The study area is located in the coastal area of Adélie Land in East Antarctica, on the ice sheet, in front of the Dumont d'Urville station located on Petrel Island. Two specific locations are considered (Fig. a). The point D5 (66.70∘ S, 139.88∘ E) is located 1.5 km from the shoreline at 170 m a.s.l., and the point D17 (66.72∘ S, 139.72∘ E) is located 6 km inland at 415 m a.s.l. At D5, the air temperature frequently exceeds 0 ∘C during daytime in summer, while winter minima rarely reach less than -30 ∘C. At D17, summer maxima are slightly below 0 ∘C, and winter minima are around -30 ∘C. The area is characterized by strong episodic katabatic winds blowing from the Antarctic Plateau to the coast that affect the two locations. These winds are responsible for snow erosion and deposition, carving sastrugi at the surface (Fig. b). Blowing-snow events are common, as reported by expeditioners . D17 is equipped with an automatic weather station (AWS) that includes sensors for temperature, wind speed and relative humidity at 2 m height; a four-flux radiometer (Kipp & Zonen CNR4 sensor) for the measurement of the incoming and outgoing shortwave and longwave radiation; two ultrasonic distance sensors (Campbell Scientific SR50) for the measurement of the snow surface height; and a FlowCapt™ instrument that estimates the drifting-snow flow from the sound generated by the impacts of the snow grains on 1 m long vertical pipes.

In addition to common meteorological equipment, two homemade spectro-albedometers called Multiband were deployed in January 2017 at D5 and D17 to monitor the snow spectral albedo (Fig. ).

The Multiband instrument is a multi-spectral-band optical radiometer operating in the visible and near- and shortwave-infrared domains. It runs autonomously in the harsh conditions of the Antarctic environment thanks to the combination of a solar panel and a battery, with interventions needed just to recover measurements, once a year. It measures the incoming and reflected radiation in order to provide the spectral albedo of the snow surface. The structure of the Multiband instrument is depicted in Fig. . Incident and reflected solar radiation are collected by two homemade cosine-shaped light collectors located on the up- and down-facing sides of the instrument's head following . Two 6 m long optical fibers (800 µm core diameter, QP800-VIS NIR, Ocean Optics) transmit the radiation inside the instrument to a rotating filter wheel containing 12 optical filters at user-selected wavelengths. The selection for this study was optimized for the albedo of snow and includes only seven different wavelengths. They are 500, 600, 800, 925, 1050, 1300 and 1550 nm (TECHSPEC^® hard-coated optical density (OD) 4.0 25 nm bandpass filters, Edmund Optics). During a measurement cycle, the wheel rotates by successive steps, and the irradiance is measured for both the incident- and reflected-light channels by two dual-band Si–InGaAs photodiodes (DSD2, Thorlabs). The measurements of the irradiance are simultaneous for the two channels, but the photodiodes are located in front of different slots of the filter wheel, which means that the order of measurement with respect to the wavelength is shifted between the two channels. For this reason, the remaining slots of the filter wheel were filled with duplicates of the filters at 500, 800, 1050 and 1300 nm in order to make some of the measurements for the same wavelength simultaneous for the two channels. Table  shows the resulting list of wavelengths of the selected filters and the order of measurement during the rotation of the filter wheel. The measurements chosen for the albedo retrieval are underlined. The rotation cycle of the filter wheel lasts nearly 30 s. A high-gain, electronic board converts and amplifies the signal from the two dual photodiodes. Then, a data logger registers the measurements (CR1000, Campbell Scientific). Since the incident and reflected measurements for three wavelengths are not simultaneous, we record the changes in the incoming radiation during the 30 s of the measurement cycle in addition. To this end, another upward-facing optical fiber located near the albedometer head transmits the environmental incoming radiation to a Si photodiode (FDS100, Thorlabs) without any spectral filter. This photodiode acquires one wideband measurement (350–1100 nm) for each step of the wheel, which is converted, amplified and recorded as the narrow-band measurements. At last, once per cycle, the dark current is measured for each dual photodiode using an obstructed window in one of the filter wheel slots. After suited artifact corrections detailed in Sect. , the spectral albedo is computed as the ratio of the reflected- and incident-irradiance spectra. The albedometer head was installed at the edge of a north–south-oriented metal gantry (Fig. ) in order to minimize gantry components or shades in its field of view. The head height at the moment of installation was of 1.6 m above the surface, which decreased on average by 18 cmyr-1. After installation, the head was leveled thoroughly. The active sensors, batteries and acquisition electronics are buried in the snow to ensure the stability of their temperature and thus of the measured signal accuracy. The Multibands at D5 and D17 ran during Antarctic summer from January to March 2017 and have been running uninterruptedly from December 2019 until today. Interventions were carried out once a year to download observations and to check the leveling of the albedometer head. The most recent one occurred in December 2021.

A Multiband measurement cycle consists of seven raw incident- and reflected-irradiance spectral measurements; 4 dark values, 1 for each band of the 2 dual-band photodiodes; and 12 wideband irradiances. Raw albedo measurements are subject to multiple artifacts and errors that need to be corrected. We developed a series of processing steps to derive accurate albedo spectra, as shown in Fig. . These steps are described here.

Raw measurement selection. Measurements acquired under unstable or low-sun illumination conditions are first discarded. These conditions are detected when the standard deviation of the 12 wideband measurements over the 30 s cycle exceeds 1 % of the mean wideband radiance. Measurements collected with solar zenith angle (SZA) larger than 70∘ are also discarded because of the reduced performance of the collectors at grazing angles and the lower interest for the surface energy budget. Finally, we had to exclude the irradiance measurements at 1550 nm because of their poor signal-to-noise ratio.

The specific surface area (SSA; m2kg-1) and the effective optical grain diameter (dopt; mm) are retrieved from the diffuse albedo spectrum for each day. The first one is the surface area of the ice–air interface of snow per unit of mass of ice. The second is the diameter of a collection of spheres having the same SSA as the actual snow. These variables are interchangeable and linked by the following relationship: 1dopt=6ρice⋅SSA, where ρice=917 kgm-3 is the ice density. Both variables are retrieved with the two-parameter model described in and derived from the analytical asymptotic radiative transfer theory (ART; ). Briefly, the direct and diffuse albedo are modeled as 2αdir(λ)=A⋅exp⁡-127(1+2cos⁡θ)2Bγ(λ)3ρiceSSA(1-g)3αdiff(λ)=A⋅exp⁡-42Bγ(λ)3ρiceSSA(1-g), where λ is the wavelength, A is an unknown scaling factor that accounts for wavelength-independent artifacts (residual calibration error or illumination fluctuations), and θ is the solar zenith angle (SZA). Here we take the absorption enhancement parameter B=1.6 and the asymmetry factor g=0.85 as suggested by . γ(λ) is the ice absorption taken from . This model is valid for albedo measured on a flat surface; vertically and horizontally homogeneous snowpack; and clean snow, which is the case in our area (after the slope correction). The two unknown parameters, A and SSA, are computed by fitting the albedo model (Eq. 3) to the daily diffuse albedo provided by the slope correction at 800, 925 and 1050 nm using a least-square minimization. The model is applied considering that the albedo is measured at 800, 925 and 1050 nm precisely, ignoring the spectral width of the filters (25 nm). Indeed, using Eq. (2) and (3) in numerical tests, we found maximum differences in albedo to be less than 0.004 when accounting or not for the spectral band width, which is a negligible error. Finally, the effective optical grain diameter is deduced from SSA using Eq. ().

Over 10 000 measurement cycles were acquired at each location (every hour during daytime and every 10 min in the hour following local noon, 03:00 to 04:00 UTC), from September to March between January 2017 and December 2021. For D5 and D17, respectively, 9.6 % and 20.1 % of measurement cycles were discarded because of unstable illumination conditions during the cycle. An additional 0.6 % and 0.5 % of the measurements were discarded after dark subtraction, inter-calibration and cosine correction because the albedo computed before slope correction contained values below 0 or above 2 for at least one wavelength. A negative albedo value suggests an incorrect dark correction (signal smaller than the dark), while values above 2 are possibly due to large slopes, considered too difficult to correct , or to snowfall obstructing the upward-looking collector. The slope correction was then applied to the remaining spectra for each day of the over 600 d at each station. To assess the quality of the fit, we computed the root mean square error (RMSE) between all the measured and modeled albedo spectra acquired during each day and filtered out the outlier acquisitions by using a z-score test with a 2σ threshold. The slope correction was then repeated without these outliers. After this step, if the RMSE >0.05, the output of the fit is definitively discarded (4.4 % and 6.2 % of the days at D5 and D17, respectively). This step helps identify and remove the measurements that may have been acquired under different sky conditions than those assigned to the whole day.

Over 600 slope-corrected, daily spectral diffuse albedo values were obtained for each station. The average RMSE is 0.021 and 0.020 for D5 and D17, respectively. To assess how such an error impacts the retrieved SSA and dopt, we computed the theoretical diffuse albedo of five SSA values between 10 and 60 using Eq. () (with A=1) and added a normally distributed random noise with a standard deviation of 0.021. SSA and dopt were then retrieved for each of the 1000 perturbed albedo spectra for each of the 5 SSA values. The resulting distributions are shown in Fig. a for the SSA and Fig. b for the optical diameter.

The standard deviations of the perturbed distributions increase with the SSA, from 26 % for a SSA of 10 m2kg-1 to 64 % for a SSA of 60 m2kg-1 with respect to the true value. The distributions of the perturbed dopt in Fig. b show a decreasing absolute standard deviation but an increasing relative standard deviation as dopt decreases. As a conclusion, we consider that the error for measured SSA above 60 m2kg-1 and the corresponding dopt of less than 0.11 mm are too large for our application, and these measurements are discarded. The final results account for 541 and 601 d with valid SSA and dopt for D5 and D17, respectively. Their final distributions are shown in Fig. a and b (in orange).

As no reference to in situ measurements of grain size are available at our site, we first rely on satellite observations for an independent assessment of the quality of our retrieval and, second, on broadband albedo measurements recorded at D17 using a conventional shortwave radiometer (CNR4).

The measured dopt is compared to dopt retrieved from images from the OLCI (Ocean and Land Colour Instrument) on board the ESA's Copernicus Sentinel-3A/B ). The retrieval is done using the algorithm SICE, the “Pre-Operational Sentinel-3 Snow and Ice Product” . Shortly, the albedo is retrieved from the calibrated and geolocated top-of-atmosphere radiances measured by OLCI at 865 and 1020 nm, and the snow grain size is computed from the albedo using the ART , as in our case, but with a largely different algorithm. Surface snow is assumed to be clean (which is often the case in Antarctica; ) and vertically and horizontally homogeneous over a flat, 300 m pixel.

The dopt measured by Multiband (MB) and retrieved with SICE for D5 and D17 (for 140 and 150 clear-sky dates according to the CSI) is shown in Fig. . At both locations, a group of SICE dopt has very low values independently of the value of the corresponding MB dopt, likely because of the presence of undetected clouds on the pixel during the satellite overpass. A statistical test is thus applied to identify outliers between low SICE dopt values. If the latter are <0.2 mm, and the difference to the MB dopt exceeds 3 times the average difference, the measurement is considered an outlier and rejected. A total of 29 and 28 outliers (in yellow) were identified for D5 and D17, respectively, with 27 common dates among them. The remaining pairs of dopt (in blue) have a Pearson’s correlation coefficient of 0.67 for D5 and 0.76 for D17. The RMSEs between the MB and OLCI dataset are 0.36 and 0.16 mm for D5 and D17, respectively. Still, we believe the RMSE estimate at D17 to be more realistic than the one at D5. Indeed, the 300 m pixel of OLCI including D5 may not be completely representative of the snow reflectance because of the recurrent presence of blue-ice areas and containers, as well as the strong average north–south slope of the area covered by the pixel (8.3 %), which is almost double the average pixel slope at D17 (4.3 %). Thus, we conclude that Multiband measurements correlate with the SICE product with an overall error of 0.16 mm RMSE that can be considered to be an upper bound, since both the satellite and in situ measurements are subject to independent errors.

A CNR4 radiometer providing shortwave broadband albedo is installed at D17. To test the quality of our measurements and the performance of our algorithm, we recalculate the broadband albedo from the retrieved dopt and compare to the one observed by the CNR4. To this end, we compute the diffuse and direct albedo between 0.28 and 2.5 µm with a step of 0.01 µm using Eqs. () and (). One broadband albedo is computed every half hour, as CNR4 measurements are available every 30 min. The spectrum shape and the direct-to-diffuse ratio of the incoming shortwave radiation are computed with SBDART for a pristine, sub-arctic winter atmosphere. Simulations were made for solar zenith angles between 43 and 70∘ and cloud optical thickness of 0 and 30 for clear-sky and overcast conditions, respectively. The broadband albedo is computed as follows: 4αBB=∫λIdir(λ)αdir(λ)dλ+∫λIdiff(λ)αdiff(λ)dλ∫λIdir+diff(λ)dλ. The resulting broadband albedo values measured at solar noon by Multiband and by the CNR4 are shown in Fig.  (solid line). The 10th–90th percentiles of the broadband albedo observed during the entire day is also shown (shaded area). The intraday variability in the albedo measured by the CNR4 is much wider, as its mean standard error is 0.061 for the CNR4 measurements and 0.008 for Multiband's. This difference, as well as the presence of non-physical albedo values above 1, is explained by sensor tilt and by the presence of a slope in the footprint of the CNR4 sensor due to the recurrent presence of sastrugi at D17 and to the relatively low installation of this sensor (<1 m). Sastrugi being transitory bedforms, this slope is dynamic, and it is not possible to correct the broadband albedo reliably from its effect, as opposed to spectral albedo . Still, it is possible to reduce the effect of the local slope by considering only the broadband albedo measured at noon (solid lines in Fig. ), which is the least affected by tilt and slope as the sun is at its highest. At noon the two time series compare favorably, with a Pearson correlation coefficient r=0.51, a mean negative bias of 0.009 of the CNR4 and a standard error of 0.060. Significant differences above 0.1 of the two albedo values are 6 % of measurements and may be explained by the presence of persistent sastrugi structures or by a tilt of the sensor. Indeed, if we exclude days with CNR4 measurements of the broadband albedo >1 (75 % of all days), potentially due to the presence of strong slopes and clear-sky conditions, the correlation of the time series of broadband albedo measured at noon increases, with r=0.62, a mean negative bias of 0.013 of the CNR4 and a standard error of 0.051. Similarly, considering only measurements acquired during days with overcast sky – which are not affected by surface slope – the correlation coefficient increases up to r=0.55, with a mean negative bias of the CNR4 of 0.023 and a standard error of 0.047.

The time series of retrieved SSA and dopt are shown in Fig. . Even though these two variables are strictly inversely related, the visual interpretation of dopt is more suitable for large grains, and conversely, the SSA is more suitable for small grains. This is the reason why both are shown. In the following, we mainly refer to dopt variations as the prevailing conditions in summer in our study area (melt and wet metamorphism) favor large grains.

The retrieved dopt varies between 0.11 and 1.7 mm at both stations, with mean values of 0.55 and 0.43 mm for D5 and D17, respectively. The variations at both locations are correlated, with a Pearson correlation coefficient between the two time series of r=0.72. Still, the dopt at D5 is larger than at D17 for 70 % of all days with valid measurements. The variations in dopt occur at the seasonal scale but also at more rapid scales. On the seasonal scale, the mean dopt is lowest in October (0.20 mm at both locations) and highest in January (0.82 mm at D5, 0.54 mm at D17). Superimposed on these seasonal variations, we observe occasional sharp decreases in dopt followed by a rapid increase with typical timescales of 1–10 d. These rapid variations usually co-occur at both locations, which confirms the geophysical nature of these changes, i.e., excluding measurement artifacts. Given this high correlation, we mostly interpret the time series at D17 in the following.

The time series of dopt measured at D17 are compared here in Fig.  with daily snowfall amount (SF), snow drift (SD), daily maximum surface temperature (Ts,max) and average surface height (SH). Snowfall is taken from the ERA5 reanalysis . The ERA5 grid point corresponding to D17 is situated at 66.75∘ S, 139.75∘ E, 3.5 km from D17, at an altitude of 430 m a.s.l. (versus 415 m a.s.l. for D17). Only snowfall values >0.5 kgm-2d-1 are taken into account to avoid considering spatially confined precipitation that the reanalysis spreads over the whole pixel. The daily amount of drifting snow is measured in situ by the FlowCapt™ instrument . Although it reproduces well the occurrence of drifting-snow events, this instrument underestimates the actual drifting-snow flow . Moreover, FlowCapt™ is mounted approximately 1 m above the ground and misses events of snow mobilization below a height of 1 m. The maximum daily surface temperature Ts,max is retrieved from the time series of the incoming and outgoing longwave irradiance measured and averaged every 30 min by the CNR4 at the AWS at D17 using the Stefan–Boltzmann law and a surface emissivity of 0.98. The daily average surface height is retrieved from SR50 distance measurements with respect to a surface located 2.25 m below the sensor for the first year of measurements (2016–2017) and 1.75 m below for the following. Days with rain events are distinguished by the presence of snowfall, daily maximum surface temperature Ts,max of 0 ∘C and daily maximum air temperature Tair,max above 0∘C.

Snowfall is reported for 34 % of the measurements. The percentage drops to 28 % if only days with SF >1.0 kgm-2d-1 are considered. December and October have, on average, the minimum and maximum number of days of snowfall per month, 19 % and 47 %, respectively. In terms of intensity, the month with minimum snowfall is still December, with a daily average of 0.50 kgm-2d-1 over the years of measurement and a limited inter-annual variability (0.32 kgm-2d-1 in 2021 to 0.64 kgm-2d-1 in 2020). February is the snowiest month, with a daily average of 3.12 kgm-2d-1 but a larger year-to-year variability (0.93 kgm-2d-1 in 2020 to 6.28 kgm-2d-1 in 2021).

Snow drift is almost ever-present (90.6 % of measurement days) according to the FlowCapt™ measurements. The few periods without snow drift last between 1 and 3 d. The median daily amount of snow transported by the wind is 510 kgm-2d-1, while the maximum exceeds 24 000 kgm-2d-1. This range of values is similar to those observed by for the month of January 2011 at D17. Its intensity is strongly variable among seasons and years. The months with the weakest and strongest mean daily snow drift (154 and 5388 kgm-2d-1) are January 2017 and September 2021. On average, December and January have the weakest mean daily snow drift (1748 kgm-2d-1); October and November are intermediate (1930 kgm-2d-1); and September, February and March are the strongest (3068 kgm-2d-1). This is compatible with the findings from that snow drift is weakest in summer and strongest in winter at D17. Still, the inter-annual variability for the same period is sometimes stronger than this inter-seasonal variability. For example, the daily mean for January 2021 is 323 kgm-2d-1 and over 10 times more for January 2020, with 3389 kgm-2d-1. found that major drift events (events whose duration exceeds the 75th percentile of all events) contribute over 70 % of the total transported snow mass at D17 on a yearly scale. Although most of these events occur during the winter season, this explains the strong variability in the mean daily snow flow for the same month among years.

The time series of the 2 complete years (2019/2020 and 2020/2021) show that the maximum daily surface temperature Ts,max is, on average, lowest in September (-17.2 ∘C) and highest in December (-0.2 ∘C). A daily maximum surface temperature Ts,max≥0 ∘C is hereinafter used as a qualitative indicator of surface melt. Days satisfying this criterion are 3 % of the total measurements in November, 48 % in December, 27 % in January, 7 % in February and none during the other months.

The surface height is measured by two sensors from 2017 to March 2021 and only one during the season 2021/2022. The variations measured by the two sensors are similar and correlated (r=0.93). The surface height decreases on 59 % of days of measurements with respect to the previous day, with an average decrease of 14.5 mm. During the remaining 41 % of measurement days, the surface height increases by 19.5 mm on average. These time series are characterized by occasional, sharp increases followed by long, slow decreases throughout all seasons.

In the conception of the Multiband albedometer, multi-band acquisitions were chosen over spectral ones in order to reduce its energy consumption. As a result, Multiband could be deployed in multiple remote locations of the Antarctic continent, where it has been acquiring reliable measurements in complete autonomy. Although only six bands are exploited for the processing, and only three are used for the grain size retrieval, Multiband observations have compared favorably with satellite-retrieved grain size and in situ measurements of the broadband albedo. Moreover, the coherence between the grain size time series measured at D5 and D17 is remarkable. The two locations, despite an altitude difference of approximately 300 m, have similar grain size values during fall and at the end of the summer season (0.33 mm at D5, 0.31 mm at D17) and clearly diverge only during the melting period in full summer (0.74 mm at D5, 0.52 mm at D17). In spite of the heterogeneity of the two locations as seen from the field, the similarity of the grain size time series provides confidence to the Multiband measurements. Moreover, it justifies the use of simulations and reanalysis at multi-kilometer resolution in this area .

Our observations depict the evolution of the snow surface grain size in the windy coastal locations of D5 and D17 in Adélie Land, East Antarctica, over five summer seasons for a total of nearly 600 d with successful retrieval. Through the comparison with the time series of the meteorological variables and snow height observed at D17, we relate the complex evolution to the multiple competing environmental drivers of the grain size, namely dry metamorphism, snow drift, snowfall and melt. Overall, our results show a seasonal cycle characterized by small grains in spring and autumn and larger grains in summer. These variations directly impact the broadband albedo with typical values between 0.82 and 0.75 (Fig. ), implying a large variation in the shortwave net radiation absorbed in the snowpack. However, our results also suggest that the evolution of the surface albedo in the presence of snow drift due to strong wind is less predictable than it would be in the absence of these drivers.

Snow drift at D17 has mainly an erosive character, as observed during 67 % of measurement days when snow drift occurred without snowfall or melt. Increases in the surface height are less frequent (33 %) but significant in terms of magnitude, with a median and maximum snow height change of +3 and +186 mm, respectively. However, these large variations in the snow height do not coincide with major changes in the grain size. Indeed, most grain size variations (83 %) range between -0.1 and +0.1 mm. Both episodes of snow accumulation and decrease in the grain size during 58 % of days with snow drift imply the deposition of small, drifting grains on the snow surface, as suggested by after the observation of similar dynamics in the Canadian Arctic. Still, even when the grain size increases from one day to another, growth happens at a very slow pace (median +0.03 mm). This noteworthy stability of the grain size challenges the expectation of growing snow grains under the action of dry metamorphism during the summer season. We think that this limited growth may also be related to the input of small grains deposited during snow drift, despite the overall erosion of the snow surface in this period. Indeed, observed similar dynamics of reduced metamorphism after the deposition of thin layers of high-SSA grains by snowfall at Dome C, which they attribute to the inhibition of the positive feedback described in . For the 28 measurements with an increase in grain size over +0.1 mm, the average grain size was large (0.63 mm) even though the average surface temperature during those measurements remained below -10 ∘C. Such large grains without melt are explained by the erosion of the snow surface, which likely occurred until a layer of very cohesive, coarse-grained snow was uncovered, as in Fig. c.

The interaction of snowfall and wind drift also has an important impact on the observed snow grain size. Indeed, days with snowfall frequently featured surface erosion (45 %) and an average grain size increase of +0.02 mm, against the average grain size decrease of -0.03 mm of days with accumulation (55 %). The erosion of the snow surface reveals that precipitation is completely swept away by the wind for almost half of the snowfall events at this location, at least for the spots under the surface height sensors. Moreover, the evolution of the snow grain size after episodes of erosion during snowfall follows similar dynamics to those of erosion alone, with a grain size increase of 56 % and a decrease of 44 % (60 % and 40 % in the case of erosive snow drift alone). mention frequent snowfall as a determinant factor to limit the length of low-albedo periods leading to surface melt. In windy areas, however, the frequency of snowfall leading to actual accumulation should be considered, as almost a third of all snowfall events result in surface erosion followed by an increase in the snow grain size. The result of these complex interactions at D17 is that the commonly expected outcome of snowfall – an increase in the surface height and a decrease in the grain size – is only met on 33 % of days with snowfall, and the average grain size change during precipitation is in fact negligible (<0.01 mm). Finally, the weak, negative correlation between the daily snow accumulation and the wind speed (r=-0.22) suggests that accumulation is more likely for more intense snowfall and lower wind speeds and lets us mark an average daily wind speed of 12 ms-1 as a rough boundary between accumulation and erosion at D17.

The evolution of the grain size in the presence of surface melt is also complex. The expected output for surface melt, which is an increase in the grain size and a moderate decrease in the surface height (due to the compaction of the snowpack), is only met on 27 % of days with melt. Indeed, 37 % of melt days saw an increment of the snow height and an average grain size decrease. However, even on days with surface erosion (63 %), the sharp increases in the grain size that would be expected because of surface melt are occasional, while the average increase is +0.05 mm only. We believe this to be due to the input small grains at the surface that contrast or limit the melt-driven grain growth. As a consequence of these interactions, the dopt is, on average, rather stable on days with surface melt, with a mean grain growth of 0.02 mmd-1 only.

The dynamics observed at D17 may compare to locations with similar characteristics in terms of snow type, temperatures, humidity and wind regimes only. However, they show a remarkable range of possible evolution paths of the snow optical properties involving snow metamorphism, snowfall and snowmelt in the presence of snow drift, some of which are regularly observed in the literature. and , for example, described the evolution of the snow SSA over one summer season at two locations over the East Antarctic Plateau. In this flat area, strong winds are less common than on the coast, and SSA usually increases during snowfall and slowly decreases afterwards. However, they both observed punctual episodes of sharp SSA decreases after snowfall due to the removal of the fresh snow layer by snow drift. Also, and observed punctual decreases in the grain size (or equivalent increases of albedo) during dry, warm summer periods at EastGRIP (central Greenland) and over the King Baudouin ice shelf (Dronning Maud Land, East Antarctica), respectively, likely due to the deposition of smaller snow grains after drift events. Inversely, our main conclusion is that the prevailing dynamics caused by snow drift at D17 are to prevent sharp changes in the snow grain size. Indeed, snow drift (1) delivers small grains onto the surface, which contrasts the action of metamorphism, either dry or wet, and (2) prevents the deposition of snowfall. The major drivers of the grain size increase – metamorphism – and decrease – precipitation – are thus inhibited, granting a relative stability of the grain size with respect to the ensemble of snowfall and surface melt events. Nonetheless, rare, sharp increases in the dopt caused by the wind were observed. They are mainly due to the erosion of the surface snow until a refrozen melt layer. It is possible that the deposition of small grains over these layers is prevented by the strong cohesion of the surface, as suggested by .

Some limitations of this study should be mentioned:

On the basis of the analysis carried out in Sect.  and , we established an upper-bound RMSE =0.16 mm for the dopt retrieved from Multiband measurements. Although this value is higher than most variations in the grain size measured from one day to another, statistical considerations remain valid. Moreover, most measurements are likely to have a lower RMSE as the satellite-retrieved dopt and in situ measurements have independent error sources.