The CloudSat cloud-profiling radar is a nadir-looking 94 GHz radar which measures the signal backscattered by hydrometeors. Radar reflectivity profiles are divided into 150 vertical bins with a resolution of 240 m, with a 1.7×1.3 km2 footprint and up to 82∘ of latitude. CloudSat has been operating fulltime since April 2006 but because of a dysfunctional onboard battery has been only able to provide daylight observations since April 2011. The satellite is characterized by a period of 16 days, so it exactly overpasses a location every 16 days. The DDU station is overpassed by a descending orbit, whereas the PE station is overpassed by ascending and descending orbits, which are less than 10 km away from each station. The CloudSat vertical bins are relative to the geoid, and depending on the altitude where the stations are located, the first exploitable bin (out of the ground clutter alteration altitude) varies significantly. Moreover, in locations where the ice does not interfere much with the radar signal (ocean and some coastal areas), the ground clutter layer is thinner and lower altitude bins can be used. We are using at the DDU station CloudSat profiles from the fourth bin, which is located at 961 m a.g.l. At the PE station the first exploitable bin is the fifth, which is located at 1043 m a.g.l. We use the 2C-SNOW-PROFILE product which retrieves profiles of liquid-equivalent snowfall rates. The product is based on assumptions on snow particle size distribution, microphysical and scattering properties which induce many uncertainties in the calculation of the relationship between radar reflectivity and snowfall rate (see Sect. 2.2).

The MRR is a vertically profiling Doppler radar operating at a frequency of 24.3 GHz (K band) with a beamwidth of 2∘ (around 50 m in diameter at a 3000 m altitude). At both stations, the resolution was set to 100 m per bin, ranging from 300 m – for the first valid available measurements – to 3000 m. However, we only consider the data up to 2500 m because of the change in the snow microphysical properties above this altitude . The MRR's raw measurement – Doppler spectral densities – is available at a 10 s temporal resolution. The collected data were processed using the IMProTool developed by . At the DDU station, the radar reflectivity derived from MRR was calibrated by comparison with a colocated X-band polarimetric radar over the period from December 2015 to January 2016 (for more details, see ). Through this calibration with the second radar, the reflectivity (at X band) is converted into snowfall rates using a Ze/Sr relation : Ze=76×Sr0.91, with Ze the radar reflectivity (in dBZ), and Sr the snowfall rate (in mm h-1). proposed a range of values of 69–83 for the prefactor and 0.78–1.09 for the exponent corresponding to a confidence interval of 95 %.

For the instrument operating at the PE station, hereafter called MRR2, the average Ze/Sr relation is given by : Ze=18×Sr1.10. The range of prefactor, 11–43, and exponent, 0.97–1.17, for this equation spans a confidence interval of 40 % due to the summation of uncertainties in particle size, shape, measurement and conversion from reflectivity Ze to snowfall rate Sr. For this study, the MRR2 data used are processed with the algorithm. Unlike , we did not calibrate the ground radar dataset with CloudSat reflectivities (1) because we want an independent evaluation of the CloudSat CPR dataset and (2) because we do not consider surface precipitation rate comparisons. The mean precipitation profiles obtained over the MRR observation periods (2015–2016 for the DDU station and 2012 for the PE station) were also used to evaluate how typical the four precipitation events are . They are obtained using the same Ze/Sr relationships as the ones introduced earlier (see Eqs. 1 and 2) and are separated into quantiles. According to , the minimum detection of both MRRs varies between -14 and -8 dBZ, corresponding to 0.00122–0.00546 mm h-1 at the DDU station and 0.00385–0.0135 mm h-1 at the PE station. However, these values correspond to theoretical cases of clear sky. Therefore we analyzed the density probability functions of the MRR1 (DDU station MRR) at three different levels to determine a minimum threshold of detectability of ground radars (Fig.  in Appendix). We used the lowest level out of the ground clutter layer (about 1200 m a.g.l.) and selected a threshold of 0.005 mm h-1 (see the vertical dashed line in Fig.  in Appendix).

A radiosonde is a meteorological device containing a set of sensors to measure the characteristics of the atmosphere from ground level to an altitude ranging from 25 up to 30 km. Parameters measured are temperature, relative humidity, wind speed, wind direction and pressure.

At the DDU station, the radiosonde system used is a METEOMODEM M10. The relative humidity accuracy is 3 % and its temporal resolution is 2 s. The temperature measurement is realized every 1 s with an accuracy of 0.3 ∘C. At the PE station, the ground receiving systems used are GRAW-GS-E and GRAW radiosondes DFM-09-QRE. Relative humidity is measured with an accuracy of 3 % and a temporal resolution of 4 s. The accuracy and the temporal resolution of the temperature measurements are 0.2 ∘C and 3–4 s.

We summarize in Table  the characteristics of the four recorded precipitation cases when both CloudSat and ground-based MRRs simultaneously record a snowfall event and when the satellite is in the vicinity of the stations. Due to the CloudSat delay of revisit, satellite overflights near the DDU station are located either less than 10 km or more than 80 km away. CloudSat tracks passing through a radius of 10 km around each station (Fig. ) were selected. Each CloudSat flyby over a station takes less than 10 s and covers a distance between 11.90 and 17.33 km. We consider that the four associated weather systems are static with regards to the CloudSat satellite overfly. However, MRRs are stationary and local precipitation patterns are typically associated with transient large-scale and mesoscale weather systems. We therefore analyzed the synoptic conditions by using radiosonde data and reanalysis (ERA-Interim) from the European Centre for Medium-Range Weather Forecasts (ECMWF) in order to determine the adequate MRR time series corresponding to CloudSat observations. We estimated a duration for which MRR observing conditions agree most with those of CloudSat using the following equation: Δtavg=ΔxsatVwind, where Δtavg represents the temporal range of the MRR observations including the CloudSat overflight dates, Δxsat is the length of the track inside the 10 km radius area over stations and Vwind is the vertically averaged wind velocity. All characteristics are shown in Table .

All CloudSat measurements were selected within a 10 km radius from each station and averaged for each vertical bin. A variance in the CloudSat retrievals is computed for the duration of each overpass (see Fig.  in Appendix).

The MRR confidence intervals are calculated using the range of Ze/Sr parameters given by for the Dumont d'Urville station and for the Princess Elisabeth station. At the DDU station, according to , for an altitude higher than 2500 m where there is a crystal dominance for precipitation, the parameterizations used for Ze/Sr conversion are not adapted anymore. That is why MRR measurements are considered and compared to equivalent CloudSat vertical bins only in the first 2500 m of the atmosphere. In contrast with the coastal areas, we would expect less riming at the PE station compared to the DDU station, while aggregates are expected to occur at the PE station given the measured large particle sizes . Also the low variability in the vertical profile of mean Doppler vertical velocity at the PE station suggests that aggregation and/or riming of particles is not frequent in this region and hydrometeor type is relatively constant in the vertical profile . Without this change in the proportion of the different hydrometeors, the ground-based Ze/Sr relationships would be still valid higher up.

Focusing on the Dumont d'Urville station, Fig. a shows a good agreement between CloudSat and the MRR's snowfall rates for each vertical level. Indeed, an averaged satellite precipitation rate at all levels is included within the 95 % MRR confidence interval. The MRR profile presents a maximum of the snowfall rate of 0.75 mm h-1 at 750 m and an inversion of the precipitation rate likely due to low-level sublimation processes, whereas the ground clutter prevents CloudSat from seeing the inversion. This precipitation event is likely generated by the passage of the second low-pressure system, as described previously using the corresponding radiosounding. According to , this precipitation event is representative of the climatology of the DDU station as it lies between the 20th and 80th quantiles (indicated by grey dashed line) with a shape similar to the average climatology in solid black line.

According to Fig. b, there is a poor concordance between the two datasets for low snowfall rate values. The MRR recorded low-level strong values until a null signal of precipitation from 1000 m upward, where CloudSat still recorded small but significant rates. An inversion of the precipitation rate at low levels is also observed under the maximum precipitation rate of 1 mm h-1 at 600 m. The strong gradient of this inversion is likely due to katabatic wind effects, which can drastically dry out atmospheric layers when blowing down from the ice cap. This event shows that the use of CloudSat for surface precipitation determination may be problematic in certain conditions for a specific event. It is also important to note that this event is an anomalous climatological event at the DDU station, in comparison with the quantiles of the vertical structure of precipitation both in terms of snowfall rate and shape.

Figure c shows a good agreement between the four lowest values of CloudSat observations and the MRR profile. Indeed, every averaged satellite measurement is included in the 40 % confidence interval, but the standard deviations indicate a large dispersion. Above this altitude precipitation rate is small and the agreement is weaker. This is similar to what is observed in Fig. b. CloudSat observes again a small signal of precipitation where MRR recorded a null snowfall rate, suggesting some limitations in the sensitivity or attenuation of the MRRs but also a satellite sensitivity for low snowfall rates. This event is an important anomalous climatological event at the PE station because the observed snowfall rates are much higher than the snowfall rates of climatology. This is caused by the passage of an atmospheric river over the station.

Figure d snowfall rates observed by both CloudSat and MRRs are quite low compared to the three other cases but the agreement remains good for the five lower satellite levels. According to , this precipitation event is representative of the climatology of the PE station with in particular the presence of virga with very low precipitation rates included between the high and low quantiles.

Figure  represents the correlation for (all data, all levels) CloudSat and MRR precipitation reports for the four events using the error bars shown in Fig. . Error bars for the MRRs are implemented by using the confidence intervals obtained with the Ze/Sr relations. Large error bars correspond to the PE station's MRR and smaller ones represent the DDU stations's MRR confidence interval. CloudSat error bars represent the variance of measurements collected along the swath. A linear regression fit between CloudSat and MRRs is performed and shows a good correlation between both datasets.

A previous study by showed that CloudSat-measured ice cloud reflectivity is 1 dB higher than an airborne cloud radar, and following a statistical evaluation with basic cloud properties and five ground-based sites a weighted-mean difference in Ze, which ranges from -0.4 dBZ to +0.3 dBZ, is observed when a period of ±1 h around the CloudSat overpass is considered. According to , CloudSat tends to observe lighter snowfall events (smaller than 2 mm h-1) in comparison with the NOAA National Severe Storms Laboratory (NOAA/NSSL) multi-radar multi-sensor (MRMS/Q3).

Figure b shows that CloudSat can report small but significant snowfall when the MRR signal is virtually zero. The shift between the two instruments is estimated in this case at +0.040±0.005 mm h-1. Looking at Fig. c for the three last CloudSat bins above 2 km height, an averaged snowfall rate of +0.033±0.003 mm h-1 is observed when the MRR signal at the PE station is null. Concerning Fig. d, a similar value of +0.030±0.001 mm h-1 is recorded by CloudSat, but this time MRR also records a similar signal of +0.029±0.008 mm h-1. This difference in measured values suggests a difference in sensitivity of the two radars even if these measured rates are above the MRR detection limit of 0.005 mm h-1 (see Sect. 2.2). This shift in snowfall rates could either be due to a strong attenuation of the MRR backscattered signal with the altitude or due to the detection of cloud water by the CPR as it is more sensitive to small atmospheric particles and clouds.

The CloudSat 2C-SNOW-PROFILE product already contains its own uncertainties estimates, calculated from hypothetical parameters such as the mass–diameter distribution of the hydrometeors, their microphysical and scattering properties. Our analysis suggests that under Antarctic (and probably polar) conditions, this uncertainty can be significantly reduced. By assuming that CloudSat and MRR snowfall rates datasets follow a normally distributed deviation from the mean, a correlation coefficient is calculated in order to establish the degree of similarity between both observations. By using the covariance of both data records, we found a correlation coefficient of 0.99, which confirms a very good agreement between both radar data (see Appendix).

For each CloudSat vertical bin, we calculated the distance of satellite measurement to the corresponding interpolated MRR observation. We averaged these values by weighting them with the MRR confidence intervals and we found a range of CloudSat uncertainties from -13 % up to +22 %.