Correspondence to: Barbara Chimani, Central Institute for Meteorology and Geodynamics, Hohe Warte 38, A-1190 Vienna, Austria. E-mail: firstname.lastname@example.org
This study presents a new gridded dataset providing absolute monthly mean temperatures across the Greater Alpine Region (GAR) of Europe at a spatial resolution of 5 arcmin (6 × 9 km in the region) from 1780 to 2008. The starting point was a set of long-term homogenized station time series. To assure the quality of the analyses back in time, when the station density decreases, missing measurements were reconstructed by an Empirical Orthogonal Function analysis that can deal with gappy data. It is shown that the reconstructed values comprise similar statistical features to the observations and that the method produces no breaks between the reconstructions and the observations. The compound anomaly dataset was then interpolated separately for two different altitude ranges to preserve anomaly gradients between high and low elevations. This allowed for the derivation of individual anomalies at each grid point in GAR. Finally, these smooth anomalies were blended with the highly resolved monthly mean absolute temperature fields, provided by a project of the European Climate Support Network. The added value of this new high resolution and long-term temperature dataset is shown and discussed using the examples of the height of the 0 °C altitude and vertical lapse rates. For the first time, these and other features are now available for more than two centuries in a topographically complex region like GAR.
Changes of air temperature through the past, the present and the future receive increasing attention as climate change and its consequences are expected to severely affect the ecosystems of the earth and human civilization. High-quality measured climate data from regular meteorological networks are the best source for studying ongoing trends and variations. General Circulation Models and Regional Circulation Models along with statistical downscaling techniques constitute the main instruments to investigate and assess possible future climate changes on the global, continental or the regional scale. An assessment of the past climate reaching far back in time often relies on proxy data. Both the assessment of the past and the estimation of the future climate need instrumental data for calibration and/or verification. To overcome the problems related to the statistical significance of climate trends and variations in respect to the given strong high-frequent climate variability, a sufficient length of the observations is a prerequisite. As such, a long-term, high-quality dataset is of major importance for any further analysis.
For the majority of scientific studies and direct practical applications, gridded data at an adequate resolution in space and time are the optimal basis. Such gridded datasets are based on observations and hence there are limitations that are mainly determined by the time-dependent station network density. Single station series usually lack the necessary representation due to the inhomogeneity of the spatial distribution typical for climate parameters. Also station networks are usually not evenly distributed in space and thus may lead to biased results.
The gridding activity which is the topic of this study aims at providing a comprehensive dataset for the temperature evolution in the ‘Greater Alpine Region’ (henceforth GAR) of Central Europe (4° to 19°E, 43° to 49°N, 0 to 4800 m asl.) at an adequate spatial resolution in respect to the given network density and to the complicated topography. The starting point is the temperature subset of the station series of HISTALP (Historical Instrumental Surface Time series for the Alps). It stands out through: (1) a dense network of 140 station time series, (2) covering periods of 100 to 250 years, (3) having been subject to intensive quality control including homogeneity testing and adjusting and (4) having an outstanding ‘early instrumental potential’ of 40 stations in 1850 and still 12 in 1780.
The temperature subset of HISTALP was described and analysed for the first time by Böhm et al. (2001). The basic paper describing HISTALP in general is by Auer et al. (2007). Recently, it was shown that in the early part of the temperature readings (around the 1860s and earlier), a systematic ‘early instrumental bias’ occurred. This was due to insufficient shading of the thermometers (Büntgen et al., 2006; Frank et al., 2007; Hiebl, 2006). This bias was successfully reduced by Böhm et al. (2010).
So far HISTALP contains monthly series, mainly because a considerable part of the early data exists in printed yearbooks or other anthologies at a monthly resolution only. The restriction to monthly series allowed for a sufficient spatial density (regarding the early instrumental period), which is necessary to successfully apply homogenizing procedures. For the scarcer long-term daily series, the necessary spatial network density is not sufficient compared with the spatial decorrelation length to allow for relative homogeneity testing—neither in the target region of our study nor elsewhere (cf., e.g. Scheifinger et al., 2003; Moberg and Jones, 2005). A careful homogenization is an indispensable pre-condition for climate time series analysis (cf., e.g. Peterson et al., 1998; Aguilar et al., 2003). The still existing lack of station density for long daily series excludes them from analyses of long-term trends and variability. For monthly series however, Auer et al. (2007) showed that the desirable correlation of at least 0.7 with a sufficient number of reference series for homogeneity testing is available back to the late 18th century in GAR.
The new dataset, which is introduced here, is a further enhancement of the HISTALP temperature dataset. We intended to create an absolute temperature grid at a spatial resolution, adequate to represent also smaller-scale features within the complicated terrain of GAR. The resolution is 5 arcmin latitude–longitude in space (approximately 9 × 6 km in GAR) and 1 month in time.
Section '2. Data' describes and discusses the two datasets that are used and combined: (1) a high resolution gridded short-term (1961–1990) dataset (containing information of 1700 single stations, see Hiebl et al., 2009) used to display the typical monthly patterns of high resolution absolute temperature fields in the region and (2) the long-term HISTALP dataset (currently 140 stations with a continuous thinning back in time).
Section '3. Reconstruction of temperature data in station mode' outlines the method used here to overcome the problem of gappy data (von Storch and Zwiers, 1999, section 13.2.8). Efthymiadis et al. (2006) successfully applied a related method to produce a gridded precipitation dataset for GAR. Here we focus on temperature, which is much less variable in space and time. The output of this section are 140 reconstructed long-term monthly temperature series each going back to 1780.
Section '4. Interpolation' deals with the interpolation of the site-based dataset to the equidistant grid of 5 arcmin resolution in latitude and longitude. A Gaussian inverse distance weighting is used to interpolate for each month anomaly fields separately for two altitude layers. The appendant vertical anomaly gradients for each grid point were then added to the stable high resolution patterns of the monthly 1961–1990 climatologies.
Finally, Section '5. First analyses of the dataset' shows the potential such a dataset may have using examples of long-term regional/local trends, analysing vertical structures of climate variability, etc.
2.1. Low resolution long-term database
The HISTALP database (Auer et al., 2007, http://www.zamg.ac.at/histalp) contains data in an area in the centre of Europe, extending from 4° to 19°E and from 43° to 49°N. It includes the complex topography of the European Alps and their wider surroundings in Austria, Switzerland, France, Germany, Italy with the Apennines and the Adriatic coast, Slovenia, Croatia, Bosnia-Herzegovina, parts of Hungary, Slovakia and the Czech Republic. The time series of the stations have been outlier-corrected and homogenized (Auer et al., 2007; Böhm et al., 2010). This dataset can be regarded as the state of the art. It is exploiting the given network potential in terms of length and density. For the present investigation, a number of 140 homogenized temperature series (contained in HISTALP) have been used (Figure 1). Three stations (Basel, Genève and Torino) provide reliable measurements back to 1760, and 35 station series start to report in the first part of the 19th century. The increase in station density from the 1850s onwards was caused by the founding of regular meteorological services.
2.2. High resolution short-term database
The temperature GAR-HRT (Greater Alpine Region—High Resolution Temperature) grids produced by a project of the European Climate Support Network (ECSN, http://www.eumetnet.eu/) provides the absolute mean temperature patterns for the climate period 1961–1990 across GAR for each month of the year. Temperature measurements of roughly 1700 stations were interpolated to a 1 km grid by a multilinear regression technique, taking into account longitude, latitude, height and the distance to the sea. Adjustments have been made to account for cold air pools, coastal and lakeshore effects as well as slopes and urban areas (Hiebl et al., 2009). Here, a version of this dataset is used that disregards the effect of urban areas, as no continuous historical information on the city sizes is available back to the late 18th century.
3. Reconstruction of temperature data in station mode
As elaborated above and displayed in Figure 2, temperature measurements are not available at all GAR sites throughout the past 250 years and there is a thinning of the network back in time. Hence, a simple interpolation strategy, based on the available temperature readings, would lead to a severe loss of spatial variance of the temperature fields over time simply because of the number of available stations. As a consequence, temporal variance would change at single grid points within the fields.
One possible approach that reduces the effects of a sparse station network is the use of an Empirical Orthogonal Function (EOF) analysis (see, e.g. von Storch and Zwiers, 1999, section 13). The standard EOF analysis describes the observations at a particular time by the sum over the products of the Empirical Orthogonal Function-patterns (henceforth EOFs) with the appendant time-dependent EOF coefficients (a.k.a. PCs) at that particular time. The main point is that measurements at many sites can often be approximated by just a few EOFs (and their coefficients). The quality of the approximation may be described by the correlation coefficient, the root-mean-square error (RMSE), etc. However, the standard approach applies to datasets without missing measurements.
Climate datasets are often gappy as observations are not always available at all sites. This applies even more to datasets that reach far back in time. However, the EOF concept can still be used. The procedure is little different from that of the standard approach. The elements of the covariance matrix are calculated in almost the standard way. However, whenever there is a missing observation in the covariance, this addend is disregarded. Such elements should not be fractioned by the total number of time steps the observations show, but by the total number minus the number of disregarded addends. The EOFs can be calculated. The coefficients cannot be found by a simple inner product of the EOFs with the measurements (anomalies), but by applying the least square method. This approach can be found in detail in von Storch and Zwiers (1999, section 13.2.8). The time coefficients of the EOFs calculated from a gappy dataset are not uncorrelated in general. This is because they are derived via the method of least squares and not by the projection of the EOFs onto the anomalies (von Storch and Zwiers, 1999, section 13.2.8). If there are no missing values, this approach leads to the projection of the EOFs onto the anomalies and the correlation disappears.
Efthymiadis et al. (2006) reconstructed station-based precipitation data for the GAR by using EOFs. However, unlike described above, they used a fixed recent time period (about the past 70 years) when all stations continuously provided measurements. Temperature anomalies are distributed much smoother in space than precipitation within GAR (Matulla et al., 2005; Auer et al., 2007) and hence the reconstruction of temperature measurements is less demanding than the reconstruction of precipitation. Absolute temperature measurements within GAR contain a lot of variability due to the complex topography, and therefore the anomalies to the climate mean (common reference period 1961–1990) were used here. The coverage of GAR by the stations provided by Böhm et al. (2010) and Auer et al. (2007) can be regarded as sufficient for temperature, and monthly temperature anomalies are regionally smooth on a monthly base (the EOFs, which were used here, simulate 84.5% of the anomaly-field variability). The EOF-based reconstructions are carried out for each month in the seasonal cycle separately, because the EOFs are changing during the course of the year. This setup allows deriving the spatial–temporal variability during the full period.
Our dataset contains 140 stations of which 12 have measurements reaching back in time to 1780 (Figure 2). Figure 3 shows the resulting time series at the example of a section of absolute monthly temperature series for Bologna.
The approach of reconstructing station data prior to interpolating the temperature anomalies to a regular grid instead of calculating time series for each grid point using EOFs of gridded analyses of the period with a dense station network was chosen for different reasons. First, the approach of reconstructing station time series before the spatial interpolation is widely used (e.g. Efthymiadis et al., 2006; using multiple regression: Xia et al., 1999; Eischeid et al., 2000; using linear regression: Widmann and Bretherton, 2000; Gyalistras, 2003). This is done to reduce the smoothing effect of the spatial interpolation. A second reason is the already existing precipitation dataset compiled by Efthymiadis et al. (2006) for the same area. Using similar methods to derive datasets of temperature and precipitation should feature a higher degree of uniformity between them. Another point is that the atmosphere in a mountainous region cannot be approximated by a simple linear function in the vertical (e.g. Hiebl et al., 2009). As different height levels would have to be treated separately, the EOF analysis would give reason to a more complicated approach. A thorough comparison of different mathematical approaches is beyond the scope of this study.
To validate the site-based reconstructions, a cross-validation was used. Therefore, the recent last 60 years were kept at all stations and the observations at the station under consideration prior to that period are disregarded. The information given at all the other stations is used to reconstruct the time series at the considered station. Then a comparison of the reconstruction to the disregarded observations is carried out. This is done by the station.
Figure 4 shows the correlation between the measured and the reconstructed anomaly time series for each station. The correlations are generally high, with a mean of 0.9. Slightly higher correlations are achieved north of the Alpine chain than in the South (not shown). The RMSE (reconstruction vs observations anomalies) for all sites is on average about 1.2 °C (Figure 5). This is tolerably small compared with the range (which is about 7 °C) and a total anomaly range of about 14 °C. Regarding the station-based RMSE, no difference can be found between stations in the north and south of the Alpine chain.
In order to detect estimated potential breakpoints when time series have been extended backwards by reconstruction, 30 year averages (1780–1809 and 1930–1959) have been calculated for each low-level station. A comparison between a group of stations having measurements in 1780–1809 and a second group, including those stations that have been reconstructed, was done. The difference between the means of those two groups is 0.04 °C in the early period and 0.5 °C in the later period with measurements for each station of both groups. As 0.5 °C is about 30% of the standard deviation averaged over the low-level stations for 1930–1959, this result is satisfying. The differences between the two groups of stations in both periods are smaller in summer than in winter.
These tests show that combining directly measured and reconstructed station series is suitable for our application. They provide a basis for the following gridding that can be regarded homogeneous in time because it contains sufficient spatial information in periods when data were scarce, like at the beginning of the instrumental period for instance.
As GAR is an area of highly complex and structured topography exhibiting considerable height differences, it is necessary to consider the elevation when interpolating temperature time series in space. Compared with the given station density of the 140 long-term series, the topography induces subgrid effects in absolute temperature that cannot be neglected. In order to account for these, the temperature field was split up into a spatially smooth but temporally varying anomaly field and a temporally stable but spatially varying mean field. Therefore, these spatial patterns were taken from the existing monthly 30 year mean fields of Hiebl et al. (2009), which are based on a set of 1700 sites for 1961–1990.
Based on existing studies about a likely decoupling of the thermal structure (e.g. Whiteman, 2000) of valley and high elevated air temperature in the mountains, the anomaly fields were calculated separately for a high- and a low-elevation subset. Instead of using a specific elevation threshold, the distinction between the lower and the upper levels is due to the subregions found by Böhm et al., 2001. The interpolation of the station anomalies was done with an inverse distance weighting scheme that uses a Gauss curve as the weighting function (w) (Schöner, 2004, Auer and Böhm, 1994) by using
with d the distance between the grid point and the station and a an additional influence factor that was set to 3 × 10−5. This resulted in a weighting of 0.3 at a distance of 200 km—a reasonable distance that is based on existing studies on the spatial decorrelation of temperature measurements in the region (Scheifinger et al., 2003; Auer et al., 2007). The weights were normalized by the interpolation method. Because of the considerable high station density in lower levels and the high correlation between temperatures series of mountain sites (Böhm et al., 2001), it was possible to determine an anomaly value at each grid point within both elevation ranges at a spatial resolution of 5 arcmin.
Figure 6(a) and (b) show an example of the low-elevation interpolation for January 1930 for a low (regarding those stations that were available in 1830) and the complete station density (127 stations, 3 reconstructed values). The importance of the higher network density is obvious, as the resulting gridded anomaly fields are quite different. Figure 6(c) shows the analyses using measurements for all the stations available in 1830 (27 stations) plus the respective reconstructions at the other sites. The added value of using the reconstructed series is to be seen using the example of the RMSE. The RMSE (comparing the analyses using observations of the complete station network with those using observations of the sparse station network of 1930 alone or together with reconstructed data) reduces from 0.57 to 0.36 °C if the reconstructed values (Figure 6(c)) are used additionally to the observations (Figure 6(a)).
The necessary vertical interpolation to gain the anomalies for each grid point was performed between two interpolated altitudes assigned to the respective high- and low-elevation anomaly fields. The interpolated altitudes were calculated as weighted means of the station altitudes, using the weights already derived for the inverse distance interpolation of the anomaly fields. The height of the nearest ECSN grid point was used as the height of the absolute temperature grid points. The absolute mean grid point temperatures were calculated using
with Tijm being the absolute temperature for each grid point ij and month m, T̄ijm the mean temperature of month m at the grid point ij, T′ the temperature anomalies for the upper and lower levels. hij is the height of the grid point in the temperature analyses while hij_lower is the described interpolated height of the grid point at the lower level.
The described procedure helped to preserve the individual temperature anomalies at different elevations, generally for mountainous areas, in particular during late autumn and winter when inversions typically produce cold anomalies at low altitudes in contrast to warm anomalies at high elevations.
The resulting dataset offers fields of absolute monthly mean temperatures (henceforth ‘grid mode’) for GAR along with time series for each grid point (henceforth ‘time-series mode’). Both modes were used for verification.
The time series of stations and neighbouring grid points are similar. The absolute mean difference between grid point and station values for all months from 1780 to 2008 ranges from 0.5 to 1 °C (e.g. station Vienna Hohe Warte: 0.5 °C, located in relatively flat lowlands—or Bregenz: 1.0 °C, located in a valley). The bigger part of the mean differences can be accounted for by the height difference between the grid point and the station (∼0.3 °C in the case of Vienna and ∼0.6 °C for Innsbruck). Largest differences between grid point and station data for these two stations range from − 2 to + 3 °C. The differences are due to unresolved subscale topography processes in the mountainous parts of GAR.
Our grid-mode dataset was compared with the highest resolved long-term temperature grid (a global dataset reaching back to 1901 with a spatial resolution of 0.5°; http://www.cru.uea.ac.uk/cru/data/hrg) provided by the Climatic Research Unit of the University of East Anglia (Mitchell and Jones, 2005). The comparison was done by interpolating our grid-mode values with the coarser CRU-grid. Although the average patterns for each decade between 1901 and 2000 are similar in winter and summer, the differences between the two datasets are smaller in winter than in summer (Figure 7). North of the Alps, the CRU dataset is warmer than the one presented here, whereas the southern parts of the alpine ranges and the mountain chains in Italy and the Adriatic Coast are colder in the CRU dataset. The differences remain almost the same over all decades. One reason for this behaviour might be the different topographies used to establish the datasets. Perhaps more dominant is the denser regional long-term HISTALP network along with the high resolution mean fields. This combination is apparently better suited to the necessities of a mountain region than the CRU setup, which is focused on global coverage.
5. First analyses of the dataset
5.1. Case studies
The dataset permits the identification of single year/ seasonal/month-based characteristics like the summer of 1816, which is known to exhibit an extreme cold anomaly, caused by the volcanic eruption of Tambora or the very warm summer of 2003 that caused a heat wave across large parts of Europe. Figure 8 compares the month July of the mentioned summers. Compared with the mean values of 1961–1990 (GAR-HRT dataset) for June to August for different locations in the area (in the Po-Valley, near Vienna, close the border between Tirol and Italy, in the east of Croatia and in the Rhone Valley), the following differences in the temperature evolution in 2003 are present: Although July is usually the warmest summer month in the greatest part of GAR, it was the least warmest in 2003, with a positive deviation from the mean of 2–3 °C. June and August were about 2 °C warmer than July. In the area of Vienna, the summer temperatures were as high (about 22 °C) as average temperatures in the Po-Valley, where they ranged from 26 to 28 °C in 2003. The temperature in the region of Lyon, in which average summer temperatures are usually the same as in the surrounding of Vienna, in June 2003 was considerably warmer. In contrast to these positive anomalies that occurred throughout GAR but the alpine range, in July 2003 the cold anomalies in July 1816 have been present in the northern part of the region mainly, sparing, e.g. the Italian lowlands and the north eastern part of GAR. July 1816 anomalies in the east of Croatia were just 0.5 °C, which is quite a moderate value for such an extraordinary summer. Anomalies have been much more dramatic in France for example, where the temperatures in the Rhone Valley ranged from 15 °C to slightly above 16 °C during that summer. On average, such temperatures are experienced there in September (15.7 °C). A comparison of July 1816 with July 2003 shows a mean difference in GAR of 4.8 °C.
The extremely cold February 1929 is shown in (Figure 9) as an example for the other tail of the distribution. In this winter, continental cold air masses reached GAR leading to exceptionally cold temperatures in the eastern parts of the region. In the lowlands of Austria, the mean monthly (February) temperature did not exceed − 9 °C. Negative monthly mean temperatures were present also in warmer regions such as Italy, the Adriatic Sea and the Coast of Croatia. The Tyrrhenian coasts of Italy, the Cote d'Azur and parts of the Rhone valley observed colder than normal—but still positive—monthly mean temperatures. Temperatures at high elevations remained relatively mild with mean anomalies of − 4.2 °C only from the 1961–1990 mean, whereas the coarse resolution subregion NE of GAR had a mean anomaly of − 10.0 °C. This anomaly ranks number 1 in the entire subregional mean time series of more than two centuries. Figure 9 underlines the described SE–NW anomaly gradient present in February 1929s low elevation temperature distribution.
5.2. Mean temperatures of altitude bands
The major advantage of the new dataset is its potential for analysing absolute temperatures. As a first example, we examine mean time series across altitudinal levels. The average temperatures for different height levels were calculated separately for the northern and the southern parts of GAR. These regions were defined according to the principal climatological subregions defined by Auer et al. (2007). The height levels are referred to by their central altitude. Here, we show three examples: 400 m (300–500 m), 1500 m (1400–1600 m) and 3000 m (2900–3100 m). The winter series for the two subregions, north and south of the Alps (Figure 10), show that cold winter anomalies, like those in 1929/1930 and 1962/1963 or during the Second World War, were most pronounced in the lower altitude bands whereas the extraordinary winter 1829/1830 was more extreme at high elevations.
Another extremely cold winter 100 years later (1928/29) showed the opposite vertical behaviour. As discussed in Section 5.1, the cold anomalies were stronger at lower elevations than in the mountains. A further notable fact is that the strong recent winter warming since the mid-1970s is more pronounced at medium and high elevations (0.5 to 0.7 °C/10a) and weakest at low elevations in the South (0.45 °C/10a). The differences between the northern and the southern parts of the area diminish with increasing height.
5.3. Vertical lapse rates
Another advantage of this dataset is its potential to analyse the temporal variability and evolution of vertical lapse rates. In Figure 11, the 11 year running means for the mean vertical lapse rates of north and south of the Alps between 3000 and 400 m are displayed for winter (DJF) and summer (JJA). It can be seen that the atmosphere over the southern Mediterranean region is, in general, less stable than over the northern parts and that the lower troposphere is less stable in summer than in winter. The high frequent variability of the lapse rate is stronger in winter than during summer and stronger in the northern parts of the region than in the south. Decadal scale anomalies are visible with weak lapse rates in the cold northern winters around 1890. Interestingly, the same can be found close to 1990 when southern and northern winters were accompanied by weak vertical lapse rates again, but now went along with very mild winters. The only significant decadal summer anomaly occurred in the cold 1810s, when vertical summer lapse rates were strongest in the entire region. This anomaly towards unstable thermodynamic conditions went along with considerably more precipitation in the Alps (Auer et al., 2005).
5.4. 0 °C altitude
Using the mean temperature gradients of the GAR-HRT analysis (Hiebl et al., 2009) and combining them with the monthly temperature anomaly gradients (calculated for each month as a necessary step for calculating the monthly temperatures) allowed the determination of the spatial distributions of 0 °C altitude for each month between 1780 and 2008.
In Figure 12, examples of time series of the 11 year running means of the 0 °C altitude can be seen for winter at six different grid points in GAR. 0 °C altitudes below sea level are of no further interest in the real atmosphere and have been truncated therefore.
The grid point located in the Po Valley (black solid) shows comparable heights as the point located east of the Swiss Jura (broken black). The continental influence causes rather low 0 °C heights close to Vienna (light-grey broken) and Budapest (light-grey solid). The distinguished zonal differences as a consequence of stronger continental type climate is well documented (cf. Auer et al., 2001). At all the grid points, the cold winters of the 1810s and around 1890 stand out due to very low 0° altitudes. Another common feature is the recent increase of the 0° altitude by approximately 600 m since the cold early 1940s. This 70 year trend is overlain by three warm and two cold decadal anomalies of 300–400 m each. Presently, the decadal maxima at a mean 0° altitude are 400 m higher than they used to be during the last decadal minimum in the early 1980s.
Figure 13 depicts an example of the grid mode showing the analysis of 0° altitude for December 1813. Just as indicated by the time series, the agreement between the Po Valley and parts of France can be clearly seen. Particularly, low 0° altitudes can be found alongside the Alpine range and over the south-east of Austria. A north–south gradient is also apparent. The most distinct spatial gradients of the 0 °C altitude appear across the Dinaric Mountains, which separate the elevated 0 °C heights near the coast (all the way up to about 4000 m asl.) from the significantly lower ones within the mountains and across the east down to < 1000 m in the Croatian–Bosnian interior. A similar pattern—but less pronounced—can be found in this month over the Italian peninsula, where the Apennine chain divides the country into the western part with higher and the eastern part with slightly lower 0 °C altitudes.
An EOF-based method that allows processing gappy data has been applied to calculate missing monthly mean temperature anomalies at stations. Thereby, a complete dataset containing monthly time series at 140 stations since 1780 for the GAR could be derived. The respective smooth monthly anomaly fields at two height regions were then blended with already existing high resolution, monthly mean fields based on about 1700 single stations 1961–1990 across the region. The resulting dataset of high resolution absolute single-month temperature fields covers the period from 1780 to 2008. The spatial resolution of 5 arcmin (5 × 9 km within GAR) is comparable to the resolution of other meteorological parameters, like precipitation for instance. As homogeneous data were used and the number of stations series (measured and reconstructed) did not thin back in time, the quality of the reconstructions was improved in comparison with the analysis using measurements alone.
First analysis of spatial and temporal temperature variability, lapse rates and the elevation of the 0 °C altitudes should serve to emphasize the scientific potential of this new dataset. It is expected to be useful in climate research as it contains lots of information across a long period of time, at a resolution sufficiently high to cover developments throughout the complex topography across the GAR.
The dataset was developed in the context of the internal project GAR-SCHNEE-200 of the Central Institute for Meteorology and Geodynamics, Vienna, funded by the Austrian Ministry for Science and Research. For the database of the station mode temperature series, we acknowledge the well-established informal cooperation of all data providers in the region. We are thankful to Dr Hermann Kuhn whose help allowed for a smooth processing of our calculations at the Helmholtz Zentrum Geesthacht (HZG) and to two anonymous reviewers whose comments helped to strengthen the study.