An attempt to explain recent trends in European snowfall extremes

The goal of this work is to investigate and explain recent trends in total yearly snow-depth and maximum yearly snow-depth from daily data in light of both the current global warming and the low-frequency variability of the atmospheric circulation. We focus on the period 1979-2018 and compare two different data-sets: the ERA5 reanalysis data and the EOBSv19 S precipitation data, where snow-depth is identified from rainfall by applying a threshold on temperature. On one hand, we show that the decline in average snow-depth observed in almost all European regions is coherent with the mean 5 global warming and previous findings. On the other hand, we observe contrasting trends in maxima. We argue that this apparent discrepancy between trends in average and maximum snow-depth comes from the subtle effects of atmospheric circulation in driving extreme events and the non-trivial relation with global warming: a warmer Mediterranean Sea may enhance convective precipitation in winter-time and trigger heavy snowfalls. We discuss the limitations of block-maxima indicators and of static identification of trends based on regional or grid-points analysis, paving the way for attributing changes in extreme snowfalls 10 via analogs-based methods.

rivers discharge. Since the snow is immediately transformed into water, SWE does not distinguish between snowfalls which produced accumulations on the ground or not. The other quantity, namely the snow-depth (SD), is a measure of the snow height on the ground and it can be affected by several problems due to gravitational settling, wind packing, melting and recrystallization. SD is a quantity of interest for its societal impact: large SD amounts correspond the snow to be removed to free 95 ground transportation infrastructures. In this paper we will therefore use daily SD and express it in cm. We now explain how to get this quantity from the different data-sets considered in this study.
-For ERA5 [80W-50E,22.5N-70N], we use the accumulated total snow that has fallen to the Earth's surface. From the ECMWF description, this quantity consists of both snow due to the large-scale atmospheric flow and convective precipitations. It measures the total amount of water accumulated from the beginning of the forecast time to the end of the 100 forecast step. This quantity is higher than the snow-depth if snow has melted during the period over which this variable was accumulated. The units given measure the depth the water would have if the snow melted and was spread evenly over the grid box. We get the snowfall from hourly data and construct the daily SD by summing up the snowfall in intervals of 24 hours. We chose ERA5 data-set as the preferential one for our study because of its physically consistency and the use of advance assimilation techniques for its compilation. Besides the nominal 0.25 • horizontal resolution, we will also 105 compute the statistic over the NUTS-2 regions, this scale being the one used by European stakeholders to assess impacts.
-For E-OBSv19.0 [40.375W-50E,25.375N-75.375N] only lands points, we do not dispose directly of snowfall, SWE or SD data. We have to infer them from daily total precipitation and daily mean temperature data. We apply a simple algorithm which consists of considering as snowfall all precipitations occurred in days where the average temperature is below 2 • C. Of course with this method we can have false positive as well as false negative events, but we have verified 110 (not shown) that results on the trends do not depend qualitatively from the threshold providing that it is chosen between 0 • C and 2.5 • C. Since we use a threshold of 2 • C, some of the precipitation would not be snowfall. In order to avoid overestimation we consider SD only 2/3 of the daily amount obtained.
We now present the climatology for the two data-sets used in this study and focus on two quantities: yearly total snow-depth SD (average 1979-2018 in Figure 1a,c,e) and the maximum yearly (block maxima) snow-depth SD from daily data (average 115 1979-2018 in Figure 1b,d,f) for the three data-sets. Despite local differences, we can remark a substantial agreement among all data-sets for the two variables considered. We remind that E-OBSv19 data are defined only for land points. The agreement between the ERA5 and the E-OBSv19.0 data-set is remarkable, with the latter showing generally lower SD, possibly due to our choice of the factor 2/3 when converting precipitation into snow. Analysing the climatology we remark that, at southern latitudes and on the plains, mean and max statistics tend coincide because the number of snow days per year is limited, i.e. all 120 snowfall is concentrated in one or few events. section we focus on the intensity of positive or negative trends regardless of their significance. As pointed out by (Altman and Krzywinski, 2017), statistical testing based on pvalues presents several limitations, and can produce misleading results even in designed experiments. Here, we privilege the physical complexity of the phenomenon, as information about pure statistical significance has already been discussed in the previous section. In Figure 5 we show the box-plots of the yearly maxima organized in two different periods (1979-1998 and 1999-2018) for the 10 regions having the largest positive (a) and negative (b) trends in maxima of SD. The insets of Figure 5 show the location of the regions with largest trends and the magnitude of the 160 trends (size of disks). Largest positive trends are located mostly in the Balkans. It is interesting to observe how boxplots and trends provide a different information: for ITF1 (Abruzzo region, in Italy) we detect the largest positive trend, but the bulk of the distribution (visualized by the colored bar in the boxplot) shifts instead to lower values. The increasing trend is therefore largely due to the two outliers. Another example is TR42 (Kocaeli, Turkey), where we have a small trends but a large positive shift in the distribution. 165 We now analyse the relation between largest trends and the long-term changes of the atmospheric patterns associated to these events. We divide the sample into two periods: 1979-1998 and 1999-2018 and consider three different atmospheric fields: the daily averaged geopotential height at 500 hPa (Z500) as a tracer of the atmospheric circulation (Jézéquel et al., 2018), the daily averaged two-meters temperature (T2M) to account for thermodynamic changes and the snow-depth (SD). We prefer 170 the Z500 variable to the sea-level pressure because the latter shows a strong variability during these events: when computing median and mean sea-level pressure fields associated to these events we observe that they do not produce similar patterns. For each region and each period, we average the fields corresponding to the days when the yearly maxima of SD are observed. We then subtract the average for the first period from that of the second one obtaining the anomaly fields displayed in Figure 6 (Z500), Figure 7 (T2M) and Figure 8 (SD). We report the results only for the 10 regions displaying the largest negative 175 (panels in the red frame) and positive trends (panels in the blue frame). For Z500 ( Figure 6) we remark positive anomalies for regions showing largest negative trends. This implies that circulation patterns associated to recent heavy snowfalls display higher geopotential heights (weaker cyclonic structure) than maximum SD events in the 1979-1998 period. It is interesting to note how the anomalies show preferentially an anti-zonal or a blocked pattern, with negative Z500 anomalies generally concentrated over eastern Europe. As one would expect in a warming climate, the T2M anomalies (Figure 7-red panels) are

Conclusions
We have analysed recent trends in yearly total and maximum snow-depth SD from ERA5 reanalysis and the E-OBSv19.0 datasets. Even though the two products show large differences in trends, we have identified a robust signal in the general decrease 210 in the yearly total snow-depth, in particular for Northern and Western Europe. For SD maxima, trends are more contrasted: negative trends persist over Western Europe, but over the Mediterranean area we identified a certain number of regions showing positive trends.
This discrepancy between average and extreme SD trends is compatible with future scenarios for winter Mediterranean precipitations. Polade et al. (2017) project an overall decrease of winter average precipitations over the Mediterranean sea, but 215 an increase of extreme precipitations and of their variability. Extremes should be favored by a warmer sea, with a larger variability of moisture and potential energy. They could also benefit from blocking patterns forcing southward movement of polar extratropical cyclones towards Europe (Liu et al., 2012). Our analysis of the atmospheric circulation associated to maxima snowfalls suggests that these blocking patterns are crucial in determining heavy snowfalls ( Figure 6). From the analysis of 500 hPa geopotential height patterns in the last 20 years, we observe more anticyclonic conditions over Western Europe associated 220 to cyclonic conditions over Eastern Europe. This explains both the negative trends over western Europe and the positive trends over the Balkans regions and Turkey. Even though this could suggest a relation between our finding and the arctic amplifi-