We performed an overland traverse in austral summer 2016/17 – a joint venture between the Coldest Firn (CoFi) project and the Beyond EPICA – Oldest Ice Reconnaissance (OIR) pre-site survey (Karlsson et al., 2018; Van Liefferinge et al., 2018) (Fig. 1). The CoFi project aims at an improved understanding of firn densification with samples from the EAP. In its framework, five firn cores have been drilled, referred to as B51 and B53 (both drilled in 2012/13) and B54, B55 and B56 (drilled on the traverse in 2016/17).

From Kohnen Station the traverse went to the former B51 drill site. Right after B51 the traverse split up and followed two different legs, to reunite at the OIR field camp at 79∘ S, 30∘ E. After accomplishing the OIR survey and drilling the firn core B54, the traverse continued to the former Plateau Station (abandoned in 1969) and then returned back to Kohnen Station.

We follow Stenni et al. (2017) using the term EAP for the region higher than 2000 m above sea level (a.s.l.).

The traverse covers a region with a 10 m firn temperature range of about -44.5 ∘C at Kohnen Station (Oerter et al., 2000) to -58.4 ∘C at Plateau Station (Kane, 1970; Picciotto et al., 1971), which belongs to the lowest firn temperatures ever recorded (cf. Dome A: -58.3 ∘C; Cunde et al., 2008).

At Kohnen Station the accumulation rate used to be 64 kg m2  a-1 in the period between 1200 and 2000 (Oerter et al., 2004, 1999) with an increasing tendency to over 80 kg m2 a-1 over the last decades (Medley et al., 2018). At Dome Fuji 27.3 kg m2 a-1 was measured (Hoshina et al., 2016). For the locations along the traverse, an accurate value is difficult to obtain. Large-scale accumulation estimated based on remote sensing techniques (Arthern et al., 2006) is assumed to be too high for the EAP (Anschütz et al., 2011). Karlof et al. (2005) determined an accumulation rate of ∼ 45 kg m2 a-1 close to location 5 (Fig. 1); Anschütz et al. (2009) published ∼ 20 kg m2 a-1 for sites between location 8 and B53 as well as OIR camp and Dome Fuji. A high interannual variability of accumulation rate is observed in several places on the EAP (Hoshina et al., 2016, 2014; Oerter et al., 2000). A 1 m deep snow profile can therefore cover a time period of about 4 years at Kohnen Station and up to 20 years on the interior plateau.

While the northern part of the traverse (Kohnen Station – B51) is more strongly influenced by synoptic activities with periodic snowfall (Birnbaum et al., 2006), the interior plateau (OIR camp to Plateau Station) is characterized by diamond dust deposition from a clear-sky atmosphere (Schwerdtfeger, 1969), which was described by Furukawa et al. (1996) as the calm accumulation zone. Wind maps (Lenaerts and van den Broeke, 2012; Parish, 1988; Sanz Rodrigo et al., 2012; van Lipzig et al., 2004) show generally low mean wind speed (around 6 m s-1) from Kohnen Station along the ice divide up the EAP, but lower values for the region around Plateau Station. Due to the prevailing Antarctic high-pressure system over the EAP and the gentle slopes, the katabatic winds reach only moderate wind speeds there. While occasionally snow storms with wind speeds exceeding 15 m s-1 can happen at Kohnen Station, this is not the case on the interior plateau.

For clarity, we define the terms used in the following paragraphs in Table 1.

Along the traverse route, vertical snow profiles were extracted using the snow liner sampling technique, also described by Schaller et al. (2016). Each vertical profile was taken using a carbon fiber tube of 1 m length and 10 cm in diameter. The liner was pushed into the snow until the liner top was level with the snow surface. Afterwards, a snow pit next to the liner was dug and the snow was cut at the liner bottom with a metal plate to take the filled liner out of the pit wall. Both ends were covered with a Whirl-Pack^® plastic bag to reduce possible contamination by touching the liner ends and air ventilation. During the sampling process, the liner was handled carefully to avoid concussions that destroy the original snow stratigraphy (e.g., not to bounce against the liner with the shovel and placing it softly into the sample box). A 1 m snow profile can be retrieved within 15 min. The liners were stored in isolated polypropylene boxes and shipped to the Alfred Wegener Institute (AWI) in Bremerhaven in a continuous cold chain. In total 144 snow profiles in different setups and total lengths were taken (Sect. 2.2.1–2.2.3). All strategies described in the following sections have been applied independently of each other.

The snow liners have been non-destructively analyzed at AWI with the core-scale microfocus X-ray computer tomograph in a cold cell (μCT), specifically constructed for snow, firn and ice cores. For technical details see Freitag et al. (2013) and Schaller et al. (2016). Before the measurement all liners were weighed. The weight of the carbon fiber tube was subtracted. The exact height of filled snow inside the liner was determined using the μCT. Then, ρL was calculated volumetrically. All liners have been measured in a 2D mode using a setup of 140 kV and 470 µA at -14 ∘C. Breaks and lost snow in the snow profiles have been spotted during the scan and corrected (set to NaN) in the μCT density profiles, which have a vertical resolution of ca. 0.13 mm (see Appendix).

For the calculation of the μCT density only the central segment of the liner is used as scattering effects at the outer parts of the liner occur. The used segment corresponds to less than half of the snow volume in the liner. Missing snow at the edges of the profile does not influence the μCT scan. Accuracy of the ρμCT can be affected by the calibration, which is done with three cuboids of bubble-free ice with different lengths in every scan individually, or at the horizontal variability on the very small scale, as the central part of the profile can have a different density than the edges.

It is generally possible that the snow profiles are subject to compression during sampling or transport. Therefore the exact snow volume determined with the μCT is rescaled to the original 1 m length (length of every single snow profile is determined individually) to avoid this potential error source. But lost snow in the liner (or at top or bottom), e.g., in non-cohesive layers (such as depth hoar layers), can lead to lower densities. Thus, ρL is also affected by errors. Conger and McClung (2009) reported that snow sampling devices with larger volumes usually result in higher precision in snow density. The volume of the snow liners (radius: 5 cm, length: 1 m) is 7855 cm3, 16 times the volume with the highest precision in their study. As the volume error among single liners is not known, we assume a 0.3 mm variation in both dimensions (length and radius), resulting in a volume error around 1.2 %. As still small parts inside the liner might not be completely filled with snow (e.g., lost snow during the transport), we estimate the undersampling error of the liner method to be less than 1.5 %. Additional error sources are the precision of the used scale (1 g or 0.03 % compared to the mean value along the traverse) as well as weight variations among the carbon tubes (< 0.1 %). The maximum relative error is estimated to be below 1.9 %.

Both ρμCT1m and ρL are in good agreement with each other (Fig. 4). The differences between the volumetrically calculated ρL and ρμCT1m are on average only 0.6 %. As the μCT density is sensitive to calibration, we consider ρL to be more accurate for a 1 m interval. Some systematically higher values in the μCT measurements can be caused by low-quality calibration in single measurements. Therefore, for the 1 m surface snow density we use ρL. For the comparison of intervals smaller than 1 m, we use the μCT-derived density ρμCT (Table 1).

Fisher et al. (1985) defined stratigraphic noise as a “random element caused by the surface irregularities”, which is present in any taken snow profile or ice core. This stratigraphic noise is mainly caused by spatially inhomogeneous deposition in combination with wind, leading to snow patches or dune structures that usually have a spatial extent of several meters. This stratigraphic noise hampers the representative (i.e., for a certain location or area) estimate of surface snow density, when not considered in the sampling strategy. To still be able to get a representative value or profile (of density or other parameters) at a given spot, several samples have to be taken at a distance, at which they are not subject to the same stratigraphic noise. For example, samples should not be taken from the same dune or snow patch, as these values cannot be considered to be spatially independent. By stacking or averaging independent samples, the stratigraphic noise is reduced. For example, this has also been performed for isotopes (Karlöf et al., 2006; Münch et al., 2016), and a common (annual or seasonal) climatic signal can be retrieved despite a high level of stratigraphic noise.

The (minimum) sampling distance between two samples was quantitatively described for snow density by Laepple et al. (2016). In a 2D high-resolution trench study at Kohnen Station they have shown that the correlation coefficient between single profiles decreases rapidly with increasing distance and settles at a constant value after 5–10 m. In the following we refer to samples taken at this distance as “spatially independent”. Consequently, we consider the multiple snow profiles at one location to provide spatially independent ρL. In the OIR trench, we assume a sampling distance of 5 m between two profiles to be sufficient.

For a representative 1 m ρloc, we aim for a relative error of less than 2 %. To test how many snow samples per location are needed for this representativeness, we calculated σH1m of ρloc. We used the maximum number of spatially independent ρL for ρloc (further called n). We did this for both the multiple liners at the traverse locations and the OIR trench. At the locations along the traverse we use all four available ρL values (n=4) to calculate σH1m of ρloc. In the trench, we created two sets of ρL, which have a minimum sampling distance of 5 m. We get two sets of seven (n=7; using different snow profiles in both sets) and calculated the mean value of both sets. We then derive the standard error (σn), which depends on the number n of ρL at a given location, by 1σn=σH1m/2;n=1, with the denominator being a varying number of snow profiles from 2 to n-1. This means, for example, when using seven profiles (like one set in the OIR trench), we are able to calculate the standard error for two to six profiles. In this way we use the maximum sample size without an artificially caused bias in the data. This can happen, for instance, by creating sets with a small sample size and picking ρL randomly. Accordingly by (a) using large volumetric samples we improve the accuracy, and by (b) using several profiles at each location we improve the representativeness of the density values derived for each location. We are therefore able to deliver a more accurate and representative density of each site, compared to previous studies.

We pooled several snow profiles for further investigation to characterize the surface density of a larger (≥ 10 000 km2) region. We chose a minimum number of 10 profiles (0–1 m) per area. We followed the classification of Furukawa et al. (1996) as well as possible and used the 3500 m a.s.l. contour line as the approximate boundary between different wind and accumulation regimes on the katabatic wind zone and the interior plateau (calm accumulation zone). This way we classified one major area “ascending plateau area” (AP) with 64 profiles, covering roughly 140 000 km2 between Kohnen Station and OIR camp and the smaller “interior plateau” (IP) with 29 profiles between OIR camp and Plateau Station (28 500 km2). We did not include the OIR trench, as this specific location would have been overrepresented. The area around B53 (28 500 km2) was treated as a separate area as it is on the interior plateau close to the ice divide (“B53 and vicinity” – 10 profiles). Additionally, we handled the area around Kohnen Station (Ko) with roughly 10 000 km2 as another separate unit (“Kohnen and vicinity” – 45 profiles). The sample availability at Kohnen Station from other studies is sufficient, several liners from other sampling programs in seasons 2015/16 (16 profiles) and 2016/17 (18 profiles) have been added to the evaluation. The areas are color-coded in the overview map (Fig. 1).

As we present density data on different scales, in this context we use the term “local” scale for distances between profiles at one location and the area around a sampling location (i.e., tens of meters, Table 1). In contrast, the term “regional” scale is used for distances between several locations (100 to 1000 km) and areas in the dimensions of the subregions defined above. For all subsets, we present a spatial distribution of ρL and ρloc.

The relative surface elevation of the OIR trench was measured using optical leveling at each profile position and in between two consecutive profiles. Additionally, at the OIR camp and Plateau Station surface roughness transects were measured. The optical level was placed at the transect starting point. The first height measurement was done at a 10 m distance to the starting point and repeated every 2 m up to a 58 m distance relative to the start, resulting in 25 measuring points per transect. In total six transects have been done at one location with 1 m lateral spacing between them.

ρL ranges in the OIR trench from 347 to 380 kg m-3. We calculated ρloc for the OIR trench (± standard deviation) with 365±10 kg m-3, which is 3.1 % higher than for the whole traverse (Sect. 3.2). σH is between 10 and 27 kg m-3 for 0.1 m sampling intervals and between 5 and 10 kg m-3 for 1 m sampling intervals (Fig. 5 and Table 4). The highest σH0.1m can be found in the top 0.3 m. σV3m of the 3 m profiles is 34 kg m-3 (Table 4).

Here we present data from Sect. 2.2.1 and 2.2.2. Along the traverse we find ρL ranging from 324 kg m-3 (pos. 22C) to 382 kg m-3 (pos. 16A). The average ρL calculated from 114 liners along the traverse is 354±11 kg m-3 (Fig. 6).

ρloc (Table 1) is calculated from multiple snow profiles (Sect. 2.2.2) at each location. At locations 21 and 1 close to Kohnen Station we find the lowest ρloc with 344 and 345 kg m-3, respectively. The highest ρloc is found at position 5 with 372 kg m-3 (Table 2). The average ρloc along the traverse is 355±8 kg m-3. To characterize the surface variability, we calculated σH1m for each location separately. The minimum σH1m is 2 kg m-3 at position 20 (and position 13 with only two profiles taken); the maximum σH1m is 15 kg m-3 at position 22.

A detailed overview of all ρloc and σH1m along the traverse can be found in Table 2, and a visualization is in the appendix (Fig. 13).

In Fig. 7 we compare the calculated σn according to Sect. 2.4. For four spatially independent snow profiles in the OIR trench, we get a value for σn of less than 1.5 % (4.9 kg m-3) relative to ρloc (355±2 kg m-3). We note that on average σn in the OIR trench is higher than the average of the four areal subsets (7.0 kg m-3 in contrast to 6.1 kg m-3 for two profiles and 5.7 kg m-3 in contrast to 5.0 kg m-3 for three profiles).

Consequently, we consider four snow profiles to be sufficient for a ρloc with σn of less than 2 %. Unfortunately, we cannot test a number of profiles higher than six. But assuming a constant σH1m, seven spatially independent profiles are needed to assure a relative σn of less than 1 %.

In the spatial density distribution of ρL and ρloc, we find similar values for Kohnen and vicinity (352±1 kg m-3), ascending plateau area (356±1 kg m-3) and the interior plateau (355±2 kg m-3) (Fig. 8). These have less than 1 % difference from the average value of the whole traverse. Only B53 and vicinity show lower density values (349±3 kg m-3, -1.7 % compared to the traverse location mean density 355 kg m-3).

Looking at the density distribution of the high-resolution μCT density profiles (for details, see Appendix), we find a normal distribution of the snow density in the first meter (Fig. 9). We see a shift towards higher densities in the OIR trench and a higher probability for lower densities in B53 and vicinity, but in general a similar distribution of density in all subregions is found.

We calculated the confidence interval (95 %) of ρL for each respective subregion (Table 3). We want to stress that the number of samples of “B53 and vicinity” is lower than recommended for this method. The mean value for the traverse is represented in all four intervals of the subregions. We note that the interval for Kohnen and vicinity just includes this value.

The snow density directly measured at the surface in general shows high spatial variability (Figs. 5 and 10). To characterize the spatial variability of density in a given area (tens of meters for traverse locations and trenches, hundreds of meters for Kohnen Station), we use the parameter σH. For a comparison we used snow liners along the traverse (liners sampled at OIR trench presented in a separate column), from Kohnen Station (Schaller, 2018) and from East Greenland Ice-core Project (EGRIP) camp site (75∘37′ N, 35∘59′ W; 2702 m a.s.l.). Shown is also σV for the respective areas, which can be interpreted as temporal (seasonal or annual) variations in density. We computed both (σH and σV) for 0.1, 0.5 and 1 m intervals each (Table 4).

The maximum height difference between the lowest (first) and highest (last) profiles in the OIR trench is 38.5 cm. The height values of each position are given in the appendix (Table 6). We find significant differences in the surface topography at both places. At OIR camp the height differences between the lowest and highest points of the measured transects are 60 % larger than the height differences at Plateau Station (Table 5). The variation in height differences between the six transects at each location is low with a standard deviation of 2.4 cm (OIR camp) and 2.0 cm (Plateau Station).

To discuss the 1 m snow density using the liner technique, we compare our dataset with data by Oerter (2008). In that study, snow pits with 20 km spacing have been dug and sampled along a small transect from Kohnen Station upstream towards B51 (comp. Fig. 1). A detailed map of the sampled region by Oerter (2008) is available in Huybrechts et al. (2007). Snow density has been measured volumetrically in each snow pit using discrete samples in 0.1 m depth intervals. We compare our results with density data from locations 1 to 4 (including single snow profiles in between) at two different depth resolutions (0.1 and 1 m). For our study, we use ρμCT0.1m and ρL. For the 1 m interval from Oerter (2008) we use the average density value of all discrete samples between 0 and 1 m.

ρ1m values from both studies are in good agreement with each other. ρ1m derived with the liner method tends to be 1 %–5 % higher than the one from Oerter (2008) (Fig. 10). A higher discrepancy can be seen in the mean density of the upper 0.1 m. While we find on average ρμCT0.1m=349 kg m-3 from liner measurements, ρ0.1m for Oerter (2008) is 293 kg m-3. The calculated σH0.1m over the whole distance is 31 kg m-3 for our study and 25 kg m-3 for Oerter (2008). Interestingly, ρ0.1m in Oerter (2008) is always lower than ρ1m, which is not the case in samples from our study. Due to the soft and unconsolidated snow at the surface, we assume that the undersampling error is higher at the surface for small sampling devices, which forces a systematic error towards smaller values (Fig. 10). Snow in greater depth has undergone sintering processes and is more coherent; therefore the undersampling error should also be smaller. Additionally, a systematic error with increasing depth in the data by Oerter (2008) cannot be excluded, as the sampling device (core cutter) might densify the snow with each interval due to the thick wall in relation to the sampling volume. In contrast to other devices, the liner method preserves the original stratigraphy of the snow column. In combination with the μCT measurement on different chosen depth intervals, this results in a density value with less uncertainty, especially for small sampling intervals at the snow surface. Despite the sampling strategy, the difference between both datasets can be caused by different weather conditions during the sampling. This affects in particular the upper centimeter of the snow column.

In the following we discuss the advantages of a 1 m snow density in contrast to smaller depth intervals. In this context we refer to the data presented in Table 4. At sites with accumulation rates higher than 100 kg m-2 a-1 (e.g., EGRIP), small sampling intervals (< 0.5 m) do not contain the seasonal or annual variability over several years (see also data by Oerter, 2008, in Fig. 10), at sites with lower accumulation (in this context < 60 kg m-2 a-1) the density might be masked by the high stratigraphic noise. Both effects can be seen in the low σV0.1m in contrast to σV1m looking at data from different sites in Table 4. Higher σV1m in snow profiles from EGRIP are caused by a clearer seasonal density cycle, which is barely or not detectable on the EAP. This can be explained with higher temperatures as well as higher accumulation rates at EGRIP. In the case of surface melting like in the year 2012 (Nghiem et al., 2012), σV1m can be even higher. We find lower σH at the surface in samples from EGRIP in contrast to EAP. This can be explained by the non-uniform deposition causing high undulations in the surface topography. We measured the topography in the form of dune heights (Table 5), which are often 30 to 40 cm high and exceed the yearly accumulation by far. Snow layers do not form as spatially consistently as at sites where the (predicted) yearly layer thickness is larger than the amplitude of dunes. This also affects the snow density as the signal cannot form homogenously over a larger distance and causes larger σH. For all presented sites, the σH0.1m is 2.4 to 4 times higher than the σH1m, which is explainable by the more comprehensive density spectrum over larger depth intervals. This high horizontal variability is mainly caused by the existing small-scale topography, in particular dunes. The variability decreases below the maximum measured dune heights of 30–35 cm below the surface. These dunes have a higher snow density (Birnbaum et al., 2010) than snow that gets deposited in local depressions due to enhanced wind packing (see Sect. 4.3). This is also visualized for the OIR trench in Fig. 5. A snow patch of low density can be seen at the surface between 0 and 5 m (horizontal distance) and rather high density between 18 and 25 m (horizontal distance) (Fig. 5c). This illustrated the need to choose a far enough distance to reduce the effect of stratigraphic noise (Sect. 2.5).

The temperature-dependent densification effect does not affect the 1 m snow density substantially. By comparing all μCT density profiles over depth, we cannot see a significant increase in density over the first meter. Also according to the model by Herron and Langway (1980), at a temperature of -43 ∘C (annual mean air temperature at Kohnen Station after Medley et al., 2018), the increase in snow density by densification from the surface to 1 m depth is 10 kg m-3. At a -53 ∘C annual mean air temperature (-10 ∘C compared to Kohnen Station) the densification is roughly 8.3 kg m-3. A temperature change of -1 ∘C would lower the densification-induced density by about 0.17 kg m-3.

In summary, due to the high snow density variability in the upper decimeters of the snowpack, we suggest the 1 m density as a feasible approach to derive the surface snow density independently of local recent weather conditions. For a representative value, at least four samples should be taken per location with the respective sampling distance. The densification of snow over the first meter is negligibly small. Furthermore, we want to advert to the time efficiency of the liner method here. A 1 m snowpack density with four samples can be obtained within 1 h. Even if a high-resolution study in a snow pit is done, a snow profile using a liner can always be added to the discrete sampling in a snow pit for comparison.

Long-term changes in temperature, accumulation rate or wind systems can also affect fluctuations in density. At Kohnen Station a 1 ∘C temperature rise per decade has been recorded by an automatic weather station, jointly operated by the Institute for Marine and Atmospheric Research (IMAU) and AWI (Reijmer and van den Broeke, 2003) over the past 20 years and discussed by Medley et al. (2018). Recent studies postulate in some areas of Antarctica, partly also on the EAP, an increase in the accumulation rate (Frieler et al., 2015; Medley and Thomas, 2019) caused by a temperature rise. However, accurate accumulation rates for the interior EAP are hard to determine and are generally overestimated (Anschütz et al., 2011).

We test the impact on surface snow density of a 1 ∘C temperature rise as well as a 15 % increase in accumulation rate at Kohnen Station. We use the surface snow density parameterization after Kaspers et al. (2004): 2ρ=7.36×10-2+1.06×10-3T+6.69×10-2A˙+4.77×10-3W, where T is the 10 m firn temperature (K), A˙ the accumulation rate (kg m2 a-1) and W the mean wind speed (m s-1).

For comparison we also use the surface snow density parameterization after Sugiyama et al. (2012), as this one has been calibrated in particular with samples along a traverse over the EAP: 3ρ=305+0.629T+0.150A˙+13.5W, with T (∘C), A˙ (kg m2 a-1) and W (m s-1) at the given location.

A temperature rise of 1 ∘C and an increase in accumulation rate of 15 % at Kohnen Station would increase the surface snow density by 1.7 kg m-3 according to Kaspers et al. (2004) and by 2.0 kg m-3 according to Sugiyama et al. (2012). According to both parameterizations, the difference in density between this study and Oerter (2008) cannot be solely attributed to these climatic changes as both potential increases are inside the error range of ρloc. Despite uncertainties in the precision of the sampling method or natural (climatic) variability, the discrepancy in surface density between both datasets can also be caused by stratigraphic noise over time. To give an example here, we compare ρloc of snow profiles from Kohnen Station taken in two different seasons at the same position. We use 17 profiles along a transect line with 0.5 m spacing from the season in 16/17, which were resampled in the season in 18/19 (both unpublished). The climatic conditions during this time span did not change significantly. ρloc(16/17) and ρloc(18/19) both have the same value and the same standard deviation 350±6 kg m-3. Although this example can give an estimate for the robustness of our density measurements using the liner method, we are not able to completely decouple the spatial variability and the temporal variability as we cannot resample the exact same position (and thus the exact same snow).

In a second test, we use an annual mean temperature of -50 ∘C (223.15 K), accumulation rate of 40 kg m2 a-1 and wind speed of 6 m s-1, which are roughly the mean values of the area covered with the traverse. While the parameterization by Sugiyama et al. (2012) is fairly accurate compared to our 1 m snow density (+5 kg m-3), keeping the temperature and accumulation rate constant we have to increase the wind speed to 9 m s-1 to reach the surface snow density along the traverse using the parameterization by Kaspers et al. (2004).

In general we conclude that several parameterizations for the surface snow density (Kaspers et al., 2004; Sugiyama et al., 2012) need further tuning for regions with low accumulation and low temperatures like the EAP. Rather, local parameterizations should be used for regions with similar environmental conditions instead of continent-wide parameterizations.

In order to overcome the sparsity of ground truth surface snow density, regional climate models and derivatives with adequate snow deposition modules are often used to obtain estimates of accumulation and surface snow density on a full regional scale. Ligtenberg et al. (2011) presented firn density averaged from surface to 1 m depth over a period from 1979 to 2011. It is forced by RACMO2.3p1 mass fluxes and skin temperature and gridded at 27 km resolution.

Compared to the firn densification model presented by Ligtenberg et al. (2011), we find systematically higher values for density on the interior EAP than the model predicts for the same locations. While ρloc spans the range from 346 to 372 kg m-3, the firn model provides a range from 308 to 332 kg m-3 (Fig. 11). Having sound statistics at these locations, we exclude the systematic bias to be caused by our observations, but rather we assume a shortcoming of the model to yield densities which are about 10 % too low. This could be caused by a multitude of reasons, e.g., model physics, spatial and temporal resolution, or forcing. As the parameterization by Kaspers et al. (2004) provides density values closer to our ground truth data than the model output by Ligtenberg et al. (2011), we suggest revising the used slope correction (Helsen et al., 2008) for the EAP.

Our observation is consistent with recent field observations on the EAP (Sugiyama et al., 2012) or snow density collections from over 2 decades (Tian et al., 2018). Sugiyama et al. (2012) found a density around 350 kg m-3 for the same depth interval (0–1 m) along a traverse between Dome F and Kohnen Station, with a similar spatial variability. Nevertheless, we cannot detect a clear trend in density along the whole traverse route. A potential reason might be the increase in elevation, distance to the coast and major Dronning Maud Land (DML) ice divide on the one hand and the decrease in temperature as well as accumulation rate (Fig. 11) on the other hand. As the sampling took 6 weeks in total (Table 2), we exclude an effect of seasonal density variability as well as a significant effect of accumulation during the traverse (as the only observed accumulation on the traverse was a few diamond dust events above 3500 m a.s.l. during the nights and some drift snow). We explain the increase in surface density along the ice divide from Kohnen Station towards B51 (Figs. 8b and 11) by smaller grain sizes due to decreasing temperature. The combination of the lower accumulation rate and longer exposition and mixing at the snow surface seems to create a higher surface snow density here. The observation of this systematic change in density is also visible in results of Sugiyama et al. (2012) and not captured by firn models. In fact, the model by Ligtenberg et al. (2011) shows the opposite trend along this traverse section (km 0–500 in Fig. 11). High density at B51 goes along with stronger dune formation than at Kohnen Station, which was observed to increase along this traverse part, and higher potential for wind packing due to lower accumulation rates. This is consistent with observations of dune formation at wind speeds exceeding 10 m s-1 (Birnbaum et al., 2010) or observation of wind-packing events (Sommer et al., 2018) causing increased snow density.

Modeled density is parameterized by wind speed, but the process of denser packing by wind scouring and redistribution over the time until the snow is finally buried might be underestimated. We assume that the modeled low density values for locations 14 and 15 (Plateau Station and B56, Fig. 11) in the calm accumulation zone are caused by the relatively low wind speed (Lenaerts and van den Broeke, 2012; Sanz Rodrigo et al., 2012), in combination with low temperatures and humidity (Picciotto et al., 1971). But the wind on the interior plateau is not strong enough to cause wind packing and sintering of snow crystals. It rather redistributes them smoothly at the surface, which also happens at low wind speeds. This process is significantly different from wind packing at high wind speeds. Thus as the sintering process is prolongated it increases the density on the long term, which also causes an increase in density variability at the surface. But as the low densities cannot be seen for the whole interior plateau region (Fig. 8d), we consider it to be a process that needs very specific settings on the high plateau than average characteristics. The abundance of wind speeds higher than 10 m s-1 might be a limiting factor in this context.

Different environmental conditions at B53 and vicinity might cause lower density here as well (Fig. 8c). High σn for subset B53 and vicinity should not be over-interpreted, as only one sampling location with four profiles is present there. Still, σloc is highest here amongst all locations with multiple liners along the traverse (compare also σH1m in Table 2). An explanation can be a different wind and accumulation regime at the distant side of the ice divide causing high heterogeneity on a very small scale.

Small fluctuations in density within the error range at nearby locations can be explained by stratigraphic noise (Laepple et al., 2016; Münch et al., 2016). Stronger variations in density, e.g., beyond 1 standard variation, can be caused by a complex interaction between wind speed and surface roughness on the small scale but also have been shown to originate from dynamic interaction of ice flow over bedrock undulations, thus altering surface slope and in turn elevation and accumulation rate on the large scale in this region (Anschütz et al., 2011; Eisen et al., 2005; Rotschky et al., 2004). For a detailed conclusion regarding the influence of bedrock topography on the density fluctuations in our data, we consider the local scale (10 m) to be too small and the regional scale (100 km) to be too large. We suggest a different sampling scale (i.e., 10 km spacing of representative density) for this purpose.

As already stated above, we cannot conclusively attribute a cause to the model behavior as we also neglected the atmospheric forcing of the firn densification models, which could explain parts of the density discrepancy between field data and modeled values. Unfortunately, it is also difficult to pin down the mechanism for the observed systematic spatial distribution of density. As the surface snow density parameterizations are mainly dependent on temperature and wind speed, the influence of both might be too high while processes acting on the snow surface like snow redistribution and packing play a major role in snow density. Obviously, a dedicated sensitivity study with a snow deposition and firn model is needed to discriminate the various processes affecting post-depositional snow metamorphism and densification. We suggest setting up a specific model test designed for the EAP and using datasets like ours and those from comparable studies as the standard against which to evaluate model outcomes.

Firn densification models are used in altimetry to convert height changes of the ice sheets to mass changes. The more accurate the modeled firn density provided by these models is, the lower the uncertainties in the calculated mass changes will be. Therefore, our presented density data can be of particular interest to improve the accuracy of ice sheet mass balances.

One way in altimetry is to use a simple density mask as an input parameter (e.g., McMillan et al., 2014; Schröder et al., 2019). In regions with a strong influence of ice dynamics, only the density of ice is used. In the remaining areas, also in large parts of East Antarctica where the ice flow velocities are low (Rignot et al., 2011), the density of firn is used. In this conversion, uncertainties in snow density have a direct impact on the result in mass. In our case, the 10 % density underestimation in previous studies can lead to a 10 % mass error (e.g., Alexander et al., 2019). Shepherd et al. (2019), in contrast, use firn or ice density by defining areas of dynamic imbalance, which depend on surface uplift or lowering in relation to firn column changes. This method is even more sensitive to uncertainties in the firn densification models, as it subtracts variations in firn density over time.

Despite the impact of density on the height-to-mass conversion, the snowpack properties can also influence the microwave penetration into the snow and therefore considerably affect the radar altimetric measurements. Generally, snow properties like density, grain size and liquid water content can influence the permittivity (Mätzler, 1996), but spatiotemporal variations in these parameters also influence the measurements (Davis and Zwally, 1993). Furthermore layering of the snowpack seems to affect the penetration depth, like shown in Slater et al. (2019) for Greenland. Interestingly, the density distribution of density (Fig. 9) does not show as much difference between the subregions as previously expected due to different accumulation rates. While we can see differences on the local scale (OIR trench), on the regional scale the vertical density distribution of the subregions is very congruent. Therefore further high-resolution studies on the vertical variability of the snowpack are needed on the EAP, especially with regard to high surface variability.

In the following we provide an idea of how the mass of the firn column depends on the choice of the surface density. Based on our findings we employ a simple quantitative calculation of the mass in the firn column with the density data presented in this study (average ρloc) using the semiempirical firn densification model by Herron and Langway (1980). We use an annual mean temperature of -50 ∘C and an accumulation rate of 40 kg m2 a-1 as input parameters. We use the two different surface snow densities ρ0(1)=320 kg m-3 (Ligtenberg et al., 2011) and ρ0(2)=355 kg m-3 (this study) and sum up the water equivalent (w.e.; based on 1000 kg m-3) in the firn column.

We calculated 59.0 m w.e. for ρ0(1) and 61.0 m w.e. for ρ0(2) in the firn column down to the firn–ice transition in 92.9 m, where scenario ρ0(2) reaches the critical density of 830 kg m-3. The calculation is in good agreement with firn density (μCT) measured in core B53 (unpublished data). This difference of +2 m w.e. corresponds roughly to an underestimation of 3 % mass for the firn column only using the modeled initial density. Other effects like an overestimation of the accumulation rate on the interior plateau are not taken into account.