AROME is the high-resolution NWP system at Météo France . Its 2.5 km horizontal resolution upgraded to 1.3 km in 2015; makes it of particular interest for forecasting intense events (like convective rains) and small-scale processes in alpine terrain, such as orographic precipitations or Foehn effects, thanks to a realistic description of the topography. AROME is a spectral and non-hydrostatic model which combines the physical package of the research model Meso-NH with the dynamical core of the non-hydrostatic version of the limited area NWP ALADIN model . A detailed description of the physics and data assimilation schemes can be found in . In particular, the precipitation phase is derived from the cloud microphysical scheme.

The implementation of AROME as an operational system is made through 30 h forecasts at the 00:00, 06:00, 12:00 and 18:00 UTC nominal analysis times, over a domain covering France. We use here the hourly forecasts issued from the 00:00 UTC analysis time, from +6 to +29 h, extracted on a regular latitude / longitude 0.025∘ grid to build a continuous forcing from 1 August 2010 to 31 July 2014 over the domain of study.

Some changes in the operational configuration of AROME occurred during the 4 years of simulations: the simulation domain was extended during summer 2012 with a modification of the topographic database. The topography from Global 30 Arc-Second Elevation dataset (GTOPO30) was used in a low-resolution version (5 km) before summer 2012 and at 30 arcsec (approximately 1 km) resolution afterwards, which led to a modification of the forcing files orography in the middle of our simulation period.

The SAFRAN analysis system provides hourly atmospheric forcing data for each of the 23 massifs of the Pyrenees (Fig. ). Within each massif, the forcing is provided by 300 m altitude steps. SAFRAN reanalyses take a preliminary guess from the global NWP model ARPEGE (from Météo France, 15 km grid spacing guess projected on a 40 km grid), complemented by available observations from automatic weather stations, manual observations carried out in the climatological network and in ski resorts and atmospheric upper-level sounding. In particular, a daily precipitation analysis is included, with a climatological guess depending on a daily determination of the general weather pattern. This determination is based on a classification of nine weather patterns, defined by Météo France mountain forecasters to be representative of the main precipitating regimes of the Pyrenees. It is made following the synoptic circulation, through the altitude of the 500 hPa geopotential level. The precipitation phase is derived from a simple threshold of 1 ∘C air temperature at 2 m above the ground. In this study, SAFRAN forcing was interpolated over the 0.025∘ grid of the domain described in Sect. , following the method described by .

Table  summarises error statistics for snow depth during the whole period of study. The number of stations available varies from year to year (from 62 to 79) because of modifications in the model topography and missing data. Scores were also computed for a constant number of stations (restricted to 46, not shown) and showed that the annual variability of the number of stations does not impact the results and the analysis exposed hereafter. These scores show a global overestimation of snow depth by AROME–Crocus with an overall bias of +55 cm, while the overall bias of SAFRAN–Crocus is +22 cm. The overall STDE reaches 70 cm for AROME–Crocus compared to 57 cm for SAFRAN–Crocus. The errors are rather high for both models, which will be explained in the next sections.

For both models, the highest STDEs are found for winters 2012–2013 and 2013–2014, two very snowy winters. In terms of spatial distribution, the positive bias and STDE decrease from west to east for AROME–Crocus, with notable errors in the western zone. In the eastern zone, AROME–Crocus and SAFRAN–Crocus STDEs are equivalent. AROME–Crocus scores are equivalent in France and Spain, while SAFRAN–Crocus behaves slightly better in France, probably due to a higher number of observations assimilated by the model. In regard to altitude, biases are constant for SAFRAN–Crocus and decrease for AROME–Crocus, which implies a higher relative bias in the [1000 m, 1800 m[ range.

Figure  shows scores for each station over the whole period of study. Almost all stations show an overestimation of snow depth, particularly for AROME–Crocus with extreme positive biases on the Atlantic foothills. The three highest biases for AROME–Crocus are given by the following three stations: Isaba El Ferial (+188 cm; massif of Navarra, western Pyrenees, Spain), Arette La Pierre Saint Martin (+209 cm; massif of Pays-Basque, western Pyrenees, France) and Soum Couy Nivôse (+229 cm; massif of Aspe–Ossau, western Pyrenees, France), all located in the vicinity of the Pic d'Anie, the first summit above 2500 m a.s.l. close to the Atlantic Ocean. These three stations also show a very high STDE (higher than 1 m). The two next highest biases are located in the north-west foothills: Gourette (+135 cm; massif of Aspe–Ossau, western Pyrenees, France) and Hautacam (+154 cm; massif of Haute Bigorre, western Pyrenees, France). This region is particularly exposed to W-NW flows due to its proximity to the Atlantic Ocean. There is thus an excessive orographic blocking on these first peaks by AROME. Except for these stations, biases and STDEs are more homogeneous in the rest of the Pyrenees.

Winter 2011–2012 was characterized by a deficient snowpack in the Spanish Pyrenees, due to dry and warm weather in the southern side of the chain . It was also characterized by a strong contrast between the French and the Spanish sides of the Pyrenees: even if the French Pyrenees exhibited a deficit of snow for most of the winter (with respect to the climate normal), the first half of February 2012 was exceptionally cold and snowy in France. The Spanish Pyrenees were far less prone to snowfalls due to the northern flow. This asymmetry (and the ensuing fall in the Spanish hydropower production in springtime) was highlighted in terms of snow cover duration in . Hereafter are shown the added value of AROME high-resolution forcing for simulating a particular meteorological contrast due to the topography and the resulting snow cover distribution.

Figure  gives an overview of the snow cover simulated by AROME–Crocus and SAFRAN–Crocus (values of SWE higher than 20 mm), compared to MODIS fractional snow cover images, on 22 February 2012. This date (selected because of clear sky conditions) is close to the end of the intense cold and snowy events in the French Pyrenees, corresponding to a maximum contrast between both sides of the Pyrenees. This contrast appears clearly on MODIS snow cover image, where snow is only present on the highest summits of the Spanish Pyrenees, on the border ridge, while snow covers most of the French Pyrenean massifs and Val d'Aran (in Spain, but on the northern side of the Pyrenean highest ridge). The absence of snow in the Spanish Pyrenean foothills is particularly well represented in the AROME–Crocus simulation, and the snow cover distribution matches observations. On the contrary, SAFRAN–Crocus simulation exhibits a rather homogeneous snow cover in Spanish massifs (despite still lower quantities than in the French Pyrenees). The snow cover spatial distribution, and particularly the snow deficit in the Spanish Pyrenees, is thus better simulated by AROME–Crocus.

This improvement in terms of snow cover may be attributed to AROME dynamical behaviour in complex topographies. showed that the snowfall deficit in 2011–2012 was more sensitive at Spanish stations exposed to southern flows, while Spanish stations more exposed to northern flows exhibited a lower negative anomaly. The snowpack was mainly constituted by N-NW flows during this season, which is confirmed by a study of SAFRAN weather patterns. We cumulated all snowfalls (from SAFRAN outputs) which occurred on the studied domain between 1 October 2011 and 22 February 2012 (date studied in Fig. ). Of the cumulated snowfall, 67 % fell during days of north to north-western flows, which correspond to two synoptic patterns: a minimum geopotential in the Genoa gulf and a maximum in Ireland, associated with N and NW flows (38 %), and disturbed NW flow with strong geopotential gradient, implying strong precipitations on the NW French Pyrenees and a Foehn effect in Spain (29 %). During the four winters 2010–2014, these synoptic conditions constituted 45 % of total snowfalls. In contrast, only 4 % of total snow quantities fell during days of south to south-western flows (against 14 % over the period 2010–2014).

The behaviour of both forcing models in such specific synoptic conditions is of particular interest. Snowfalls from AROME and SAFRAN were cumulated from 1 October 2011 to 22 February 2012. They are represented in Fig.  along a NW/SE cross section, as well as cumulated positive ΔSWE from measurements of three stations close to the transect. Orographic blocking is visible on the windward sides, with a maximum snowfall immediately upstream of the highest summit whereas a Foehn effect in Spain implies a drastic drop of snowfalls immediately behind the highest ridge. The orographic shield of the Haute Bigorre first high summits leads to fewer snowfall than upstream for the same altitude (approximately 4 times less). This windward / leeward distinction within a massif is not simulated by SAFRAN, since two points at the same altitude and within the same massif get the same amount of snowfall. The difference between both forcings is marked at Esera (Spanish massif), where the orographic shield and resulting dry weather is not represented enough by SAFRAN compared to AROME. Such differences are even more marked when filtering only cumulated snowfalls occurring by N-NW flows (not shown). AROME simulations are in good agreement with the two Spanish stations, which are located at an altitude close to the model's topography. SAFRAN snowfalls are too low at the station closest to the border but in good agreement at the second Spanish station. Observations for France are in better agreement with AROME than with SAFRAN but still higher than both simulations. This may be due to the difference of altitude with the models. This study emphasises the added value of AROME dynamics, which allow us to better take into account mesoscale orographic effects.

The comparison between AROME–Crocus, SAFRAN–Crocus and MODIS snow cover distribution is extended to two entire winters: 2011–2012 (characterized by an average deficit of snow) and 2012–2013 (extremely high amount of snow). Table  summarises two metrics (ASSD and Jaccard index) that evaluate the match of simulated and observed snow covers in different domains. AROME–Crocus scores are better than SAFRAN–Crocus scores for the whole Pyrenees (higher Jaccard index and lower ASSD for both seasons). This is also true for the Spanish, central and eastern domains, whereas scores are equivalent for France. SAFRAN–Crocus performs better in the western Pyrenees. The seasonal evolution of scores over this domain (not shown) indicates that both models have equivalent skills during the accumulation season, while SAFRAN–Crocus performs better during the melting season. This result is consistent with the results of Sect. : AROME–Crocus strongly overestimates snow quantities in the western Pyrenees, which results in a later presence of snow on the ground in the springtime.

Figure  shows the evolution of daily ASSD and Jaccard index for winter 2011–2012 over the whole Pyrenees (within SAFRAN massifs). Both scores attest that AROME–Crocus improves the representation of the spatial snow cover distribution compared to SAFRAN–Crocus until late March. SAFRAN–Crocus shows a slightly better agreement than AROME–Crocus after late March, i.e. at the beginning of the melting season due to the overestimation of snow quantities by AROME–Crocus. On 22 February 2012 (date studied in the previous section, Fig. ), J= 0.61 and ASSD = 1.22 pixels for AROME–Crocus, while J= 0.40 and ASSD = 2.09 pixels for SAFRAN–Crocus, which quantifies the better agreement seen in Fig. .

The STDE of daily ΔSD indicates the ability of the model to forecast (or analyse) the appropriate daily evolution of snow depth. This score was computed for AROME–Crocus and SAFRAN–Crocus. It is equal to 7 cm (and bias equal to 0 cm) for both models, with low spatial variation. STDE is slightly higher during the most snowy winters (8 cm in 2012–2013 and 2013–2014 against 6 cm in 2010–2011 and 2011–2012). This is the first complementary information to global scores that indicate that, despite an overall overestimation, AROME–Crocus gives similar results compared to SAFRAN–Crocus in terms of daily snow depth variations.

A classification by category of the increase (accumulation) and decrease (ablation and settling) of SD gives a better view on the behaviour of the models. The categorical frequency distribution of ΔSD is plotted in Fig. , according to eight accumulation categories, two decrease categories and one “no variation” category [-0.2 cm, 0.2 cm[. Small daily accumulations (between 0.2 and 10 cm per day) are overrepresented by both models, while the occurrence of medium and high daily accumulations (more than 10 cm per day) is underestimated by both models. However, the frequency of medium and high accumulation events predicted by AROME–Crocus is systematically closer to the observations than SAFRAN–Crocus. There is also a clear discrepancy between both models and observations for the strong decrease category, largely underestimated by both AROME–Crocus and SAFRAN–Crocus.

In terms of quantities, the categorical sums of ΔSD (not shown) indicate that SAFRAN–Crocus strongly underestimates the high accumulation quantities. AROME–Crocus is closer to observations for these categories (particularly for the [10 cm, 20 cm[ category, the main contributor to the snow accumulation). It is counterbalanced by an overestimation of small accumulation quantities, since an underestimated strong accumulation event is counted in the smaller accumulation category. The sum of all accumulation categories shows an overall underestimation of snow accumulation by both models: the total sum of observed accumulations is 904 m, against 857 m for AROME–Crocus (-5 %), and 753 m for SAFRAN–Crocus (-17  %). The largest difference concerns the category of strong decrease, globally missed by both models. Since AROME–Crocus and SAFRAN–Crocus underestimate accumulations, the strong decrease category becomes the main contributor to the overall overestimation of snow depth: the positive bias shown in Sect.  is not due to an excess of snowfall but to an insufficient snow depth decrease. Total decrease quantities are more pronounced for AROME–Crocus than SAFRAN–Crocus as a logical consequence of more marked accumulations. Plotting the cumulated ΔSD by altitudinal range (under 1800 m, between 1800 and 2200, and above 2200 m) highlights a similar behaviour of both models, except for a stronger underestimation of high accumulations by SAFRAN–Crocus at the lowest altitudes (not shown).

In order to isolate the specific behaviour of AROME–Crocus in the Atlantic foothills, ΔSD categorical distribution is plotted in Fig.  for the three stations near Pic d'Anie, where the positive bias was found to be the highest in Sect. . In contrast to its general behaviour, AROME–Crocus strongly overestimates accumulations, particularly strong accumulations. At the same time, strong decreases are also underestimated, which results in a rather high positive bias.

The performance of models for daily snow accumulations is further studied thanks to the ETS, computed for threshold categories (Fig. ). Scores are similar for AROME–Crocus and SAFRAN–Crocus. The ETS is almost 0.40 for the “all accumulations” category (more than 0.2 cm) and is under 0.10 for high accumulations (more than 40 cm). SAFRAN–Crocus has a better ETS for small accumulations, but the ETS of AROME–Crocus is better for all accumulations over 10 cm, except for extreme accumulations (more than 60 cm). However, the very small sample size for this category (47 observed events) makes impossible any reliable interpretation. A distinction by altitudinal range shows equivalent ETS for AROME–Crocus and SAFRAN–Crocus above 1800 m, and higher ETS for AROME–Crocus for medium and strong accumulations under 1800 m (not shown).

A complementary information on winter precipitation comes from the network of gauges in the French Pyrenees (red dots in Fig. ). Daily accumulations of precipitation (rainfall plus snowfall, cumulated from 06:00 to 06:00 UTC) from the forcing models are then directly compared to precipitation gauges measurements for days with a maximum temperature of 2 ∘C in order to reduce the proportion of rainfall amongst precipitation. Most of these observations are assimilated in SAFRAN reanalyses, while they are not taken into account in AROME forecasts. Figure  shows cumulated precipitation by category for both models and observations (right) compared to cumulated ΔSD at the same stations (left). Contrary to ΔSD, AROME overestimates precipitation measured by gauges (+73  %). The optimal interpolation basis of the SAFRAN analysis system should mathematically not be biased on the assimilated observations over a long period. The slightly positive bias obtained in this study (+17 %) may be linked to the fact that some assimilated observations are not included in our evaluation dataset and/or to differences between the climatological guess and the mean precipitation amount of the 4 years under study. The strong overestimation of AROME is particularly notable for the largest amounts. The different distribution of precipitation and ΔSD for AROME, with a higher proportion of strong precipitation than of strong snow accumulations, may be due to settling effects: the stronger the snowfall, the stronger the snowpack settles under its own mass, which shifts the distribution to the left.

The overestimation of precipitation by AROME compared to precipitation gauges seems to be an apparent paradox, as we highlighted an opposite behaviour in terms of snow accumulation. This theoretical discrepancy can be explained by the quality of precipitation gauge measurements. The undercatch of solid precipitations by gauges, mainly due to wind effects on falling snowflakes trajectories, is well known and very variable. This issue is investigated by the WMO Solid Precipitation InterComparison Experiment e.g.. There is no undercatch correction applied to these manual measurements, which implies that real precipitation amounts can be underestimated in the observations under windy conditions. The difference between accumulation and precipitation errors also involves modelled snow density; this issue is discussed in Sect. .

A major part of model positive bias in SD is due to the underprediction of strong SD decreases. Consequently, the understanding of models biases implies a more developed study of ablation processes. Strong decreases, more than 10 cmday-1, can be related to ablation processes such as melting or wind-induced erosion, which need to be studied separately. To this end, two diagnostics have been applied to identify such processes. Melting snow days (MSD) correspond to days when the snow upper layer temperature is equal to melting point at 12:00 UTC, in SAFRAN–Crocus outputs (there are no snow surface temperature measurements available). Wind-blown snow days (BSD) are identified at automatic weather stations only, where 10 m wind measurements are available. BSD correspond to days when 10 m wind speed exceeds 8 ms-1 during more than 10 min but no melting is diagnosed (only dry snow can be drifted). This value is based on the estimate of wind threshold for dry snow transport by . These criteria are obviously quite rough, but a comparison with snow depth plots is quite satisfactory. As an illustration, the diagnosed days are reported in Fig.  together with the snow depth evolution measured and simulated by AROME–Crocus, at the Maupas automatic station (massif of Luchonnais, central Pyrenees, France), where blowing snow events are known to be frequent. For instance, a good example of BSD occurred on 14 December 2012 with a 60 cm snow depth drop. MSD happen generally after April 2013 and are associated with decreasing snow depth.

To quantify the impact of wind-blown snow events on the performance of models, the cumulated ΔSD for AROME–Crocus and observations are plotted in Fig. , for BSD and all days, with a finer categorisation of SD decreases. This study is restricted to seven automatic stations measuring wind speed and SD (mean altitude: 2203 m.a.s.l). For observations, BSD contribute to all decreasing rates, in the strongest proportion for high decreasing rates (more than 20 cmday-1). For AROME–Crocus, BSD do not contribute to the strong ablation categories but to small ablation and accumulation categories in the same proportions. Cumulated ΔSD for high decreasing rates is equal to -1106 cm for all observations and equal to -781 cm for BSD only (excluding MSD), while it is equal to 0 cm for AROME–Crocus in both cases. It means that wind-blown snow is the main contributor (71 %) to this category, the remaining contribution coming from MSD or other processes.

Similarly, the cumulated ΔSD is plotted in Fig.  for MSD and all days, at all SD stations. Very strong melting (more than 20 cmday-1) is sometimes observed but never predicted. Strong melting (between 10 and 20 cmday-1) is much underrepresented by models, while melting of less than 10 cmday-1 is overrepresented. Cumulated ΔSD for high decreasing rates (more than 20 cmday-1) is equal to -7741 cm for all observations and equal to -3215 cm for MSD only, while it is equal to -41 cm for AROME–Crocus in both cases. Melting snow represents 42 % of this category, the remaining contribution coming from BSD or other processes. The behaviour of SAFRAN–Crocus is similar to AROME–Crocus for BSD and MSD (not shown). The simple diagnostics for BSD and MSD may miss some wind-blown snow or melting events.

Consequently, the underestimation of strong decreasing rates comes mainly from ablation processes: on the one hand, from wind-blown snow events which are not represented by models, as they are small-scale processes; on the other hand, from an underestimation of strong snowpack melting (more than 10 cmday-1). Other reasons for very high decreasing rates can be the strong settling after an intense snowfall or a rain-on-snow event, but it probably constitutes a limited part of this category.