Reconstruction of daily snowfall accumulation at 5.5 km resolution
over Dronning Maud Land, Antarctica, from 1850 to 2014 using an
analog-based downscaling technique

Abstract. The surface mass balance (SMB) over the Antarctic Ice Sheet displays large temporal and spatial variations. Due to the complex Antarctic topography, modelling the climate at high resolution is crucial to accurately represent the dynamics of SMB. While ice core records provide a means to infer the SMB over centuries, the view is very spatially constrained. General circulation models (GCMs) estimate its spatial distribution over centuries, but with a resolution that is too coarse to capture the large variations due to local orographic effects. We have therefore explored a methodology to statistically downscale snowfall accumulation, the primary driver of SMB, from climate model historical simulations (1850–present day) over the coastal region of Dronning Maud Land. An analog method is set up over a period of 30 years with the ERA-Interim and ERA5 reanalyses (1979–2010 AD) and associated with snowfall daily accumulation forecasts from the Regional Atmospheric Climate Model (RACMO2.3) at 5.5 km spatial resolution over Dronning Maud in East Antarctica. The same method is then applied to the period from 1850 to present day using an ensemble of ten members from the CESM2 model. This method enables to derive a spatial distribution of the accumulation of snowfall, the principal driver of the SMB variability over the region. A new dataset of daily and yearly snowfall accumulation based on this methodology is presented in this paper (MASS2ANT dataset, https://doi.org/10.5281/zenodo.4287517, Ghilain et al. (2021)), along with comparisons with ice core data and available spatial reconstructions. It offers a more detailed spatio-temporal view of the changes over the past 150 years compared to other available datasets, allowing a possible connection with the ice core records, and provides information that may be useful in identifying the large-scale patterns associated to the local precipitation conditions and their changes over the past century.



Introduction
In the context of the global climate warming, polar ice sheets have increasingly gained attention, due to the threat of a massive sea level rise at the global scale (Garbe et al. , 2020). While the Greenland Ice sheet is eroding at an increasing speed both from the base and the surface (Lenaerts et al. , 2019), the Antarctic Ice Sheet is sometimes viewed as subject to a mitigation snowfall accumulation potentially related to climate warming due to the uncertainty on the process at play, the large interannual variability in the region and the unknown history of the local SMB.
Indications of a recent increase of SMB in Dronning Maud Land have been found at some locations (Lenaerts et al. , 2013;Schlosser et al. , 2016;Medley et al. , 2018;Philippe et al. , 2016), but a stationary or decreasing trend has been found elsewhere (Thomas et al. , 2017;Vega et al. , 2016;Altnau et al. , 2015;Schlosser et al. , 2014). Most of the coastal ice core drilling sites https://doi.org/10.5194/essd-2021-12  Verification of the performance of the downscaling scheme trained on 11 years (1979)(1980)(1981)(1982)(1983)(1984)(1985)(1986)(1987)(1988)(1989)(1990) of ERA-Interim data and RACMO2.3 daily snowfall and applied on ERA-Interim for a 10 years period (1991)(1992)(1993)(1994)(1995)(1996)(1997)(1998)(1999)(2000). The monthly time series comparison over a location of maximum of annual snowfall accumulation shows a high degree of accuracy, while the spatial comparison of the accumulated snowfall (in mm) over 1996 on Princess Ragnhild Coast in presence of ice rises (eg Derwael Ice Rise -DIR -, Lokeryggen Ice Rise -LIR and Hammaryggen Ice Rise -HIR) illustrates the high degree of fidelity of the analog method (top right) in reproducing the RCM (bottom right) accumulation patterns, especially in the West-East difference of accumulation around ice rises (Kausch et al. , 2020). large statistical gain. As expected, the spatial and temporal variations of RACMO2.3 have been preserved (Figure 2 shows an example of time series comparison and a comparison over Princess Ragnhild Coast characterized by the presence of ice rises near the coast). The root mean square difference on a daily basis over the validation period ranges from 10% in large 140 accumulation areas near the coast to 15% in the inner regions. The advantage of the analog method is that it allows one to identify the major types of weathers delivering precipitation and their occurrence over time in the form of PCs. This could be of interest for the analysis of a frequency change in the precipitation over long time periods.
The downscaling method was then repeated over the ensemble of the ten climate runs from the CESM2 model over the period 1850-2014. Before the downscaling could be applied, we first needed to make the PCs compatible with the reanalyses.

145
A simple bias correction based on the linear regression of EOFs (Feudale and Tompkins , 2011;Yu et al, 2018), followed by a quantile mapping, transforms the CESM2 original PCs into "Reanalysis-like" PCs. The operation is done after the verification    EOFs maps from ERA-Interim and ERA5 are stored in 2 separate files. At last, PCs time series from ERA-Interim and ERA5 fields are stored in 2 files.

170
As expected, daily fields differ among the different members of the downscaled CESM2 results. When a high precipitation event over DML can be present on a specific day in member 7, it may not be the case for all the other members ( Figure 5).   While a large disparity is observed on daily maps, the effect is much smaller on the annual accumulation ( Figure 6). We illustrate this by a daily map (members 1 to 8) and an annual accumulation over the entire domain of the downscaled CESM2 member 1, trained with ERA5 ( Figure 5).  For each date, the PCs are provided, and associated to the EOFs for each meteorological field. An analysis of the association between snowfall regimes with the synoptic patterns is then possible. In Figure 7, we show an illustration of the first 9 EOFs for Surface atmospheric pressure over the GCM domain considered (40°S-90°S, 10°W-60°E), as well as the time series of the associated PCs, showing their various amplitudes.

Comparison with ice core measurements 180
The annual variation of the surface mass balance is estimated from ice cores retrieved from a set of locations over the coastal DML. The length of the records differs between the different sites and can cover more than 150 years. The annual accumulation computed from the downscaled daily snowfall is compared to SMB from the different ice core records from Thomas et al.  (2017), after a conversion of ice core estimates to water equivalent height ( Figure 1, Figure 8). A good overall match between the downscaled snowfall and the ice cores is found, as well as a large bias reduction compared to the use of CESM2 without 185 downscaling. which we only consider the reconstruction using ERA-Interim (data available at https://earth.gsfc.nasa.gov/cryo/data/antarcticaccumulation-reconstructions). The downscaled CESM2 member 1 averaged over the total period seems in agreement with the averaged reconstruction based on the ice core records, especially to reconstruct the gradient from the coast to inner land, but 200 depicts much more details related to topography ( Figure 10).

Uncertainties
The uncertainties in the daily snowfall estimates compiled into this dataset can arise from 1) the choice and accuracy of the model sources (regional climate model, re-analysis, "historical" climate model), and 2) the accuracy of the downscaling https://doi.org/10.5194/essd-2021-12 Open Access

Sampling in choice of analogs
A simple way to assess the effect of the sampling of the analogs on the estimation uncertainty is to use the bootstrap techniques.
We have created a distribution of 100 ensembles of analogs, with the 20 analogs randomly drawn with repetition at the ice core sites (Figure 11). The downscaled CESM2 member 1 with 100 different analogs draws represented at the ice core sites. The 210 uncertainty is between 5 and 10 % of the mean value.   Figure 10. The coast-to-inland gradient of the downscaled CESM2 member 1 averaged over the total period seems in agreement with the averaged reconstruction based on the ice core records of (Medley and Thomas , 2019), but depicts much more details that could be useful in detailed analysis of ice core records representativity (same color scale).

Choice of reanalysis used for training
The choice of reanalysis used for training the analogs database has an effect on the downscaling obtained from CESM2. The choice of reanalysis influences the result, at equivalent training scores (Figure 12). This is one of the reasons for extending the database to ERA5 as training reanalysis instead of ERA-Interim. More independent reanalyses could be used to further assess 215 the uncertainty. The difference CESM2 downscaling using ERA-Interim or ERA5 in the training reveals an uncertainty of 25% in average at the ice core sites.

Choice of "historical" GCM runs
CESM2 simulations consist of 10 members, corresponding to different initializations (Danabasoglu et al. , 2020). The inherent model uncertainty is therefore impacting the results and is quantified here (Figure 13). The scatter of the 10 members of CESM2 downscaling using ERA-Interim in the training reveals an uncertainty of 20% (standard deviation) and up to 40% at some ice core sites, maybe a direct consequence of the CESM2 members not being temporally correlated. another recognized RCM used for polar climates at high resolution. The newest simulations over Antarctica at 30 km resolution (Agosta et al. , 2019) have shown similarities with RACMO2 at 27 km resolution, but revealed localized differences over the ice sheet related to a more realistic sublimation of falling snowfall in comparison with RACMO2 (Gallée et al. , 2013). A simulation at the same resolution with another RCM optimized for polar regions, like MAR, could be of great interest to better 230 frame the uncertainty linked to RCM physics. The uncertainty linked to the choice of RCM has not been evaluated in this study and could be envisaged to extend the database if new RCM simulations at such resolution are available in the future.

Conclusions
We propose here a reconstruction of snowfall evolution over Dronning Maud Land, Antarctica, at 5.5 km resolution using an analog-based downscaling technique. This technique has allowed us to exploit the detailed spatio-temporal estimation of  Figure 13. The scatter of the 10 members of CESM2 downscaling using ERA-Interim in the training reveals an uncertainty of 20% (standard deviation) and up to 40% at some ice core sites.
balance over the region, its evolution and its association to synoptic weather conditions. The method can be easily replicated with new RCM and GCM simulations.

Data availability
The files of the dataset (the annual snowfall, the PCs and the EOFs) are available on Zenodo platform (http://doi.org:10.5281/ zenodo.4287517). However, due to size limitations, only 2 daily snowfall files out of 20 have been stored there, the whole set 245 is available on request to the contact author.