a. Air temperature

Figure 2 shows the RMS values of the difference between measurements (43 sites, some with sometimes sporadic data) and SAFRAN analyzed fields for the annual mean air temperature over the 44 yr for the two datasets (WITH represented by the solid line and WITHOUT by the broken line). No constraint is applied to the analysis scheme and the presented RMS values therefore include the observation errors of the concerned sites, which increase the results. It is difficult to separate the relative share of these two errors. As a general indication, the values of RMS temperature observation error used in our area of the Alps range from 1° to 1.5°C depending on the site and result from our own monitoring and experience. These values are close to those suggested, for instance by Fuentes and Heimann (1996). The WITH set exhibits higher quality because of the additional information of the test sites, but the RMS difference between the two sets is low, from 0.1°C in 2000–09 to 0.3°C in the 1960s. The SAFRAN analyses are globally better at the end of the millennium because of the improvement of the snow and weather network with an increasing number of meteorological observations. For extreme values (details not shown here), the RMS values vary from 0.1°C in 1996 for WITH simulations for a site in the southern Alps to 4.7°C in 1986 for WITHOUT simulations for a site in the central Alps. The minimum (TN) and maximum (TX) daily air temperatures have also been compared. The bias is generally positive for TN (mean value of 1.2°C for WITH and 1.9°C for WITHOUT set) and negative for TX (mean value of −1.0°C for WITH and −1.4°C for WITHOUT set).

The relative decrease in performance over the 1980s has been identified to a lack of information in the databases, mainly due to the demise of the French manual observation network that had not yet been compensated for by the deployment of the new snow weather observation network in mountainous areas—which clearly produced a positive impact during the 1990s. The automatic observation network was also, at that time, only in its infancy, with data difficult to integrate in the analysis scheme. As described farther on (and also visible in Fig. 9), the 1980s are also representative of a net change in the warming temperature trend (Trenberth et al. 2007), especially concerning the daily minima that are not used explicitly by SAFRAN. This phenomenon could also partially explain the bad RMS values of that critical period if we could establish the sensitivity of this particular parameter with respect to the final result in the framework of a reduced observation network; however, this point is still under investigation.

The presented results are also an indirect evaluation of SAFRAN accuracy according to the information used and of the sensitivity of the analysis scheme to its input data. When the observation network is sparse, as during the 1980s, the analysis error is close to the guess-field error. On the other hand, a denser observation network implies less analysis error in the vicinity of the observation errors. As a whole, the magnitude of these differences is very close to those obtained by Quintana (RMS value of about 1.5°C), who used SAFRAN over all of France with a majority of low-elevation areas as previously indicated (Quintana-Seguí et al. 2008).

Some detailed results concerning seven selected sites (one for each French Alps department) are shown in Table 5 and concern TN–TX and mean daily temperature (TM) verifications. The values correspond to RMS differences and bias differences. Results of the comparisons for the WITH and WITHOUT experiments are indicated in a similar way. The results of the comparisons are obviously better for WITH experiments (because of the use of the additional observations in the analysis scheme), except for Luceram, a site recently set up in the southern French Alps (detailed results not shown) with few available data and for which insertion in the SAFRAN data decreases analysis performance. Except for this particular site, the bias is always positive for TN and negative for TX, as already mentioned. This supports our previous idea concerning the possible weaknesses of the diurnal cycle analyzed by SAFRAN (underestimation of the amplitude, but less error on the mean temperature value) and the possible improvements that could be achieved by using explicitly the information on these extremes. This also shows that the analysis process is not trivial and that the characteristics of each site have to be carefully taken into account, which is not yet the case for Luceram. The same analyses were performed using only winter data and both bias and RMS show similar values (not shown here).

b. 24-h precipitation

As 24-h precipitation is not a continuous daily parameter, direct comparisons are not easy. We therefore used two types of comparisons.

• For the sites previously selected in each French Alps department (seven sites, already used in Table 5), objective comparisons are computed between the observed and the WITH and WITHOUT corresponding SAFRAN analyzed quantities only when the daily quantity is higher than 0.2 mm. Table 6 shows this comparison using different statistical parameters such as the average of observed data and SAFRAN analyzed data, standard deviation of the difference, and the correlation coefficient. As for temperature, the Luceram site shows the worst results for all statistical parameters, which reinforces our previous doubts concerning the use of these data in the analysis scheme. However, the weak differences between WITH and WITHOUT results are representative of the good stability of the SAFRAN scheme, especially in relation to precipitation, for which characteristic spatial lengths are smaller than for temperatures.

• For all the 43 sites, contingency tables of daily precipitation have been compiled to compare observations (rows in the table) and SAFRAN analyzed data (columns). Seven classes from near 0 (≤0.2 mm) to high precipitation (>40 mm) were defined. Table 7 shows the result in terms of percentages in the different classes for the WITH and WITHOUT simulations. For the WITH results, the value of the Hansen and Kuiper Skill Score (Wilks 1995) of 0.607 as well as the percentage of well-classified cases of 68.2% (based on the diagonal elements of Table 7) are globally quite satisfactory. Concerning the WITHOUT experiment, the comparisons between the two sets show that the SAFRAN analyses without using the data of the additional sites are slightly worse with a Hansen and Kuiper skill score of 0.575.

In both experiments, SAFRAN analyses give values lower than measurements, especially for high values of precipitation. Despite this, these validations globally show the ability of SAFRAN to reproduce the mountain meteorological climatology of precipitation since 1958 but with a slight bias generally due to local effects. However, the real-time operational runs show that this does not affect the results for climatological purposes (Martin et al. 1994; Martin 1995; Quintana-Seguí et al. 2008).

c. Analyzed temperature and precipitation trends

Before discussing the final SAFRAN output analyzed on the massif scale in terms of temporal trends for precipitation and temperature, a quick overview of the observation series will be provided to evaluate the modeled results. The purpose is to assess the possibility of drawing conclusions with only the modeled fields at locations or areas where no observations are available. For this, 21 observation sites with more than 10 000 data (two observations per day) were selected from the initial list of 43 sites previously used (Table 4). For example, Fig. 3 shows the annual observed temperature for representative locations of the three main geographical areas: Nice for the southern Alps, Annecy for the northern Alps, and Villard-de-Lans for the central Alps. All locations exhibit a clear temperature increase over the past 40 yr of about 1.5°C for Annecy and Nice and 1.1°C for Villard-de-Lans (with higher variability). These results, required for the same period as the modeled results, are relevant for the last 45 yr, but cannot be extrapolated to longer periods such as the whole century. They exclude particularly the important 1940s and 1950s decades, as can be seen when comparing Figs. 3c and 3d for the Annecy series.

Figure 4a exhibits the mean temperature trend for 21 observation sites and the corresponding SAFRAN analyzed values downscaled at these points. As explained, it is difficult to simulate precise geographical locations with the modeled results, which do not take into account small-scale orographic effects at these locations. In addition, some observation sites have also been greatly influenced by surrounding urbanization, as is probably the case for Megeve. However, the mean air temperature trends for the 21 sites are 0.025°C yr⁻¹ for the observed data and 0.028°C yr⁻¹ for the SAFRAN simulated data. The results are therefore of the same order of magnitude even if the analysis overestimates the values for many points such as those in the Vercors massif.

Similar remarks can be made for the precipitation trends shown in Fig. 4b, especially in the northern Alps where both SAFRAN analyses and observations show a small temporal increase, often overestimated by the model. Trends are rather weak in the southern Alps for this parameter. As very few observation series cover the full temporal period and as the observations are above all representative of the winter season, these values are difficult to interpret both in time and spatially. The mean precipitation trends, presented in Fig. 4b, are +1.6 mm yr⁻¹ for the observed data and +2.6 mm yr⁻¹ for the SAFRAN simulated data. Note the clear latitudinal difference with a positive trend both for the observations of the northern Alps (those from about 1 to 30 on the x axis in Fig. 4b) and the SAFRAN results, and no real trend in the south. However, even if the positive trend values are consistent with those of Fig. 9, they are not statistically significant. Moisselin et al. (2002) point out the lack of consistency and of significance of most of the observed precipitation series over the southeast of France and these data are the main SAFRAN inputs. It is therefore impossible to draw valid conclusions on the scale of the observation site concerning these trends. However, considering cross validations only, which is our purpose here, we observe a consistent positive trend for precipitation in the northern Alps and no trend in the southern Alps, both for observed data and SAFRAN results.

a. SAFRAN air temperature climatology

The climate of the French Alps is temperate (Fig. 5) with annual mean air temperature at 1800 m MSL varying from 3.4°C in the north (Chablais massif) to 5.1°C in the south (Mercantour massif) near the Mediterranean Sea. This latitudinal variability is globally consistent with that observed for France as a whole at lower elevations (annual mean air temperature of 12.9°C for the city of Toulouse in the south and 10.0°C for the city of Lille, 850 km to the north). The variations are slightly higher in winter (from −1.4° to +0.4°C; Fig. 5, top right panel) than in summer (from +8.3° to +9.9°C; Fig. 5, bottom right panel). This low variability with latitude over these two seasons is partially due to the fact that results shown concern near surface conditions at a constant midaltitude elevation, which implies a partial influence of the more smoothed free atmosphere conditions. In addition, the strongest latitudinal gradient occurs mainly during the intermediate seasons according to the latitudinal variations of the polar front over France.

b. SAFRAN precipitation climatology

The climate of the French Alps is mainly determined by a northwesterly atmospheric flow as can seen in Fig. 6, which shows annual mean precipitation at the massif scale. This influence is visible both in summer and winter (rhs of Fig. 6), with more convective precipitation in summer. Year-to-year variability of annual precipitation can be very high (commonly 100% for annual data and much more seasonally) and regional trends exist (next Fig. 9b). Frei and Schär (1998), along with Beniston (2005), also insist on the dynamic interaction between weather systems and mountains and on the influence of sea moisture, especially during southerly conditions. At 1800 m MSL, the maximum annual precipitation amounts to nearly 2000 mm in the northwestern foothills (particularly Chartreuse and Aravis), and decreases to less than half that amount toward the southeast (831 mm for Queyras). A secondary maximum is located in the extreme southeast associated with the occurrence of northward Mediterranean flows. Note the small difference between summer and winter in the massif precipitation distribution despite the differences in meteorological patterns and types of precipitation. The snow fraction is about half in the northwest and only one-third in the south. In this respect, the three southernmost massifs (Ubaye, Alpes-Azuréennes, and Mercantour) get less snow than the overall driest massif (Queyras); Ubaye receives only an average of around 280 mm of snow water equivalent, as compared with 944 mm of total precipitation.

c. Mean vertical gradients

Four main areas of the Alps (Fig. 7a) regrouping the different massifs of Fig. 1 have been determined by expert meteorological clustering. The first split separates areas of greatest dissimilarity and divides roughly the northern from the southern areas. This line does not really run W–E, but rather SW–NE. Note that Haute-Maurienne in the central east has a pronounced southern influence. The second split separates the northwestern foothills from the central ranges. Whereas Belledonne, Beaufortin, and Mont-Blanc have a rather foothill character, Vercors (the southernmost foothill massif) resembles more the central massifs. The third split divides the southern Alps, notably with Ubaye included in the far south. These subdivisions can be considered logical, except perhaps for Belledonne, which has close ties to Chablais-Mt Blanc (rather than to its immediate neighbors), and Dévoluy that is closely related to Queyras-Parpaillon.

For each area (Fig. 7a) thus representative of the snowpack conditions, low-atmosphere vertical gradients have been computed for the near-surface temperature (e.g., the massif averaged air temperature at 2 m at different surface elevations) and for the annual rainfall. As shown in Figs. 7b,c, these gradients are very linear over the entire area. Note that this is not imposed by the analysis scheme (Durand et al. 1993) and results directly from the processing of the observations and the ERA-40 fields. From north to south, the mean near-surface vertical temperature gradient varies from −5.0° to −5.5°C (1000 m)⁻¹. These rates are very close to those computed by Rolland (2003) over the Italian Alps. The annual vertical rainfall gradients exhibit a larger latitudinal dependence with respective values from north to south of 294, 195, 172, and 178 (1000 m)⁻¹.

a. Temperature trends

SAFRAN filtered daily temperatures over several areas are presented in Fig. 9a together with the filtered NAO index. The mean values over the entire Alps (black curve) exhibit the classical shape of the last years characterized by a plateau until the 1970s, followed by a more pronounced increase of about +1°C. All the different regional areas show the same features modulated by the latitudinal variability and temporal smoothing. These characteristics have already been pointed out by Trenberth et al. (2007) for a larger spatial scale but with the same magnitude for the temperature trend, and are mainly due to the increase of the daily minimum temperatures as quoted by Moisselin et al. (2002) and Beniston (2005). On a large scale, Prömmel et al. (2007) have identified a relationship between strong westerly flows across the North Atlantic and a positive NAO index, which results in a correlation between NAO and air temperature that is quite large and positive in the north of the Alps but smaller in the south. Beniston and Jungo (2002) widely studied the relationship between NAO index and the temperature and pressure fields in Switzerland with a particular emphasis on the increased values since the 1980s. They successfully linked high-value NAO periods over the entire Alps to high pressure blocking events accompanied by vertical circulation inducing compressional warming, subsiding velocities, and decreasing cloudiness and thus positive temperature anomalies. Scherrer et al. (2006) have also shown an enhanced occurrence of blocking-type high pressure systems over Europe and its link with NAO. According to Fig. 8a, the correlation between filtered NAO index and temperature over our working area is about 0.7, which corroborates the previous results and indicates a mutual influence on this finer scale. The latitudinal variability is consistent with the discussion in Prömmel et al. (2007).

Detailed results (not shown here) show an overall rise of about +1.5°C for Chablais (the northernmost French massif) over the last 30 yr. The winter half-year increase was almost +2°C and the summer increase about +1.5°C with a constant very limited variation but higher variability during late summer. All foothill massifs (Chablais–Vercors) including Mont-Blanc, Beaufortin, and Belledonne show in general the same behavior. Chartreuse is the most extreme massif with a net winter rise of almost +2.5°C. The Mercantour massif is well representative of the central and southern massifs. The most striking difference to the northern massifs is a strong temperature decrease in early winter (−2°C) since the mid-1980s followed by only a slight increase in midwinter but an increasing trend in late winter (up to +3°C), which implies only a slight increase (+0.5°C) on the scale of the overall winter season. All central and southern massifs roughly follow this pattern with Haute Tarentaise-Vanoise-Maurienne being the least distinctive and Queyras-Parpaillon-Ubaye being the most pronounced.

b. Precipitation trends (rain and snow)

Figure 9b shows the same results for precipitation. As in several other studies, no clear temporal trend or clear relationship with the NAO index can be found for any of the concerned areas that exhibit only a clear latitudinal variability between the northern and southern Alps. A linear fitting procedure was performed for the curves representative of the northern and southern areas in relation to the validation process presented in Fig. 4 and the very small trends observed in the northern observations. However, these indications of a possible small positive trend in the north and of a very flat shape in the south are not statistically significant and allow no conclusions to be drawn.

These results could appear to be contrary to other studies such as that of Quadrelli et al. (2001), who showed, over a much larger area of the Alps (about 15 times bigger than ours), a clear negative correlation between NAO index and the first component of an EOF decomposition of the winter precipitation field. In fact, the numerous differences with our experiment, in particular our study area for which the main climatological precipitation features are more represented by their second EOF component and the use of yearly precipitation at midaltitudes (1800 m MSL) over 44 yr, make comparisons and conclusions difficult because of these scale and decomposition effects.

However, all these considerations and study comparisons lead to questions, especially for this precipitation parameter presenting high local variability and links with large-scale circulation patterns that are difficult to state. Prömmel et al. (2007) mention the northern deviation of the westerly winds in the southern Alps together with decreased precipitation amounts. In their study of accumulated new snow totals (a quantity relatively well related to precipitation) over the Swiss Alps, Scherrer and Appenzeller (2006) observe a correlation of their first orthogonal mode with surface pressure anomalies over southeastern Europe. Quadrelli et al. (2001), over their large area, show a good correlation between their first precipitation mode and north–south fluctuations of the Atlantic midlatitude westerlies, whereas their second mode, much less correlated to NAO, is more influenced by northwest flows. Indeed, the latter meteorological situations are basically the rainiest over our working area (Fig. 6) during winter, whereas the summer season is more influenced by convection. These two points can partially explain our very weak correlation with NAO.

Detailed results (not shown here) show flat mean shapes in Chablais both for winter and summer seasons but with larger interannual and interseasonal variations. The Grande Rousses massif presents a significant increase during the summer period (∼70 mm per decade) and is one of the only massifs to show a small positive trend, whereas Mercantour shows a small negative trend especially during the winter season. However, Chablais presents two extreme values for the last two winters (2001 and 2002) and Mercantour includes three very high values during recent winters (1997, 1998, 2001) while its snowfall rises to a high point at the end of the 1970s before dropping.

c. Annual distribution

The annual distribution of monthly mean temperature and precipitation at 1800 m MSL is presented in Fig. 10. The previously seen marked temperature increase is clear in Fig. 10a (left: entire Alps; top right: northern Alps; bottom right: southern Alps) both in winter and summer seasons. The winter season exhibits fewer cold events, begins a bit later in the north, and ends earlier in the south. The summer season becomes clearly warmer over a longer time. The transition period between winter and summer temperatures appears to be decreasing (as shown by the “green” area) in all regions, which denotes shorter intermediate seasons, consistent with the present personal feelings of many inhabitants.

The precipitation (Fig. 10b, same areas as in Fig. 10a) does not exhibit any temporal structure and we see mainly the latitudinal gradient as well as the main features of the southern Alps: dry in summer and winter and storms in autumn. Particularly in winter, year-to-year variability can be very high and appears to have increased even more in recent years. While the last decade is generally marked by low precipitation, some outstanding maximum years clearly stand out. The early winters of 1997, 1998, and 2001 have beaten all records in the south, but below 2000 m MSL precipitation fell predominantly in the form of rain. The midwinters of 1995 and 1999 brought record precipitation in the north falling as snow down to 1000 m MSL and the far south received large snow amounts down to low levels in 1993 and 1995. The year 2001 was an outstanding year for late-winter record snowfalls throughout the French Alps except in the far south (Mercantour, Alpes-Azuréennes) and Haute-Maurienne in the east. Even if it is standard to split the Alps into a northern and southern part, the central massifs, in particular, can show major deviations. Particularly in early summer, these central massifs can differ considerably from both northern and southern massifs (results not illustrated here), showing a strong increase (up to 100% over the whole period for Grandes-Rousses and Pelvoux).

d. Vertical trends

At different elevations (from 600 to 3600 m MSL), Table 8 presents, among other features, the Spearman’s rank correlation coefficient “ρ” computed for the daily near-surface SAFRAN analyzed temperatures over the entire area of the Alps for the new 47-yr period. This coefficient is simply a special case of the Pearson product-moment coefficient in which the data are converted to rankings before calculation, and has been widely used by many authors such as Moisselin et al. (2002) for trend detection. Here, its vertical variation shows a clear positive increase with time especially at midelevations (1500–2000 m MSL). The corresponding significance has been evaluated through a Student’s t test with a 95% confidence interval [the corresponding t values are presented in Table 8 in the t(ρ) column]. As the t threshold corresponding to our sample size is about 2, all levels except the highest (3600 m MSL) present a significant positive near-surface temperature increase over the limited study period. Even though Spearman’s method does not require the assumption that the relationship between the variables is linear, many studies (such as Trenberth et al. 2007 and references therein) have computed linear fits, but generally over longer periods. We have also determined such fits at the different elevations despite the 47 available years, which implies results only representative of this period. The limited accuracy of the linear assumption is visible through the values of the square of the correlation coefficient (column R² in Table 8) between raw and fitted values where only midelevation values are of little significance. However, all vertical levels (except the highest) exhibit a positive linear trend (the a column) corroborated by their 95% confidence interval (±a column) with a clear emphasis at midelevation and a weaker signal higher.

Discussion of these results is hampered by the characteristics of the analyzed temperature, which here is representative of the near-surface conditions but at different mountainous elevations. It is therefore the “subtle” result of surface and free atmosphere conditions with the interaction of the orographic features and effects such as sun occultation or meteorological-induced circulation. The highest elevations can therefore be assumed to be more representative of the free atmosphere conditions, which implies a reduced, or very weak, positive temperature trend. At low elevations, the temperature trend is superimposed on other phenomena such as valley effects, boundary layer processes, local observation site characteristics, and less sun radiance, which introduce noise in the positive signal. A complementary explanation of this vertical variability can be found in the behavior of the NAO index for which the fluctuations are linked to pressure field anomalies. Over a large part of our study period, the observed positive NAO fluctuations (Fig. 9) are thus representative of more frequent high pressure situations and of the induced vertical temperature inversions for which the tops are generally located within these midelevations. These phenomena could also be increased by the winter snow cover decrease at these elevations and by increased summer dryness (not presented here).

The mean values obtained at midelevations correspond to those given by Trenberth et al. (2007) but with a larger confidence interval, mainly due to our short time series. Rebetez and Reinhard (2007) find a slightly higher value (0.057°C yr⁻¹) for 12 Swiss stations over the 1975–2004 period. Beniston and Jungo (2002) also determined an altitudinal variation of temperature anomalies with minimum values at low elevations. These results can also be compared to the observed trend values in Fig. 4, which well illustrate the variability in our mountainous area.

The similar study for precipitation (not shown here) does not show any significant results for our area over the same considered time period.

e. Link between temperature and precipitation trends

Looking at snow precipitation trends in the light of temperature trends reveals that in the north, falling temperatures are associated with slightly rising snowfalls (early winter) and rising temperatures cause diminishing snowfalls (midwinter–early summer). Constant late-summer temperatures show no impact on snow precipitation trends, as would be expected. However, the example of Mercantour in the far south shows that strongly dropping early winter temperatures do not necessarily result in increasing snowfalls, since total precipitation is also decreasing. At Grandes-Rousses, in the central part, we see that strongly rising late-winter temperatures have hardly any effect either on snow or rain precipitation, but a strong early summer temperature increase is accompanied by a very strong rainfall increase and a slight snowfall decline. Finally, near-constant late-summer temperatures are accompanied by a strong positive rainfall trend but have no effect on snowfall.

Beniston (2003) studied the possible impacts of these climatic trends in mountainous areas on hydrology, snow conditions, glacier vegetation, and tourism; he mentions more particularly several research works with SAFRAN-Crocus. Some elements of our study are in common with those of Beniston, especially the uncertainties concerning precipitation and the NAO–temperature link.