Radar and ground-level measurements of precipitation collected by EPFL during the ICE-POP 2018 campaign in South-Korea

Abstract. This article describes a four-month dataset of precipitation and cloud measurements collected during the International Collaborative Experiments for PyeongChang 2018 Olympic and Paralympic winter games (ICE-POP 2018). This paper aims to describe the data collected by the Environmental Remote Sensing Laboratory of the École Polytechnique Fédérale de Lausanne. The dataset includes observations from an X-band dual-polarisation Doppler radar, a W-band Doppler cloud profiler, a multi-angle snowflake camera and a two-dimensional video disdrometer (https://doi.pangaea.de/10.1594/PANGAEA.918315, Gehring et al. (2020a)) . Classifications of hydrometeor types derived from dual-polarisation measurements and snowflake photographs are presented. The dataset covers the period from 15 November 2017 to 18 March 2018 and features nine precipitation events with a total accumulation of 195 mm of equivalent liquid precipitation. This represents 85 % of the climatological accumulation over this period. To illustrate the available data, measurements corresponding to the four precipitation events with the largest accumulation are presented. The synoptic situations of these events were contrasted and influenced the precipitation type and accumulation. The hydrometeor classifications reveal that aggregate snowflakes were dominant and that some events featured significant riming. The combination of dual-polarisation variables and high-resolution Doppler spectra with ground-level snowflake images makes this dataset particularly suited to study snowfall microphysics in a region where such measurements were not available before.


view of the particle. Horizontal wind induces a horizontal displacement of the particles, such that the superposition of the one-dimensional sections can lead to distorted particles. This issue is thoroughly investigated with numerical simulations in Nešpor et al. (2000). The two orthogonal two-dimensional projections yield to a three-dimensional shape information, which can be used to compute the equivalent drop diameter and the aspect ratio. This makes it possible to compute the raindrop size distribution (DSD). Since the vertical distance between the two light sheets is known, the particles' fall velocities can 5 also be computed. 2DVD data can also be used for snowfall microphysics studies. Brandes et al. (2007) derived the particle size distribution (PSD) from 2DVD data in Colorado. Huang et al. (2010) and Huang et al. (2015) used a 2DVD to derive radar reflectivity-snowfall rate relations. Finally Grazioli et al. (2014) used 2DVD data to develop a supervised hydrometeor classification method.

MXPol
First, the noise floor is determined from the raw power following the method from Hildebrand and Sekhon (1974). Then, the polarimetric variables are computed based on the backscattering covariance matrix following Doviak and Zrnic (1993). The computation of K dp is based on an ensemble of Kalman filters as detailed in Schneebeli et al. (2014).

15
To monitor the stability of the radar signal, a radar target simulator (RTS, Schneebeli and Leuenberger) developed by Palindrome Remote Sensing GmbH was installed during the campaign. Unfortunately, due to technical issues during the campaign, the data could not be used for calibration of the radar. However, we conducted dedicated calibration measurements with the RTS in July 2018 just after the ICE-POP 2018 campaign. The results showed that the reflectivity measurements have errors smaller than 1 dBZ. 20

Hydrometeor classification
The dual-polarisation observables were used to feed the hydrometeor classification from Besic et al. (2016). The centroids of all four polarimetric variables used for the classification have been trained on MXPol data from various field campaigns in the Swiss Alps, in Ardèche (France), in Antarctica and on the present dataset in Korea. Recently, Besic et al. (2018) developed a demixing approach of this hydrometeor classification, in which the proportion of hydrometeors for each radar sampling volume 25 is estimated, instead of one dominant class. The classes are crystals, aggregates, light rain, rain, rimed ice particles, vertically aligned ice, wet snow, ice hail and high-density graupel and melting hail. This approach is essentially built upon the concept of entropy, as defined in Besic et al. (2016), which reflects the uncertainty with which a hydrometeor class is assigned to one sampling volume. This de-mixing method has the advantage of revealing the spectrum of hydrometeors present in the observed precipitation. The classification was applied to all RHIs of the precipitation events shown in Fig. 4. Only the data above 2000 30 4 https://doi.org /10.5194/essd-2020-134 Open Access Earth System Science Data Discussions Preprint. Discussion started: 9 July 2020 c Author(s) 2020. CC BY 4.0 License. m a.s.l. have been selected for the hydrometeor classification shown in Sect. 4, because of ground echoes contamination and partial beam filling below this altitude.

Differential reflectivity bias correction
For a correct interpretation of Z DR , the offset introduced by the existence of differences in amplitude in the horizontal and vertical channels needs to be subtracted. This calibration can be achieved by analysing Z DR values in a specific subset of 5 the range gates of the vertical PPI, which were performed at least once per hour during the whole campaign. Unfortunately, MXPol is affected by extremely high Z DR values in the low gates, probably caused by issues on the transmit-receive limiter. Therefore, a classical calibration procedure such as the one described in Gorgucci et al. (1999) cannot be applied. Instead, we decided to select the range interval used for the correction among the upper gates, unaffected by the issue. The first step of the calibration procedure was the removal of data with signal to noise ratios lower than 5 dB or ρ hv < 0.95. For each PPI, we also 10 removed the range gates in which we encountered at least one non-valid Z DR measurement, to avoid introducing a bias caused by some angles being over-represented. Subsequently, we computed, for each range gate, the standard deviation of the Z DR distribution over the whole campaign duration. This standard deviation is remarkably constant more than 1 km above the radar, while its magnitude increases rapidly in the closest gates, due to the issue mentioned before. After computing the median of these values in the top 25 % of the range gates, we impose a maximum threshold of 0.1 dB on the absolute difference between 15 the standard deviation at each range gate and the median value. The median of all Z DR values from the range gates satisfying the condition is 2.66 dB with 50 % of the values within 0.32 dB. This median value of 2.66 dB was subtracted from all Z DR measurements to get the corrected Z DR dataset.

MASC
The raw data from the MASC are stereographic photographs of hydrometeors and measurements of fall velocities. Praz et al.

20
(2017) developed a hydrometeor classification and riming degree estimation of MASC pictures based on a multinomial logistic regression model. More recently, Hicks and Notaroš (2019) used convolutional neural networks to classify MASC snowflake images. In this paper, we will use the algorithm from Praz et al. (2017) to classify the MASC data collected during ICE-POP 2018. Schaer et al. (2020) developed a method to classify MASC images as blowing snow, precipitation or a mixture of those. This makes it possible to filter the results and minimise the influence of possible blowing snow. Even though a double 25 fence windshield was present during the ICE-POP 2018 campaign, 31% of the particles were classified either as a mixture of precipitation and blowing snow or pure blowing snow. In the present dataset, all particles are retained, but the information needed to filter out blowing snow particles is added. As explained in Schaer et al. (2020) a threshold of 0.193 on the normalised angle ψ can be used with ψ < 0.193 corresponding to pure precipitation. The results of the hydrometeor classification shown in Section 4 correspond to pure precipitation only. Preprint. Discussion started: 9 July 2020 c Author(s) 2020. CC BY 4.0 License.

WProf
The raw data from WProf were saved without any filtering, in the form of raw Doppler spectra. The spectra are then dealiased with an algorithm based on the minimisation of the spectral width at each range gate, similar to Ray and Ziegler (1977). This method assumes aliasing up to one folding, which is sufficient for the Nyquist intervals considered here (Table 1). From the dealiased spectra, the noise floor was determined using the method from Hildebrand and Sekhon (1974). The moments (V D , 5 Z, σ v , skewness and kurtosis) are then computed from the dealiased spectra above the noise floor.

Atmospheric gas attenuation
To estimate the attenuation due to atmospheric gases, we used the recommendations from the International Telecommunication Union (2013). Figure 2 shows the histograms of dry air, water vapour and total two-way attenuation (up and including 10 km) at W-band from 30 November 2017 to 31 March 2018, which corresponds to the period during which radiosoundings are

Sensitivity
To visualise the sensitivity of WProf and MXPol, Fig. 3 shows the empirical joint distributions of range and reflectivity values during all precipitation events of the ICE-POP 2018 campaign. The minimum measured reflectivity values represent the sensitivity. A threshold on the signal to noise ratio of 0 dB was applied on MXPol and WProf data in all figures presented.

20
For WProf (Fig. 3a) we can clearly see the effect of the three vertical chirps on the minimum detectable reflectivity. One can see that WProf has a higher sensitivity than MXPol at all range gates.  Table 4 shows the amount of data collected by each instrument.

Precipitation events
The measurement time from MXPol does not take into account the repositioning of the antenna between each scan, which typically takes the same time as the scan averaged over the whole cycle. This is why the measurement duration from MXPol is about half that from WProf, which measured continuously. The number of triplets captured by the MASC indicates the number 5 of sets of three pictures captured by the three cameras. For each picture, the classification selects one particle that is in focus.
The maximum rate of images is 2 Hz, hence only two hydrometeors can be identified every second. The 2DVD measures continuously at a rate of 34.1 kHz and can identify multiple particles in its sampling area, unlike the MASC. This explains why the number of particles captured by the 2DVD is two order of magnitudes greater than the number of triplets of the MASC. 2017   10 The 25 November 2017 event has the third-largest precipitation accumulation, but the second-largest mean precipitation rate (see Fig. 4). Figure 5 shows a strong westerly flow associated with a broad upper-level trough. Analysis of backward trajectories (not shown) revealed that the moisture was pumped from the Yellow Sea and lifted over the topography leading to a broad cloud and precipitation system.  and it observed almost exclusively raindrops, which are classified as small particles (Praz et al., 2017). The riming degree is above 0.25 (dotted line), because in MASC pictures of raindrops, only the reflection from the flashlights is visible and appear as small white spots, which is interpreted as rime by the classification. Hence the degree of riming in case of rain should not be used or be manually set to zero. Figure 6f shows the DSD computed from 2DVD data. The largest raindrops are observed during the most intense precipitation period (11:00-13:00 UTC) and correspond to the highest vertical extension of the cloud.

28 February 2018
The 28 February 2018 event stands out as the most intense of the whole campaign, in terms of accumulation and mean precipitation rate. At 00:00 UTC (not shown) a prominent PV streamer on eastern China and a low-pressure system eastward over the Yellow Sea are present. The PV streamer intensifies the surface cyclone and by 12:00 UTC 28 February (Fig. 7) the system is fully developed with the warm front passing over PyeongChang and leading to the observed precipitation. Note that At 00:00 UTC 28 February the nimbostratus (i.e. precipitating cloud associated with the warm front) can be observed above 2000 m, while fog is forming below 1000 m (Fig. 8a). Between the fog and the nimbostratus base, a dry layer is present where the precipitation from the nimbostratus sublimates to form virgas. At 03:00 UTC precipitation reaches the ground and lasts until 16:00 UTC. As temperatures at MHS are between 0 and 2 • C before 06:00 UTC, the liquid water attenuation of the melting snowflakes can lead to underestimation of the reflectivity measurements of both MXPol and WProf. After 06:00 UTC, 5 the temperatures at MHS are below freezing and hence there is almost no liquid water attenuation (attenuation from SLW droplets can be neglected). One can notice a region of embedded convection between 07:30 and 08:00 UTC and turbulence around 4000 m between 08:00 and 10:00 UTC (Fig.8b). During the passage of the front aggregates prevail, while from 06:00 to 08:00 UTC rimed particles dominate (Fig.8d). The class aggregates also contain rimed particles, which explained why the period dominated by rimed particles in Fig. 8d (06:00 to 08:00 UTC) is not visible in the MASC classification as graupel particles. However, it is clear from the degree of riming, that rimed particles are present during this period. The 04 March 2018 event has the second-largest precipitation accumulation. Figure 9 shows the synoptic conditions. There is a strong south-westerly flow advecting significant moisture from the Yellow Sea, as can be seen by the integrated vapour transport (IVT) fluxes reaching 1000 kg m −1 s −1 and a low-pressure system located south of Korea. This large moisture transport leads to widespread precipitation over the Korean peninsula with a maximum over the centre of South Korea. The 20 equivalent potential temperature shows the presence of warm and humid air reaching the cyclone centre. The large sea-level pressure gradient on the eastern Korean coast suggests the presence of strong south-easterly winds close to the surface (i.e. in geostrophic balance the wind is proportional and perpendicular to the pressure gradient). This south-easterly flow impinging the Taebaek mountains from the East Sea might have been orographically lifted and participated in an enhancement of the observed precipitation.
25 Figure 10a,b shows the reflectivity and Doppler velocity from WProf. The beginning of the event is dominated by rain with a melting layer around 2500 m which appears clearly from the Doppler velocity. The melting layer abruptly drops to the ground level (i.e 789 m a.s.l. at MHS) at 14:00 UTC, as temperatures quickly dropped below 0 • C (not shown). The cloud contains mainly crystals and aggregates (Fig. 10d).

07 March 2018
The 07 March 2018 event was the fourth most important in terms of precipitation accumulation (see Figure 4), but was the longest one because of a shallow precipitating system which lasted 12 hours after the main part of the event. On 07 March 00:00 UTC an upper-level trough is moving eastwards from China. Korea is under the influence of a ridge and clear sky conditions dominate. As the trough moves, moist unstable air from the Yellow Sea is advected over Korea and precipitation 5 sets in. Starting from 07 March 15:00 UTC a low-pressure system develops south of the Korean peninsula and the trough becomes a broad PV streamer. The precipitation intensity increases until the PV streamer passes over Korea. At 18:00 UTC the precipitation weakens, while the low-pressure system is further intensifying on the eastern flank of the PV streamer and reaches Japan with more intense precipitation than observed in Korea. The key differences between this event and the 28 February are that the cyclone formed more to the east with a less pronounced PV streamer and that the locations of both features were not 10 appropriate for a mutual intensification, as was the case on 28 February. This suggests that the timing and respective positions of the PV streamer and the low-pressure system during the 28 February event were key ingredients for its intensity. Figure 12a,b shows the reflectivity and Doppler velocity from WProf. The nimbostratus cloud associated with the surface cyclone generates precipitation, which starts around 10:00 UTC and lasts until 04:00 UTC. A shallower precipitating system brings again precipitation from around 08:30 UTC to 19:00 UTC. This event has some similarities with the 28 February: they 15 are both associated with a surface cyclone at the eastern flank of a PV streamer and they both feature a nimbostratus cloud followed by a shallower precipitating system. The latter is also associated with graupel particles (Fig.12e). The radar-based classification (Fig. 12d) shows mainly crystals and some aggregates. Since only values above 2000 m are considered and the precipitating cloud after 06:00 UTC 08 March is below 2000 m, no hydrometeors are present in the classification of Fig. 12d after 06:00 UTC 08 March. 20

Conclusions
In this article we presented a four-months dataset of cloud and precipitation measurements by an X-band polarimetric radar, a W-band Doppler cloud profiler, a multi-angle snowflake camera and a two-dimensional video disdrometer in the PyeongChang region in South Korea during the ICE-POP 2018 campaign.
The campaign was characterised by mostly cold, dry and windy weather. However, four major precipitation events took place 25 and contributed to 68% of the total precipitation accumulation over the campaign (25 November 2017 to 15 March 2018). We presented the meteorological conditions and data from these four events. The event with the largest precipitation accumulation (i.e. 28 February 2018) was characterised by an upper-level cyclonic enhancement due to the presence of a PV streamer, which led to a mature frontal system and intense precipitation. This event is further described in Gehring et al. Preprint. Discussion started: 9 July 2020 c Author(s) 2020. CC BY 4.0 License. study snowfall microphysics, thanks to the synergy between dual-polarisation and spectral information, as well as snowflake photographs.
Future studies could use the data presented in this paper together with other measurements from ICE-POP 2018, which will be publically released. This includes radar data at X, K-u and K-a band and is particularly suited for microphysical studies with multi-frequency measurements.