In order to study the megadune areas, we combined three main datasets (satellite images; meteorological data from reanalysis products; digital elevation model, DEM) to create a method for the automatic detection of snow glazed surfaces. We tested two methods, one by creating five sample transects in the EAIIST area (Fig. 1c) and visually identifying thresholds of albedo, thermal brightness temperature (BT) and SPWD to discriminate between glazed surfaces and surrounding snow. The five transects were created in different areas of the megadune field, and they show relatively different wind directions and topographic aspect and slope, with the aim of representing the widest possible range of SPWD values. For the second method, i.e. a maximum likelihood supervised classification, we created 30 polygons in the glazed-snow area and 30 for firn, which were used to train the classification algorithm.

The study includes four main processing steps: Landsat 8 OLI image processing for the calculation of NIR albedo, extraction of thermal BT from Landsat thermal band 10, SPWD calculation from ERA5 and satellite sastrugi-based wind direction, estimation of the surface velocity and migration of megadunes using feature tracking (2014–2021), and comparison of GPR–GPS measurements from 1999 with the REMA DEM from 2014 (specific strip on the area). The first three steps were the basis of the automatic detection of the glazed-snow areas.

On average, in the five analysed transects NIR albedo ranges from 0.81 to 0.86 (α) in the upwind area (snow sastrugi) and from 0.73 to 0.81 (α) downwind (glazed surfaces), with differences inside the transects of about 0.07 (α) with a maximum value of 0.1 (α). The maximum contrast of NIR albedo between glazed surfaces and snow sastrugi usually occurs at springtime (October–November) and decreases during the summer season (Fig. 2). Our remote sensing observations agree with previous analysis that pointed out that in late summer, radiative cooling of the uppermost surface layer leads to formation of surface frost, by condensation of local atmospheric vapour onto the snow surface; this gives the glazed surface a more diffuse specular reflection than in spring and changes its appearance in albedo and BT (Scambos et al., 2012, and references therein). Along the transects, the correlation of NIR albedo from the different images is high (R2: up to 0.99) during the spring season (24 November 2013, 27 December 2013) and decreases by the end of the summer and in comparison with the following years, with an R2 of 0.7 only after 2 years (17 December 2015) and up to 0.6 after 6 years (December 2019). A similar decrease in correlation occurs from the comparison of the SPWD and NIR albedo from 2013 (R2: 0.66) to 2019 (R2: 0.39).

For the thermal BT, we observed an intra-seasonal trend on all transects: in fact, while thermal BT remains ≥ 244 K during the middle of the summer (11 and 27 December 2013), it decreases moving away from the summer solstice. Temperatures range between 238 and 240.5 K on 25 November 2013, 26 d from the solstice. The difference increases on the date farthest from the solstice, 28 January 2014 (38 d from the solstice), with the lowest values ranging between 236 and 239 K. The BT varies up to 1.5 K for each transect but up to 3–4 K within individual images. Intra-annually, the difference between glazed surfaces and snow is higher during the spring (max 1 K in November) and tends to decrease over time, becoming lower than 0.5 K at the end of summer (January), where differences between the two surfaces are hardly detectable and the correlation between the two parameters frequently decreases drastically. These differences are directly correlated to the ones observed in NIR albedo, as a higher quantity of energy is absorbed on glazed surfaces.

High correlations are found between NIR albedo and thermal BT with an R2 of up to 0.67 (95 % confidence interval) and between SPWD vs. NIR albedo and thermal BT (R2=0.44 and 0.57 at 95 % confidence interval, respectively) calculated along all the transects (Fig. 3). The comparison between thermal BT and SPWD shows the same pattern observed for the NIR albedo but is proportionally inverse with respect to SPWD (Fig. 3), with higher temperatures corresponding to the glazed part of downwind areas of the dunes and conversely, lower values related to snow sastrugi in the upwind zones.

Based on the transects, the variability in NIR albedo at a seasonal (spring–summer) to pluriannual scale is observed to be twice as large in the snow accumulation area on the upwind flank and the bottom of the leeward flank (Fig. 2) compared to the glazed-surface NIR albedo (0.7 % compared to 0.3 % NIR albedo variability), which remains more stable and more highly correlated at a seasonal (spring–summer) and pluriannual scale.

The analysis of sastrugi direction using seven Landsat scenes from the spring and summer months during the period 2013–2020 show small differences in direction within each image and in repeated imagery (< 5∘), confirming the stability in direction of sastrugi landforms and thus the persistence of katabatic wind.

The comparison of the results of wind direction obtained using sastrugi direction by satellite (resampled using bilinear interpolation) and ERA5 present similar values for both areas, with a lower difference in the EAIIST area (differences of 1∘ in average values) compared to It-ITASE (9–14∘; see Table A3).

At the regional scale (30 km spatial resolution), the entire megadune field has an average SPWD of 1.2 m km-1, when calculated using sastrugi-based wind direction, and 1.1 m km1 when using ERA5, in agreement with previous studies (e.g. Frezzotti et al., 2002b). To distinguish between leeward (glazed surface) and windward flanks of the dunes for the two sites, the SPWD based on sastrugi was further resampled to 120 m using bilinear interpolation. For the SPWD on megadunes at a local scale (100 m), we found a mean value of 5.6 ± 1.0 m km-1 for the leeward side and negative SPWD values, with a mean of -4.2 ± 1.6 m km-1 on the windward flanks.

The application of the supervised classification and the self-defined-threshold methodology on the different combination of the analysed parameters for the 2013–2014 season and on the 17 December 2015 scene showed contrasting results (Table 2).

In fact, even if the general accuracy in distinguishing firn and glazed surfaces was on average high (mostly > 80 %), with the highest values on the scene from 17 December 2015 (up to 94 % and 98 % in the combination of NIR–SPWD for supervised and self-defined thresholds, respectively), lower values and higher differences were found when comparing the accuracy in detecting only glazed snow. There, the inclusion of SPWD in the classification significantly improves the ability to detect these surfaces, with a doubling of the accuracy during the season 2013–2014 (between 0 % and 30 % for single parameters or combination without SPWD to around 50 % of season average in the other cases and up to 55 % for the NIR–BT–SPWD combination). The increase is also observed for the 17 December 2015 scene, from 78 % of NIR alone to 83 % for NIR–SPWD (which becomes 95 % in the case of NIR–SPWD for the self-defined-threshold methodology, the highest calculated value of accuracy in detecting glazed snow). Summarizing, the impact of the inclusion of the SPWD is evident, both applying the supervised classification and self-defined-threshold methodology, as the SPWD doubles the accuracy compared to the cases with no SPWD. We further observed issues in using BT in these approaches because of its high heterogeneity at a high scale (i.e. entire Landsat scene), as BT presents strong variations across the image; the strong improvement in using a scene (i.e. 17 December 2015) with no interferences (e.g. fog, clouds, “low” SZA) and close in time to the REMA DEM (stripes from January to October 2016), leading to a significant increase in accuracy of the method (from highest values of ∼ 50 % of scenes from season 2013–2014 to over 80 %); and the higher accuracy when applying a self-defined-threshold approach.

The absolute megadune movement calculated using feature tracking on optical satellite image pairs spans from 3 to 6 years and presents small differences in the two study areas, ranging between 10.5 and 14.2 m a-1 overall, with no detected significant trends over time. The average values are similar when using different datasets (Landsat 8 OLI and Sentinel-2) but with Sentinel-2 velocities showing slightly lower average values compared to Landsat 8. Moreover, due to the slightly higher spatial resolution (10 m vs. 15 m), the number of features tracked using Sentinel-2 is an order of magnitude higher than that of Landsat 8 OLI (Table 1), even if the number of pixels is higher (with a ratio of 1.4) in Landsat scenes. The direction of the migration does not differ much across the different datasets, showing opposite values compared to wind direction. The second method used to calculate the migration velocity is the ridge vectorization and tracking for the same image pairs. This method shows slightly higher velocities (16.7 ± 3 m a-1) than IMCORR tracking of megadune features (11.9 ± 3.6 m a-1) at the EAIIST site for the period 2013–2019 but within the error.

At the D6 It-ITASE site, five GPR–GPS transects were surveyed on megadunes (Frezzotti et al., 2002b); the comparison between GPS elevations (3 January 1999) and the REMA DEM (2 February 2014) provides information about the relative change in elevation at high resolution (decametre level) of the megadunes during the past 15 years. On the five transects, we observe an almost stable elevation in correspondence with glazed surface/leeward flank, whereas the maximum difference in elevation (from +1.2 to +1.9 m, with an average maximum value of +1.4 m) occurs always in the snow accumulation/upwind flank on the correspondence of the trough (Fig. 4). By projecting the transects along the prevalent wind direction (239∘), based on the surrounding sastrugi orientation, we were able to evaluate the megadune migration using the relative change in elevation. Using the crest/trough position of each dune, we calculated an average displacement of 11 ± 5.2 m a-1 from all transects (Fig. 4). The migration of the dunes is evident in all transects with the upwind migration of the crest over the upstream flank and of the trough on the upstream flank of the previous megadune. In contrast, the glazed surfaces on the downwind flank remained generally stable in elevation over time (Fig. 4) but are clearly buried at the upstream flank foot and migrate at the crest. At the D6 site, Vittuari et al. (2004) measured an ice velocity of 1.46 ± 0.04 m a-1 with a direction of 97∘ using repeated GPS measurement between 1999 and 2001. The closer value of MEaSUREs ice flow at the D6 site is 2.2 ± 1.1 m a-1 with a direction of 89∘, in agreement with GPS measurements. At the EAIIST site, MEaSUREs data show an ice flow of 6.1 ± 3.4 m a-1 with a direction of ∼ 65∘. Both velocity directions agree with the topographic slope at the site. Applying Eq. (1) for the calculation of megadune migration components, we obtained a sedimentological migration of 18.4 ± 6.7 m a-1 (229∘) at EAIIST and 15.4 ± 4.7 m a-1 (246∘) at It-ITASE using Landsat 8 OLI data and 16.0 ± 7.3 m a-1 (230∘) at EAIIST and 13.6 ± 4.9 m a-1 (245∘) at It-ITASE with Sentinel-2.

In this study, we showed the possibility of calculating the SPWD based on wind direction from ERA5 and Landsat-derived sastrugi. At both investigated sites, the direction of the wind from ERA5 at a velocity higher than 11 m s-1 was found to be closer to the direction of sastrugi surveyed by satellite. The small difference between the two datasets could be correlated to the formation of sastrugi, as wind speed ≥ 11 m s-1 was previously reported by Kodama et al. (1985) and Wendler et al. (1993) to be required for the formation of sastrugi. While the EAIIST site shows similar average directions to ERA5, in the other study area (It-ITASE) a slightly higher difference was found between the two datasets for wind velocity slower than 11 m s-1. The direction retrieved from Landsat is strongly dependent on high-velocity prevailing winds (katabatic winds), which shape the sastrugi and direction, while ERA5 also takes into account wind directions other than the katabatic. In addition, this difference could be caused by the different spatial and temporal resolutions between the satellite and ERA5 (30 m vs. 30 km, scene-based vs. average of 20 years), as well as inaccuracies in the ERA5 wind direction. The larger difference in wind direction using the various datasets (ERA5, sastrugi detected by satellite, sastrugi measured on the field) at the It-ITASE site could also be attributed to the higher variability in the katabatic wind direction; in fact, this site is at the northern limit of megadune field (Fig. 1), and a relatively high variability in katabatic wind direction (> 10–15∘) could be among the threshold factors that does not allow for the formation of megadunes in the northern part. However, with the aim of applying this methodology at a large scale using ERA5 data, e.g. the whole continent, the differences between the two sources can be significant (e.g. at the It-ITASE site) and could produce errors in the SPWD calculation. Therefore, the use of sastrugi could be a more accurate way to interpret the prevalent wind direction with high wind speed (≥ 11 m s-1) compared to ERA5.

The SPWD is the only parameter that could be considered almost constant at a 10 km scale, in consideration of the stability of the direction of the katabatic wind, driven mainly by surface slope and the Coriolis force. In contrast, albedo and above all thermal BT change both temporally, annually and during seasons, and spatially across the satellite scene. In fact, NIR albedo significantly varies because of surface changes of up to 0.1 α and between the beginning, the middle and the end of the summer season in relation to the SZA by ±0.01–0.02 α. Frezzotti et al. (2002b) pointed out the presence of huge sastrugi (up to 1.5 m in height) located on the windward flank and alternation of sastrugi (up to 40 cm) and glazed surfaces located at the bottom of the interdune area. The observed change on NIR albedo and BT on the windward flank is correlated to the sastrugi formation and deterioration during the season and their relative change in shadow (Warren, 1982). In addition, thermal BT varies from a higher temperature near the summer solstice to lower values in late spring and summer, in the range ±5–10 K. In both cases, the differences between leeward flanks where glazed surfaces are located and windward flanks of megadunes are not high enough to overcome the seasonal variability and thus a constant range for albedo and temperature is impossible to determine. Spatially, the satellite-derived NIR and thermal BT show large variability inside the same satellite images, in particular for thermal BT, but strong correlation among the two parameters up to an R2 of 0.99 along each transect. The observed variabilities could be related to the different illumination condition and meteorological conditions with development of surface hoar crystal due to fog and under calm sunny weather with a downward as well as an upward vapour source to the near-surface layer. The growth of surface hoar crystals dramatically changes the snow structure, specific surface area and density, as well as surface roughness, leading to significant changes in albedo and therefore surface temperature (Gallet et al., 2014). For these reasons, different thresholds can not only be required when investigating an entire tile of Landsat in the case of self-defined-threshold methodology but also can explain the inability of supervised classification based only on BT to distinguish between glazed snow and firn. With BT, in fact, the difference in values across the images is higher than the mean spectral difference between these two surfaces at these wavelengths. For the same reasons, the classification approaches including BT present, in most cases, lower accuracy than the other cases, suggesting against using it to map glazed snow at a large scale. On the other hand, NIR albedo does not show this sort of problem and instead demonstrates good ability in distinguishing between the two surfaces, with a good accuracy even when used alone (especially in the 17 December 2015 case, Table 2). Additionally, it is evident how the implementation of SPWD is fundamental in aiding the detection of glazed snow, together with the NIR band, particularly by improving the detection even when the other parameters present more uncertainties (owing to e.g. SZA and fog). In fact, in all the combinations where SPWD is included, the accuracy in the distinction between glazed snow and firn shows more constant results and minimum and maximum accuracy across the analysed season is similar. Finally, our results point out the importance of using satellite images with no interferences (e.g. clouds, fog, high SZA), with the aim of automatically detecting glazed snow, as the accuracy of the method drastically increases (17 December 2015 compared to the other cases) and also that the self-defined-threshold approach provides better results in terms of accuracy than the supervised classification. Nevertheless, even if the self-defined-threshold model shows a higher accuracy, supervised classification allows for overcoming the issue of defining accurate thresholds across a certain season, providing good accuracy results especially on good quality images, as calculated on 17 December 2015.

By using the classifications with the highest accuracy in the EAIIST area achieved based on the scene from 17 December 2015 (NIR–SPWD combination), where approximately 75 % of the area is covered by megadunes, we could calculate that the glazed surfaces cover around 43 % of this specific dune area, i.e. ∼ 80 % of leeward flanks.

In conclusion, the detection of glazed surface/ablation area by satellite images is challenging because of differences in illumination and meteorological conditions (cloud cover, blowing snow, etc.) among different satellite images. Nevertheless, the high-resolution digital terrain model (REMA) allows for calculating an SPWD with unprecedented detail, similar to the resolution of optical satellites (Landsat 8–9, Sentinel), and significantly improves the detection of glazed/ablation surfaces at 10 m scale across the continent; therefore, it could significantly improve research on the SMB of the Antarctic Ice Sheet.

The absolute position of the megadune crest and trough are driven mainly by two processes: snow ablation/accumulation processes and ice sheet surface flow. GPS and GPR profiles along the It-ITASE traverse show the presence of paleo-megadunes buried up to the investigation depth of 20 m (Frezzotti et al., 2002b). Analysis of the D6 firn core allowed for detecting the Tambora eruption signal (1816 CE) at 15.36 m depth with an average snow accumulation of 36 ± 1.8 mm w.e. a-1, whereas an average value of 29 ± 7 mm w.e. a-1 of spatial variability in SMB at D6 site was evaluated by GPR calibrated using accumulation at three firn cores (Frezzotti et al., 2005). The elevation changes during 15 years observed using GPS and REMA show a relative increase in accumulation on the windward flank with the maximum value at the trough compared to the glazed-surface area from 29 to 46 mm w.e. a-1 with an average value of 34 mm w.e. a-1, using a density of 360 kg m3 in the first 2 m. This value is very close to the estimated change in accumulation in the megadune area from 7 to 35 mm w.e. a-1 provided by Frezzotti et al. (2002b) using the variability in GPR internal layering at the megadune site. The minimum value represents a decrease in accumulation of up to 75 % or more on glazed surfaces. The relative stability of glazed surfaces with respect to elevation change and NIR albedo confirms the extremely stable SMB low value of the glazed surfaces with respect to accumulation areas, due to the long-term hiatus in SMB forced by wind scouring processes.

Using the isochrone distance of 1.5–1.8 km between the 180-year-old paleo-crest detected by GPR and the recent crest from GPS observations (1998–1999 CE), we can evaluate the windward migration of the paleo-megadune crest at about 8–10 m a-1. This vector from field observations summed with an ice flow from GPS of 1.46 ± 0.04 m a-1 with a direction of 97∘ produced an absolute migration of 10.3 m a-1 with a direction of 214∘. This value is in very good agreement with absolute migration calculated using the elevation comparison between GPS and REMA (11 ± 5.2 m a-1) and with satellite tracking (from 11.4 to 14.0 m a-1), in particular with Sentinel-2 images (11.4 m a-1, Table 1). At the D6 site, the movement components show different intensity with an order of magnitude of difference: 1–2 m a-1 for ice flow vs. 13.6–15.4 m a-1 for sedimentological migration. The components present nearly opposite directions: 97∘ for ice flow and 245∘ for sedimentological migration. The results allowed us to calculate all the components of migration and to conclude that for a megadune with a wavelength of 3 km we could calculate an absolute migration of approximately 10 m a-1. This burying process of snow on glazed surfaces takes about 300 years, with a overlap of the crest to the trough and glazed to sastrugi surface as observed by GPR internal layering. These results are strongly in accordance with Courville (2007), who determined a burial rate of 330 years based on a firn core drilled in 2003/04 and migration rates of approx. 12 m a-1 (from AVHRR data; Advanced Very High Resolution Radiometer) at a field located at 80∘47′ S, 124∘29′ E in the megadune region of the EAIIST site.

The megadune migration on the upwind part observed by elevation change and tracking is also confirmed by the comparison of NIR and BT along the studied transects. These parameters remain relatively stable during the observed time on the glazed surface on the leeward flank, whereas the positive SMB upwind flank and bottom trough area change significantly not only at a pluriannual scale but also at a seasonal scale. Hence, we observe a general overestimation of sedimentological and absolute migration using remote sensing with a mean difference of +1.9 m a-1 for Sentinel-2 (uncertainties of 19 % for sedimentological migration and 10 % for absolute migration). Using Landsat 8 OLI images, larger differences were found, with an average overestimation of 3.8 m a-1. This difference could be caused by spatial variability in processes: with remote sensing we analysed a much wider area, as opposed to in situ observations which were acquired in transects on a limited section of the megadune field. Finally, the spatial resolution and geolocation (Mouginot et al., 2017) could affect the satellite data, as demonstrated by the worse results obtained using Landsat images at 15 m spatial resolution against 10 m resolution of Sentinel-2.

The ice velocity of MEaSUREs is based on SAR images and is in very good agreement with GPS measurement, and the tracking methods of IMCORR using optical images and crest displacement are in agreement with the migration of morphologies observed from the comparison between change in elevation by GPS and REMA. Based on our analysis, the sedimentological processes are analogous at the It-ITASE and EAIIST sites. At the second site, a faster ice-flow motion was observed by MEaSUREs, and the velocity of absolute migration is reduced by almost 35 %, compared to the initial sedimentological-migration velocity.

The ice velocity based on SAR images presents a phase centre that penetrates up to 10 m on dry and cold firn (Rignot et al., 2001) and provides information on ice flow and not surface features. In contrast, using feature tracking on optical images (Landsat and Sentinel-2), it is possible to estimate the absolute migration (migration + ice flow) of surface features that could be significantly different from ice flow as for the megadunes.