Climate variability and change in the Greater Alpine Region over the last two centuries based on multi-variable analysis

Transcription

1 INTERNATIONAL JOURNAL OF CLIMATOLOGY Int. J. Climatol. (2009) Published online in Wiley InterScience ( Climate variability and change in the Greater Alpine Region over the last two centuries based on multi-variable analysis Michele Brunetti, a * Gianluca Lentini, b Maurizio Maugeri, b Teresa Nanni, a Ingeborg Auer, c Reinhard Böhm c and Wolfgang Schöner c a Institute of Atmospheric Sciences and Climate, via Gobetti, 101 I Bologna, Italy b Department of Physics via Celoria, 16 I Milan, Italy c Central Institute for Meteorology and Geodynamics, Hohe Warte 38, A-1190 Vienna, Austria ABSTRACT: An extensive analysis of the HISTALP database is presented with the aim of giving a comprehensive picture of secular climate variability and change in the Greater Alpine Region (GAR, 4 19 E, N). The HISTALP database encompasses 242 sites and concerns temperature, pressure, precipitation, cloudiness, sunshine duration, vapour pressure and relative humidity. The analyses are based on four regional mean records representing different GAR low-level areas and on an additional mean record representing high-level locations. The first goal of the paper is to give an overview of the seasonal and annual records for the different variables, aiming to highlight both variability on decadal time scale and long-term evolution. Then it focuses on trend and correlation analysis. Trends are presented both for the period of common data availability for all regional average series and for moving windows that permit studying the trends over a wide range of timescales. Correlations among the different variables are presented both for the regional average series and for their high-pass-filtered versions. The analyses, beside highlighting a warming that is about twice as large as the global trend, also show that the different variables have responded in different ways to this warming and that the mutual interactions linking the different variables are often present only at specific temporal scales and only in parts of the GAR and in defined seasons. In spite of this complex behaviour, which may also be due to some residual inhomogeneities still affecting the data, the analyses give evidence that the HISTALP database has an excellent internal consistency and show that the availability of a multi-variable database turns out to be very useful in order to evaluate the reliability of the reconstruction of each variable and to better understand the behaviour and the mutual interactions of the different variables. Copyright 2009 Royal Meteorological Society KEY WORDS homogenised series; Alpine region; temperature; pressure; precipitation; cloudiness; sunshine duration; vapour pressure; relative humidity Received 26 September 2008; Revised 10 December 2008; Accepted 11 December Introduction The Alps constitute the most relevant topographic ridge of Europe, and influence atmospheric circulation over a wide range of scales. As a consequence of its complex geography, the Alpine region exhibits a variety of different climates, ranging from maritime influences (both from the Mediterranean Sea and the Atlantic ocean) to continental features (such as the plains of Eastern Europe and the inner Alpine valleys), and from low elevation to mountain climates. The Alpine region is not only interesting because of its complex geography but also for its exceptional availability of high-quality secular data records, especially as far as the early instrumental period (pre-1850) is concerned. Activities of the past years have considerably improved the availability and quality of long-term climate data * Correspondence to: Michele Brunetti, Istituto ISAC CNR Via P. Gobetti, 101 I BOLOGNA ITALY. m.brunetti@isac.cnr.it and metadata for the Alps and their wider surroundings (4 19 E, N), an area that will be referred to as GAR (Greater Alpine Region, covering approximately km 2 ; Figure 1); such activities allowed the construction of a huge multi-variable database, called HISTALP (Auer et al., 2007). The HISTALP database has already been extensively analyzed for temperature and precipitation (Böhm et al., 2001; Auer et al., 2005; Brunetti et al., 2006b), whereas only a few very preliminary results have been presented concerning other important variables such as pressure, cloudiness, sunshine duration, vapour pressure and relative humidity, that have been added to the database only in the past years. The state of the art of the HISTALP database is widely described by Auer et al. (2007) in which, besides its setting up and homogenisation, its potential applications have also been presented and explored. The Auer et al. (2007) paper may be regarded as the logical part 1 of the present paper, which aims to analyze the HISTALP database giving a comprehensive picture Copyright 2009 Royal Meteorological Society

2 M. BRUNETTI ET AL. Figure 1. The station network in the GAR (dots) and the resolution of the gridded version of the HISTALP dataset. The figure also shows the regions corresponding to the four low-level CRSMs. of secular climate variability and change in the GAR, by means of an analysis of a wide range of relevant meteorological parameters (temperature, pressure, precipitation, cloudiness, sunshine duration, vapour pressure and relative humidity). The truly innovative and original aspect of the HISTALP database lies in the fact that, on one hand, HISTALP contains a wide range of meteorological variables and, on the other hand, that it covers a large geographical domain encompassing 12 European nations. In this sense, the HISTALP database represents a significant novelty and therefore it is not so easy to find, in the international scientific literature, papers dealing with comparable databases. Actually, as far as the GAR is concerned, there are other interesting papers in addition to the ones presented by the research team that contributed to the construction of HISTALP, even though they do not make use of all HISTALP variables: among them, it is worth citing Efthymiadis et al. (2006) and van der Schrier et al. (2007). The former aims at constructing a min resolution precipitation grid for the GAR, and the latter provides a discussion on European Alpine moisture variability for the same time period. The simultaneous analysis of a wide spectrum of meteorological variables, as the ones included in HISTALP, is significant for better investigating the atmospheric processes that modulate and trigger the variability and trends shown by the single meteorological parameters and for understanding the mutual interactions linking the different variables (Gaffen and Ross, 1999; Kaiser, 2000; Wang and Gaffen, 2001; Huth and Pokorná, 2005; Beniston, 2006). Moreover, the availability of a multi-variable database allows assessing the consistency between the behaviour of the different variables, thus permitting to increase the confidence in the results of the analyses (Gaffen and Ross, 1999; Kaiser, 2000; Wang and Gaffen, 2001; Huth and Pokorná, 2005). A further benefit of the availability of secular records for a wide range of variables concerns the possibility of comparison between models and observations (Matulla et al., 2005). The paper is organised in six sections. After the introduction, Section 2 presents the HISTALP database and describes the regional average series used in the paper. Section 3 gives an overview of the seasonal and annual records for the different variables aiming to highlight both variability on decadal time scale and longterm evolution. Sections 4 and 5 respectively focus on trend and correlation analysis. Trends are presented both for the period of common data availability for all regional average series and for moving windows that permit studying the trends over a wide range of periods and timescales. Correlations among the different variables are presented both for the regional average series and for high-pass-filtered versions of them, as high-pass-filtered records often allow for a better evaluation of the common variances among variables. Finally, Section 6 gives some conclusive remarks. 2. Data This study is based on HISTALP, one of the best regional secular multi-variable databases available in the world; in particular we use the version released in

3 CLIMATE VARIABILITY IN THE GREATER ALPINE REGION June 2006, consisting of records updated to At present time, it is available on the ALP-IMP web site ( The HISTALP database has become progressively consistent in the frame of different international projects, from ALOCLIM (Auer et al., 2001) and ALPCLIM (Böhm et al., 2000) to ALP-IMP (Böhm, 2006), and many national projects. In the meantime, rigorous quality check and accurate homogenisation have been undertaken, also thanks to a remarkable metadata availability. The HISTALP database is widely described by Auer et al. (2007) that give full details on data and metadata availability, time evolution of network density, inhomogeneity and outlier detection and adjustment, gap-filling, gridding and clustering of the records into climatologically homogeneous subregions. The HISTALP database encompasses 242 sites (Figure 1) and concerns temperature, pressure, precipitation, cloudiness, sunshine duration, vapour pressure and relative humidity. Actually, the last variable is not independent of the others, as atmospheric thermodynamics allows expressing relative humidity as a function of temperature, pressure and vapour pressure. However, as the HISTALP database does not include individual observations, but monthly averages (from daily means), relative humidity also has to be considered as an independent variable. Temperature, pressure and precipitation records display the best data availability in terms of both space and time coverage. Cloudiness and sunshine duration have a lower density and some areas have no data; vapour pressure and relative humidity have more severe limits and cover only few parts of the GAR so far. The climate information in HISTALP is not only stored in station mode, but also as gridded data and Coarse Resolution Subregional Mean series (CRSMs). The gridded data have been obtained by means of a modified Gaussian-weighted inverse distance interpolation (for further details, see Auer et al. (2007)). They consist of monthly temperature, pressure and precipitation anomalies, expressed in terms of differences from the normal values for temperature and pressure (additive anomalies) and in terms of ratios with respect to the normals for precipitation (multiplicative anomalies). The grid resolution is 1 degree for both latitude and longitude (Fig. 1). Also the CRSMs consist of monthly additive (for temperature, pressure, vapour pressure and relative humidity) or multiplicative (for precipitation, cloudiness, and sunshine duration) anomalies with respect to the averages. The CRSMs were constructed in order to give a synthetic description of GAR variability and longterm trends, by condensing most of the climatic signal of the 242 sites of the database in a few areal mean records. The clustering of the GAR records in a few climatologically homogeneous subregions was obtained by a spatial optimisation of the results of different single-variable regionalisation attempts, each one based on a Principal Component Analysis applied to the station records. The results are discussed in detail by Auer et al. (2007); they give evidence that most variables show a favourable subdivision into four horizontal subregions of approximately similar size, roughly corresponding to the scheme northwest (NW) northeast (NE) southwest (SW) southeast (SE) (Figure 1) and an additional cluster of high-elevation sites. So, the CRSMs have been obtained by clustering HISTALP stations into five groups. In order to reduce the entity of the data to be managed and to allow a more schematic description of the results, the analyses presented in this paper are based only on such CRSMs. Compared to working with the single station or gridded series, the averaging procedure used to produce the CRSMs can also be expected to further reduce possibly undetected inhomogeneities of the station or grid point records. CRSMs are available at monthly, seasonal and yearly resolution, with seasons defined according to the scheme DJF, MAM, JJA and SON and with winter dated according to the year in which December falls. Years are defined by the period January December. The analyses presented in this paper concern seasonal and yearly records. Table I shows the CRSMs considered in this paper and gives evidence of the starting year corresponding to each variable. Moreover, it defines the acronyms that will be used for the different areas. Besides the five basic CRSMs, for some applications five additional combined subregional mean series (Low (L), North (N), South (S), West (W), East (E): see rows 6 to 10 in Table I) were constructed in order to provide further information. Such combined series were also used to study the differences between the W and the E mean subregions (the W E pair is expected to be particularly useful for detecting oceanic versus continental features), as well as between the N and the S subregions (such pair is expected to investigate Central European versus Mediterranean features), and between the high (H) and the L ones (the H L pair is deemed useful for detecting possible elevation-dependent signals). The five basic CRSMs (see lines 1 to 5 in Table I) were calculated as means of the respective station mode series, whereas the five combined CRSMs were calculated as shown in Table I. The length of a combined subregional mean series was always set to that of the shortest of the corresponding basic CRSMs. For temperature, pressure, cloudiness and sunshine duration all ten CRSMs were constructed, whereas for precipitation H CRSMs was not constructed, because, as described and referenced more in detail in Auer et al. (2005), wind exposed summit sites, with a high solid precipitation share, were excluded from the homogenised version of the HISTALP precipitation dataset. For vapour pressure and relative humidity, due to lowest spatial coverage, CRSMs were produced only for NW, NE, SE, H, L and N, with L actually being representative only of three out of four regions.

4 M. BRUNETTI ET AL. Table I. The five basic CRSMs and their five linear combinations. HISTALP-CRSM-subregion code Starting years of CRSM-series acr. Combined CRSMs Subregion Air pressure Temperature Precipitation Sunshine duration Cloudiness Relative humidity Vapour pressure NE Northeast NW Northwest SE Southeast SW Southwest n n H High n L (NW+ NE + SW + SE)/4 Low N (NW+ NE)/2 North S (SW+ SE)/2 South n n W (NW+ SW)/2 West n n E (NE+ SE)/2 East n, the CRSMs is not available;, the combined CRSMs was calculated even though not all CRSMs were available 3. Overview of seasonal and annual CRSMs for the different variables Before going into the details of trend and correlation analysis, in this section we present an overview of the evolution of the HISTALP variables, aiming both to give evidence of the behaviour of the different variables in the various subregions and to investigate the consistency among all the CRSMs. The discussion is based on Figures 2 to 8 that show the annual CRSMs of HISTALP variables, together with Gaussian lowpass filters (GLPFs), superimposed on the records, in order to give evidence of (1) decadal variability (GLPF with σ = 3 years and window of 31 years, hereinafter GLPF3) and (2) long-term evolution (GLPF with σ = 15 years and window of 61 years, hereinafter GLPF15). The figures also show the curves obtained by subtracting the GLPF15 from the CRSMs and by smoothing the residuals by means of 3-year σ GLPFs. Such records (RES3s) give further information on decadal variability and are more effective for the comparison of variables showing different long-term evolution. Most of the signals discussed in this section will be described more in detail in the next sections, in which long-term trends and correlations among variables are discussed Temperature Temperature CRSMs analysis (Figure 2) is performed starting from 1774, in order to have full data coverage at least for low-level subregions. GAR temperature evolution is discussed extensively in Böhm et al. (2001) and Auer et al. (2007). They give evidence that all CRSMs are similar in year-toyear and decadal variability as well as in long-term trend, allowing the whole area to be studied by means of only one series: the mean over the entire GAR. Such a record (see GLPF3 of L CRSMs in figure 2) displays annual mean values starting from lower values before 1790, followed by two relative maxima in the 1790s and the 1820s, interrupted by a sudden cold period in the 1810s. After the 1820s there is a gentle trend toward lower temperature values through two main minima in the 1850s and around 1890 with a relative maximum in the 1860s. The whole 20th century is characterised by rising temperatures toward a first maximum in the late 1940s and a second one in the last 15 to 20 years, which is the principal maximum of the 250-year period covered by the record. The seasonal analysis of L CRSMs (figures not shown) gives evidence that, beyond some common features, there are also remarkable differences. Concentrating only on the long-term features, all seasonal series initially decrease to a minimum, then show a positive trend until 2005, but the location of the minimum is different: around 1890 for winter, in the 1850s for spring, in the 1910s for summer and autumn. Moreover, the evidence of the initial decreasing trend is highest in spring and summer, lower in autumn and lowest in winter. Actually, at the present time, more analyses of the data of the 18th century and of the first part of the 19th century are in progress. They started from the conclusive results of the ALP-IMP project (Böhm, 2006) that displayed, especially in summer, a significant disagreement between instrumental data and temperatures estimated by means of tree-ring analysis (Frank et al., 2007). Such analyses seem to suggest that the summer temperature data of the early period (i.e. before about 1860) may be slightly overestimated. The warmest years in the period are 1994 and 2000, both with an average GAR anomaly of +1.8 K relative to the period. Other remarkable warm years are 2002 and 2003 (+1.7 K and +1.6 K respectively). It is, however, worth noticing that even if most of the warmest years concern the past 15 to 20 years, some years of the early period such as 1822, 1794, 1797, 1834 and 1811 also show yearly mean temperatures over 1 K above the average. These values may, however, at least partially, be an effect of the fact that the summer data of the early period may be slightly overestimated. At seasonal level, the

5 CLIMATE VARIABILITY IN THE GREATER ALPINE REGION Figure 2. The upper and lower graphs show the annual temperature anomalies for the five basic CRSMs and for the L CRSMs (thin lines), together with the corresponding Gaussian low-pass filters (continuous thick line: GLPF3; dashed thick line: GLPF15). The intermediate graphs show the corresponding RES3s. They are obtained by subtracting the dashed thick line series from the continuous thin line series and by low-pass filtering the residuals by means of the GLPF3. Figure 3. as in Figure 2, but for pressure.

6 M. BRUNETTI ET AL. Figure 4. as in Figure 2, but for precipitation. Figure 5. as in Figure 2, but for cloudiness.

7 CLIMATE VARIABILITY IN THE GREATER ALPINE REGION Figure 6. as in Figure 2, but for sunshine duration. Figure 7. as in Figure 2, but for vapour pressure.

8 M. BRUNETTI ET AL. Figure 8. as in Figure 2, but for relative humidity. most remarkable warm event is indeed summer This well-known warm event (see e.g. Schär et al., 2004; Ansell et al., 2006; Brunetti et al., 2006a) characterised all the Alpine region with the highest anomaly in NW (+4.8 K) and the lowest one in NE (+3.9 K). The coldest years are 1829, 1864 and 1871, with an anomaly of 1.5 K, followed by 1816, 1805, 1879 and 1940, all with an anomaly of 1.4 K, whereas the lowest seasonal anomaly is winter 1829 (from December 1829 to February 1830), with 5.1 K (Munzar, 2000; Casty et al., 2005). So, the comparison with past temperatures shows anomalies, such as the one of winter 1829, that seem to give evidence of more extreme seasons than summer Actually the anomalies from the normals are probably not the most effective data for estimating the exceptionality of past warm and cold events, as anomalies depend both on year-to-year variability and long-term changes. Moreover, the seasonal dependence of year-to-year variability causes the comparison among events of different seasons to be not completely significant, as in the Alpine area winter shows higher variability than summer. So, in order to allow a more significant comparison among the most extreme seasonal temperatures, we also analysed seasonal subregional series obtained by first detrending the records and then by normalising all detrended records by their standard deviations, calculated over the period of temperature data availability for all-gar subregions ( ). The detrended records were simply obtained by subtracting from the data the GLPF15 shown in Figure 2. Such smoothing curves allow removing from the records running normal values defined for periods of the order of 50 years. The results of the analysis of such records highlight the exceptionality of the summer 2003 event: it has a value of 4.2 (i.e. it is 4.2 summer standard deviations above the summer normal value corresponding to the year 2003), which is 1.8 summer standard deviations above the second warmest season. As far as the coldest events are concerned, there is another event with a very remarkable anomaly: spring 1785 (3.6 spring standard deviations below 1785 spring normal). Other remarkable cold events are autumn 1912 (3.2 autumn standard deviations below 1912 autumn normal) and the already discussed winter 1829 ( 2.8 winter standard deviations). The well-known summer 1816 event (1816 is known as the year without a summer ; it followed some years characterised by particularly strong volcanic eruptions, the most violent one being the Tambora (Indonesia) in April 1815) has only the 7th rank with 2.6 summer standard deviations below 1816 summer normal, even if it turns out to be the most important summer cold event.

9 CLIMATE VARIABILITY IN THE GREATER ALPINE REGION 3.2. Pressure As for temperature, CRSMs analysis (Figure 3) is performed for pressure also, starting from 1775, in order to have full data coverage at least for low-level subregions. Also, for pressure the low-level CRSMs show very good agreement both in terms of year-to-year and decadal variability (GLPF3), whereas the agreement with the high-level CRSMs is less visible. Such a behaviour is due to the fact that station level pressure anomalies depend on both air pressure at sea level and air virtual temperature, as virtual temperature regulates pressure vertical lapse rate. Such virtual temperature dependence gives a rather small contribution to low-level CRSMs, as the average over subregional station heights ranges from about 50 meters for SW to about 400 meters for NW and NE, with the SE stations being characterised by an average level of about 250 meters. On the contrary, such an effect is much more important for high-level stations whose heights are up to about 3500 meters. In order to give a quantitative estimation of this effect we assumed a sea-level pressure of 1013 hpa and a wide variety of possible combinations of temperature and relative humidity and we evaluated, by means of the hypsometric formula, the sensitivity of pressure at 400 meters to virtual temperature changes: the results give evidence of a sensitivity of hpa/k. So, for low-level CRSMs, such an effect is rather low in relation to the variability of the records (see Figure 3). It has, however, to be considered if small signals, such as the ones concerning long-term evolution, are taken into account. By averaging all low-level CRSMs, we obtain a record that captures most of the GAR sea-level pressure year-toyear and decadal variability. Such a record (see GLPF3 of L CRSMs in Figure 3) starts with a relative maximum in the 1770s, then the values decrease, reaching minima in the 1800s and 1840s, interspersed with relative maxima around 1820 and in the 1830s. After the minimum of the 1840s there is an approximately 100-year period with a tendency to display negative anomalies (the most evident minima concern the 1870s, the 1910s and the 1930s), even though in this period there are maxima such as the one in the 1860s. After the 1930s minimum there is a more or less continuous tendency to increase, with remarkable maxima in the late 1940s and around The last maximum is also the absolute one, with values about 1 hpa above the long-term average. Therefore, the most evident long-term features of annual L pressure are a tendency to decrease in the initial period (i.e up to the 1840s) and a tendency to increase in the last one (i.e. from the 1930s). As far as seasons are concerned (figures not shown), the spring series is very similar to the yearly one, even if it displays its absolute maximum in the late 1940s and displays a distinct minimum around 1890 that is not so evident in the yearly record. Summer has a long and very high maximum in the initial period, followed by rather stable values from around Autumn has a long minimum between 1835 and 1885, with oscillating values in the previous and following years. Finally, winter has a tendency to increasing values until about Afterwards it decreases until around 1970 and then increases sharply, reaching its absolute maximum around the year 1990, with values of about 3 hpa above the 20th century mean. Some interesting similarities are evident from a comparison between GAR temperature and pressure records, both in the decadal timescale (interesting examples concern the relative maxima in the 1820s, in the 1860s and in the late 1940s) and in the year-to-year variability (this being also highlighted by a positive correlation between pressure and temperature Figures 17 and 18). It is very interesting to notice some single seasonal events also; one example above all is summer 1816: it is the coldest summer, and also the summer with the lowest pressure anomaly ( 2.7 hpa). On the contrary, there are also examples highlighting important differences as the one concerning the 1990s, when temperature displays a strong increase, whereas pressure decreases. So, the results of the comparison between temperature and pressure variability seem to suggest that GAR temperature evolution may have been influenced, at least partially, by variations in atmospheric circulation (see also Auer et al., 2007) Precipitation Precipitation CRSMs (Figure 4) cover the period for all low-level subregions, whereas H CRSMs is not available. GAR precipitation evolution is extensively discussed by Brunetti et al. (2006b). They give evidence that it is not possible to describe GAR precipitation by means of an average over all CRSMs. In fact, besides some similar features, such as during the two wet decades at the beginning of the 19th century, the minimum in the 1860s, the maxima in the 1910s and 1930s and the minimum in the late 1940s, there are many differences in long-term behaviour as well as in year-to-year and decadal variability, the most relevant of them consisting of a marked split in about the past 140 years between a wetting trend in Northern regions (in particular, in NW) and a drying one in Southern regions (in particular, in SE). On a seasonal basis (figures not shown), some relevant features characterizing all subregions are the spring and summer minima in the 1860s and in the late 1940s, the winter minimum around 1990, the autumn rather high values in the initial three decades, followed by a long-term drying until the 1980s, and the evident autumn increase in the two recent decades. On the contrary, there are many important features concerning only parts of the GAR. So, even if the average over GAR is not representative of any of the subregions, there are some interesting all-gar features that may well reflect that precipitation displays a continental scale background signal also. The coherence of precipitation in the different GAR subregions and the wider European-North Atlantic precipitation patterns is discussed by Efthymiadis et al. (2007).

10 M. BRUNETTI ET AL. By considering the average over the whole GAR, the year with highest precipitation is 1910, with 126% of the 20th century mean precipitation, whereas the year with lowest precipitation is 1921 (67% of the 20th century average). The comparison between GAR temperature and precipitation records gives evidence of a tendency to display an opposite behaviour in summer: it is evident in both the decadal and year-to-year variability (also highlighted by a highly significant negative correlation Figures 17 and 18) but not in the long-term evolution. An interesting example of the opposite behaviour of temperature and precipitation concerns the very cold summers of the 1810s that display, especially in Southern regions, rather high precipitation. On the other hand, precipitation shows a much stronger link with pressure, highlighted both by highly significant negative correlations (see Figures 17 and 18) and by opposite behaviour of the RES3s in all seasons and all subregions. However, also in this case, the link does not concern long-term evolution. The pressure precipitation link is particularly evident where all- GAR precipitation features are present. A very significant example is the absolute minimum in spring precipitation in the late 1940s that corresponds to the absolute maximuminspringpressure Cloudiness Cloudiness CRSMs are shown in Figure 5. As discussed in section 2, cloud cover data are not distributed homogeneously throughout the GAR; more specifically, they are not available for France. The common period for all subregions is from 1878 to As for precipitation, it is not possible to describe GAR by means of the average over all CRSMs, because for cloudiness also, besides some common features, there are many differences in long-term behaviour as well as in year-to-year and decadal variability. The most relevant difference concerns long-term evolution of western and eastern subregions: the first ones present a tendency to increase over the whole covered period (the same behaviour is also observed for H CRSMs), whereas the latter ones show a tendency toward a decrease from about Among the all-gar features characterising yearly cloudiness CRSMs, we mention the initial minimum (at about 1870), the two maxima in the 1910s and 1930s, the minimum in the 1940s, the maximum at about 1980 and the minimum around At seasonal level (figures not shown) some interesting features are the winter absolute minimum around 1990, the spring and summer distinct minimum in the late 1940s and the strong increase in the past 20 years in autumn. As far as single relevant events are concerned, we observe that the years showing greatest cloudiness negative anomaly are 2003 and 1921, together with 1874, 1949 and It is worth noticing that all such low cloudiness events correspond to minima in precipitation amounts too. The good agreement between GAR cloudiness and precipitation does not concern only extreme years, but characterises the entire records. It concerns both year-to-year and decadal variability (correlation is high Figures 17 and 18 and the RES3s display a number of common features) and it is present in all seasons and subregions. However, if long-term evolution is considered, there are also important differences as the one of SW, where precipitation shows a tendency to decrease, whereas cloudiness shows an increase. As far as the comparison with temperature and pressure is concerned, the results are similar to the relations of precipitation with such variables, the link between cloudiness and temperature being slightly more evident and the one between cloudiness and pressure slightly less evident. However, also in this case, there is no evidence of any relation in long-term evolution Sunshine duration Sunshine duration CRSMs are shown in Figure 6. Similar to cloudiness data, sunshine duration data are also not available for the whole GAR, in this case being missing for Italy. The common period for all subregions is from 1886 to Also for sunshine duration, as for cloudiness, there are important differences in the longterm evolution of the different subregions, especially before the 1920s, where NE, SE and NW CRSMs show a tendency to decrease, whereas SW and H CRSMs do not show such a tendency. After the 1920s the longterm evolution is more consistent, even though H CRSMs increases in the past 20 to 25 years are larger than those of L CRSMs. The difference between long-term evolution of H and L sunshine duration CRSMs is highest in summer, lower in spring and autumn and not present in winter. Besides the differences in long-term evolution, there is good agreement concerning both year-to-year and decadal variability. Such an agreement is highlighted both by high significant correlation among sunshine duration CRSMs in all seasons and by the presence of several common features in the RES3s, as the maximum in the 1890s, the minima in the 1910s and in the 1930s, the maximum in the late 1940s, the minimum at about 1980 and the strong increase in the last 25 years, especially due to 1980s and the past 10 years. As far as single relevant events are concerned, we observe that the years with greatest sunshine duration anomalies are 1921 and So, we also find years that are extremes for cloudiness. As between cloudiness and precipitation, between sunshine duration and cloudiness also such a relation does not concern only extreme years but characterises the entire records. It concerns both year-to-year and decadal variability (correlation is high and the RES3s show a number of common cloudiness minima (maxima) and sunshine duration maxima (minima)) and is present in all seasons and subregions. However, also in this case, if long-term evolution is considered, there are significant

11 CLIMATE VARIABILITY IN THE GREATER ALPINE REGION disagreements. The most remarkable one concerns subregion H and, though less evident, subregion SW that shows a similar long-term increase in cloudiness and in sunshine duration Vapour pressure Vapour pressure data are available for the Northern and Eastern parts of the GAR and for the H subregion (Figure 7). The longest record is NE starting in 1837 and the shortest is H starting in The analysis of vapour pressure CRSMs gives evidence of a good agreement with temperature records, both for year-toyear and decadal variability and for long-term evolution. The agreement is excellent in winter (seasonal figures not shown) for all low-level CRSMs. Beside the good agreement with temperature, vapour pressure does not show any other significant link with other variables, not even in cases where such variables are significantly linked with temperature (e.g. cloudiness in summer). So, it is not surprising that the analysis of the absolute maxima and minima of vapour pressure CRSMs does not highlight most of the events shown by other variables. Nevertheless, the strong link with temperature is once more underlined by the fact that the absolute maximum in the seasonal vapour pressure CRSMs concerns summer 2003 (2.1 and 1.8 hpa above the 20th century average for N and H CRSMs respectively) Relative humidity Like vapour pressure data, relative humidity data are also available for the northern and eastern parts of the GAR (Figure 8). The longest records are NE and SE starting in 1860 and the shortest is H starting in Considering year-to-year and decadal variability, relative humidity turns out to be the variable with the lowest spatial coherence. Such behaviour is particularly evident in autumn and winter: it is not only a consequence of the well-known vertical decoupling between the boundary layer and the free atmosphere but is also due to the rather low spatial coherence of low-level records. In spring and summer the consistency among the different CRSMs (seasonal figures not shown) is higher, even though in these seasons there are only a few common features, the most interesting one being the summer decrease in the past 30 to 40 years. It is also worth noticing that the late 1940s (period in which the other variables exhibit the more consistent minima and maxima) show a minimum in low-level relative humidity as well. Such a feature is, however, not present for H subregion that, on the contrary, has its absolute relative humidity maximum in this period. So the late 1940s seem to be a dry period in the lowlands but one of typically increased convective moisture transport towards the mountain observatories at high elevation. Another period with a coherent relative humidity pattern at low level and a different behaviour at high level concerns the past 20 to 25 years. In this period, at low elevation the vapour pressure increased by approximately 0.8 hpa, but such an increase was not enough to balance the temperature increase of more than 1 K. Therefore relative humidity decreased considerably. On the contrary the smaller increase of about 0.4 hpa at high elevation was sufficient to keep relative humidity rather stable (or only slightly decreasing) in the mountains, due to the lower saturation vapour pressure in the colder air at high elevations. As far as long-term evolution is considered, NE and SE give evidence of a 20th century decreasing tendency (mainly due to spring and summer), whereas such a signal is not present for NW and H subregions, even though these subregions (especially the NW) also show a slight tendency to decrease in the past 30 years. The 20th century decrease may suggest that, at least in part of the GAR, there is a reduction in relative humidity linked to global warming. However, such a conclusion is not completely supported by the data, as relative humidity shows its strongest decrease around 1970, where temperature is still rather stationary. As far as the comparison with other variables is concerned, for low-level subregions there is some evidence of common features with precipitation and cloud cover and opposite ones with sunshine duration. However, such signals are not very clear and concern only some periods of the records. On the contrary, for the H subregion, the direct relation with cloud cover and the inverse relation with sunshine duration are much more evident. 4. Trend analysis The first step in trend analysis was the estimation of trends over the longest common period of data availability for all variables ( ). Trends were estimated by least square linear fitting, and their significances were evaluated with the Mann Kendall non-parametric test (Sneyers, 1990). The results are shown in Table II. The clearest signal is the positive and highly significant temperature trend, highlighted by slopes of the annual CRSMs ranging from 1.1 to 1.6 K per century (for SE and NW, respectively). The largest warming trends are found in winter and summer. As far as spatial distribution is concerned, the trend of the Western subregion is slightly larger than that of the Eastern subregion, and the trend of the Northern subregion is slightly larger than that of the Southern subregion. However, such differences are in many cases within one standard deviation. The only other variable that shows a positive slope in all subregions and in all seasons is vapour pressure. In this case, however, some subregions have no data available and not all seasonal trends are significant. It is worth noticing that for vapour pressure also all available yearly CRSMs show highly significant positive trends. Also, pressure and relative humidity present a rather coherent picture of seasonal and subregional trends, with positive slopes for pressure and negative slopes for relative humidity. For pressure the strongest increase concerns spring (all subregional trends are highly significant), whereas in the other seasons there are only a

12 M. BRUNETTI ET AL. Table II. Annual and seasonal trends for the five basic CRSMs and their linear combinations. All trends have been calculated by least square linear fitting considering the period. Each box shows the trends (together with its standard deviation), according to the scheme that is shown in the box at the lower right corner. Only values with significance level greater than 90% are indicated; for lower values of significance only the sign of the trend is indicated. Bold numbers indicate trends with significance level higher than 99%, numbers in italics indicate trends with significance level higher than 95%. Significance levels have been evaluated by means of the Mann Kendall non-parametric test.

13 CLIMATE VARIABILITY IN THE GREATER ALPINE REGION few highly significant trends. It is interesting to notice that there is a North South gradient in pressure trends, whereas Western and Eastern subregional trends are very similar. It is also worth noticing that the pressure trend of yearly H CRSMs is more than twice the trend of yearly L CRSMs. Such an effect is the consequence of the strong temperature increase in the considered period. Also for relative humidity the high-level trend (positive slope for yearly CRSMs) is different from the low-level one (negative and significant). Such a difference turns out to be maximum in summer, where L CRSMs have a highly significant negative trend and H CRSMs displays a 95% level significant increase. The difference in high-level and low-level relative humidity trends is discussedbyaueret al. (2007): they underline that such a behaviour has to be considered together with the fact that in the GAR there is nearly identical long-term vapour pressure increase for high and low elevations, in spite of the significantly lower water vapour content (in an absolute sense) of the (colder, less dense) high-elevation air masses. Thus, they conclude, the slope of approximately 0.5 hpa per century of both high- and low-level vapour pressure records tells us of a more effective moisture transport towards the Alpine peaks compared to the less intense one (in relative terms) into the valleys, basins and plains. In other words, according to Auer et al. (2007), at higher levels the moisture transport appears to be able to balance the drying potential of the 20th century warming, causing relative humidity to remain relatively stable. On the contrary, such an effect seems not to be present in the valleys, basins and plains and therefore a long-term relative humidity decrease has occurred in correspondence with the 20th century warming. As far as precipitation, cloudiness and sunshine duration are concerned, the spatial patterns of the trends are more complicated. There is evidence of a North South signal in precipitation, with a negative trend in Southern subregions and a positive trend in the Northern ones (see Brunetti et al. (2006b) for more details). Moreover, there is evidence of an East West signal in cloudiness, with a negative trend in Eastern subregions and a positive trend in the Western ones, mirrored by an analogous trend in precipitation, even though the significance of the signal is lower for this variable. The same pattern, with the same sign, is unexpectedly evident for sunshine duration also. In this case, however, the significance is low. As far as cloudiness and sunshine are concerned, the greatest inconsistency seems to be the corresponding high-level, highly significant positive trend for both variables Running trend The slopes, obtained by applying linear regression to the records, give information that is not sufficient to describe the complex time evolutions highlighted by Figures 2 to 8. So, we subjected the CRSMs also to running trend analysis. This technique was applied as in Brunetti et al. (2006b): for each subregion, both for annual and seasonal series, slopes were estimated within windows whose widths range from 20 years up to the entire series lengths, running from the beginning to the end of the series. The results are shown in Figures 9 to 16, where window widths and the central years of the windows that the trends refer to are represented on y and x axes, respectively. Slopes are represented by the colours of the corresponding pixels. Pixels are plotted only for trends significant above 90% level. Significances are evaluated by the Mann Kendall non-parametric test (Sneyers, 1990). This type of running trend analysis is an innovative method (first introduced by Brunetti et al. (2006b)) for an in-depth investigation over different subperiods, allowing to visualise climate trends on a wide range of timescales. So, Figures 9 to 16 capture the whole spectrum of significant trends present in the series thus providing a complete quantitative description of the peculiarities observed in Figures 2 to 8 and giving evidence of which features are most important in terms of trends at all timescales Temperature Figure 9 shows the results of running trend analysis applied to L yearly temperature CRSMs. The results are very similar for the high-level subregion and for the single low-level subregions also (figures not shown). Considering timescales longer than 50 years, the most relevant result is that the two-phase behaviour highlighted by Figure 2 turns out to be characterised by significant trends both as far as the decrease before around 1890 and the increase in the following period are concerned. The same results applies to seasonal records as well; in particular, spring and autumn present the same trend pattern of the year, whereas some differences are evident in summer and winter, with the former displaying negativetrend pixels that last for some more decades (up to about 1900), and the latter displaying positive-trend pixels starting some decades earlier (around 1860) and having no significant negative-trend pixels in the first part of the series. On shorter timescales, trends are strongly influenced by decadal variability, and periods of significant negative (positive) trends can also be present in correspondence with a clear long-term increase (decrease). Some interesting peculiarities are the negative-trend pixels around the 1800s, the 1830s, the 1870s and the 1950s, and the positive-trend pixels around the 1790s, the 1890s, the 1910s, the 1940s and the past 15 to 20 years. These features are also clearly evident in some seasons, in particular in spring, while other seasons present only some of them or present some other features that are not evident on a yearly basis. It is also interesting to notice that the strong positive trend characterising the last decades is higher in summer and spring and lowest in autumn Pressure As for temperature, for pressure also the clearest result of the running trend analysis applied to yearly L CRSMs

14 M. BRUNETTI ET AL. Figure 9. Running trend analysis for the L temperature annual and seasonal CRSMs. The y axis represents the window width, and the x axis represents the central year of the window over which the trend is calculated. Only trends having significance greater than 90% are plotted. Significance levels have been evaluated by means of the Mann Kendall non-parametric test. is that the two-phase behaviour highlighted by Figure 3 turns out to be characterised by significant trends as far as both the decrease in first period and the increase in the recent period are concerned. In this case, however, the trends are significant only if the considered windows include the first decades and the last decades of the series for the decreasing and the increasing tendencies respectively (Figure 10). The difference between the trends of the two parts of the record is particularly evident in spring and in summer (with the dominant influence of a final increasing phase in spring and the dominance of an initial decreasing phase in summer), whereas in autumn the initial decreasing phase is not so clear, trends being rarely significant. As far as winter is concerned,

15 CLIMATE VARIABILITY IN THE GREATER ALPINE REGION Figure 10. as in Figure 9, but for the L pressure annual and seasonal CRSMs. the situation is remarkably different, the trend being significantly positive both in the first and in the last parts of the series. Even though the long-term trends of temperature and pressure are, at timescales longer than 50 years, rather similar the variations of pressure seem to be too large (about 1 hpa) to be solely a temperature-induced bias, because of the fact that HISTALP records are not obtained from sea-level data. As far as shorter timescales are considered, the most interesting feature are the strong positive-trend pixels around Such a feature is mainly due to winter. If N and S yearly CRSMs are compared, the same twophase trend of L CRSMs is evident (see NW/NE and

16 M. BRUNETTI ET AL. Figure 11. as in Figure 9, but for the E cloudiness annual and seasonal CRSMs. SW/SE in Figure 3). However, both the initial decrease and the final increase are stronger in the Northern than in the Southern subregions. Such a behaviour causes the difference between N and S CRSMs (N S CRSMs) to exhibit negative-trend pixels before about 1860 and positive-trend pixels in the following years (figure not shown). Also in this case, the magnitude of the trends seems to be too large (the yearly N S CRSMs long-term increase in the past 150 years is about 0.6 hpa) to be only a temperature-induced bias due to the slightly higher level of the Northern stations. The two-phase behaviour in the N S record is present at seasonal level also, the increase of the N S record in the second part of the series being particularly evident in spring, summer and

17 CLIMATE VARIABILITY IN THE GREATER ALPINE REGION Figure 12. as in Figure 9, but for the W cloudiness annual and seasonal CRSMs. autumn and the initial decrease being particularly evident in winter. All the previous discussion concerns low-level data. Running trend analysis of H CRSMs is not discussed here because high-level pressure is strongly dependent on the temperature of the air masses beneath the highelevation sites: we prefer to focus our analysis on records that represent sea-level pressure as effectively as possible Precipitation GAR precipitation trends were widely studied by Brunetti et al. (2006b). The most relevant feature is the opposite

18 M. BRUNETTI ET AL. Figure 13. as in Figure 9, but for the W sunshine duration annual and seasonal CRSMs. long-term behaviour of Northern and Southern subregions, with a tendency toward wetter conditions in the North and lower precipitation in the South. This is particularly true after the 1850s, the long-term precipitation trend of Northern subregions before this date being negative too. We do not present figures here as full details on yearly and seasonal results are given for the different GAR subregions in the Brunetti et al. (2006b) paper.

19 CLIMATE VARIABILITY IN THE GREATER ALPINE REGION Figure 14. as in Figure 9, but for the E sunshine duration annual and seasonal CRSMs Cloudiness and sunshine duration The most interesting feature shown by the results of running trend analysis of yearly cloudiness CRSMs is that the different long-term evolution of the Eastern subregion and the Western and the high-level subregions highlighted by Figure 5, turn out to be characterised by significant trends. In fact the Eastern subregion (Figure 11) shows positive-trend pixels before about 1920, and negative-trend pixels in the following period, even though there are some exceptions, such as a few

20 M. BRUNETTI ET AL. Figure 15. as in Figure 2, but for the L relative humidity annual and seasonal CRSMs. positive-trend pixels in the 1960s; on the contrary, the Western (Figure 12) subregion has a dominance of positive trends, with the only exception of the most recent decades, where in general no significant trends are present and where the only pixels highlighting negative trend can be observed. The positive trends are even more evident for the high-level subregion (figure not shown), that has always positive trend for windows up to 40 years, with the only exception of those centered in the 1940s. Such different behaviour is mainly due to a clear cloudiness maximum in the Eastern subregion around 1920 that is neither evident in the Western subregion nor in the high-level one (Figure 5). It causes both the higher Eastern subregion to exhibit positive trends in

21 CLIMATE VARIABILITY IN THE GREATER ALPINE REGION Figure 16. as in Figure 2, but for the L vapour pressure annual and seasonal CRSMs. the first part of the record and the Eastern subregion to exhibit negative trends after about The Eastern subregion negative trends after 1920 are mainly due to winter and autumn, seasons in which, after around 1950, the Western subregion has some negative-trend pixels. It is, however, interesting to highlight that in the past 20 to 30 years autumn cloudiness CRSMs have positive-trend pixels both in the Western and in the Eastern parts of the GAR and in high-level subregion as well. Also, the results of running trend analysis of yearly sunshine duration CRSMs give interesting information, allowing to better describe the inconsistencies between cloudiness and sunshine duration long-term trends that have been highlighted by the comparison of Figure 5 and

22 M. BRUNETTI ET AL. Figure 6 and by Table II. In fact, they give evidence that, at longer timescales, in the Western subregion, there is positive trend for sunshine duration (Figure 13) as for cloudiness and that such inconsistent behaviour is even clearer for the high-level subregion (figure not shown). On the contrary, in the Eastern part of the GAR, sunshine duration (Figure 14) shows the same two-phase pattern of cloudiness, but with opposite signs. The results of running trend analysis of the seasonal CRSMs give evidence that the inconsistency between Western subregion cloudiness and sunshine duration trends is mainly due to winter, whereas in the other seasons there is a general opposite behaviour between cloudiness and sunshine duration. On the contrary, for the high-level subregion, such an inconsistency is mainly due to summer. An interesting peculiarity highlighted by running trend analysis on sunshine duration series is the decreasing trend between the 1950s and the 1980s, followed by a positive trend from the 1980s to today. This is also clear in the seasons, particularly in summer, and matches the global dimming and the recent brightening described for other regions of the globe in the international scientific literature (Stanhill and Cohen, 2001; Wild et al., 2005) Relative humidity and vapour pressure The most interesting feature highlighted by the results of running trend analysis of yearly relative humidity CRSMs is that the difference in the long-term trends of low and high-elevation subregions (see Figure 8) turns out to be characterised by significant trends, the former displaying almost only negative-trend pixels (Figure 15) and the latter generally having positive-trend pixels before around 1950 and negative-trend pixels in the following period (figure not shown). The same results are evident also for the high-level seasonal records, even though the time of the change from positive to negative trends is different for each season. It is interesting to notice that, on shorter timescales, the trend of yearly L relative humidity CRSMs is almost always significantly negative, with the exception of two periods, centred around 1920s and 1950s, where generally no significant trend pixels are evident and where there are also a few positive-trend pixels. As far as vapour pressure is concerned, high and low elevation stations are characterised by the same trend for all timescales, trends being almost always positive at all timescales, with the exception of two periods, centred around 1920s and 1950s, where negative-trend pixels and no significant trends, respectively, (Figure 16) are found. These are the same periods in which relative humidity has no negative-trend pixels. It is very interesting to observe that the periods with non-decreasing relative humidity and with decreasing vapour pressure are the only periods of the 20th century that do not show increasing temperatures. 5. Correlation among variables The last step of our research was the study of the correlation among HISTALP variables, with the aim of highlighting the agreement in their high-frequency (i.e. year-to-year) variability. It was performed considering both seasonal and yearly records. The correlation among the different variables was analysed both considering the original CRSMs (ORs) and high-pass-filtered series (HPFs) (i.e. residuals from the GLPF15s), as using such residuals from the long-term behaviour avoids getting too high (low) estimations of the agreement in high-frequency variability in presence of coherent (non-coherent) long-term trends. Moreover, first-difference series (FDs) were also used, in order to further check whether the correlations are affected by potential non-adjusted inhomogeneities. In fact, let us suppose that a non-adjusted inhomogeneity (consisting in a step function that adds a constant error to the data from a starting time t 1 to an ending time t 2 ) still affects a series: the GLPF15 (and, then, also the residuals from it) will be affected by this error in the neighbourhood of t 1 and t 2. On the contrary, the first differences are affected by the inhomogeneity only as far as the data corresponding to t 1 and t 2 are concerned. Another situation in which the FDs may be better than the HPFs in assessing the high-frequency common variability among variables is when there are strong variations in the slopes of the trends. Such a situation is, for example, present for some variables at the beginning of the 1980s. Correlations analysis was performed over the period , for which all variables are available for all the subregions. Moreover, the temporal stability of the results was investigated by calculating the correlations (in a running way approach) within 30-year windows, moving through the period of common data availability of all possible pairs of variables. This information permits to better characterise the correlation among the different variables and it can also be used as an indirect tool to check the quality of the data of the most problematic periods, normally corresponding to the most ancient ones. The results of correlation analysis of the HPFs are presented in Figures 17 and 18, the first concerning the period of data availability for all variables, the latter referring to the moving windows. The most evident correlation is indeed the one between cloudiness and sunshine duration. Such correlation is present in all subregions (yearly values range from 0.74 (SW) to 0.87 (H)) and in all seasons (seasonal values range from 0.82 (SW-summer) to 0.95 (H-winter)). Moreover, it is persistent through the whole common period spanned by the two variables ( ), the seasonal correlation coefficients resulting within the following ranges for all 30-year windows: 0.78 to 0.98 for NW; 0.71 to 0.98 for NE; 0.60 to 0.95 for SW; 0.55 to 0.96 for SE; 0.86 to 0.98 for H. It is, however, worth noticing that the most recent data show a better correlation than the older ones, the correlation being lower than 0.80 for all windows

23 CLIMATE VARIABILITY IN THE GREATER ALPINE REGION Figure 17. Correlation coefficients among the different variables (both on yearly and the seasonal bases) for the five basic CRSMs. They have been calculated considering the HPFs and considering the period. T01, temperature; P99, pressure; R01, precipitation; N11, cloudiness; SU1, sunshine duration; H01, relative humidity; H11, vapour pressure. centred after about 1960, whereas before 1960 there are windows with correlation around 0.6. The less correlated windows concern the southern subregions: they are probably an effect of rather noisy CRSMs caused by very poor station density, especially for sunshine duration. The cloudiness sunshine duration correlation is not so evident if the ORs are considered instead of the HPFs (figure not shown). Such an effect is particularly clear for the yearly H CRSMs that displays a correlation coefficient of 0.52 for the ORs, whereas the correlation coefficient of the HPFs is 0.87 (at seasonal level

24 M. BRUNETTI ET AL. Figure 18. Running correlation coefficients among the different variables for (a) the year, (b) winter and (c) summer. They have been calculated considering the HPFs and considering 31-year running windows. The pixels correspond to the central points of the 31-year windows. Only correlation coefficients having significance greater than 90% are plotted. T01, temperature; P99, pressure; R01, precipitation; N11, cloudiness; SU1, sunshine duration; H01, relative humidity; H11, vapour pressure.