3.1. Means of Snow Cover Metrics over the GAR
Along the Alpine chain, at elevations between 1500 and 3000 m a.s.l., the average snow duration (LOS) over the 2000–2019 period was 202 ± 3 days. This value was obtained when calculating the mean LOS value over all the grid points in this elevation interval (bright green pixels in
Figure 2a—the mean elevation of these grid points is 2050 m a.s.l.) for every year and then performing the mean over the 20-year period, while the uncertainty was expressed as the standard error of the 20-year mean. The highest values were found for glaciers, whose average LOS was 360 ± 4 days (blue pixels in
Figure 2a). The very high LOS values of glaciers depend on the fact that, for them, the snow period is often very long but, also, from the fact that MODIS does not distinguish between snow-covered and snow-free glaciers. The other mountain regions in the GAR outside the main Alpine chain, i.e., the Jura, Vosges, Black forest, Bavarian forest, Dinaric Alps, Cevennes and Apennines, with elevations between 1000 and 1500 m a.s.l., also showed a high average LOS, with a value of 106 ± 4 days (see yellow pixels in
Figure 2a). In contrast, low-lying areas (elevations lower than 500 m a.s.l.) south of the Alps, such as the Po plain and the mouth of the Rhone in Southern France, located in the southwestern (SW) region, had an average LOS under 10 days (red pixels in
Figure 2a). However, in these areas, the LOS does often not reflect a continuous snow cover period but only the number of snow cover days over an entire year.
Aside from the Po plain and the mouth of the Rhone area, a slightly higher duration of the snow season was observed in the other low elevation areas in the GAR. This is the case of the eastern side of the Alps, particularly in the northeastern (NE) region, which had the highest average LOS for grid points with elevations lower than 500 m a.s.l. (45 ± 4 days) and at all elevations (74 ± 4 days). The snow cover duration was also relatively high in the northwestern (NW) region (60 ± 2 days considering all grid points and 25 ± 2 days for elevations lower than 500 m a.s.l.), while the southeastern (SE) and SW region had very similar values, with 49 ± 2 days for all points, 11 ± 1 days at elevations <500 m a.s.l. in the SE region compared to 48 ± 2 days at all elevations and 13 ± 2 days at elevations <500 m a.s.l. in the SW region.
The start of the snow season (SOS) was, on average, around day 33 ± 2 of the hydrological year at elevations between 1500 and 3000 m a.s.l. of the Alpine chain (see bright green pixels in
Figure 2b), day 63 ± 2 of the hydrological year for the other mountain regions in the GAR at elevations between 1000 and 1500 m a.s.l. and gradually earlier for higher-relief areas (day 3 of the hydrological year on average at elevations >3000 m a.s.l.; see the blue pixels in
Figure 2b). In contrast, in low-relief areas (elevation <500 m a.s.l.), the SOS started, on average, on day 97 ± 2 of the hydrological year (orange pixels in
Figure 2b), even if, in these areas, SOS does often not correspond to the start of a continuous snow cover period. As observed with the LOS, the NE region showed the earliest SOS on average considering all elevations (day 76 ± 2 of the hydrological year, compared to 86 ± 1 for the NW region, 92 ± 1 for the SE region and 93 ± 1 for the SW region, which includes the Po plain).
The end of season (EOS) generally occurred around day 235 ± 2 of the hydrological year at elevations between 1500 and 3000 m a.s.l. of the Alpine chain and day 169 ± 2 of the hydrological year for the other massifs in the GAR at elevations between 1000 and 1500 m a.s.l. (yellow pixels in
Figure 2c). Above 3000 m a.s.l., the EOS occurred, on average, on day 346 ± 1 of the hydrological year, while, at elevations below 500 m a.s.l., such as across the Po plain and in the mouth of the Rhone area, the average EOS occurred on day 128 ± 2 of the hydrological year (red pixels in
Figure 2c). Additionally, in this case, at low elevations, the EOS does often not correspond to the end of a continuous snow cover period. For the EOS, a slightly later value was observed for the NE region (day 150 ± 3 of the hydrological year compared to day 145 ± 1 of the hydrological year in the NW region). In the SE region, snow disappeared earlier than in all other regions (average EOS: day 141 ± 1 of the hydrological year in the SE region and 142 ± 1 in the SW region).
3.2. Factors Explaining the Spatial Variability in the Snow Cover
The main geographical variable determining the LOS is elevation. The strong relationship between LOS and elevation is evident from
Figure 3, where we represented, for each 5-m interval of the elevation range of the GAR, several quantiles of the distribution of the corresponding LOS average values in the period 2000–2019. Considering the medians (violet points) of the distributions of the LOS values in the different 5-m elevation intervals, the relationship between LOS and elevation showed a very strong correlation between 1000 and 3000 m a.s.l. (Kendall’s Tau = 0.99). The same plot showed a similar behavior for different percentiles of the distributions of the LOS values in the 5-m elevation intervals, even if the range in which the relationship between the LOS percentile and elevation was linear changes among the percentiles. At an elevation about 100 m a.s.l., a jump in the average LOS occurred; this was caused by a number of lakes/rivers and humid areas that were not properly filtered by the water masks in MOD10A1/MYD10A1 and the additional mask from MCD12Q1 and were observed as snow-covered for up to 80 days by the MODIS snow products (see
Figure 3).
Figure 3 also gives evidence that the values of LOS corresponding to a certain elevation have a large variability, with ranges often covering an interval larger than 100 days, suggesting that LOS depends not only on elevation but, also, on other climatic characteristics.
To highlight the rather wide range of LOS values corresponding to a given elevation, we represent in
Figure 4 the difference between the average LOS of each grid point in the period 2000–2019 and the average LOS of all grid points falling in the same 5-m elevation interval.
The contrast between the grid points on the northern and southern sides of the Alps is striking; in the NE region, 91% of grid points have a higher LOS compared to those in their respective 5-m elevation range, whereas, in the NW region, the percentage of grid points with a higher LOS is much lower, 33%, owing to the inclusion of warm areas, e.g., in Southern France. In the SW and SE regions, 82% and 80% of the grid points have a lower snow duration compared to those at similar elevations. In contrast, the Po basin is rather homogeneous, with differences lower than 10 days (positive and negative) compared to areas at the same elevation. In
Figure S1, we further represent the spatial distribution of the LOS grid point values, expressed in terms of the percentiles in the distribution of the GAR LOS values falling in the corresponding 5-m elevation intervals.
A clear distinction is evident between the points north of the Alps and south of the Alps, particularly in the NE region, where 82% of grid points have a percentile above the median and 46% above the 75th percentile, owing to the inclusion in this area of rather cold regions, such as the NE plains and the Balkan Peninsula north of the Danube River. In the NW region, 70% of the points have a percentile below the median, which is observed as a balance between the areas of Switzerland and Germany, with a larger number of points falling above the 50th percentile and a larger number of areas where the percentiles are lower the median, particularly in France towards the mouth of the Rhone. This latter pattern is also observed in the SW region, where 16% of grid points fall below the 5th percentile with respect to the elevation and 61% below the 25th percentile. In the SE region, an interesting difference is observed in the eastern part of Po basin north of the Apennine chain, where the LOS of the grid points is above the 75th percentile of the points located in the same elevation range (see
Figure S1). Elsewhere in this region, most grid points fall below the median with respect to elevation: 10% below the 5th percentile and 53% below the 25th.
To investigate the geographic and topographic drivers of the snow cover variability in the GAR, we performed a linear regression on each metric using elevation, slope, aspect, shading, latitude and longitude as the independent variables. Latitude and longitude were included as the proxies for climatic variability in the area. This regression explains a high percentage of the variances in all three metrics, with the highest R
2 for the LOS (0.91, see
Table 2). Looking at individual variables, the most important factor overall is elevation, which alone explains up to 85% of the variability in the LOS and 86% in the EOS, while, for the SOS, a lower percentage of the variance is explained. The other variables appear less important. Moving from the GAR to the four subregions, the explained variance generally increases for all three metrics, with the LOS and the NW region having the highest R
2 (0.94) and SOS in the NE region, the EOS in the SW region the lowest ones (0.89).
In order to better highlight the effect of these other variables, we considered the relation of the LOS, SOS and EOS with elevation, and we removed the effect of elevation from the data by subtracting the average value of the LOS, SOS and EOS in their respective 5-m elevation intervals. We subjected these residuals to the same multivariate analysis we performed for the LOS, SOS and EOS, obtaining an explained variance of 48%, 52% and 36%, respectively. Considering individual variables, latitude and longitude can explain most of the variance in the residuals of the three metrics, especially the SOS, with 32% of variance explained by latitude and 15% by longitude, while the two variables contribute to explaining 26% and 15% of the LOS. Neither aspect nor slope or shading can explain significant amounts of variance in these residuals in the GAR (max. 3% SOS explained by aspect), which indicates that the higher values we obtained for slope and for shading in
Table 2 were actually caused by the fact that these variables have a significant correlation with elevation (Kendall’s Tau is 0.75 for aspect and 0.03 for slope, both significant at the 99% confidence level).
We then repeated the analysis concerning the residuals using 1000-m elevation ranges, as well as different LOS classes (LOS < 180 days, 180 < LOS < 330 and LOS > 330).
If we consider separate 1000-m elevation ranges, latitude appears to explain most of the residual variance, up to 34% for the LOS between 1000 and 2000 m a.s.l., and generally more for the SOS compared to the EOS and LOS (see
Table 3); in general, the importance of all the variables tends to decrease with elevation, although there are some exceptions, e.g., aspect, which explains 18% of the variance in the SOS between 2000 and 3000 m a.s.l. As concerns slope, the variable explains a minimal amount of the variance, except for the EOS below 1000 m a.s.l., which might suggest a longer permanence of snow on sloping surfaces in this elevation band. Considering all variables in a multivariate regression, the maximum amount of explained variance is 57% for the SOS between 1000 and 2000 m a.s.l. Among the four subregions, the NW one shows the highest amount of explained variance between 1000 and 2000 m a.s.l. for all three metrics, ranging between 78% (SOS) and 68% (LOS) (see
Table S1).
Considering the different LOS classes, a similar pattern was observed, as latitude explains most of the residual variance, up to 35% for the SOS for the grid points with the average LOS < 330 days (see
Table 4), and the importance of the variables tends to decrease with the increasing average LOS. The multivariate regression showed a higher amount of explained variance below 180 days of the average LOS and the highest explained variance for the SOS (53%). Considering the different regions, the NW region is the one with the highest amount of residual variance explained below 180 days of the average LOS (58% in SOS), while the SE region is the one with the lowest amount (17% of the SOS for grid points with an average LOS < 180 days) (see
Table S2).
3.3. Interannual Variability in Snow Cover Metrics and Trend Analysis
The standard deviations of the grid point LOS records in the 2000–2019 period (
Figure 5a) show higher values (>20 days) for the high-relief areas on the southern side of the Alps, compared to the northern side. This is also highlighted when standardizing the standard deviation for elevation by subtracting the average standard deviation in 5-m elevation intervals (see
Figure S2). The southern side of the Alpine chain thus shows a lower average LOS but a higher interannual variability (see
Figure 4 and
Figure S2). In contrast with high-relief areas, grid points located in low-relief areas in the southern regions (Balkan Peninsula, eastern Po basin and mouth of the Rhone) have a generally low standard deviation (<10 days), although this is often caused by the fact that the LOS of these points is 0 over multiple years. The lowest values are those of glaciers, whose standard deviation is generally close to 0. Compared to the plains south of the Alps, the northern ones generally have LOS grid point records with larger standard deviations, and the NE region shows the highest average standard deviation of the LOS records of 24 days, whereas, in the other regions, the average standard deviation of the LOS records is rather similar, ranging between 16 and 17 days.
Besides the standard deviation of the LOS grid point records, it is also interesting to present the ratios between these values and the corresponding average values (
Figure 5b), which shows rather high values in many areas, as 4% of the pixels have higher LOS standard deviations than the LOS average. The errors of the average LOS grid point values are indeed lower because of the averaging over the 20-year period. However, these rather high values suggest that a LOS climatology with a 20-year period of observations might still have some uncertainty caused by the high year-to-year variability of the LOS grid point records. This problem particularly concerns low-elevation areas that often do not have a continuous snow cover period.
In order to complete the presentation of the standard deviations of the grid-point LOS records, we present in
Figure 5c several quantiles of the distribution of the grid point ratios between the LOS standard deviations and LOS mean values in each 5-m interval of the GAR elevation range. Here, a decrease in this ratio is evident with the increasing elevation, with values above 1 for the 99th percentile of the distribution of the ratio below 1000 m a.s.l. and all percentiles nearing 0 above 2500 m a.s.l., indicating a higher and important role of the interannual variability of the LOS records on the ratios at low elevations than at high elevations. The figure also shows the same jump at elevations about 100 m a.s.l. observed in
Figure 3, which appears to be caused by the erroneous classification of some water bodies.
We further examined the interannual variability of the snow cover metrics over the period of investigation, 2000–2019, by calculating the average of the different snow cover metrics for each year over the GAR and considering the individual regions and elevation bands. In the GAR, a high year-to-year variability was observed in the LOS, ranging from 41 days in 2006 to 85 days in 2005 (see
Figure 6). Considering the four subregions, the NE region always showed the highest LOS, ranging between 43 and 110 days, while the SW and SE had the lowest (32–67 days). A similar high variability was observed for the SOS, which ranged over the GAR between day 72 of the hydrological year in 2010 and day 96 in 2006 (see
Figure S3), and the EOS, which was observed to vary between day 134 of the hydrological year in 2001 and day 161 in 2004 (see
Figure S4). No significant trends were observed for these metrics over the GAR, owing to the large interannual variability.
Considering the separate 500-m elevation bands, we observed that the lowest values of the LOS occurred nearly always in 2006 for all subregions and elevation bands below 3000 m a.s.l. (see
Figure 6 and
Figure 7). In contrast, years with above-average LOS were more heterogeneous among the subregions and elevations, with no prevalence of a specific year. An above-average LOS was observed for example in 2005 below 2000 m a.s.l. in the NE and SE regions, in 2012 between 1000 and 3000 m a.s.l. in most regions and in 2003 between 2000 and 3000 m a.sl. in all the subregions (see
Figure 7). A high year-to-year variability was observed particularly below 500 m a.s.l., where the average LOS can range between 5 days (SW region in 2016) and 84 days (NE region in 2005) and between 1500 and 2000 m a.s.l., with a range of 87 days (SW region). The variability decreased with the elevation; between 3000 and 3500 m a.s.l., the LOS ranged between 316 days (SW region, 2017) and 359 days (NE region, 2000), while, above 4500 m a.s.l., the range was approximately 5 days in the NW region. Between 3000 and 4000 m a.s.l, significant trends at the 90% and 95% confidence levels were observed for the NW, NE and SW regions. Between 3000 and 3500 m a.s.l., the trends ranged between −5.4 and −6.3 days per decade, while the trends between 3500 and 4000 m a.s.l. were much lower in magnitude (see
Table 5).
Compared to the LOS, a more heterogeneous pattern with respect to the years of maxima and minima was observed for the SOS and EOS (see
Figures S5 and S6). In general, 2006 appeared as the year with the peak SOS for most regions and elevation bands in agreement with the low values found for the LOS, while 2003 and 2008 showed below-average SOS values. Significant positive trends (90–95% confidence level) for the SOS were observed between 3000 and 4000 m a.s.l. (with the only exception the SW region), ranging between +0.1 and +1.6 days per decade, with a greater magnitude in the lower elevation range, as for the LOS (see
Table 5). For the EOS, 2006 and 2010 were observed as years with below-average values in most regions and elevation bands, while an above-average EOS was seen in 2012. No trends reached 90% significance for this metric, although most of them appeared negative above 2500 m a.s.l., while the opposite was true below this elevation (see
Table 5).
To complete the analysis of the trends in the snow cover metrics, we also calculated them at the grid cell level. Among all grid points with significant trends, 91% showed negative trends (see
Figure 8 and
Figure 9), with an average of −17 days per decade, while a small peak in positive trends was observed around 4050 m a.s.l. Grid cells with trends significant at the 95% confidence level appeared concentrated mostly at elevations above 2500 m a.s.l., (77% of the total points with significant trends) because of the larger signal-to-noise ratio at the higher elevations (see
Figure 5c). The fraction of the cells with significant trend peaks between 3000 and 3300 m a.s.l., where about 20% of the grid cells had negative trends. At higher elevations, the fraction of the cells with significant trends decreased, because a lower variability was observed, as most of the grid points had (near-) permanent snow cover throughout the period. At elevations below 2500 m a.s.l., the frequency of the grid points with significant trends was much lower, with an average fraction of 2%.
Spatially, grid cells with significant negative trends appeared distributed mostly in the NE areas, on the northern side of the Alps and in the Po basin, with values <−10 days decade
−1. In contrast, a small number of significant positive trends (>5 days decade
−1) were seen along the Dalmatian and Southern French coasts, the Apennine chain and in France at the northwest margin of the GAR (see
Figure 9). The spatial patterns were similar for the SOS (see
Figure S7), while, as seen in
Table 5, the amount of pixels with significant trends was much lower for the EOS than for the other metrics. Significant negative trends in this metric were mostly found on the northern side of the Alps, as well as in the Po basin, while a small cluster of pixels with negative trends (<−5 days decade
−1) was observed at the E margin of the GAR. Pixels with significant positive trends were instead observed on the Dalmatian and Southern French coasts (see
Figure S8).
Thus, in low-elevation areas, particularly in the NE regions, the decrease in LOS appears to be caused by an increase in the SOS (>10 days decade
−1) rather than a decrease in the EOS, which is of lower magnitude (see
Figures S7 and S8).