm, and a velocity range of 0.2-20 ms −1 with 32 velocity bins.
The VertiX uses a magnetron transmitter with a maximum peak power of 23 kW with a wavelength of 3.2 cm. The diameter of the antenna is approximately 1.2 m with a 1.8 • of 3 dB beamwidth. The reflectivity, radial velocity, and power spectrum are calculated every 2 s with a vertical resolution of 45 m. The MRR observes the reflectivity, radial velocity, and power spectrum in the vertical incidence with a frequency-modulated continuous-wave (FM-CW) at 24 GHz (K-band, 1.25 cm wavelength). It has an antenna diameter of 0.6 m, a beamwidth of 2 • , a time resolution of 10 s, and a vertical resolution of 200 m.
Quantifying of calibration bias was performed using the observed 36 rainfall events shown in Table 2. The pairs of 2DVD and VertiX (2-V) were available only for eight events and those of PARSIVEL and MRR (P-M) were available for all 36 events. The reflectivities from the 2DVD and PARSIVEL are unbiased and either the 2DVD or PARSIVEL can be first used to calibrate the VertiX or MRR. For simplicity, we used the 2-V pair only when this pair was available, otherwise, the P-M pair was used.
The cloud characteristics were analyzed from June 2014 to May 2016 for 2 years (the total days of available data are 504 days). Figure 2 shows the monthly percentage of available data and monthly cloud occurrence. The number of available data was derived from the threshold of burst power and was smaller than the actual cloud radar operation rate. During the period, the cloud radar operated and provided data of an average of 58%, with a maximum of 87% and a minimum of 14% ( Figure 2). June, July, and August had a high availability of 78%, and November, December, January, February, and March had a low availability due to the radar hardware failures. The cloud occurrence was derived from only the available data. These statistics showed the poor representativeness of the cloud characteristics for May, November, and December. Table 2. 36 rain events used to derive the calibration biases of the cloud radar reflectivity. The P-M and 2-V indicate the pairs of the particle size and velocity (PARSIVEL) and the micro rain radar (MRR), and the two-dimensional video disdrometer (2DVD) and the vertically pointing X-band radar (VertiX), respectively. The cloud characteristics were analyzed from June 2014 to May 2016 for 2 years (the total days of available data are 504 days). Figure 2 shows the monthly percentage of available data and monthly cloud occurrence. The number of available data was derived from the threshold of burst power and was smaller than the actual cloud radar operation rate. During the period, the cloud radar operated and provided data of an average of 58%, with a maximum of 87% and a minimum of 14% ( Figure 2). June, July, and August had a high availability of 78%, and November, December, January, February, and March had a low availability due to the radar hardware failures. The cloud occurrence was derived from only the available data. These statistics showed the poor representativeness of the cloud characteristics for May, November, and December.

Quality Control and Reflectivity Calibration
The non-meteorological echoes (e.g., insects and noises) were removed from the cloud radar data through using a simplistic quality control procedure. If the hardware burst power was less than 75 dB, the data were treated as noise and were removed. Data below a 300 m altitude were also removed by assuming ground clutter contamination. The data set was divided into two periods, namely the cold season from December to April, and the warm season from May to November, depending on the activity of insects and the height of the bright band. During the warm season, the height of the bright band was typically higher than 4 km and insect echoes were more frequent than in the cold season. The following thresholds were then applied for removing the non-meteorological echoes, in particular, radio interference and noise along the ray direction: Z < −40 dBZ or LDR > 0 dB for the cold season and Z < −30 dBZ below 4 km or LDR > −5 dB for the warm season. The additional threshold (Z < 0 dBZ and LDR > −15 dB below 4 km) was applied to remove insect echoes during the "warm" season. Finally, the remaining non-meteorological echoes were removed through despeckling in the vertical direction (ray direction) by eliminating the cloud layer thinner than 150 m.

Quality Control and Reflectivity Calibration
The non-meteorological echoes (e.g., insects and noises) were removed from the cloud radar data through using a simplistic quality control procedure. If the hardware burst power was less than 75 dB, the data were treated as noise and were removed. Data below a 300 m altitude were also removed by assuming ground clutter contamination. The data set was divided into two periods, namely the cold season from December to April, and the warm season from May to November, depending on the activity of insects and the height of the bright band. During the warm season, the height of the bright band was typically higher than 4 km and insect echoes were more frequent than in the cold season. The following thresholds were then applied for removing the non-meteorological echoes, in particular, radio interference and noise along the ray direction: Z < −40 dBZ or LDR > 0 dB for the cold season and Z < −30 dBZ below 4 km or LDR > −5 dB for the warm season. The additional threshold (Z < 0 dBZ and LDR > −15 dB below 4 km) was applied to remove insect echoes during the "warm" season. Finally, the remaining non-meteorological echoes were removed through despeckling in the vertical direction (ray direction) by eliminating the cloud layer thinner than 150 m. Figure 3 shows the time-height profiles of observed Z and LDR before (left panel) and after (right panel) applying quality control on 17 June 2014. Deep precipitation echoes (over 13 km) were observed from 0000 to 0500 UTC, followed by cloud echoes above 6 km and around 4 km and non-meteorological echoes below 3 km. A bright band appeared at 4-5 km with an LDR > −15 dBZ. The non-meteorological echoes below 3 km showed a larger LDR. Non-meteorological echoes were well removed by applying the quality control procedure (Figure 3b,d). However, some thin clouds and the edges of clouds were also eliminated by the continuity check. We applied this quality control to the reflectivity data of the entire case (from June 2014 to May 2016), and the quality controlled data were used for the further analysis of cloud properties.
Remote Sens. 2020, 12, x FOR PEER REVIEW 7 of 29 Figure 3 shows the time-height profiles of observed Z and LDR before (left panel) and after (right panel) applying quality control on 17 June 2014. Deep precipitation echoes (over 13 km) were observed from 0000 to 0500 UTC, followed by cloud echoes above 6 km and around 4 km and nonmeteorological echoes below 3 km. A bright band appeared at 4-5 km with an LDR > −15 dBZ. The non-meteorological echoes below 3 km showed a larger LDR. Non-meteorological echoes were well removed by applying the quality control procedure (Figures 3b and 3d). However, some thin clouds and the edges of clouds were also eliminated by the continuity check. We applied this quality control to the reflectivity data of the entire case (from June 2014 to May 2016), and the quality controlled data were used for the further analysis of cloud properties. The reflectivity of the cloud radar was absolutely calibrated by comparing it with the MRR and VertiX that were calibrated with the 2DVD and PARSIVEL. First, we examined the dependency of radar reflectivity in different wavelengths with the exponential DSD [34]. The reflectivity (Z) can be calculated from DSD using Equation (1) by assuming the spherical shape where D is the particle diameter, N(D) is the number concentration of particles for a given D, and dD is the particle diameter bin or interval. However, the measured reflectivity depends on the particle shape and scattering characteristics. We applied the T-matrix simulation [35] for the three wavelengths (Ka for the cloud radar, K for the MRR, and X for the VertiX) with the condition shown in Table 3 and for the exponential DSDs [34]. The simulation results are shown in Figure 4. The x-axis is the reflectivity from Equation (1) with the exponential DSDs [34], and the different lines are from the T-matrix simulation for the wavelengths of Ka, K, and X-bands. The reflectivity of the cloud radar was absolutely calibrated by comparing it with the MRR and VertiX that were calibrated with the 2DVD and PARSIVEL. First, we examined the dependency of radar reflectivity in different wavelengths with the exponential DSD [34]. The reflectivity (Z) can be calculated from DSD using Equation (1) by assuming the spherical shape where D is the particle diameter, N(D) is the number concentration of particles for a given D, and dD is the particle diameter bin or interval. However, the measured reflectivity depends on the particle shape and scattering characteristics. We applied the T-matrix simulation [35] for the three wavelengths (Ka for the cloud radar, K for the MRR, and X for the VertiX) with the condition shown in Table 3 Remote Sens. 2020, 12, 2242 8 of 27 and for the exponential DSDs [34]. The simulation results are shown in Figure 4. The x-axis is the reflectivity from Equation (1) with the exponential DSDs [34], and the different lines are from the T-matrix simulation for the wavelengths of Ka, K, and X-bands. Canting angle of raindrops Gaussian distribution with mean μ = 0° and standard deviation σ = 10° The reflectivity values from different wavelengths (X, K, and Ka-band) were almost identical when Z < 12 dBZ (Figure 4a). The difference in reflectivity values among different wavelengths became significant as Z > 20 dBZ due to the different scattering responses (Figure 4b). The Ka reflectivity was the largest when Z = 25 dBZ and the difference between the Ka and X-band was more than 1.5 dB. The difference between the K and Ka reflectivities was relatively smaller but became significant when Z > 33 dBZ due to the Mie effect. When comparing X-Ka and K-Ka reflectivities for calculating calibration values, we selected the reflectivity values smaller than 12 dBZ to avoid this difference among different wavelengths caused by the Mie effect. Thus, the calibration bias of the cloud radar was derived by comparing reflectivity smaller than 12 dBZ. , and X-band (VertiX) by assuming the Marshall-Palmer DSDs [35]. The x-axis is the reflectivity values from Equation (1), and the y-axis is derived from the scattering simulation with the condition shown in Table 3.
Second, the two vertical pointing radars (VertiX and MRR) were calibrated by comparing them with the ground-based disdrometers (2DVD and PARSIVEL). The correlation and mean bias of the reflectivities between the VertiX (MRR) and the 2DVD (PARSIVEL) were calculated at different altitudes. The altitude with the highest correlation was determined as the optimum altitude from which to derive the mean bias (calibration bias) for each rain event and was repeated for all the rain events listed in Table 2. , and X-band (VertiX) by assuming the Marshall-Palmer DSDs [35]. The x-axis is the reflectivity values from Equation (1), and the y-axis is derived from the scattering simulation with the condition shown in Table 3.
The reflectivity values from different wavelengths (X, K, and Ka-band) were almost identical when Z < 12 dBZ (Figure 4a). The difference in reflectivity values among different wavelengths became significant as Z > 20 dBZ due to the different scattering responses (Figure 4b). The Ka reflectivity was the largest when Z = 25 dBZ and the difference between the Ka and X-band was more than 1.5 dB. The difference between the K and Ka reflectivities was relatively smaller but became significant when Z > 33 dBZ due to the Mie effect. When comparing X-Ka and K-Ka reflectivities for calculating calibration values, we selected the reflectivity values smaller than 12 dBZ to avoid this difference among different wavelengths caused by the Mie effect. Thus, the calibration bias of the cloud radar was derived by comparing reflectivity smaller than 12 dBZ.
Second, the two vertical pointing radars (VertiX and MRR) were calibrated by comparing them with the ground-based disdrometers (2DVD and PARSIVEL). The correlation and mean bias of the reflectivities between the VertiX (MRR) and the 2DVD (PARSIVEL) were calculated at different altitudes. The altitude with the highest correlation was determined as the optimum altitude from which to derive the mean bias (calibration bias) for each rain event and was repeated for all the rain events listed in Table 2.

Classification of Cloud Types and Their Properties
The cloud was first detected to classify cloud types using a 1 min reflectivity profile. An average of 10 bins was performed in the vertical direction to produce a 1 min reflectivity profile with 150 m spacing from 300 m above the ground level. The cloud was detected by the existence of at least three consecutive values in a 1 min reflectivity profile. Thus, the cloud depths shallower than 450 m were ignored. The base and top of the cloud were then determined by the lowest and highest height of the observed clouds.
Similar to the classification criteria of Kollias et al. [3], the cloud types were classified into rain (RC) and non-precipitating (high, middle, and lower) clouds with the existence of the observed reflectivity value at 300 m height. The non-precipitating clouds were categorized into high (HC), middle (MC), and low (LC) clouds based on the height of the cloud base, H b , and top, H t , (Table 4). If the cloud base height was 6 km or higher, it was classified as HC, if the cloud base height existed between 2 and 6 km, it was MC, LC had a H b between 300 m and 2 km. The RC was classified into the deep (RainDP) and shallow (RainSH) clouds according to their vertical extent using a threshold of 2 km. The classification of "shallow (RainSH) and deep (RainDP)" was applied only for precipitating clouds at the ground (measurable echoes at H b = 300 m in this study) and the vertical extent. Certainly, RainDP included precipitating clouds with multi-layer clouds. Middle and/or high clouds existed with the shallow layer of precipitating clouds, and they were not considered in the category of shallow (RainSH) in this study. Table 4. Classification criteria of cloud types. The H b and H t indicate the height of the cloud base and the top.

Cloud Type Criteria
High cloud (HC) After classifying the different cloud types, the frequency of their occurrence and reflectivity were investigated. The monthly occurrence was defined as the ratio of the number of detected clouds to the total number of available observations (clouds in Figure 2). The seasonal frequency of cloud occurrence was calculated as the ratio of cloud data in each type to the total cloud data for a given season. Furthermore, the contoured frequency by altitude diagram (CFAD) was the frequency distribution of reflectivity according to the altitude and was used to analyze the cloud vertical structure [37]. The CFAD was obtained in an interval of 2 dBZ and the vertical distance of 150 m and was then normalized by the maximum frequency. The vertical mass flux, MF, was also calculated from Equation (2) [38]: where the mass content of clouds, M c is in kg m −3 , and the mass-weighted velocity V c m s −1 is derived from the radar Doppler velocity V D with V c = 0.85 V D . The mass content of clouds was calculated from Equation (3): where Z is in mm 6 m −3 . The vertical air velocity (V air ) was also calculated by the V t -Z technique [39] to link with possible microphysical processes. Here, the V t was the reflectivity-weighted terminal fall velocity. This technique provided the best accuracy in the upper layer although the overall accuracy was not the best. The V t was calculated using a power-law relationship between V t and Z [16,39,40] which was derived from an individual event: The a and b were derived as 0.88 (0.95) and 0.07 (0.16) for HC on 9 August 2015 (RainDP on 5 July 2014) case. Using the retrieved Equation (4) for each case, the V air can be calculated from Equation (5).

Reflectivity Calibration
We performed the reflectivity calibration for the 36 rain cases listed in Table 2. Figures 5 and 6 show the case of 2 July 2014 as an example. The time series of radar reflectivity obtained from the 2DVD, PARSIVEL, VertiX, MRR, and cloud radar showed higher Z in the VertiX and lower Z in the cloud radar when compared with either the 2DVD or PARSIVEL ( Figure 5). Here, the measurement height of the MRR and cloud radar (VertiX) was 400 m (472 m) above ground level to avoid possible contamination by the ground clutter. The Z values from the cloud radar were rarely higher than 20 dBZ and flattened when Z(2DVD) > 25 dBZ due to strong attenuation by precipitation in the cloud radar. Thus, higher reflectivity should not be used when deriving the mean bias of the cloud radar. The correlation between the 2DVD and PARSIVEL was approximately 0.95, and the mean bias was 0.26 dB (not shown). Thus, either the 2DVD or PARSIVEL can be used as the ground truth.
The a and b were derived as 0.88 (0.95) and 0.07 (0.16) for HC on 9 August, 2015 (RainDP on 5 July, 2014) case. Using the retrieved Equation (4) for each case, the Vair can be calculated from Equation (5).

Reflectivity Calibration
We performed the reflectivity calibration for the 36 rain cases listed in Table 2. Figure 5 and Figure 6 show the case of July 2, 2014 as an example. The time series of radar reflectivity obtained from the 2DVD, PARSIVEL, VertiX, MRR, and cloud radar showed higher Z in the VertiX and lower Z in the cloud radar when compared with either the 2DVD or PARSIVEL ( Figure 5). Here, the measurement height of the MRR and cloud radar (VertiX) was 400 m (472 m) above ground level to avoid possible contamination by the ground clutter. The Z values from the cloud radar were rarely higher than 20 dBZ and flattened when Z(2DVD) > 25 dBZ due to strong attenuation by precipitation in the cloud radar. Thus, higher reflectivity should not be used when deriving the mean bias of the cloud radar. The correlation between the 2DVD and PARSIVEL was approximately 0.95, and the mean bias was 0.26 dB (not shown). Thus, either the 2DVD or PARSIVEL can be used as the ground truth.  Table 2).
The mean bias in 2DVD-VertiX and PARSIVEL-MRR was first calculated to derive the calibration bias of the VertiX and MRR (upper panels in Figure 6). These pairs were highly correlated in particular with moderate and heavy rain. The scatter increased in light rain. The mean bias of the VertiX (MRR) was approximately 7.7 dB (−0.7 dB). This bias was applied to the VertiX and MRR which were subsequently compared with the cloud radar. The scatter plot of the two pairs (cloud radar vs. VertiX, and cloud radar vs. MRR) showed a systematically lower Z from the cloud radar in the higher Z values. The scattering simulation showed a slightly higher Z in the cloud radar than the MRR (VertiX) when Z < 37 dBZ (25 dBZ) ( Figure 4). Thus, the systematic discrepancy in Z with an increasing Z was not attributed to the scattering properties but the strong attenuation by precipitation. The Z difference due to the precipitation attenuation and scattering properties can be minimized by deriving the mean bias when Z < 12 dBZ. The mean bias of the cloud radar was about  Table 2).
The mean bias in 2DVD-VertiX and PARSIVEL-MRR was first calculated to derive the calibration bias of the VertiX and MRR (upper panels in Figure 6). These pairs were highly correlated in particular with moderate and heavy rain. The scatter increased in light rain. The mean bias of the VertiX (MRR) was approximately 7.7 dB (−0.7 dB). This bias was applied to the VertiX and MRR which were subsequently compared with the cloud radar. The scatter plot of the two pairs (cloud radar vs. VertiX, and cloud radar vs. MRR) showed a systematically lower Z from the cloud radar in the higher Z values. The scattering simulation showed a slightly higher Z in the cloud radar than the MRR (VertiX) when Z < 37 dBZ (25 dBZ) ( Figure 4). Thus, the systematic discrepancy in Z with an increasing Z was not attributed to the scattering properties but the strong attenuation by precipitation. The Z difference due to the precipitation attenuation and scattering properties can be minimized by deriving the mean bias when Z < 12 dBZ. The mean bias of the cloud radar was about −0.5 dB (−0.2 dB) with the VertiX (MRR). The different calibration paths (2DVD-VertiX-cloud radar and PARSIVEL-MRR-cloud radar) provided a similar calibration constant.    Figure 7). When there was no derived calibration value, the calibration value derived from the nearest day was used.  Figure 7). When there was no derived calibration value, the calibration value derived from the nearest day was used.

Seasonal Variation of Cloud Properties
The characteristics of the cloud occurrence were investigated in this section. The average annual cloud frequency was approximately 35.9%. The black bar in Figure 2 is the monthly cloud occurrence (normalized with the available data of the month) averaged from June 2014 to May 2016. The highest cloud occurrence, approximately 50%, showed in June when excluding the months with available data of lower than 30% (that is, May, November, and December were excluded due to the representativeness). The cloud was least frequent in March, less than 13%. Missing data existed over the short period of time (2 years). Therefore, the overall analysis did not represent reliable climatology. Furthermore, the quite significant data were missing due to unstable hardware conditions. Thus, given this, the overall seasonal variation of cloud properties in this section should be considered. Diurnal cycle and variation of vertical structures that were less affected by the missing data are shown in the next section. Remote Sens. 2020, 12,

Seasonal Variation of Cloud Properties
The characteristics of the cloud occurrence were investigated in this section. The average annual cloud frequency was approximately 35.9%. The black bar in Figure 2 is the monthly cloud occurrence (normalized with the available data of the month) averaged from June 2014 to May 2016. The highest cloud occurrence, approximately 50%, showed in June when excluding the months with available data of lower than 30% (that is, May, November, and December were excluded due to the representativeness). The cloud was least frequent in March, less than 13%. Missing data existed over the short period of time (2 years). Therefore, the overall analysis did not represent reliable climatology. Furthermore, the quite significant data were missing due to unstable hardware conditions. Thus, given this, the overall seasonal variation of cloud properties in this section should be considered. Diurnal cycle and variation of vertical structures that were less affected by the missing data are shown in the next section.
The frequency of cloud occurrence was approximately 15.6% (43.5% of the yearly overall cloud occurrence) in summer, 8.6% (23.9%) in autumn, 5.2% (14.4%) in winter, and 6.5% (18.1%) in spring (Table 5). This result is similar to previous studies that analyzed the cloud frequency in East Asia. According to Cai et al. [41], the occurrence of clouds in summer and autumn was, in general, greater than in spring and winter over China, based on the cloud-aerosol lidar with orthogonal polarization (CALIOP) data. Lee et al. [23] also showed that the occurrence of clouds in summer accounts for the largest proportion of the yearly frequency in Seoul, Korea. Furthermore, the occurrence by cloud types showed the highest value in RainDP (11.8%; 32.9% of the yearly average cloud occurrence), followed by HC (9.3%; 26.0%), MC (7.4%; 20.6%), RainSH (4.4%; 12.0%), and LC (2.9; 8.2%). That is, approximately 45% of clouds were accompanied by precipitation, and the RainSH was approximately 27% of the precipitation. The previous analyses with the shorter period (2-3 months) showed a similar result that the LC containing precipitating clouds was the most frequent (Oh et al. [24] and Ye et al. [25]). The frequency of cloud occurrence was approximately 15.6% (43.5% of the yearly overall cloud occurrence) in summer, 8.6% (23.9%) in autumn, 5.2% (14.4%) in winter, and 6.5% (18.1%) in spring (Table 5). This result is similar to previous studies that analyzed the cloud frequency in East Asia. According to Cai et al. [41], the occurrence of clouds in summer and autumn was, in general, greater than in spring and winter over China, based on the cloud-aerosol lidar with orthogonal polarization (CALIOP) data. Lee et al. [23] also showed that the occurrence of clouds in summer accounts for the largest proportion of the yearly frequency in Seoul, Korea. Furthermore, the occurrence by cloud types showed the highest value in RainDP (11.8%; 32.9% of the yearly average cloud occurrence), followed by HC (9.3%; 26.0%), MC (7.4%; 20.6%), RainSH (4.4%; 12.0%), and LC (2.9; 8.2%). That is, approximately 45% of clouds were accompanied by precipitation, and the RainSH was approximately 27% of the precipitation. The previous analyses with the shorter period (2-3 months) showed a similar result that the LC containing precipitating clouds was the most frequent (Oh et al. [24] and Ye et al. [25]). In Figure 8, we examined the characteristics of the seasonal frequency of cloud occurrence by cloud type, and we normalized the frequency with the seasonal average occurrence in Table 5. The cloud occurrence of HC and RainDP in summer was the largest (Figure 8 and Table 5). In particular, HC reached approximately 35% of the summer clouds and was as low as 6.6% in winter. The RainDP occurred in more than one third of the seasonal clouds except for winter. In particular, the RainDP and HC were approximately 68% in summer, which is similar to previous findings over China (Zhang et al. [42]). In the summer, the deep convection usually occurs due to the surface heating by the strong solar radiation [43][44][45]. Furthermore, the dynamic instability increased during the summer rainy seasons and precipitation frequently occurred. The clouds developing by deep convection and/or dynamic instability were possibly vertically extended and did not show any clear or no cloud regions in the vertical direction (this will be explored later). RainSH and LC most frequently occurred in winter (29% and 16%, respectively). Most of these shallow clouds are related to cloud development by the air-sea interaction during the cold air outbreak in winter in the Korea Peninsula [46][47][48]. When the Siberian high expands to the Korea Peninsula during the winter, oceans are relatively warmer than the cold air, leading to a shallow layer of conditionally unstable air with abundant moisture provided by the oceans. Convective but shallow clouds develop over the oceans and move to the land, leading to snow, which is the dominant process generating shallow clouds (RainSH and LC) in winter. The frequency of MC was higher in the cold season, namely 19 In Figure 8, we examined the characteristics of the seasonal frequency of cloud occurrence by cloud type, and we normalized the frequency with the seasonal average occurrence in Table 5. The cloud occurrence of HC and RainDP in summer was the largest (Figure 8 and Table 5). In particular, HC reached approximately 35% of the summer clouds and was as low as 6.6% in winter. The RainDP occurred in more than one third of the seasonal clouds except for winter. In particular, the RainDP and HC were approximately 68% in summer, which is similar to previous findings over China (Zhang et al. [42]). In the summer, the deep convection usually occurs due to the surface heating by the strong solar radiation [43][44][45]. Furthermore, the dynamic instability increased during the summer rainy seasons and precipitation frequently occurred. The clouds developing by deep convection and/or dynamic instability were possibly vertically extended and did not show any clear or no cloud regions in the vertical direction (this will be explored later). RainSH and LC most frequently occurred in winter (29% and 16%, respectively). Most of these shallow clouds are related to cloud development by the air-sea interaction during the cold air outbreak in winter in the Korea Peninsula [46][47][48]. When the Siberian high expands to the Korea Peninsula during the winter, oceans are relatively warmer than the cold air, leading to a shallow layer of conditionally unstable air with abundant moisture provided by the oceans. Convective but shallow clouds develop over the oceans and move to the land, leading to snow, which is the dominant process generating shallow clouds (RainSH and LC) in winter. The frequency of MC was higher in the cold season, namely 19.6%, 21.9%, and 27.4% in spring, autumn, and winter, respectively.   Table 5.

Diurnal Variation of Clouds
The diurnal variation of cloud occurrence is shown for the cloud type and season (Figure 9). The HC had a diurnal variation with a high occurrence between 1800 and 0900 LST, and a lower frequency at 0900-1600 LST, except for winter. The diurnal variation of the altitude of the maximum frequency was significant in spring and autumn. In summer, the HC occurred in the wide range of the altitude. The MC in spring and winter had a similar pattern with a minimum occurrence during the day and a peak occurrence during the night. However, a maximum frequency appeared from 0300 to 1200 LST in autumn.
Remote Sens. 2020, 12, x; doi: FOR PEER REVIEW www.mdpi.com/journal/remotesensing cycle peaks appear in the morning (0300-0900 LST) and the late afternoon (1600-2000 LST), respectively, in summer and winter over Korea. This maximum precipitation in the late afternoon is explained by ground heating by solar radiation and subsequently increased instability in the atmosphere. However, the frequent precipitation in the morning is related to the instability associated with radiative cooling in the upper atmosphere and complex processes of precipitation [50]. The weak diurnal variation in LC and RainSH was present (not shown).  Figure 10 shows the vertical profiles of the averaged reflectivity for the four seasons. The summer clouds had the highest reflectivity and the winter clouds had the lowest reflectivity at all heights. In summer and autumn, the average reflectivity increased sharply at approximately 4.5 km, A clear single peak appeared in RainDP from 1200 to 1800 LST in autumn, however, this feature was inconsistent in other seasons. The high frequency appeared during the entire day in summer and two main peaks occurred from 0300 to 1000 LST and from 1400 to 2000 LST, which is consistent with the two peaks of the maximum frequency of precipitation in the early morning and late afternoon by Lim and Kwon [49]. Lee and Seo [50] also showed similar results that the diurnal and semidiurnal cycle peaks appear in the morning (0300-0900 LST) and the late afternoon (1600-2000 LST), respectively, in summer and winter over Korea. This maximum precipitation in the late afternoon is explained by ground heating by solar radiation and subsequently increased instability in the atmosphere. However, the frequent precipitation in the morning is related to the instability associated with radiative cooling in the upper atmosphere and complex processes of precipitation [50]. The weak diurnal variation in LC and RainSH was present (not shown). Figure 10 shows the vertical profiles of the averaged reflectivity for the four seasons. The summer clouds had the highest reflectivity and the winter clouds had the lowest reflectivity at all heights. In summer and autumn, the average reflectivity increased sharply at approximately 4.5 km, indicating that a bright band was observed near this altitude. The bright band appeared near 2 km in winter. However, the bright bands were unclear in spring. There was no significant change of reflectivity with heights below 4 km in summer, but a significant decrease in reflectivity was noticeable below 1.5 km in spring and autumn. indicating that a bright band was observed near this altitude. The bright band appeared near 2 km in winter. However, the bright bands were unclear in spring. There was no significant change of reflectivity with heights below 4 km in summer, but a significant decrease in reflectivity was noticeable below 1.5 km in spring and autumn. The slope of the average reflectivity can be related to the growth of cloud particles. The rapid increase in reflectivity above 7 km (roughly −15°C with an assumption of pseudo-adiabatic lapse rate from the melting layer) in summer and autumn indicated the significant growth of ice particles. A smaller gradient of reflectivity appeared between 5 and 7 km, specifically in summer, indicating a rapid growth above 7 km (roughly −15°C layer) and a relatively slower growth below 7 km. Further thorough studies are expected to understand this change in the growth rate related to the microphysical processes. However, this dramatic change of growth rate was less prominent in spring and winter; there was almost no change in the slope at 7 km in spring and a slight change at 4 km in winter.

Characteristics of Vertical Profiles
The vertical structure of clouds was examined as CFADs of reflectivity and Doppler velocity for each season (Figure 11). The thickness of clouds was the largest in summer (approximately 14 km, Figure 10b), and the clouds were the thinnest in winter (Figure 11d). Furthermore, clouds occurred most frequently in summer and least in winter. The peak of frequency appeared at the two heights (the HC and LC region) for all seasons with a lower frequency in MC. The upper peak shifted to the lower layers from warm to cold seasons. The two modes (precipitation mode in Z > 0 dBZ and cloud mode in Z < −10 dBZ) were evident except for in summer. The cloud mode was dominant in spring and winter, while the precipitation mode was dominant in summer. The precipitation mode was linked with clouds in MC and HC, showing the downward increase in reflectivity in this dominant precipitation mode. This increase was the largest in the layer of 8-12 km in summer and was linked with the gradual increase (toward larger negative values) in Doppler velocity. A similar trend occurred in the different seasons to a smaller degree (the least increase in spring). Furthermore, the Doppler velocity became almost constant in the layer of 6-8 km in summer. This constant Doppler velocity with increasing reflectivity indicated the existence of a significant updraft in this layer. A The slope of the average reflectivity can be related to the growth of cloud particles. The rapid increase in reflectivity above 7 km (roughly −15 • C with an assumption of pseudo-adiabatic lapse rate from the melting layer) in summer and autumn indicated the significant growth of ice particles. A smaller gradient of reflectivity appeared between 5 and 7 km, specifically in summer, indicating a rapid growth above 7 km (roughly −15 • C layer) and a relatively slower growth below 7 km. Further thorough studies are expected to understand this change in the growth rate related to the microphysical processes. However, this dramatic change of growth rate was less prominent in spring and winter; there was almost no change in the slope at 7 km in spring and a slight change at 4 km in winter.
The vertical structure of clouds was examined as CFADs of reflectivity and Doppler velocity for each season (Figure 11). The thickness of clouds was the largest in summer (approximately 14 km, Figure 11b), and the clouds were the thinnest in winter (Figure 11d). Furthermore, clouds occurred most frequently in summer and least in winter. The peak of frequency appeared at the two heights (the HC and LC region) for all seasons with a lower frequency in MC. The upper peak shifted to the lower layers from warm to cold seasons. The two modes (precipitation mode in Z > 0 dBZ and cloud mode in Z < −10 dBZ) were evident except for in summer. The cloud mode was dominant in spring and winter, while the precipitation mode was dominant in summer. The precipitation mode was linked with clouds in MC and HC, showing the downward increase in reflectivity in this dominant precipitation mode. This increase was the largest in the layer of 8-12 km in summer and was linked with the gradual increase (toward larger negative values) in Doppler velocity. A similar trend occurred in the different seasons to a smaller degree (the least increase in spring). Furthermore, the Doppler velocity became almost constant in the layer of 6-8 km in summer. This constant Doppler velocity with increasing reflectivity indicated the existence of a significant updraft in this layer. A similar but less prominent signature was shown in the Doppler velocity in autumn. The dramatic increase in Doppler velocity (toward negative values) was shown in the height of 3 to 5 km for spring, summer, and autumn, indicating the existence of a melting layer. The melting layer was the highest in summer and less prominent in winter, with some signature below 2.5 km.
Remote Sens. 2020, 12, x FOR PEER REVIEW 16 of 29 Remote Sens. 2020, 12, x; doi: FOR PEER REVIEW www.mdpi.com/journal/remotesensing similar but less prominent signature was shown in the Doppler velocity in autumn. The dramatic increase in Doppler velocity (toward negative values) was shown in the height of 3 to 5 km for spring, summer, and autumn, indicating the existence of a melting layer. The melting layer was the highest in summer and less prominent in winter, with some signature below 2.5 km. Seasonal CFADs were analyzed according to cloud types, HC, MC, and LC ( Figure 12 and Figure  13). The seasonal variation of the HC was prominent with the highest height in summer (10-12 km) and the lowest in winter (6-8 km). The dramatic growth appeared in the layer of 8-12 km in summer and of 9-10 km in autumn and was linked with the rapid increase in Doppler velocity in this layer (upper panels in Figure 13). Furthermore, this layer matched the maximum increase in the mass flux (upper panels in Figure 14). The largest mass flux was shown in summer. To further investigate the possible growth processes, we selected a HC case in the summer (9 August, 2015) and derived the vertical air velocity. A similar dramatic growth was shown in the CFAD of reflectivity ( Figure 15). Both reflectivity and Doppler velocity rapidly increased with decreasing heights. However, the terminal fall velocity did not change (almost constant) with heights. A significant upward motion was shown throughout the HC regions, in particular above 9 km altitude. The updraft reached up to 0.9 m s −1 . This strong updraft led to the rapid growth of ice crystals by the depositional process and possibly the riming process in this layer during summer.
No significant increase in Doppler velocity and small mass flux was shown in winter. The reflectivity values were also smaller in winter. Furthermore, the CFAD of Doppler velocity was Seasonal CFADs were analyzed according to cloud types, HC, MC, and LC (Figures 12 and 13). The seasonal variation of the HC was prominent with the highest height in summer (10-12 km) and the lowest in winter (6-8 km). The dramatic growth appeared in the layer of 8-12 km in summer and of 9-10 km in autumn and was linked with the rapid increase in Doppler velocity in this layer (upper panels in Figure 13). Furthermore, this layer matched the maximum increase in the mass flux (upper panels in Figure 14). The largest mass flux was shown in summer. To further investigate the possible growth processes, we selected a HC case in the summer (9 August, 2015) and derived the vertical air velocity. A similar dramatic growth was shown in the CFAD of reflectivity ( Figure 15). Both reflectivity and Doppler velocity rapidly increased with decreasing heights. However, the terminal fall velocity did not change (almost constant) with heights. A significant upward motion was shown throughout the HC regions, in particular above 9 km altitude. The updraft reached up to 0.9 m s −1 . This strong updraft led to the rapid growth of ice crystals by the depositional process and possibly the riming process in this layer during summer.
No significant increase in Doppler velocity and small mass flux was shown in winter. The reflectivity values were also smaller in winter. Furthermore, the CFAD of Doppler velocity was narrower in the upper layer and slowly broadened with decreasing heights, indicating the broadening of snow particles with different velocities, likely different mass and size of particles.   Unlike HC, the height of the maximum reflectivity frequency appeared at similar heights (5-7 km) in MC, but the difference in reflectivity still existed with the largest value in summer and the smallest in winter (second row in Figure 12). Furthermore, the rapid growth appeared in the upper region of MC (HC region) in all seasons. This was also shown in the rapid increase in the Doppler velocity (see the second row in Figure 13) and mass flux (see the second row in Figure 14) in this upper region of MC. Different to HC, the reflectivity of MC rapidly decreased with heights in the lower layer (MC regions), which was linked with a near-constant Doppler velocity at heights between 4-7 km, except for in summer. The mass flux was nearly constant or decreased in this lower layer. Furthermore, the gradient of the Doppler velocity in summer was nearly constant in the layer of 6-8 km, although a rapid increase was shown in the lower and upper parts, indicating an updraft, as shown in the CFAD of the Doppler velocity in summer in Figure 13. Remote Sens. 2020, 12,    LC showed a more dramatic change of reflectivity with heights. The rapid growth was shown in the upper layer, and a dramatic decrease appeared in the lower layer. This dramatic decrease was linked with the possible evaporation of precipitation particles, mostly raindrops, so that the cloud and precipitation particles did not reach the surface. The warm season showed two peaks of the frequency (upper and lower layers), while the cold season had a continuous frequency distribution. The peak of the frequency appeared in smaller Doppler velocity ranges (−1-0 ms −1 ) due to dramatic evaporation in the lower layer. Similar to MC, updraft regions were present in the layer of 6-7 km in summer.
RainDP showed a rapid increase in reflectivity, in particular, in the upper HC layer during spring, summer, and autumn, while the winter RainDP had the largest gradient of reflectivity in the MC layer of 2-5 km (Figure 16), indicating the rapid growth of RainDP in the HC and MC regions. Unlike the CFAD in Figure 11, the frequency gradually increased with decreasing heights, with a dominant peak in the reflectivity range of 10-20 dBZ, indicating a continuous growth of clouds and precipitation particles in the MC regions instead of the strong evaporation shown in LC. The summer RainDP was the deepest and the winter RainDP showed the smallest depth. In particular, the summer RainDP had some clouds reaching 10 dBZ at 10 km. Furthermore, there was a signature of updraft (constant Doppler velocity) in the layer of 6-8 km with rapid increases in Doppler velocity above and below the layer. These overall features in summer can be explained by the frequent vertical development of strong subtropical clouds due to synoptic forcing and vertical wind shear [51].  The constant Doppler velocity was further examined for a RainDP case on 5 July, 2014. A similar analysis as in Figure 15 was applied for this case. The distinctive change of Z gradient was shown in the layers above and below 8 km. Similar to HC, the dramatic growth was linked with the rapid increase in the Doppler velocity, and a significant updraft that reached up to 1 m s −1 (Figure 17). The Doppler velocity decreased in the layer below 8 km with decreasing heights and slightly increased in the layer below 7 km. The clear signature of updraft was shown in the layer of 6-8 km (see Figure 17d). The physical causes of this updraft region centered at 7 km (about −15 • C) are not certain and require further investigation.  Unlike HC, the height of the maximum reflectivity frequency appeared at similar heights (5-7 km) in MC, but the difference in reflectivity still existed with the largest value in summer and the smallest in winter (second row in Figure 12). Furthermore, the rapid growth appeared in the upper region of MC (HC region) in all seasons. This was also shown in the rapid increase in the Doppler velocity (see the second row in Figure 13) and mass flux (see the second row in Figure 14) in this upper region of MC. Different to HC, the reflectivity of MC rapidly decreased with heights in the lower layer (MC regions), which was linked with a near-constant Doppler velocity at heights between 4-7 km, except for in summer. The mass flux was nearly constant or decreased in this lower layer. Furthermore, the gradient of the Doppler velocity in summer was nearly constant in the layer of 6-8 km, although a rapid increase was shown in the lower and upper parts, indicating an updraft, as shown in the CFAD of the Doppler velocity in summer in Figure 13.
LC showed a more dramatic change of reflectivity with heights. The rapid growth was shown in the upper layer, and a dramatic decrease appeared in the lower layer. This dramatic decrease was linked with the possible evaporation of precipitation particles, mostly raindrops, so that the cloud and precipitation particles did not reach the surface. The warm season showed two peaks of the frequency (upper and lower layers), while the cold season had a continuous frequency distribution. The peak of the frequency appeared in smaller Doppler velocity ranges (−1-0 ms -1 ) due to dramatic evaporation in the lower layer. Similar to MC, updraft regions were present in the layer of 6-7 km in summer.
RainDP showed a rapid increase in reflectivity, in particular, in the upper HC layer during spring, summer, and autumn, while the winter RainDP had the largest gradient of reflectivity in the MC layer of 2-5 km (Figure 16), indicating the rapid growth of RainDP in the HC and MC regions. Unlike the CFAD in Figure 11, the frequency gradually increased with decreasing heights, with a dominant peak in the reflectivity range of 10-20 dBZ, indicating a continuous growth of clouds and precipitation particles in the MC regions instead of the strong evaporation shown in LC. The summer RainDP was the deepest and the winter RainDP showed the smallest depth. In particular, the summer RainDP had some clouds reaching 10 dBZ at 10 km. Furthermore, there was a signature of updraft (constant Doppler velocity) in the layer of 6-8 km with rapid increases in Doppler velocity above and below the layer. These overall features in summer can be explained by the frequent vertical development of strong subtropical clouds due to synoptic forcing and vertical wind shear [51]. The constant Doppler velocity was further examined for a RainDP case on 5 July, 2014. A similar analysis as in Figure 15 was applied for this case. The distinctive change of Z gradient was shown in the layers above and below 8 km. Similar to HC, the dramatic growth was linked with the rapid increase in the Doppler velocity, and a significant updraft that reached up to 1 m s -1 ( Figure 17). The Doppler velocity decreased in the layer below 8 km with decreasing heights and slightly increased in the layer below 7 km. The clear signature of updraft was shown in the layer of 6-8 km (see Figure  17d). The physical causes of this updraft region centered at 7 km (about −15℃) are not certain and require further investigation.  The RainDP could contain not only vertically developed single-layer precipitation clouds (top height ≥ 2 km) but also multi-layer clouds with shallow precipitation clouds below, and different layers of clouds above, due to the definition of the RainDP in this study. The RainDP was divided into the vertically continuous single-layer RainDP and multi-layer RainDP. The latter was identified by the no cloud area of at least 500 m layers in the vertical direction. Over 70% of RainDPs were the continuous single-layer RainDP ( Figure 18). In particular, the single-layer RainDP was over 80% in summer and winter. This suggested that most RainDPs were single-layer precipitation clouds with strong vertical connections in the development of clouds and precipitation growth. The multi-layer RainDP was composed of RainSH below 2 km, and MC or mostly HC (lower panels in Figure 18). The RainDP could contain not only vertically developed single-layer precipitation clouds (top height ≥ 2 km) but also multi-layer clouds with shallow precipitation clouds below, and different layers of clouds above, due to the definition of the RainDP in this study. The RainDP was divided into the vertically continuous single-layer RainDP and multi-layer RainDP. The latter was identified by the no cloud area of at least 500 m layers in the vertical direction. Over 70% of RainDPs were the continuous single-layer RainDP ( Figure 18). In particular, the single-layer RainDP was over 80% in summer and winter. This suggested that most RainDPs were single-layer precipitation clouds with strong vertical connections in the development of clouds and precipitation growth. The multi-layer RainDP was composed of RainSH below 2 km, and MC or mostly HC (lower panels in Figure 18).
The RainSH showed different characteristics with the seasons (Figure 19). Common rapid increases in reflectivity (Doppler velocity) were shown in summer, while the other seasons showed constant or decreasing reflectivity (Doppler velocity) due to evaporation. An updraft of larger than 1 m s −1 occurred at a 1.5 km height in summer. In particular, the reflectivity gradient was about 20 dBZ km −1 in the layer of 1.5 km (Z = −20 dBZ) to 0.5 km (Z = 0 dBZ), indicating that the growth of precipitation particles was the most efficient in summer. The updraft appeared in the lower layer in the different seasons and no systematic increases in the Doppler velocity were shown.  The RainSH showed different characteristics with the seasons (Figure 19). Common rapid increases in reflectivity (Doppler velocity) were shown in summer, while the other seasons showed constant or decreasing reflectivity (Doppler velocity) due to evaporation. An updraft of larger than 1 m s −1 occurred at a 1.5 km height in summer. In particular, the reflectivity gradient was about 20 dBZ km −1 in the layer of 1.5 km (Z = −20 dBZ) to 0.5 km (Z = 0 dBZ), indicating that the growth of precipitation particles was the most efficient in summer. The updraft appeared in the lower layer in the different seasons and no systematic increases in the Doppler velocity were shown.

Discussions
The updraft region in the 6-8 km layer (1-3 km above the melting layer) was prominent for MC, LC, and RainDP, particularly in the summer. This region coincides with constant Doppler velocity (its rapid increase above and below this layer) and the dramatic change of the vertical gradient of reflectivity. When we consider the pseudo-adiabatic lapse rate, the temperature of this region is

Discussion
The updraft region in the 6-8 km layer (1-3 km above the melting layer) was prominent for MC, LC, and RainDP, particularly in the summer. This region coincides with constant Doppler velocity (its rapid increase above and below this layer) and the dramatic change of the vertical gradient of reflectivity. When we consider the pseudo-adiabatic lapse rate, the temperature of this region is approximately −20 to −7 • C. One possible cause of this updraft could be because of the dynamical instability such as strong wind shear or thermal instability. However, it is uncertain why these instabilities occur in the same layer. On the other hand, the melting could induce significant mesoscale circulation that could happen in the same height from the melting layer. Further thorough research will create a better understanding of this systematic updraft.
Recently, comprehensive studies on the retrieval of vertical air velocity in the vertically pointing radars have been explored [16,39,40,52]. Several methods were evaluated and their accuracies are documented. The growth rate and corresponding microphysical processes highly correlate with the vertical air motion. Furthermore, the high-resolution rawinsonde, ceilometer, and radiometer observation can further provide information on the liquid water path, environment air, and cloud base. Thus, further investigation of the microphysical processes is desired regarding the derived vertical air motion and other instrumentations.
This research is limited to the cloud data of only two years, and the cloud radar was somewhat unstable during the observation period. This study intends to provide a baseline for understanding cloud statistics and vertical structures and related microphysical processes. Further comprehensive studies are recommended with a long-term data set.

Conclusions
A ground-based cloud radar can be used to analyze cloud properties and the vertical distribution of clouds and help to understand the processes in cloud formation and development. In this study, we analyzed the occurrence and vertical structures of clouds for two years using cloud measurements from the Ka-band cloud radar at the Boseong site. The reflectivity of the cloud radar was first calibrated using two vertically pointing radars (VertiX and MRR) and two disdrometers (2DVD and PARSIVEL). The VertiX and MRR were compared with reflectivity derived from disdrometric measurements and the cloud radar reflectivity was subsequently compared with the calibrated reflectivity from the VertiX and MRR by considering the scattering effect and attenuation by precipitation. The non-meteorological echoes (insects, noise, etc.) were removed from the calibrated reflectivity of the cloud radar using the different thresholds of Z and LDR in warm and cold seasons and despeckling in the radial direction. The clouds are classified into the five types of precipitating (RainDP and RainSH) and non-precipitating clouds (HC, MC, and LC) by following the criteria of Kollias et al. [3]. We investigated the cloud occurrence and vertical structure of reflectivity, the Doppler velocity, and mass flux and their linkage with dominant microphysical processes for the five cloud types.
The average frequency of cloud occurrence was 35.9% over the observation period with the maximum frequency of 50% in June. The RainDP occurred most frequently (11.8%), followed by HC (9.3%), MC (7.4%), RainSH (4.4%), and LC (2.9%). The precipitating clouds (RainDP and RainSH) were almost 45% of the cloud occurrence. In particular, the HC and RainDP were frequently observed in summer and autumn due to strong surface heating and dynamic instability. However, RainSH, LC, and MC were dominant in the winter season. In particular, RainSH and LC were approximately 45% of the winter clouds and mostly due to cloud development by the air-sea interaction during the cold air outbreak.
We also examined the growth of cloud particles using the slope of the average reflectivity. The vertical gradient of reflectivity (with decreasing heights) above 7 km (about −15 • C layer) was larger and becomes smaller at 5-7 km, indicating that falling ice crystals grew faster above 7 km. Furthermore, this implies different growth modes or microphysical processes in the layer below and above 7 km. This change of vertical gradient was less significant in winter and spring.
The CFADs of reflectivity showed that the summer clouds are the thickest, and winter clouds are the thinnest. The peak of frequency appeared in the two heights (the HC and LC regions) with a low frequency in the MC region. Furthermore, the precipitation and cloud modes were shown in the LC region except for summer. The precipitation mode had a vertical linkage with MC and HC and showed a downward increase in reflectivity and Doppler radial velocity. However, the cloud mode was discontinuous with a lower frequency at the 2-4 km heights. A significant updraft dominantly exists in summer in the layer of 6-8 km (about 2-2.5 km above the melting layer).
The HC had a prominent seasonal variation with the highest height in summer and the lowest in winter. The rapid increase in reflectivity, the Doppler radial velocity, and mass flux in the layer above 7 km in the warm seasons (summer and autumn) indicated the possible growth of ice crystals by the riming processes in addition to depositional growth. A similar feature appeared in the autumn to a lesser extent. In contrast with HC, the peak frequency height of MC appeared in a similar height for all seasons with significant evaporation below this height. The Doppler velocity of MC rapidly increased in the layer above 8 km due to rapid growth, while it remained nearly constant below 8 km except for in summer. The mass flux was almost constant or decreased in this lower layer. The similar rapid growth in the upper layer and evaporation in the lower layer appeared in LC with a dramatic increase and reduction in reflectivity with decreasing heights. The evaporation was also supported by the frequency peak in the smaller Doppler velocity range (−1-0 m s −1 ) near the ground. However, summer and autumn showed the two peaks in the upper and lower layers with a lower frequency in the MC region.
RainDP showed the continuous growth of precipitation particles in the entire layer, in particular in the upper HC and MC regions. The frequency of the precipitation mode increased toward the ground. RainSH showed the rapid increases in reflectivity and Doppler velocity in summer only. The precipitation growth was extremely efficient with the reflectivity gradient of about 20 dBZ km −1 . The updraft larger than 1 m s −1 is present for all seasons and is stronger than that in the MC or HC regions.