At Dome C (75°06^′ S, 123°20^′ E), we excavated a 50 m long and 1.5 m deep straight snow trench in which 35 vertical snow profiles were sampled over three weeks in December 2019. The trench was perpendicular to the main orientation of the wind and the sampling was carried out on the wall face sheltered from the wind, which is southerly on average at Dome C . Individual samples were collected directly from the wall in 12 mL Corning tubes, and the 2D horizontal and vertical coordinates of the sample were recorded. We used a laser level to establish a common reference height for all of the profiles, materialized by a horizontal string (visible in Fig. b and d). Throughout the article, we refer to the absolute depth, z, of each sample as the vertical distance to this common height reference, in contrast with the (regular) depth usually defined as the vertical distance to the local surface. The surface topography was rather flat (50 cm < z < 60 cm) in the left section of the trench (0–35 m) and a 30 cm high dune (minimum z= 20 cm) was present on the right section (35–50 m) (Fig. a). Vertical profile sampling took place every 2 horizontal meters along the whole trench and in addition finer horizontal resolutions (from 10 cm to 1 m) was applied near the left end of the trench (between 0 and 5 m), on the ascending slope of the dune (between 38 and 44 m), and on top of the dune near the right end of the trench (around 48 m). Along this trench, the profiles are labeled P followed by their horizontal position in meters measured from the left end of the trench. Most profiles were sampled with a 3 cm vertical resolution and four profiles, P0, P0.3, P47.7 and P48, had a higher vertical resolution of 1.5 cm. P0 was replicated 30 cm to the left. Similarly, P0.3 was replicated 20 cm to the left, and P48 was replicated 30 cm to the right. Because the decorrelation length is roughly 1 m (Appendix, Fig. ), it is possible to mix snow from profiles 10 or 30 cm away, respectively. Here, some of the snow samples of high resolution profiles did not have sufficient amount of snow to realize the entire span of chemical measurements (anions and cations), so in this case, we used extra snow from the replicate profile to complement. For simplicity, we kept the labels P0, P0.3 and P48 for these composite profiles. This short horizontal spread does not impact the results as we later focus on the coarser 2 m horizontal resolution.

After sampling, the snow samples in the Corning sampling tubes were parafilmed, and kept frozen during shipment to laboratories in Europe. The samples were analyzed by ion chromatography at the Alfred Wegener Institute (Bremehaven, Germany) and the Institut des Géosciences et de l'Environement (Grenoble, France) for the concentration of anions (SO42-, NO2-, NO3-, MSA, Cl-, Br-, oxalate, F-, acetate and formate) and cations (Li+, Na+, NH4+, K+, Mg2+, and Ca2+).

We also deployed a Snow Micro Pen to estimate snow hardness. This measurement was carried out after the chemical sampling, with the probe applied 20 cm behind the exposed wall of the snow trench. Measurements were realized every 25 cm along the trench (200 profiles). The vertical resolution is extremely fine (0.004 mm) but the measurements were averaged at 1 cm resolution to obtain a macroscopic significance. The snow hardness probe was set to penetrate up to 150 cm, but it stopped earlier for several profiles due to too hard layers, leaving a few incomplete profiles. The Snow Specific Surface Area (SSA), a measure of the grain to pore interface surface area relative to the ice mass (unit: m2kg-1) was measured using ASSSAP . Snow density was measured using a scale and squared metal sampler tool with a known volume. Both SSA and snow density were measured at 3 cm resolution along the same profiles as the chemistry samples.

The measurements carried out on each vertical profile were combined to create a 2-dimensional image for each chemical and physical tracer in the upper 1.5 m of snow (Sect. ).

In order to derive accumulation from the snowpack, we constructed an age model for the trench. Instead of computing an age model for each profile independently, we proceed in two steps: in the alignment step (this section), we relate each depth profile in the trench to the depth of a reference profile by a squeeze and stretch function. The tie points for this transformation are obtained by matching features such as peaks and crests in the signal of three selected tracers: SO42-, methanesulfonic acid (MSA), and SSA. In the dating step (next section), we compute an age model for a single reference profile. The age model for the entire trench is then obtained by applying the transformation of the alignment step to the reference age model.

To select tie points across the pits, we developed a versatile alignment software in Python hereafter referred to as Alice (ALignment interface for ICE cores). This software allows the user to select tie points across a variety of chemical and physical tracers. We used the three aforementioned tracers, with SO42- chosen as the primary tracer. Sulfate SO42- is considered as a being preserved in snow , showing only moderate peak broadening except over timescales of thousands of years, due to diffusion . Using Alice, a profile to be aligned is viewed against a previously selected reference profile. As the trench was sampled at the beginning of summer, before significant metamorphism had occurred , we used the high SSA values (> 40 m2kg-1) of recently deposited winter snow to match the topmost portions of the profiles and to detect hiatus in snow deposition. We have performed a sensitivity test to ensure that this was also an appropriate threshold for the trench dataset. We counted all SSA values in the trench dataset (26 evenly distributed profiles) above a certain threshold, for threshold values ranging from 20 to 60 m2kg-1. We see a clear transition around 38–42 m2kg-1, where the datapoint count increases sharply under 38 m2kg-1, indicating a longer persistence of such SSA values during grain coarsening. This confirms that 40 m2kg-1 is a sweet spot to identify fresher snow (Supplement, Fig. S1). The rest of the profile is aligned by matching sulfate peaks, first with the largest peaks and then refined with sub-features. If the sulfate peak matching is ambiguous, MSA profiles were consulted. We note that MSA is well preserved in the first 100 cm depth, despite MSA disappearing at higher depth .

We chose to use the single raw profile P0 as the reference profile because of its higher resolution (1.5 cm instead of 3 cm) and the presence of several sulfate features that are suitable as an alignment target. Despite these advantages, we observed that P0 features a lower SSA in its top snow than other profiles, and that on some profiles up to 3 cm of upper snow could not be matched to this reference. We interpret these facts as a missing snow layer at the top of P0. To solve this issue, we completed the 3 cm upper part of P0 with the mean surface snow properties in the 3 cm upper part of all profiles over the trench. This permits us to align all profiles with respect to the reference up to the surface (Supplement, Sect. 3.1). Similarly to this missing snow layer at the surface, in any single raw profile, there will be missing layers that can hinder alignment with other profiles. Yet, we chose to stick with aligning against a single reference profile because a pre-averaged profile is too smooth to obtain confidence in identifying successive peaks. We realized sensitivity tests about the impact of aligning on a single reference profile, as described in the Discussion Sect. .

To summarize, the alignment routine goes as follows: (1) Display a profile to align P on top of the reference profile R leveling the top of the profiles; (2) looking at SSA values, evaluate whether top of the profile P contains fresh snow, or on the contrary is subject to an annual hiatus, and choose the topmost tie point (Fig. a and d); (3) align the rest of the peaks with sulfate, starting with main crests and peaks, and refining with sub-features (Fig. b and e). (4) in case of ambiguous peak matching, switch to MSA profiles to resolve uncertainties (Fig. c and f).

We construct an age model for the reference profile. This age model is then propagated to the aligned profiles in the entire trench. Based on previously published snow accumulation rates (8 cmyr-1, ), we did not expect to reach the current depth of the Pinatubo sulfate level (about 2 m depth), deposited in Antarctica in 1992 , so we have no volcanic horizon in the trench to serve as a dating point. Instead, our age model relies on similarities between the trench reference sulfate profile and sulfate concentrations in previous snow pits dug in the Dome C area in the 2000–2020 period. We were able to align these sulfate profiles together (Fig. ) and considered that the top of each of these snow pits is an certain absolute dating point on the reference profile.

From the set of 22 snow pits dug at Dome C in the last 20 years we selected 13 that could be aligned with the reference profile of the trench (the others showed no remarkable feature in common with the reference in their upper 15 cm section). This 9/22=41 % discarding ratio is comparable to the observations of , where they showed that a volcanic peak only has 60 % chance of being present in any two given cores. To construct the age model of the reference profile, we chose to only keep the top of each snow pit as an absolute tie point, as it is the date known with the highest confidence. Despite this accuracy in sampling date, we take into account the fact that the top of each snowpit might have been subject to erosion, as up to 20 % of profiles might have snow missing for a given year . We do this as we deal with peak matching uncertainty: error bars were assigned to the tie points depth on the reference profile. The error ranges between 5 and 10 cm depending on the quality of the match in sulfate signals. Small errors (5 cm) are assigned to the tie points just before a large and recognizable feature such as the broad sulfate peak in the 90–120 cm depth (December 2005), but conservative errors (10 cm) are assigned to tie points in a series of similar peaks, or in a section with little overlap, such as at depth 140–150 cm, corresponding to December 2000. This later translates to an uncertainty on the age which contains the uncertainty on the age of snow in the upper section of the snowpit (very likely less than a year). The alignment was realized for the 13 profiles, giving 13 dating points on the reference profile. However only six dating points were kept for the age model due to several snow pits being dug within the same month, resulting in duplicate tie points (Table ). For a comprehensive use of all the data, we ensured that the error bars associated with the dating capture the variability observed across these duplicates, which represents the fact that snow accumulation can be uneven in the upper sections of profiles of a given year. We used a different approach to obtain a dating point for the bottom of the reference, where alignment with an older snow pit is not possible due to the lack of overlap. The snow pits covering the Pinatubo horizon were associated with a linear age model between the date of sampling and 1992. This age model was then transferred to the reference after alignment (Fig. b and d, upper x-axis). The mean of these age models is used to get a temporal tie point for the bottom of the reference profile. The error bar in dating for this tie point is the standard deviation of tie point age among age models. Figures for the alignment of all snow pits, including duplicates, are presented in the Appendix (Fig. ).

In the dating routine, for visual aid, the snow pit to be aligned and the trench reference profile have a depth offset corresponding to the accumulated snow between times of sampling. The depth offset was calculated using 8 cm snow per year .

The accumulation rates we calculated in this study are compared to other available datasets. First, the changes of elevation measured every year around Dome C since 2005 in the GLACIOCLIM stake network which can be converted into accumulation. This network consists of 50 stakes located 2 km south to the Concordia station (75°12^′ S, 123°32^′ E). The stakes are spaced approximately 40 m apart, arranged in a 1 km × 1 km cross. The stakes are measured every year. Second, we compare our results with 3 years of high resolution accumulation data estimated from the elevation maps measured by an autonomous surface laser scanner called Rugged LaserScan (RLS), which was operational from January 2015 to December 2015 (period 1) and from January 2016 to December 2017 (period 2) . The timeseries has daily resolution. The spatial setup of the instrument between both periods was different (total scanned area of 40 and 110 m2 and horizontal resolution of 2 and 3 cm respectively). We only use the second period for detailed spatial analysis as its spatial extent is closer to that of the trench, and the continuous 2 years of data are necessary to capture important features of the accumulation dynamics.

The elevation change obtained from both the stake network and the laser scanner is converted into accumulation rate in mass equivalent using constant density value. Following the literature, we use a reference surface density of 320 kgm-3 . In addition, we present the results obtained with a density of 295 kgm-3, which seem to be better suited according to our observations. This latter value is based on the trench density profiles (Fig. ) and on other published data reporting density around 300 kgm-3 .

We also compared our results with the outputs of the fifth generation ECMWF Reanalysis (ERA5) estimating the accumulation as the difference between the total precipitation variable (mmw.e.) and the total evaporation and summed over a year. The output of the reanalyses is selected at the grid point nearest Dome C (75°00^′ S, 123°25^′ E). ERA5 accumulation rates shows good agreement with observations in Antarctica at the regional scale .

Sulfate concentration shows a high profile-to-profile, and intra-profile variability (Fig. ), characterized by a mean value of 72 ngg-1 and a standard deviation of 20 ngg-1. Global patterns of high and low concentrations are present across the trench, with small to high differences (±10–30 cm) in absolute depth. For instance, around z=80 cm, sulfate concentrations above 100 ngg-1 were measured in almost all profiles. Similarly, local minimum concentrations of about 50 ngg-1 were measured in almost all profiles around z=140 cm. The deepest coherent layer which can be identified in the aligned trench, corresponding to a minimum sulfate concentration of 50 ngg-1, is located around z=180 cm. There were about 15 measurements on the reference profile P0 outside of the ±1 standard deviation range, which provided a convenient target for alignment with the other profiles. We refrained from interpreting these cycles as seasonal signal, and focused on identifying global multi-annual patterns between profiles. These patterns suggest that the chemical tracers can be used to obtain tie points between the depth profiles in the trench (Sect. ). From these tie points, isochrones are drawn across the trench, as represented by the gray dashed lines in Fig. a. We assume that these isochrones are associated with past accumulation events , and therefore represent past configurations of the surface topography. An inspection of the density profile against depth shows that density is about 10 % lower in the upper 6 cm section of the trench compared to the full profiles (295 kgm-3 vs. 328 kgm-3), although no significant trend is found in the average density profile up to 1.5 m depth (Appendix, Fig. ). This indicates that deformation of surface features is less than 10 % during the burial process. Note that the dating of the isochrones is carried out in Sect. .

The chemical tracer tie points are used to align the profiles together. The aligned trench shows an increase in inter-profile correlation (r‾=0.3) (Fig. S3) compared to the leveled trench (r‾=0.1) (Fig. ), as expected since sulfate is used to align the trench. It is noteworthy that in horizontal averages, phase differences in the signal (phase noise) due to different deposition depths have been significantly reduced in the aligned trench compared to the leveled trench (Appendix, Fig. ). The use of an independent proxy is necessary to validate the alignment, which is provided by snow hardness measurements (Sect. ). The horizontal decorrelation length of the sulfate signal (Appendix Sect. A6) is 1.0 m in the leveled trench and 1.4 m in the aligned trench. This supports the use the evenly 2 m spaced out subset of profiles (Fig. a, bold labels) when computing horizontal averages of tracer or accumulation rates, in order to avoid local scale correlation of the time series.

The age model obtained on the P0 reference profile is shown in Fig. . The dating is calculated by linear extrapolation of eight temporal tie points over the depth of the reference profile as detailed in Sect. .

The first dating point is on top 1.5 cm of the reference profile, which corresponds to the most recent snow deposition before the date of sampling (December 2019), with no error, since there is no accumulation hiatus at the surface in the reference profile (Sect. 3.1).

The next six dating points were obtained by alignment of the reference profile to snow pits dug in the Dome C area over the past 20 years. The depth uncertainty of dating points (horizontal error bars in Fig. ) leads to an uncertainty in the snow age at a given reference depth (vertical error bars) ranging from ±0.5 years for the January 2019 tie point to ±1.5 years for the January 2012 tie point. The average limited spread of the individual age models (doted colored lines in Fig. ) has a mean standard deviation of 0.7 years between models and gives confidence in the accuracy of the trench age model at ±1 year.

The last tie point at the bottom of the reference profile is obtained from the mean over the six linear age models transferred from the aligned snow pits reaching the Pinatubo horizon (Table ). The standard deviation among the age models gives an error estimate. They indicate that the snow layer buried at 151.5 cm depth was deposited in early 1999 ±1 years. From the surface snow deposited in June 2019, the trench thus archives about 21 years of snow accumulation, with a mean elevation increase of 7.2 cmyr-1.

We evaluate how closely the dated isochrones in the trench (Fig. c) can be interpreted as past surface topography using snow penetrometry as an independent variable not used for the alignment. The dated isochrones drawn on top of the penetrometry measurements are shown in Fig. a. We use the case of the mid-2015 isochrone as a validation of the past topography because it follows closely a clear boundary between a softer and an harder layer detected by penetrometry: the Pearson correlation coefficient between the hardness boundary and the isochrone is r=0.67, p<0.05, with only 4 cm average vertical distance between the two lines. The origin of the hard layer detected in mid 2015 is likely the signature in the snowpack of a remarkable meteorological event described in , as discussed in Sect. .

In order to reconstruct accumulation, the stratigraphy based on the isochrones is combined with density measurement to convert observed elevation change in cmyr-1 of snow into accumulation in kgm-2yr-1. The mean trench accumulation time series is calculated as the mean of the 26 profiles evenly distributed across the trench (2 m apart). The mean density measured over the trench is 328 kgm-3 with a standard deviation of 42 kgm-3. The surface snow (depth < 6 cm) has density of 293 ± 28 kgm-3, 10 % lower than the average. Overall, we observe no detectable densification in the first 1.5 m depth, which justifies that the distance between isochrones is representative of the accumulation depth at deposition time. The mean annual accumulation of the 26 trench profiles only differs by 4 % on average when using mean density of 328 kgm-3 compared to full density profiles in the trench, while the difference in total accumulation for the period 2001–2019 are negligible (< 0.1 %) when using average vs. spatially resolved densities (Sect. ). Although the reference profile age model reaches as far as early 1999, the starting year for the accumulation reconstruction is set to 2001, since is it the first year identified on all of the profiles. The resulting time series is shown in Fig. . It spans from early 2001 to late 2019 with a mean accumulation of 23.9 ± 1.5 kgm-2yr-1. The error is estimated as the combination of the uncertainty of the depth of the 2001 layer (135 ± 3 cm) and the dating uncertainty around ±1 year in 2001.

GLACIOCLIM stake network elevation change are converted to accumulation rate using a constant density value ρsurf. The reference density value is ρsurf= 320 kgm-3. Based on the density observed in the surface snow in the trench, we propose a second value of density of ρsurf= 295 kgm-3 to convert snow elevation change to accumulation rate. With a density of 320 kgm-3, the accumulation rate estimate is 26.7 ± 0.3 kgm-2yr-1 (for 2004–2019). A density of 295 kgm-3 yields 24.6 ± 0.3 kgm-2yr-1 (for 2004–2019). The ERA5 reanalysis time series for the 2001–2019 period is 24.9 kgm-2yr-1 precipitation and 2.6 kgm-2yr-1 evaporation rate, resulting in 22.3 kgm-2yr-1 net accumulation rate. We note that the net accumulation from ERA5 does not include potential contribution of snow drift, contrary to the trench and stakes data. Turning to spatial variability, taking 2σ/n as the spatial envelope where σ is the standard deviation across profiles and n is the number of profiles, we obtain 2σ/n= 4.5 kgm-2yr-1 for the trench (26 profiles, 50 m scale) and 2σ/n= 6.5 kgm-2yr-1 for GLACIOCLIM stakes (50 profiles, 1 km scale) (5.8 kgm-2yr-1 with ρsurf= 295 kgm-3). Finally, the inter-annual standard deviation in accumulation rates is quite comparable between the three datasets, with 4.4 kgm-2yr-1 for the trench (2001–2019 period), 5.7 kgm-2yr-1 for GLACIOCLIM stakes with ρsurf= 320 kgm-3 (2004–2019 period) (5.2 kgm-2yr-1 with ρsurf= 295 kgm-3) and 3.8 kgm-2yr-1 for ERA5. Restricting to the 2004–2019 period does not change the results by more than 3 % for the trench and ERA5. While the amplitudes are also similar, there is a slight difference in phase (Fig. ), which could be attributed to local spatial variability of the 50 m long trench. This point is discussed in the Appendix .

Annual accumulation hiatus, near zero accumulation or erosion over a year, is frequently observed near Dome C . We investigate whether we can detect it in the reconstructed trench accumulation. In order to describe accumulation hiatus from a final accumulation product (snow pits) in which, by definition, only strictly positive accumulation is recorded, we have extended the definition of hiatus to a snow accumulation less than 3 cm over a given period of time. This 3 cm threshold corresponds to the smallest accumulation amount we can detect with the trench resolution. For example, the 2012 and 2013 isochrones in Fig. b are in a very narrow range of depths (< 3 cm) around profile P32, indicating an accumulation hiatus at this location for the year 2012. We consider all pairs of isochrones and compute the fraction of profiles where hiatus are occuring. We bin average this hiatus fraction by the time period separating isochrones in each pair. The same analysis is repeated on the elevation profiles of the RLS data, with a threshold of 3 cm accumulation. The hiatus fraction as a function of time period for the trench and RLS is shown in Fig. . The 19 years record of accumulation in the trench, with 3 cm vertical resolution, equivalent to at least 6 months temporal resolution, is compared to the 2 years time series of the RLS (period 2) with daily resolution. There are 26 evenly distributed accumulation time series 2 m apart in the trench, compared to 110 elevation change timeseries of 1 m2 averages of the RLS (110 m2 scanned area).

The hiatus probability is computed for the trench isochrones at monthly resolution for periods from 6 to 36 months. We did not consider periods shorter than 6 months for the trench as it is under our resolution limit. Hiatus probability reaches 42 % at 6 months (15 % standard deviation across accumulation events), and 5 % at 1 year (5 % standard deviation). We found no hiatus (0 %) extending over 2 years.

For the upper section of the trench (depth < 20 cm), a second computation of hiatus statistics is possible based on SSA values. SSA is decreasing for older snow and the SSA values corresponding to snow older than 1 year can be deduced from the dated trench. We found empirically a threshold value between 35 to 37 kgm2 for 1 year old snow. The fraction of profiles with SSA values under this threshold is 3/26 and 6/26 respectively, resulting in a annual hiatus probability of 10 %–20 % for the year 2019, higher than the estimation with the chemistry dating. The details of this alternative approach is provided in the Supplement (Sect. S1). The same computation with the 2 years SSA threshold value of 23 m2kg-1 yields 0 profile with 2 years accumulation hiatus, agreeing with he chemistry dating.

These results are compared against the RLS dataset. The hiatus probability is computed for daily elevation maps for all periods between 1 to 730 d. The results in the RLS area are similar with an agreement at 6 months, a marked decrease with only a slightly higher (non statistically significant) probability of hiatus in the range 6–18 months. No accumulation hiatus over 2 years or more are detected. There is 41 % mean hiatus probability at 6 month, with a 15 % standard deviation across accumulation events, and a 10 % probability at 1 year, with a 10 % standard deviation. This is close to the 5 ± 5 % estimate in the trench, considering that the observation period is 19 years for the trench and 2 years for RLS. We note that there are 2 annual accumulation events within the 19 years of the trench record that show a 20 % hiatus fraction.

Isochrones in Fig. a show that the snow accumulation does not lead to a flatter surface with time, but instead the persistence of several years of some topographic features over time is observed. For instance, the 10 cm high bump around P10 took 5 years before being erased. The larger 30 cm high dune located in the 35–50 m section of the trench seems to persist throughout the accumulation record (about 20 years). We therefore hypothesize that accumulation and topography are relatively decoupled at the meter scale. In order to quantify this phenomenon, we define the following measure of local topography. At a given point on an isochrone, we define Δz as the difference between absolute depth z of the isochrone at that point, and the mean absolute depth values of the two neighboring points (at +2 m and -2 m). Δz measures the curvature, it is negative for bumps and positive for holes.

Figure  shows the distribution of snow accumulation against Δz. While the highest accumulation amounts (> 10 cmyr-1) are more often associated to holes as expected, the bulk of the distribution is very spread out. We perform a linear regression of snow accumulation against Δz and express the slope in filling fraction of the holes (cmyr-1cm-1 = %yr-1). This gives a positive slope of 15 %yr-1, and a determination coefficient of r2= 0.08 (p-value < 0.05). This linear slope value of 15 %yr-1 is very low compared to 100 %yr-1, the value expected for the annual filling of holes. It means that a hole on the surface (resp. bump) receives only 15 % more (resp. less) annual accumulation than its surroundings. The low determination coefficient also shows that there is no consistent relationship between topography and accumulation. This supports the hypothesis of a strong multi-year persistence of the meter-scale topography. While these findings might seem contradictory with general observations in snow accumulation dynamics, where wind-driven snow transport is generally expected to smooth surface topography , one must recall the unique setting of Dome C, where the very low accumulation rate (8 cmyr-1) is of the same order as the micro-topographic features 10–20 cm high. This coupled with frequent wind speeds above 10 ms-1 can cause the light snow to be removed and will limit the tendency for snow to accumulate in depressions.

The dated stratigraphy constructed on the trench data by inter-profile alignment and dating of the reference profile was successful in recovering sub-decadal characteristics of accumulation patterns at Dome C both in mean accumulation values, spatial and temporal variability (Sect. ). However, these accumulation estimates are constrained by limitations arising from the uncertainties caused by the inter-profile alignment (Sect. ), as well as the reference, single profile dating (Sect. ).

One intrinsic limitation of the method is the number of recognizable features in the sulfate signal. There are about 15 peaks or crests higher or lower than the ±1σ range on P0, which is reflected in the fact that 14.5 tie points per profile were used on average for the alignment, or one every 10 cm. Compared with the 21 years present in the cores, this suggests that there is no exploitable traces of seasonal variability in the sulfate signal at Dome C. This was tested by sampling four profiles at 1.5 cm resolution instead of 3 cm, namely P0, P0.3, P47.7 and P48. The average number of tie points for the 3 high resolution profiles beside the reference profile P0 was 13. This means that with 3 cm vertical resolution, we can effectively resolve the stratigraphy exploiting the full significant sulfate signal, and that the related uncertainty is effectively the precision related to sampling resolution.

The sampling resolution of 3 cm in the trench is to be compared to the annual mean accumulation of 7.8 ± 1.4 cmyr-1 (23.9 ± 4.4 kgm-2yr-1, Sect. ), meaning our precision is 3/8=4.5 months on average and 3/6=6 months for the thinner layers. To evaluate the impact on accumulation rate estimates, we have performed a sensitivity test on the alignment by shifting isochrones by a random height of up to 3 cm around our final alignment. The result is an uncertainty of ±0.8 kgm-2yr-1 on accumulation, or approx 3 % (Sect. ).

In addition, we estimated the global uncertainty of the dating procedure of the trench (Appendix, Fig. ) to roughly ±1 year, evenly affecting all profiles (Sect. ). This was confirmed by the comparison between the dating of the trench and the penetrometry hard layer detected in March 2015, and attributed to exceptional meteorological condition around a strong wind event. The April 2015 sulfate isochrone is co-localized within 4 cm vertical distance with a boundary between higher and lower snow hardness (Fig. c). documented a remarkable meteorological event in March 2015, when wind exceeding 7 ms-1 removed about 2 cm of fresh snow from the surface. This type of high wind speed would typically lead to harden crust, therefore the 2015 snow hardness transition could be the signature of this event. Another independent validation arises from the isotopic composition measurements where the maximum δ18O values identified in the trench in summer 2014 can be found in surface snow isotopic composition time series in January 2014 and minimum values of d-excess identified in the trench in summer 2015 can be found in surface snow isotopic composition time series in January 2015 . Overall, we conclude that the trench stratigraphy gives local accumulation with a 6 months inter-profile resolution, and ±1 year global uncertainty.

Beyond the temporal resolution limits of this method, we also cannot obtain negative accumulation by studying the final accumulation product in the snowpack. By contrast, the GLACIOCLIM stakes network exhibits negative annual accumulation values for 10 % of the stake measurements. While negative accumulation is out of reach with the trench approach, we investigated low accumulation years (near hiatus) and found similar statistics for low accumulation events (< 3 cm) between the trench and RLS datasets (Sect. ). Here again, we show that our accumulation estimates are reliable when considering layers representing at least 6 months of accumulation.

We have chosen to work with a reference profile as a target for the alignment of all the profiles in the trench, instead of pair by pair alignment. Beyond the obvious gain in time in the alignment (n vs. n(n-1)/2 pairs to align, where n=35 is the number of profiles) and dating (only one profile to date), we also gain in coherence as all features recognizable on the reference were matched precisely. For the same reason, we chose to work with a single raw reference profile instead of a trench stack. Indeed, the reference profile is used as a target for the dating relative to older snow pits, which are mostly single, non-replicated profiles. By using a single raw reference, we were unable to match 9 out of the 22 snow pits used for dating (due to the anomalies in the reference and in the snow pits themselves), but we keep a high confidence for succesful identification of the remaining 13 snow pits. This 60 % ratio is consistent with the two ice-core volcanic peak identification rate of . Reference profile alignment however comes with some caveats, and the dating of the reference remains the main source of uncertainty.

First, the reference profile can itself be subject to hiatus. In Sect. , we found that annual accumulation hiatus occurs in 5 % of the profiles on average, and 6 monthly hiatus about 40 % of the time. Comparing the sulfate records in five 100 m long ice cores, found a probability of 30 % of missing one single volcanic event in a single core, which is consistent considering most of the deposition takes place over less than a year . Accumulation hiatus on the reference profile lead to missing minima or maxima in its sulfate record. Such a missing feature would be present in 95 % of the trench profiles and inaccurately aligned on the reference. However, with our approach, these effects are automatically corrected for on average. For instance, if year 2015 was missing in the reference, but years 2014 and 2016 were clearly identified, the accumulation for 2014–2015 and 2015–2016 may be inaccurate, but the accumulation of the two years period 2014–2016 would still be correct. Sensitivity tests on the choice of reference profile, using P48 instead of P0 as an alignment target, indeed show such variations, changing the intensity or phasing of inter-annual accumulation within a 5 kgm-2yr-1 error bar (Fig. c). Similarly, user induced variability of inter-annual accumulation reconstruction using the same reference, which will inevitably impact assignment of ambiguous features, is contained within a 5 kgm-2yr-1 error bar (Fig. c). We have considered the alternative of using a stacked profile as the reference to get a more representative average sulfate profile, but precisely because of stratigraphic noise, this produces a very smooth profile (Appendix, Fig. a) on which it becomes impossible to recognize any feature at the inter-annual scale.

Second, the uncertainty of dating of the reference profile must also be considered. When aligning the reference to older snow pits, we estimated a typical error of ±5 cm in the depth associated to the top of the snow pit which propagates into ±1 yr in the age depth scale of the stratigraphy (Sect. ). We have performed a sensitivity test with variations in the age model of the reference with ±1 yr variations, and found variations in accumulation as high as 5 kgm-2yr-1, i.e. about 20 % in individual annual accumulation rates. (Sect. ). The uncertainty on the total accumulation of the 19 years record is 6 % (Sect. ). Previous studies of snow accumulation rates have used seasonal signal in chemical tracers, in particular sulfate, to resolve annual layers. This has been successfully applied to higher accumulation sites such as DML as well as low accumulation sites such as Dome F . Layer counting is based on sulfate peaks observed in austral summer due to the biogenic activity . We have not tried this approach in the present study, as it is not applicable to our coarse resolution (3 cm). The higher resolution (1.5 cm) profiles P0 and P48 do exhibit high frequency seesaw variability with very variable amplitude, but peak counting results in 15–25 peaks depending on counting strategy, and it is unclear whether it could be interpreted as annual layers.

To test the sensitivity of the reference profile, we performed another alignment and dating procedure with the profile P48, and the results are presented in the Appendix. This alternative reference alters the accumulation reconstruction at the annual scale, with discrepancies of the same order as those resulting from the 20 % dating uncertainty. The mean accumulation in the period 2001–2019 is within the error bars of the first alignment using P0, and the overall shape of the mean accumulation time series is also conserved (e.g. high value around 2012, low value around 2009, low values for 2014–2016, Fig. c). Therefore, we argue that, within conservative error values on the dating alone, our alignment method yields a robust and reproducible accumulation reconstruction that is able to capture its inter-annual variability with good accuracy.

Different accumulation time series based on the trench and other lines of evidence (Sect. ) differ slightly in mean values and inter-annual variability, but overall agree within ranges of uncertainties.

The mean accumulation rates are 23.9 ± 4.5 kgm-2yr-1 (mean value for the period ± spatial envelopes) with a 6 % uncertainty for the trench (2001–2019) and 22.9 kgm-2yr-1 for ERA5 (Sect. . The lower value of ERA5 is just within the 6 % uncertainty of the trench snow accumulation rate. The mean accumulation is 24.6 ± 6.0 kgm-2yr-1 (with a 1 % uncertainty) for GLACIOCLIM stakes (2004–2019) with the surface density ρsurf= 295 kgm-3, or 26.7 ± 6.5 kgm-2yr-1 with ρsurf= 320 kgm-3. The conversion of GLACIOCLIM stakes elevation to accumulation rate relies on the choice of a constant density, traditionally set as ρsurf= 320 kgm-3 , which considers a linear regression over a 5 m density profile. Other studies recommend a density for surface snow of 300 kgm-3 instead . This is more in line with the direct measurements of surface density or of the value measured in the trench of 295 kgm-3 (depth < 6 cm). This lower value of surface density removes the 10 % gap that would otherwise arise between GLACIOCLIM stakes and trench snow accumulation rates (Fig. b). The reason for the overestimation of surface density could be due to metamorphism of surface snow upon burial in the firn. Lighter snow only constitutes the superficial 5–10 cm layer effectively considered by the stake readings, and a linear regression over more than a meter depth will over-estimate this lower surface density. We note that the mean density in the trench in the depth range 0–1.5 m is 328 kgm-3, and a linear regression on the 1.5 m gives a surface density value of 320 kgm-3 which agrees with the density profile of (Appendix, Fig. ).

We note that found a mean elevation change of 7.3 ± 0.2 cmyr-1 for the 2011–2023 period, using a stake farm of 13 poles situated 800 m south-west of Dome C, which is closer to the trench site than the GLACIOCLIM stake farm by about 1 km. This is equivalent to 23.4 ± 0.6 kgm-2yr-1 with ρsurf= 320 kgm-3 and 21.5 ± 0.6 kgm-2yr-1 with ρsurf= 295 kgm-3. The mean accumulation rate in the trench is 22.4 kgm-2yr-1 for the period (2011-2019), 6 % accurate, which would be in agreement with the stake farm with either values of ρsurf.

In addition, the trench accumulation rate lies within the spatial envelope of the GLACIOCLIM stakes network with either value of ρsurf as the GLACIOCLIM stakes cover a 1 km × 1 km. Discrepancies due to decameter-scale variability of snow deposition needs to be taken into account, and the 50 m long trench could be a specifically low accumulation area at Dome C.

We compare the amplitudes of mean annual accumulation time series between the trench and the GLACIOCLIM stakes. The maxima displayed by the trench accumulation time series have an amplitude comparable to the maximum accumulation annual values documented by direct measurement methods and reanalysis data. The peaks however are out of phase by ±1 year for the year 2011, and ±2 years for 2006 and 2008 (Fig. ). This could be due to errors in the dating of the trench chemical stratigraphy, which brings uncertainties of the order of ±1 year. However, such dephasing of ±1 year is observed within the GLACIOCLIM network itself (for the 2009 or 2013 peaks for instance), when considering stakes at the opposite sides of the 1 km × 1 km cross (Fig. ). Therefore we cannot rule out that the differences in phase are not due to the intrinsic spatial variability of accumulation at km scale, as the sites of the stake farm and the trench are approximately 2 km apart.

Overall, having multiple trenches at and around Dome C would be necessary to be able to separate the spatial variability at the different scales (decimetric to kilometric) which seems to affect snow accumulation. Alternatively, simultaneous sampling of snow-pits at the same location and spatial scales as existing stake farms would allow to re-do the comparision presented here factoring out the dependency in spatial scales.

suggests that more than 90 % of the isotopic signal archived in ice cores is noise, suggesting that it is necessary to average more than 10 ice cores 5 to 10 m apart to retrieve a coherent local isotopic signal in interior sites in Antarctica , and minimize noise due to accumulation variability. We evaluate the impact of accumulation variability on sub-decadal reconstruction from ice-cores drilled at Dome C. We have showed that the 30 cm dune observed in the 35–50 m section of the trench is persistent over 20 years. Taking two profiles from the trench as an example of two ice-cores drilled several meters apart, sitting on top of different configurations of past dunes, we see that they would have a depth offset of up to 30 cm that propagates over more than a 20 year long section of the cores. We quantify the age difference that snow samples at the same depth could have for this pair of hypothetical ice cores. Computing the distribution of years at a given depth based on our trench age model, the 98 % quantile is 3.4 years. With our high resolution alignment, this mismatch is reduced to ± 6 month. (Sect. ).

For multi-decadal or centennial records of longer ice-cores dug at Dome C, traditional alignment relies on major volcanic peaks horizons giving tie points every 50 years on average . Our approach suggests that this volcanic peak matching can be taken as a preliminary alignment, and be further refined within 50 year windows using background sulfate signal. have applied a similar peak matching method to electrical conductivity profiles in replicate EDC cores, with a correlation based automatic match routine, and have found 6600 tie points for a 44 kyr record, or one tie point every 7 years on average. Our work suggest that the precision could reach one tie point every 1.5 years on average by analyzing the detailed chemistry record and potentially improving automatic matching routine to reach the accuracy of manual matching. Concerning high resolution dating, while there will be no additional dating point to further refine the dating beyond volcanic horizons, as was done here with the 20 years snow pit dataset, we have showed that simple linear interpolation of the accumulation rate produces a timescale within the ±1 year uncertainty. By combining 2 or more replicate cores at high resolution, this open the way of interpreting climatic signal in Dome C ice-core at higher frequencies, potentially reaching annual resolution.