Measurements of precipitation in Dumont d ’ Urville , Terre Adélie , East Antarctica

The first results of a campaign of intensive observation of precipitation in Dumont d’Urville, Antarctica, are presented. Several instruments collected data from October 2015, including a polarimetric weather radar (MXPol), a Micro Rain Radar (MRR), a weighing gauge (Pluvio), and a Multi-Angle Snowflake Camera (MASC). These instruments collected the first model-free measurements of precipitation in the region of Terre Adélie (Adélie Land), including of precipitation microphysics. Microphysical observations during the austral summer 2015/2016 showed that, close to ground level, aggregates are 5 the dominant hydrometeor type, together with small ice particles (mostly originating from blowing snow), and that riming often occurs. Contamination of the Pluvio measurements in windy conditions is observed and partly removed through synergistic use of MRR data. The yearly accumulated precipitation of snow (300m above ground), obtained by means of a local conversion relation of MRR data, trained on the Pluvio measurement of the summer period, is estimated to be 815mm of water equivalent, with a confidence interval ranging between 739.5 to 989mm. Climatological data obtained from satellite-borne radars, and the 10 ERA-Interim reanalysis of the European Center for Medium -Range Weather Forecasts (ECMWF) both provide lower yearly totals: 655mm for ERA-Interim, while 679mm for the climatological data over DDU. ERA-Interim seems to overestimate the occurrence of low-intensity precipitation events especially in summer, while visual observations conducted at the research stations all year long seem to underestimate it. Overall, this manuscript provides insightful examples of the added values of precipitation monitoring in Antarctica with a synergistic use of in-situ and remote sensing measurements. 15


Introduction
The ice sheets of Antarctica contain about 90% of the world's ice and thus its evolution has potential impacts at a global scale.
It conditions the evolution of the sea level height (Rignot et al., 2011;DeConto and Pollard, 2016), and the radiative budget of the lower atmosphere.In this context, the quantification and prediction of the surface mass balance (SMB) of the Antarctic ice cap is a pressing scientific topic of investigation in order to understand if the continent is losing or gaining ice, and at what rate (Vaughan et al., 1999;Lenaerts et al., 2016).
Precipitation is an important component of the SMB as it represents, together with vapour deposition, the only net input of water and ice at the continental scale (Krinner et al., 2007).Precipitation is unfortunately also very difficult to monitor at high latitudes.The major problems hampering classical measurement techniques in Antarctica are: in the interior, the sparsity of human installations over a very large area, the extremely low temperatures and low precipitation amounts, and on the coasts the very strong katabatic winds blowing from the interior.Additionally, the complex logistics of Antarctic installations causes further difficulties and limitations for measurements to be conducted.
Until recently, information about precipitation was obtained indirectly by analysing moisture transports, glaciological surfacebased observations (Bromwich, 1990) and reanalysis based on numerical weather prediction models (Bromwich et al., 2011).
Additionally, long-running but qualitative human observation records of clouds and precipitation have been collected in some scientific stations with staff dedicated to meteorological measurements (e.g.König-Langlo et al., 1998).Recent research proposed a climatology of precipitation over a large part of the continent (Palerme et al., 2014(Palerme et al., , 2016) ) by exploiting the potential of the profiling radar on-board the CloudSat satellite, which is able to sample large horizontal areas but limited by the inability to measure precipitation at altitudes below a so-called "blind-range" above ground (1200 m above the surface for CloudSat).
In order to validate and to improve the performance of the models, and to constrain satellite-based measurements, it is necessary to establish and maintain in the medium to long term some in-situ observation sites, instrumented with precipitation measurement devices as autonomous and accurate as possible.There is therefore the need for accurate measurements of precipitation, including at the very local scale (Frezzotti et al., 2004;Schlosser et al., 2010;Welker et al., 2014).A recent effort in this direction was the establishment of an observatory in the escarpment zone of Dronning Maud Land, East Antarctica (Gorodetskaya et al., 2015).The synergy of in-situ and remote-sensing measurements allowed very first statistics of cloud and precipitation (Gorodetskaya et al., 2014(Gorodetskaya et al., , 2015) ) which showed that a few intense precipitation events govern the SMB in the area; measurement combinations have also been used to evaluate the quality of satellite-based precipitation products (Maahn et al., 2014) provided by CloudSat.It has been shown that the blind range of CloudSat, in the area of the measurements, can lead to an underestimation of precipitation amount on the order of 10% and an underestimation of the occurrence frequency on the order of 5%.The installation in DML can be considered the first well documented observatory in Antarctica that included precipitation measurements from remote sensing and in-situ instruments.An earlier effort involved co-located measurements of precipitation using radar and precipitation gauge, and was conducted at the Showa1 Japanese station (Konishi et al., 1998), but very limited information about the outcome of those measurements is yet available in the literature.A more recent effort is currently taking place in the McMurdo base, in the framework of the AWARE project Witze (2016), starting from November

2015.
In this work we present the results of an intensive observation campaign during the austral summer 2015-2016 and a first year of precipitation measurements conducted in the French base Dumont d'Urville, Terre Adélie, from November 2015, until November 2016 (and still ongoing).The data were collected in the framework of the APRES3 project (Antarctic Precipitation, Remote Sensing from Surface and Space, see http://apres3.osug.fr).We provide statistics of precipitation quantity and occurrence, and we compare them with model reanalyses and with the visual observations collected by the French meteorological office (Météo France) all year long.The main scientific objectives of this work are to contribute to a better quantification of precipitation in Antarctica (also by evaluating the products of numerical weather models) and to underline the innovative and promising aspects of the data collected until now, that may serve as an example for long-term monitoring of precipitation in other Antarctic regions.The paper is structured as follows: Sec. 2 describes the precipitation measurements, Sec. 3 lists the most relevant results, which are discussed and put into perspectives in Sec.4.Section 5 provides the summary and conclusions of the paper.

Methods
We present here data collected in a coastal location of Antarctica: the station Dumont d'Urville (DDU).The base is situated in Terre Adélie, −66.6628S, 140.0014E, (41 m above sea level), on a coastal location highlighted in Fig. 1.This region is located at the transition between the Antarctic continent and the Southern Ocean, where the terrain, which slopes downward from the inner continent to the coast, meets the ocean.

Climate and operational measurements
The climate at DDU is relatively mild in terms of temperatures, with minima rarely below -30 • C, and maxima above 0 • C in January and December, as illustrated in Fig. 1 (b).On the contrary, the wind regime is more extreme: in the low layer of the atmosphere the dominant winds are katabatic coming from the inner continent, and the dominant wind origins are always between 90 • (East) and 180 • (South), as illustrated in the wind rose of Fig. 1 (b).Because of the intensity and persistence of the winds, which are able to reach hurricane force, Terre Adélie has been often described as the windiest place on planet Earth (e.g., Wendler et al., 1997).Standard measurements of atmospheric variables (temperature, wind speed, wind direction, relative and specific humidity, atmospheric pressure) are collected regularly all year long by the French meteorological service (Météo France), and a balloon radiosounding is launched daily at 00UTC.Balloon soundings have been regularly conducted since 1956 at DDU. Visual observations of cloud, precipitation, and present weather are recorded as well.It is worth noting that at this location the visual observations are the only daily in-situ archive of past precipitation occurrence in DDU.

APRES3 Instruments
Several instruments were deployed at DDU, starting from November 2015.The instruments were deployed as illustrated in Fig. 2, and they are listed in Table 1.A Micro Rain Radar (MRR) was installed within an existing radome and has collected uninterrupted measurements since the 22 November 2015.This radar system is used to vertically profile precipitation with a resolution of 100 m at height levels ranging from 341 m to 3141 m above sea level2 .The processed data were collected with a temporal resolution of one minute.The potential of the MRR to monitor polar regions has already been highlighted by the works of Maahn et al. (2014) and Gorodetskaya et al. (2015).The simplicity of its deployment and operation makes it an attractive tool for long-term measurements in places with complex logistics and with limited possibility of support, in the case of instrumental failures.The raw K-band reflectivity measurements collected by the MRR were first processed with the method proposed by Maahn and Kollias (2012), then converted to X-band reflectivities and in a third step to snowfall 5 intensities.Additional information about the processing of the MRR data is provided in Sec.2.2.1.
A second radar, named MXPol (Mobile X-band dual-Polarization) collected measurements in the months of December 2015 and January 2016.This system, described in Schneebeli et al. (2013) and in Scipion et al. (2013), is a scanning dualpolarization Doppler radar.During its operation period at DDU, it was mainly collecting data at 75 m radial resolution and a maximum radial distance of 30 km, mostly conducting different types of scans within a repeating scanning sequence of 5 10 minutes: (i) Plan Position Indicator (PPI) scans, i.e. quasi-horizontal slices of the atmosphere, (ii) Range Height Indicator (RHI) scans, i.e. vertical slices of the atmosphere, and (iii) static vertical profiles, as the ones performed by the MRR.A depolarization lidar (e.g.Del Guasta et al., 1993), deployed at a distance of about 200 m from MXPol, collected data in December 2015 and January 2016, as a test-bed for future long-term installation of a similar device.Lidar measurements allow for discrimination of the phase of the tropospheric clouds and detection of the occurrence of supercooled liquid water, and they complement the observations of ground-based radars, that are often not sensitive to these particles.An example is given in Fig. 3, where the timeseries of MRR reflectivity, lidar signal and depolarization ratio are shown for the 15 th December 2015.
Supercooled liquid water appears in the lidar data as a layer of enhanced signal and low depolarization ratio (e.g.Del Guasta et al., 1993;Hogan et al., 2003), often when no MRR signal is visible.On the contrary, when precipitation occurs, (around 04 UTC, and from 14 to 24 UTC) the lidar signal gets fully attenuated in the lowest 500 m while the MRR is still able to sample the vertical precipitation column.
A weighing precipitation gauge (Pluvio 2 , manufactured by OTT) was deployed from November 2015 to January 2016.
This instrument provides the liquid water equivalent of snowfall falling within its measurement area at a time resolution of one minute.To avoid excessive contamination of precipitation signals by blowing snow, the Pluvio 2 was installed at a height of about 3 m above ground and its inlet was protected with a standard wind fence designed by the same manufacturer as the instrument.It must be noted that this wind shield is not sufficient to avoid the adverse effect of strong wind (frequently occurring at DDU).
Located close to the weighing gauge, a multi-angle snowflake camera (MASC) was deployed, also during the period from November 2015 to January 2016.This instrument collects high-resolution stereoscopic photographs of snowflakes in free fall, while they cross its sampling area (Garrett et al., 2012), thus providing information about snowfall microphysics and particle fall velocity.The MASC was using three identical 2448 x 2048 pixels cameras (with common focal point) with apertures and exposure times adjusted to trade off between the contrast on snowflakes photographs and motion blur effects, and a resolution of about 33 µm per pixel.The cameras are triggered when a falling particle crosses two series of near-infrared sensors.A detailed description of the system and its calibration can be found in Garrett et al. (2012),and Praz et al. (2017).To complete the set of in-situ measurements, a weather station (Vaisala Weather Transmitter WXT 520) was installed close to the Pluvio 2 and the MASC, to sample the environmental conditions in the close proximity of their measurements, as illustrated in Fig. 2.

Pre-processing of MRR data 5
The MRR was co-located with MXPol for the period of the summer campaign 2015/2016 (See Table 1).The purpose of the former instrument at DDU is long-term monitoring, which involves exposure to the extremely windy winter conditions.It was decided, in order to avoid failures during the winter when no member of the scientific team is on-site, to install the MRR inside an existing radome previously used in the base for satellite communications, as shown in Fig. 2.Although this installation ensures protection and easy access to the instrument, it adds an unknown amount of attenuation to the measurements.For  b : For the rest of the time this instrument was (and is) deployed on the Antarctic continent, about 5 km away from DDU.
c : The co-located weather station of Météo France is providing data all year long, uninterruptedly.
is close to linear (ρ 2 ≈ 0.88), and almost equivalent to a simple offset subtraction, we can hypothesize that eventual non-Rayleigh effect, due to cm-size snowflakes, were similar at the two frequencies and the following conversion has been applied to MRR data: where Z X (Z K ) [dBZ] is used to indicate reflectivity at X-(K-) band, and is the measure of uncertainty of the linear relation with respect to the scatterplot of Fig. 4 (whose standard deviation of the residuals is 1.9 dB).It is worth mentioning that Z K is originally obtained with the method of Maahn and Kollias (2012), who proposed an improved and innovative processing chain for MRR data collected in snow.Once mapped to X-band, reflectivity can be converted to snowfall rate S rate by means of Z-S power laws available in the literature.For example, the six relations proposed by Matrosov et al. (2009), and listed in Table 2 can be used.These relations were obtained by combining two different snowflake size distribution datasets, and three different mass-to-size relations.The error component of Eq. 1, and the large variability of the Z-S relations, lead to very uncertain retrieval of snowfall rate.For this reason, we optimized a local power law, by fitting its two parameters in the Z-S space given by the MRR measurements at the lowest available height and the Pluvio 2 measurements collected close to the ground, during the Summer period 2015/2016.The parameters (intercept and exponent) of the power law are obtained by means of nonlinear least square estimation.The local relation, also listed in Table 2, takes the form of Z = 76 S 0.91 .In order to mitigate the difference in sampling volume of the two instruments, it has been derived for hourly data.The 95% confidence intervals for the two parameter are 69-83 (prefactor) and 0.78-1.09(exponent) respectively.
Local-DDU Z = 76 S 0.91 * : In parentheses, the way the relations were numbered in Matrosov et al. (2009).

Pre-processing of Pluvio 2 data
It has been observed that occasionally the values of equivalent water of the Pluvio 2 show a "phantom" accumulation (similar to that reported by World Meteorological Organization, 2014).In such cases, no precipitation was observed by the researchers that were present on site and no precipitation signal was visible in the MRR data but the content of the Pluvio 2 bucket was increasing.In order to discard these cases, we combined the information coming from remote sensing (MRR) and in-situ data (Pluvio 2 ).More precisely, time steps when no signal was recorded by the MRR at its lowest available gate (300 m above ground level), are considered "precipitation free" and any increase in the cumulative precipitation records of the Pluvio 2 is thus related to external contaminations.The assumption is that precipitation is extremely unlikely to completely develop in the lowest 300 m of the atmosphere.An example of the behaviour of this simple censoring filter can be found in Fig. 5 (a).From    this case due to clear-sky blowing snow, and we hypothesize that it is caused by wind-induced vibrations of the instrument.It must be noted that this simple pre-treatment cannot compensate the contribution of snowfall mixed with blowing snow, when the positive contribution of blowing snow and precipitation, and the negative contribution due to wind-induced loss of catching efficiency occur together.

Additional data
Due to the lack of both short-and long-term precipitation measurements, net precipitation estimates in Antarctica have been obtained from numerical weather prediction models (e.g., Cullather et al., 1998;Schlosser et al., 2010).Among the available model-based products, the ERA-Interim global reanalysis provided by the European Center for Medium-Range Weather Forecasts (ECMWF) is taken as a reference as it is considered to provide the best representation of precipitation variability (Bromwich et al., 2011;Palerme et al., 2014), and the best agreement with satellite-borne measurements (Behrangi et al., 2016;Palerme et al., 2016).ERA-Interim reanalysis is used here for this reason, and because of its global coverage and easy access.The analyses at 00 UTC and 12 UTC, and forecast time steps of 6, and 12 h are used in the present work for the grid point which is the closest to DDU.The spatial resolution of ERA-Interim is 0.75 • ×0.75 • .To quantify precipitation, the model variable "tp" (total precipitation) is used here.

Microphysical observations during Summer 2015/2016
The period between November 2015 and January 2016 was heavily instrumented with devices that are able to provide microphysical information about precipitation, thus microphysical aspects are better documented during the summer months.While a complete investigation of the dominant microphysical processes and the small-scale dynamics of precipitation in this region is beyond the scope of this paper, it is worth to investigate an important microphysical parameter: the hydrometeor type.
Hydrometeor types have been recorded near the ground level (about 2.5 m above ground) by the MASC instrument through classification of individual particle pictures with the recently developed method of Praz et al. (2017), able to classify individual hydrometeors into six classes (and melting snow), and to assign to them a continuous riming degree index ranging from 0 to 1, with 1 corresponding to fully developed graupel.The riming degree is a textural information obtained by supervised classification originating from a manually-labelled training set including almost 3400 images, as detailed in Praz et al. (2017).
The choice of the available classes is based on the widely used scheme of Magono and Lee (1966).Because the instruments are deployed at a height lower than 3 m above ground level, both precipitation and blowing snow particles are recorded and classified.
A second classification method is obtained from the polarimetric data of MXPol, that can be converted into hydrometeor measurements with an hydrometeor classification algorithm (Grazioli et al., 2015).This algorithm was developed by partitioning a large number of radar observations into spatially coherent clusters by means of data mining techniques, and then to assign to each cluster a dominant hydrometeor type by means of scattering simulations, interpretation of polarimetric signatures, and comparison with in-situ data.Despite the drawbacks of being an indirect method and not being able to retrieve at near-ground  (2015).Noise in the classification at the lowest elevation angles is due to ground clutter.Range gates closer than 2 km with respect to the radar location have been censored to allow reliable polarimetric variables to be computed.Elevation angles larger than 45 • have been censored as well in order to limit the geometric reduction of the intensity of polarimetric signature with increasing elevation angles (Ryzhkov et al., 2005).
heights (because of ground clutter contamination in the radar data), it has the advantage of providing hydrometeor types over large domains and at different height levels, as shown in Figure 6. Figure 6 illustrates PPI and RHI scans of the hydrometeor classification for a case where all its ice-phase hydrometeor classes are observed.This classification method discriminates pure snowfall into three categories: crystals, aggregates, and rimed particles.The classification obtained with the MASC and the method of Praz et al. ( 2017) is summarized in Figure 8.At ground level, the majority of the particles (54%) are classified as "small", indicating hydrometeors being too small for their geometry and texture to be properly captured by the MASC.This proportion is three times higher than similar measurements collected in a wind sheltered location in the Swiss Alps, while the proportion of the other hydrometeors is similar among the two different locations (not shown here).The occurrence of strong katabatic winds being a major difference between the sites, it can be assumed that the large majority of these "small particles" observed at DDU is associated with blowing snow.During blowing snow events with strong winds (identified from visual and MRR observations on site), the number of images collected by the MASC is very large.The majority of those being classified as "small particles", this results in a large percentage of this hydrometeor type in the final statistics.Also from this classification based on the MASC we observe that riming occurs.in fact, 11% of the particles are fully rimed (graupel), all the other hydrometeor types, have a riming degree ranging mostly from 0.1 to 0.5, and sometimes larger than 0.5 for the aggregates.While the outcomes of the classification from MXPol and the MASC are not directly comparable because of the differences in measurement height, sampling volume, and available classes, it must be underlined that radar measurements are very sensi-tive to the size of the hydrometeors.Thus, a few large aggregates within a radar sampling volume, will dominate and overcome the signal coming from smaller hydrometeors.This can partially explain the different proportion of aggregates observed by MXPol (about 40% at a 400 m height), and by the MASC (19%).A second contribution to this difference may be low-level mechanical breakup of the aggregates (e.g.Vardiman, 1978).A third, and very likely, contribution is the contamination of blowing snow in the MASC measurements, namely in the "small" hydrometeor class.If, assuming that most of the "small" particles originate from blowing snow, they are removed from the statistics, then aggregates account for 41% of the hydrometeors, a value much closer to the 40% obtained with the classification of MXPol.

One year of MRR precipitation data
The MRR instrument collected precipitation data uninterruptedly, covering the evolution of precipitation over the entire year.
It therefore offers an interesting ground-based (but remotely sensed) set of data to compare with model-based data and with available human observations.Figure 9 shows the estimates coming from the MRR and other available sources of information over a year of measurements.As expected, the agreement of the local MRR relation with the Pluvio 2 is good over the summer period (Dec-January), during which the relation was obtained.In this period also the estimate of ERA-Interim provides a total cumulated precipitation within the envelope of values of the optimized Z-S relation, even though the curves shows some differences in precipitation occurrence.The optimized Z-S relation provides estimates that are close to the B90A relation of Matrosov et al. (2009).
The months with the highest accumulated precipitation were the late fall and winter months of May and June, and the month of September.Seasonally3 , Summer was the driest season, contributing only 11% of the yearly total, compared to values close to 30% for spring, 34% for fall, and 25% for winter (Table 3).The ERA-Interim totals of each month of the comparison period are within what could be observed in the period 1995-2015, with the exception of September that was the snowiest since 1995.

Precipitation occurrence
Long term precipitation data records in Antarctica are often only visual observations of precipitation occurrence.For this reason, comparing precipitation occurrence measurements is a way to better understand the quality of this source of information.
For the year 2015-2016, we can compare in terms of occurrence the information coming from ERA-Interim, Pluvio 2 , MRR, and the visual observations archived by Météo France.We deal at first with occurrence at the daily scale, and we define it for the MRR and ERA-Interim as precipitation exceeding a given threshold over a given duration.A threshold of 0.07 mm over 6h was proposed by Palerme et al. (2014), and we thus take a value of 0.28 mmd −1 as a first guess.However, the choice of a unique threshold is delicate, and we apply also a minimum (maximum) threshold of 0.001 mmd −1 (1 mmd −1 ) to cover any value that appears reasonable to assume.Figure 10 shows the number of days with precipitation recorded during each month of the measurement period.As a past reference, also shown is the historical record of precipitation occurrences from visual daily observations for the preceding years (1981 to 2015) in green with variability range.The year under investigation had an extremely dry January, and an extremely snowy September (in term of occurrence), while the other months are within the range of past occurrences. 10 Overall, ERA-Interim mostly overestimates precipitation occurrence with respect to the MRR, especially in Summer, while the visual observations underestimate it.For January and December, when the Pluvio 2 was in operation, it is in agreement with the MRR.Given the measurement correction principle based on false detection described in Sec.2.2.2, this implies that no miss-detection is evident.
A good example of the overestimation of occurrence by ERA-Interim is shown in Fig. 9, bottom.The period between the 15 10 th and 25 th of January is seen as dry by the MRR and Pluvio 2 , while several low-intensity precipitation events appear in the ERA-Interim time series.The overestimation of occurrence compensates the underestimation of the most intense snowfall events, such that at the end of January the total accumulated precipitation of ERA-interim gets close to the one of the Pluvio 2 .
As a result, the contribution of lower snowfall rates to the total accumulated snowfall is much larger for ERA-Interim, with respect to the measurements collected by the MRR and Pluvio 2 , and this difference is particularly pronounced in the summer 20 period (as shown in Fig. 11).
Figure 12 shows, at a 6 h time scale (here we consider quantitative precipitation, thus we focus on a higher temporal resolution), the evolution of precipitation occurrence as a function of a given average precipitation intensity threshold, for the full Visual observations provided by by Météo France are not limited to precipitation, and several present weather codes are archived.At DDU, those are SYNOP codes belonging to the group 7wwW1W2.In the codes recurring at DDU, three types of phenomena are mostly documented: clouds (codes 1-3), blowing snow (codes 36-39), and snow (codes 22 and 71).We consider here the codes related to snow and blowing snow and we disregard the observations of clouds.The observations are conducted during each day, on average every 5h (with higher frequency during day hours), and we compare them here with 5 the MRR measurements.This is shown in Table 4, where the MRR observations at 300 m height are compared with the visual observations of snow and blowing snow.Given the intrinsic difference between those observations, it is not possible to take one as an overall reference in the confusion matrix.However, it can be assumed that visual observations are better at reporting blowing snow, because they are conducted at the ground level, while MRR measurements at 300 m above ground are better at reporting snowfall.The most interesting outcomes of this comparison are the following ones.First, there is a good correspondence between occurrence of snow according to the MRR and visual observations of snow.Second, blowing snow occurrence is not well captured by the MRR.When visual observations report blowing snow, the MRR mostly does not report any occurrence.44.4% of the visual reports analysed here correspond to cases where blowing snow has been observed at the ground level, but no valid signal has been recorded by the MRR.This comparison is to a certain degree dependent on the threshold used to discriminate "dry" and "precipitation" in the MRR data.However, similar results have been obtained for various threshold levels (not shown here).
This result, as discussed in the next section, is not unexpected because blowing snow rarely exceeds heights of 100 m above terrain (Palm et al., 2011).

Microphysical observations
The microphysical observations, collected during the austral summer 2015-2016, and illustrated in Figs.7 and 8 suggest that even at this location on the Antarctic coasts, riming is an important microphysical process.From radar retrievals, close to the ground level, about 10% of precipitation is rimed.According to the MASC classification, 11% of the hydrometeors are fully rimed (graupel) and most of the other hydrometeor types have a degree of riming greater than 0, in particular aggregates that tend to be larger and easier targets for riming (e.g.Houze and Medina, 2005).The presence of riming indirectly shows that mixed-phase clouds are often occurring and that supercooled liquid water is available in the regions of precipitation formation.
This has been documented in the past at this location by Del Guasta et al. (1993), and it could be observed also in the test data collected with the depolarization lidar (see Table 1, and Fig. 3) during the summer period.
At the ground level the large majority of hydrometeors recorded by the MASC were small particles of non-discernible habit and non-definable riming degree (given the MASC resolution), with an occurrence three times higher than similar measurements conducted in Alpine locations.This is probably the signature of the significant contribution of blowing snow to the near-ground snow flux, that is particularly effective in recirculating small and light particles (e.g.Mann et al., 2000;Gordon et al., 2009), but it could, in a minor part, be the result of the fragmentation of aggregates in the low-level of the atmosphere, where strong katabatic winds blow.

Blowing snow and wind effect
The contribution of blowing snow was visible in the observations collected with the MASC, generating very large number of small hydrometeors, and strong winds were affecting Pluvio 2 measurements, generating phantom accumulations probably due to the vibrations of the weighing gauge.As illustrated in Fig. 5, periods with wind speeds exceeding roughly 15 ms −1 at the proximity of the inlet of the instrument generate phantom accumulations of precipitation that are removed if the MRR does not receive any signal at the same time at its lowest gate.The lowest gate of the MRR (300 m in this case) is considered to be high enough to be above the height of any wind-blown snow layer (Gordon et al., 2009;Scarchilli et al., 2010;Palm et al., 2011), which rarely exceeds 200 m of vertical development.
This combination of ground-based and in-situ instruments should be proposed again, and maintained for at least a full year to cover all the seasons.The main limitation of the filter of Pluvio 2 data is that it cannot detect cases when blowing snow occurs together with precipitation.In the case of our measurements, the total accumulated precipitation of the Pluvio 2 in the summer period drops from 66 mm to 52 mm, after the censoring is performed.The removed portion is then 21 % of the total raw accumulated precipitation.
Regarding blowing snow, the comparison between the occurrence of signal at the lowest MRR gate and the visual weather reports conducted by Météo France, summarized in Table 4, confirms that the current MRR configuration does not capture the occurrence of blowing snow events.Lower range gate spacing should be employed (lower than 100 m used in the measurements shown here) if blowing snow is of interest.

Precipitation quantification
The quantification of precipitation in the coastal regions of Antarctica remains a difficult task, affected by significant uncertainty.This study provides some estimates that help contextualize the information available until now.Fig. 9 and the summary in Table 3 shows that MRR estimates of total accumulated precipitation at the yearly scale can diverge significantly (from 484 to 1581 mm) if a range of standard Z-S relations is used, while the use of a local Z-S relation, calibrated in the Summer season, allows for a significant reduction of this range of values (from 740 to 989 mm).In this case, however, an important assumption is made: i.e. that this relation can be considered representative for the other seasons.
ERA-Interim provides a yearly estimate of 655 mm for the measurement period from November 2015 to November 2016, about 10% lower than the lowest estimate obtained with the local Z-S relation.It must be underlined once more that the estimates provided by the MRR at DDU correspond to a minimum height of 300 (±50 m above ground corresponding to a range gate spacing of 100 m).This, together with the large grid of ERA-Interim, may contribute to the differences observed with the MRR.Interestingly, ERA-Interim initially is in agreement with the MRR precipitation values until March 2016, while degree of riming in the present case.The synergy between remote-sensing and in-situ instruments has the potential to improve the quantification of snowfall amounts in conditions where strong winds affect ground based measurements, even though much remains to be done in cases when precipitation and blowing snow at the same time.The installation and long-term operation of a similar combination of instruments should be conducted again, at DDU and at other locations in Antarctica.Efforts will be devoted to develop a better long-term constraint for radar-based snowfall estimations by means of in-situ measurements of precipitation in synergy with microphysical observations and retrievals, because the relation used in this study was built on summer data only.Future work should also focus on better discriminating between snowfall and blowing snow, on the validation of satellite-based snowfall retrievals since it is of great interest to monitor the entire Antarctic continent, and in further validating ERA-Interim reanalyses and other weather and climate models.

Data availability
All the relevant observations collected in the framework of the APRES3 project will be made available as soon as possible on the website of the project (http://apres3.osug.fr),or upon direct request to the authors.

Figure 1 .
Figure 1.(a) Map of Antarctica with digital elevation model (domain south of 60 • S).A green filled circle locates the station Dumont d'Urville (DDU).(b) Temperature statistics in Dumont d'Urville, based on data collected at 1 minute time resolution in the period 2011-2015.The red bars locate the mean value and the black error bars are used to highlight the 1% and 99% quantiles.Overlayed: Wind rose (origin and intensity) statistics.

Figure 2 .
Figure 2. Main instruments deployed at DDU over the time period ranging from November 2015 to November 2016 (the MRR is however still collecting observations at the time of publication).

Figure 3 .
Figure 3. Example of a timeseries (time-height image) of MRR data and lidar data for the 15 th December 2015.

10
this reason the co-located MXPol measurements collected during the summer period are used to map the radome-affected reflectivity data provided at K-band (MRR) into X-band reflectivities.The scatter plot in Figure4shows the comparison of reflectivity values measured by the MRR and by MXPol for data collected during the period of co-location of the instruments.Because overall the relation between the two sets of measurements

Figure 4 .
Figure 4. Scatter plot of reflectivity values at 9.41 GHz (X-band, measured by MXPol) and at 24.3 GHz (K-band, measured by the MRR) during the summer campaign 2015/2016.The data correspond to time-steps when both radar were profiling (PPI and RHI scans of MXPol do not contribute).

5
the end of October 2015 until the end of January 2016, about 14 mm of liquid water equivalent snowfall have been removed, corresponding to about 21% of the uncensored data.

Figure 5 .
Figure 5. (a) Time series of Pluvio 2 untreated data, censored data (taking the MRR measurements as occurrence indicators), and their difference.(b) Wind speed measured in the near proximity (≤ 2 m) of the Pluvio 2 inlet, at the temporal resolution of one minute.

Figure 5
Figure 5 (b) shows the evolution of wind speed in the near proximity of the Pluvio 2 inlet and illustrates that the most intense phantom accumulations occur when the strongest wind peaks are observed.Because the cameras of the co-located MASC were not triggered by hydrometeors during the censored time-steps, we ruled out the possibility that phantom accumulation is in 10

Figure 6 .
Figure 6.Example of a PPI and two RHI scans collected by MXPol on the 29 th around 0014 UTC.The variable displayed in the image is the hydrometeor classification, obtained with the method of Grazioli et al. (2015).Noise in the classification at the lowest elevation angles is due Figure 7 illustrates the statistical distribution of those three classes for the period of operation of MXPol, as a function of height.Below 2000 m, the proportion of the three hydrometeor types it is relatively constant, with about 10% of rimed snowflakes, 40% of aggregates and 50% of crystals.With increasing height and getting closer to the cloud top, aggregates and rimed snowfall rapidly disappear while crystals constitute the dominant hydrometeor class.

Figure 7 .
Figure 7. Fraction of occurrence of different hydrometeor types as a function of height above ground, obtained over the period of operation of MXPol, with the hydrometeor classification algorithm of Grazioli et al. (2015).

Figure 8 .
Figure 8. Pie-chart of the hydrometeor types classified by the MASC instrument in the period from the 21 st November 2015 until the endof January 2016.The histograms shows the distribution of a riming index, ranging from 0 (unrimed) to 1 (fully rimed), for the particles of each hydrometeor class.The riming index is undefined for "small" particles, i.e. particles that are too small to be identified as a particular hydrometeor class.The classes of the chart are: small particles (SP), columnar crystals (CC), aggregates (AG), planar crystals (PC), graupel (GR), combination of columnar and planar crystals (CPC), as described inPraz et al. (2017).

Figure 9 .
Figure 9.Time series of accumulated snowfall liquid water equivalent.The relations are obtained from Pluvio 2 (in blue, for availability periods), censored from phantom precipitation, MRR (in grey the curves corresponding to the relations of Table 2, and ERA-Interim data (in black).Top panel: data corresponding to the year of measurements, from November 2015 to November 2016.Bottom panel: data corresponding to the the Summer campaign 2015/2016, from November 2015 until February 2016.

Figure 11 .
Figure 11.Cumulative contribution of increasing snowfall rates to the accumulated snowfall for the full year (top panel), and summer period (bottom panel).The time scale of snowfall intensity is 6 h also for MRR and Pluvio 2 data, to be consistent with the temporal resolution of ERA-Interim.

Figure 12 .
Figure 12.Precipitation occurrence as a function of the threshold on precipitation intensity for the full year (top), and summer period (bottom) respectively.The precipitation occurrence time scale is 6 h.

Table 1 .
Non-exhaustive list of the instruments deployed at DDU in the framework of APRES3.Only the APRES3 instruments with a certain relevance for precipitation monitoring are listed here.
a : Frequency Modulation Continuous Wave.

Table 4 .
Contingency table between occurrence of near-ground (300 m) MRR signal (columns) and visual observations of snow and blowing snow conducted by Météo France (rows).A threshold on the MRR data of 10 −2 mmh −1 is used to discriminate between "dry sky" and "precipitation".The elements of the contingency table are normalized to the total number of observations, and sum up to 100%.