Regional centennial precipitation variability over Germany from extended observation records


Correspondence to: S. Brienen, Deutscher Wetterdienst, Frankfurter Str. 135, 63067 Offenbach am Main, Germany. E-mail:


The data base of daily precipitation over Germany has been recently extended by digitizing additional historical hand-written observations. The extension from 65 to 118 stations has increased the density of the available station network to a degree which allows both meaningful regional analyses and less error-prone trend and return level estimates. In this article first results of the examination of the precipitation behaviour in the winter and summer season throughout the entire 20th century are presented. To assess the spatial scale of similarities or spatial coherence of several precipitation indices from the 118 stations, Principal Component Analysis is used. The extracted leading patterns (six in winter, nine in summer) resemble nicely regions of different geographical characteristics and prevailing wind directions which puts some credibility on the quality of the newly digitized observations. The long-term linear trends of the regionally averaged time series differ substantially between precipitation indices, regions, seasons, and sub-periods. For the whole century, significant increases are found for most of the intensity-related indices in the Southern part of the country in winter and in some of those indices in the Rhineland/Sauerland and Alpine regions in summer. The two halves of the 20th century, are, however, characterized by partly opposite trends in the precipitation indices, and different regions are affected by these changes. An analysis of trend robustness by means of 30-year moving trends indicates a low stability of the trends during the 20th century, which is partly caused by interdecadal variability of the precipitation characteristics. For 100-year return levels, changes between the estimates obtained from the entire century and from the first and second 50 years differ in particular in the extent of their confidence intervals. In consequence, the availability of long precipitation records is very important for practical applications, and further extension of data records is recommended.

1. Introduction

A series of disastrous floods and droughts in different parts of the world at the transition from the 20th to the 21th century raised growing public and scientific interest in the quantitative assessments of possible changes in precipitation characteristics. Numerous studies related to precipitation trends were published during the last two decades and most of their findings are summarized in the IPCC-AR4 (2007) report.

Many regional studies from the European continent suggest that in central and northern Europe precipitation has significantly increased in winter (DJF), while in summer (JJA) no significant changes were detected up to the end of the 20th century (Schönwiese and Rapp, 1997; Osborn et al., 2000; Haylock and Goodess, 2004; Schmidli and Frei, 2005; Zolina et al., 2008). These changes are accompanied by an increasing number of days with moderate and heavy precipitation (above the 75th and 95th percentiles, respectively) as well as an increasing contribution of heavy precipitation events to the total precipitation in the second half of the 20th century with the strongest increase in the last three decades (e.g. Klein Tank and Können, 2003; Haylock and Goodess, 2004; Zolina et al. 2004, 2005, 2009, 2010). The changes in precipitation extremes are, however, much less coherent than for temperature (e.g. Frich et al., 2002; Klein Tank and Können, 2003; Groisman et al., 2005; Alexander et al., 2006). For a reliable assessment of the changes in extreme precipitation events (e. g. with a return period of 10 years and beyond) the existing data base is, however, still insufficient as stated in the IPCC report (2007, e.g. Chapter 3.8).

Up to now, long-term analyses of the behaviour of precipitation indices derived from daily precipitation records are rare. The variability of heavy and extreme precipitation in centennial European records were studied by Zolina et al. (2005, 2009) and Moberg and Jones (2005) using the European Climate Assessment (ECA) data collection (Klein Tank et al., 2002). In their collection, long-term gap free time series are only available for a few stations. Based on 22 locations (17 in Germany) for the period 1900–2002, Zolina et al. (2005) showed positive trends in Central and Eastern Europe in winter with spatially incoherent signs and significances in the 95th percentile values and in the contribution of very wet days (above the 95th percentile) to total precipitation. Also in summer the spatial pattern of the trends and their significance are inconsistent. Based on eight selected long-term precipitation records (mainly from Germany), Moberg and Jones (2005) assessed changes in five precipitation indices over Central Europe in the period 1901–1999. They found significant upwards trends in all wetness indices in winter (DJF) during the 20th century. However, the centennial-long increasing tendency is superimposed by decadal-scale variability in precipitation characteristics. In the warm season (JJA) no significant changes could be detected. Moreover, Moberg and Jones (2005) concluded that the data base for reliable assessments of the changes in precipitation behaviour and extremes over Europe is insufficient and that more digitized daily data are needed for reliable analyses.

Investigations with higher spatial density but a shorter time period have been conducted, e.g. by Hundecha and Bárdossy (2005) and Zolina et al. (2008). Hundecha and Bárdossy (2005) examined precipitation indices in the southern part of the Rhine river basin for winter and summer in the period 1958–2001 and reported an increase in heavy precipitation in winter of about 20% and a decrease of more than 6% in summer. Similarly, Zolina et al. (2008) found mainly changes in precipitation intensity indices in the winter season in a more extensive analysis of changes in precipitation amounts and of heavy precipitation over Germany during the period 1950–2004 based on 2125 rain gauges.

A few analyses of centennial long changes in daily precipitation exist for single locations such as Hohenheim, Karlsruhe and Hamburg (Wulfmeyer and Henning-Müller, 2006; Schlünzen et al., 2010). These studies indicate qualitative similar changes: increasing total precipitation, number of rainy days and days with precipitation amounts above 10 mm as well as an increasing contribution of heavy precipitation events to total precipitation during the last three decades compared to the centennial long time series. However, only the winter trends are significant and also more pronounced than the summer trends. A brief description of the long-term changes in extreme precipitations over Germany using 11 daily records for the period 1901–2000 and 54 records for the period 1941–2000 can be found in Grieser and Beck (2003).

It can be seen that the daily precipitation behaviour over Germany in the second part of the 20th century is extensively analysed by, e.g. Hundecha and Bárdossy (2005), Grieser et al. (2007) and Zolina et al. (2008), while the few studies of centennial long daily precipitation records are limited to individual observation sites or to a small set of stations (e.g. Wulfmeyer and Henning-Müller, 2006; Schlünzen et al., 2010; Grieser and Beck, 2003).

The same situation is found for the more detailed investigation of precipitation extremes using extreme value analysis, e.g. by fitting a Generalized Extreme Value (GEV) Distribution to the seasonal or annual maximum precipitation amounts. For the second half of the last century, precipitation extremes are analysed by Fowler and Kilsby (2003) or DeGaetano (2009), for example. In the study of Fowler and Kilsby, 1-, 2-, 5-, and 10-day accumulated annual maxima have been evaluated over the period of 1961–2000 in the United Kingdom (UK). DeGaetano (2009) analysed a slightly longer period (1950–2007) over the United States. Both studies also examine changes in the return levels throughout the investigated periods, and both find considerable differences within the period, with mainly increasing return levels with time. In order to enhance the sample size in view of the short time period, Fowler and Kilsby (2003) use spatial pooling, while DeGaetano (2009) applies a resampling procedure. An analysis of German daily precipitation maxima was conducted by Grieser et al. (2007), although restricted to the fit of a Gumbel distribution (as one of three cases of the GEV) and only for the second half of the 20th century. Rodda et al. (2010) recently showed for a lower number of UK rainfall stations with longer records that in the period 1911–2006 in different parts of the country strong positive as well as negative changes (up to ± 20%) in return levels are observed.

In summary, evidence for increased intensity or increased frequency of heavy precipitation events or for the contribution of very wet days to the total precipitation is mainly limited to the past 30–50 years or to very low spatial coverage. Thus, the extended data set used in our study allows for a more detailed investigation of precipitation characteristics both in better spatial coverage as also for a longer time period.

As mentioned above, it is a well known fact that the data base for reliable assessments of changes in the weather extremes and especially in heavy and extreme precipitation is insufficient (e.g. IPCC, 2007; Moberg and Jones, 2005). In view of these deficiencies, the WMO established in 1999, in the framework of the Global Climate Observing System (GCOS), a program to rescue climatological data (Data Rescue, DARE, en.html, accessed 8 July 2012). All Member States were invited to save their data archives on electronic media and to digitize the observational and associated meta-data as well as the not yet digitized material from yearbooks (Tan et al., 2003). Some countries (e.g. USA, Canada and Australia) have already digitized systematically their climatological data (Compo et al., 2006; Page et al., 2004). Many European countries are following these examples, e.g. those in the framework of the WMO-supported project MEDARE (Rescue and Digitization of Climate Records in the Mediterranean Basin,, accessed 8 July 2012).

In this context, the Meteorological Institute of the University of Bonn initiated a national project called KLIDADIGI of rescue and digitalization of German historical daily climate records at the end of 2005, which was established by the German Meteorological Service (DWD). In close co-operation with DWD we collected, digitized and quality-controlled daily precipitation data sets of 118 stations covering the period 1901–2000. The new data set will allow a more robust assessment of changes in heavy and extreme precipitation over Central Europe. For comparison, the widely used European data set of daily precipitation records—the ECA&D data set (Klein Tank et al., 2002)—contains only 17 stations in Germany with such long records.

On the basis of the substantially extended daily precipitation data set from 65 to 118 stations (see Section '2. Data and methods'), our analysis is focused on the regional assessment of the changes in precipitation characteristics during the period 1901–2000, the comparison of changes in the first and second part of the 20th century, and the robustness of trend estimates. We also re-evaluate regional trends in heavy precipitation over Germany for the second part of the 20th century which are found in earlier studies (e.g. Hundecha and Bárdossy, 2005; Zolina et al., 2008).

While in most of the above mentioned studies precipitation variability and changes are examined at single stations, in this study changes in precipitation characteristics are investigated on the scale of sub-regions, which are derived by means of Principal Component Analysis (PCA) in S-mode including a Varimax-rotated solution. The PCA enables to detect the spatial scale of similarity in precipitation characteristics. This is useful for: (i) comparisons of modelled and observed precipitation, (ii) selection of station networks for data homogenization, (iii) priority of data extension by digitalisation and (iv) reduction of inhomogeneities.

Different rotated PCA solutions have already been successfully applied to regional analyses of monthly and daily precipitation variability and changes in several other studies (e. g. White et al., 1991; Bonell and Sumner, 1992; Widmann and Schär, 1997; Brunetti et al., 2004, 2006a, 2006b). For example, Widmann and Schär (1997) classified daily precipitation records from Switzerland for the period 1901–1990 by means of un-rotated and rotated PCA. The purpose of this study was, in parallel to the trend analysis, the homogenization of the precipitation series. According to the Varimax-rotation, three sub-regions in Switzerland were detected. The authors stated that the regional trends were best characterized by objective regionalization using rotated PCA and that in this way the inhomogeneities of the data are ‘rapidly’ reduced.

This article is subdivided in four parts: In section '2. Data and methods' the data, the derived precipitation indices as well as the methods of clustering of the station-based indices for trend estimation and for extreme value analysis are presented. The interannual variability of the indices, the linear trends for the complete, the first and the second part of the 20th century and their robustness and also return levels of maximum precipitation are analysed in section '3. Results'. In section '4. Summary and conclusions', the results are summarized and discussed.

2. Data and methods

2.1. Data

The DWD operates a very dense precipitation measurement network, which until the end of the 20th century consisted of about 4500 gauges. Currently (in 2012), the observational network is reduced to 2000 stations. Only very few digitally available records of daily measurements, however, cover a period of 100 years and longer with few gaps. Most of these stations are located in the southern part of the country.

The DWD precipitation stations are equipped with HELLMANN rain gauges with a reading accuracy of 0.1 mm. All rain gauges are protected against evaporation and in the cold season a snow cross is placed in the gauge funnel. In the 1990s the tipping-bucket rain gauges and since about 2005 weighing rain gauges with similar dimensions as the Hellmann gauge are introduced to the precipitation network. The accuracy of these automatic precipitation measurements is 0.01 mm/min. The observational time was at the beginning 7:00 local time, from 1 January 1979 the observational time changed to 7:30 CET, and since April 2004 the observational time is 6:50 CET (5:50 GMT). These changes in observational time and instrumentation are less pronounced than re-locations. Furthermore, the accuracy of the measurements depends more strongly on the location of the gauges and on the special diligence of the observer than on the used gauge type (Hellmann, 1906). A station is continued after relocation with the same identification number if the horizontal distance is lower than 5 km, the altitude difference lower than 50 m and both locations belong to the same river catchment.

The systematic and documented quality assurance of the data began in 1979 with the introduction of IT-supported verification methods. Since 2012, the 10 min observations are also checked and if necessary corrected. Before 1979, the article measurement protocols were checked manually with neighbouring stations in real time, i.e. in the following month. For the German precipitation data the systematic gauge measuring error due to evaporation, wetting losses, and wind drift is not corrected; only 15% of the stations have the necessary wind observations.

In the newly collected data set, the previously available 65 centennial long daily precipitation records have been increased to 118 and also the station distribution has been homogenized spatially by filling the gaps in data availability especially in West-Central Germany (Figure 1). The additional hand-written reports have been digitized in the framework of a joint project between the Meteorological Institute of the University of Bonn and DWD (Mächel et al., 2009). In this project, emphasis was put to the western part of the country, which explains the still existing voids in the eastern part. Extending the record availability in the eastern part of the country is the focus of another still ongoing initiative.

Figure 1.

Orography in Germany (m) and location of the 118 stations of the extended data set. Old stations are marked in grey, the newly digitized stations in blue

The used records have been extensively quality controlled especially regarding outliers and so-called cumulative measurements. These are precipitation amounts ‘summed up’ by the observer over several days (specified as dry days) and reported only on one day. Because such artificial values contaminate especially the intensity related precipitation indices an own software was developed to their detection. The detected nearly 300 erroneous values are corrected by comparison with 2–3 neighbouring stations and by checking additional observer information in original measurement protocols. The identified accumulated amounts at the target station are spread to the number of rainy days reported at the 2–3 neighbouring gauges according to the percentage ratio of the precipitation amount at individual wet days. All detected errors are well documented. Finally, to examine the spatial coherence of the data and to detect ‘outlier’ gauges, PCA in S-mode and Varimax rotation is also applied to daily precipitation records. If such a gauge is not attributed to the same PC as the next neighbours, it can be assumed that the measured records are erroneous or inhomogeneous. In this case we found no such ‘outlier’ gauges. Nevertheless, our manifold quality control of the data cannot ensure an absence of non-detected errors.

The station records have not yet undergone the tedious homogenization exercise, as few daily data sets actually have (see also Wijngaard et al., 2003), because no adequate homogenization methods for daily precipitation records or precipitation indices are available at this time (see, e.g. COST Action ES0601 ‘HOME: Advances in homogenization methods of climate series: an integrated approach’;, accessed 8 July 2012). The results of the above mentioned PCA suggest, however, a high degree of spatio-temporal coherence of the presented data.

Our analysis is confined to the time period 1901–2000 in order to minimize gaps, although several records are longer than 100 years (e.g. 1880–2008). The individual station records in the final data set have gaps of less than 1000 d in total (2.7%), which are mostly found at the end of the Second World War (1945 and beginning of 1946).

2.2. Methods

2.2.1. Precipitation indices

The daily precipitation time series are characterized by means of precipitation indices following the recommendations of the ‘Expert Team on Climate Change Detection and Indices’ (ETCCDI, e.g. Peterson et al., 2001; Nicholls and Murray, 1999). These and similar indices have been applied in several other studies (e.g. Semenov and Bengtsson, 2002; Klein Tank and Können, 2003; Hundecha and Bárdossy, 2005; Moberg and Jones, 2005; Schmidli and Frei, 2005; Bachner et al., 2007).

The indices used in this study (Table 1) are based on the definition of a wet day as a day with a precipitation amount of at least 1 mm. Thus we avoid difficulties in the recording of light precipitation. In addition we use the maximum daily precipitation of each season (MAX) for further investigation of heavy precipitation by means of extreme value analysis (EVA; see Section '2.2. Methods').

Table 1. Abbreviations, units and description of the selected precipitation indices
SUMmmTotal precipitation/season
FREQ%Wet day frequency
CDDNumber of daysMaximum number of consecutive dry days
INTmm/dMean wet day intensity
Q90mm/d90th percentile value
R90pTOT%Precipitation amount (in % of total precipitation) from days with RR ≥ Q90
RX5daymmMaximum precipitation during five consecutive days
MAXmmMaximum daily precipitation

The selected indices are computed for summer (JJA) and winter (DJF) from the empirical distributions of daily rainfall in order to avoid problems with the typically long tails of daily precipitation observations, which are often not adequately covered by fitting theoretical distributions to the empirical data sample (e.g. Vlček and Huth, 2009; Koning and Franses, 2005).

2.2.2. Clustering by Principal Component Analysis

To assess the spatial scale of similarities in the precipitation characteristics and to minimize the influence of possible artificial trends and non-detected biases in individual precipitation indices, the stations are grouped according to their common temporal variability. The objective clustering is performed by means of a PCA in S-mode in the Varimax-rotated solution (e.g. Widmann and Schär, 1997, Brunetti et al., 2006a, 2006b). The selected Principal Components (PCs) represent groups/regions of similar precipitation variability. Since the PCA is a standard method in meteorological and climatological studies, we desist from its detailed description. More detailed information about the different applications and interpretations of the PCA can be found in Compagnucci and Richman (2008), Richman (1986, 1987), Jolliffe (1987, 2002), and Wilks (2006).

Prior to the PCA, the individual indices are computed for the warm and cold seasons (JJA for summer and DJF for winter) and standardized for the whole time series of each station individually. The standardization, i.e. the subtraction of the mean and the division by the standard deviation, is necessary to ensure similar ranges of variability of the different indices. The standardized indices are arranged into a common vector of length 700 (7 indices without total precipitation times 100 years) for each of the 118 stations. Thus, the input matrix to the PCA contains 118 columns and 700 rows. The PCs are then derived from the symmetric 118x118 matrix, which contains the correlation coefficients between the 118 index time series. After some experiments, 6 rotated PCs for winter and 9 for summer are extracted, which capture all stations and explain about 70% of the spatio-temporal variance of the analysed indices in winter and nearly 60% in summer.

The output of the PCA contains for each extracted PC a vector (so-called eigenvector) of loadings (coefficients or weights) for individual stations and for each extracted PC a standardized PC time-series (scores or amplitudes). In other words, the loadings correspond to the contribution of the individual stations to the total variance (eigenvalue) of the respective PC. In comparable studies and also here, the loadings (elements of the eigenvectors) represent the spatial patterns or regions. To simplify the interpretation of the patterns the loadings can be rescaled according to Wilks (2006) by multiplying them with the square root of the eigenvalue. Accordingly, the loadings are expressed in correlation coefficients of the original data at individual station with the time-series of the corresponding principal components (scores).

When we truncate the minimum correlation coefficient (loadings) at 0.33 similar to Richman and Gong (1999), and attribute the stations to those PCs with whom they are correlated most, the resulting station clusters form coherent, clearly distinguished regions, which is not a necessary outcome. These regions resemble well the natural geographical patterns (Figure 2). It is noteworthy that very similar, well distinguishable, regions can be obtained from the rotated PCA solution derived from the Spearman rank correlation matrix of the analysed indices. As can be seen from Figure 1, the spatial coverage of stations with century-long records has been improved considerably with the newly digitized data. Some of the regions, especially in the centre of Germany, are only very sparsely filled with old stations only and we therefore cannot compare regional precipitation statistics computed with and without the newly digitized stations.

Figure 2.

The regions obtained from the rotated PCA analysis for (a) summer (JJA) and (b) winter (DJF)

The regions are named according to their location within the country; abbreviations of the region names have an additional acronym ‘-s’ for summer and ‘-w’ for the winter season, since the regions differ between both seasons. Corresponding PC numbers are indicated in brackets (Table 2). The regions coincide well with orographic terrain features, the distance to the coast, and the prevailing wind directions (Figures 1 and 2). In summer, Northern Germany (N-s) down to about 52.5°N corresponds to PC1. Central and Western Germany is covered by PC3, PC7 and PC6 (NwC-s, RS-s and WC-s). The southern part (south of 50°N) is split up into four smaller regions, roughly following the Schwarzwald (Black Forest) and Schwäbische Alb (Sw-s, PC4), Alpenvorland (Foothills of the Alps, Sa-s, PC5), Bayerischer Wald (Bavarian Forest, SeC-s, PC2) and Franken (Franconia, SwC-s, PC8). The two stations in Eastern Germany (Berlin and Potsdam) form a region of their own (NeC-s, PC9). In winter, the northern region is almost the same as in summer, but related to PC2 instead of PC1. PC1 describes now a broader region in South Central Germany (SC-w), roughly between 49 and 50°N. The stations south of this region (S-w) are represented by PC3. Two small regions (PC5 and PC6) are situated in Central Western Germany (RS-w and WC-w) and another larger region at about 52°N between the Niederrhein (Nether-Rhine) and Eastern Germany (NC-w, PC4).

Table 2. Regions for summer and winter ordered from north to south and approximative correspondence between the two seasons. The corresponding PC-numbers are denoted in brackets
 Summer (JJA) Winter (DJF)
N-sNorthern Germany (PC1)N-wNorthern Germany (PC2)
NwC-sNorthwestern Central Germany (PC3)NC-wNorthern Central Germany (PC4)
NeC-sNortheastern Central Germany (PC9)  
RS-sRheinland (Rhineland) and Sauerland (PC7)RS-wRheinland (Rhineland) and Sauerland (PC5)
WC-sWestern Central Germany (PC6)WC-wWestern Central Germany (PC6)
SwC-sSouthwestern Central Germany (PC8)SC-wSouth Central Germany (PC1)
SeC-sSoutheastern Central Germany (PC2)  
Sw-sSouth Western Germany (PC4)S-wSouthern Germany (PC3)
Sa-sAlpenvorland (Foothills of the Alps, PC5)  

2.2.3. Trend analysis

The trend analysis is based on composite time series (PC scores, transformed to non-standardized values) of individual indices for the nine (six) selected regions in summer (winter) in order to enhance the trend robustness. These PC time series represent regional weighting averages of the individual indices which differ slightly from arithmetic regional means. For example, the PC time-series of the Q90 index in the region RS-s in summer correlate with the time-series of the arithmetic regional mean of the ‘original’ Q90 index by 0.971 and the respective MAX indices by 0.975. The PC time series can be interpreted as the ‘climate signal’ while the deviations from the arithmetic mean correspond to the noise. We avoided, however, computing trends for an average of all 118 stations, because these are not sufficiently representative for the whole area of Germany (cf Figure 1).

For most indices, the trend analysis is performed using the non-parametric Kendall's tau estimator of the trend slope (Kendall, 1975). Since the use of this estimator is problematic for the wet day frequency (FREQ) and the maximum number of consecutive dry days (CDD) of a season due to frequent occurrences of identical values, the ordinary least squares regression approach (see, e.g. Wilks, 2006) is used instead. For all indices, the significance of trends is assessed by the non-parametric Mann-Kendall test (Mann, 1945; Kendall, 1975) which accepts also non-normal data.

2.2.4. Extreme value analysis

In order to investigate more closely heavy and in particular extreme precipitation, i.e. the tails of the empirical distributions, the return levels of daily maximum precipitation are estimated using EVA (see e.g. Coles, 2001). To this goal, a GEV Distribution with shape parameter ξ, location parameter µ, and scale parameter σ is fitted to the time series of the regionally averaged seasonal maxima (index MAX). The goodness-of-fit is assessed with the Kolmogorov–Smirnov test. The shape parameter ξ determines the type of the distribution: with ξ = 0 the distribution is Gumbel-type with an exponential tail; ξ > 0 indicates a Fréchet-type distribution with a so-called heavy tail; finally ξ < 0 hints at a Weibull-type distribution with an upper limit. From the estimated three parameters of the GEV (indicated by the hats in the following equation), the return levels xp of maximum seasonal precipitation are estimated:

equation image(1)

The uncertainty of the return levels, provided in terms of the 95% confidence intervals, is assessed using bootstrap resampling (Efron and Tibshirani, 1993).

2.2.5. Variability throughout the century

In addition to the estimation of trends and return levels for the whole 20th century, we also want to investigate the trend and return level variability throughout the century. Thus we also address the robustness of the estimates. To this end, we first divide the whole 100 year period in two halves, since many previous studies concentrated on roughly the second part of the century due to lack of data. Secondly, we compute linear trends for time windows of 30 year length moving through the century in steps of 1 year beginning with 1901–1930, 1902–1931 and so on. In this way, we obtain trend time series for the regionally averaged indices of length 70/71 (winter/summer) sub-periods (e.g. Brunetti et al., 2006b; Lupikasza, 2010). According to Lupikasza (2010) we define trend stability using the percentage of significant trends S (at the 90% level) over these sub-periods as:

  • unstable trends: 0%≤S < 15%

  • poor trend stability: 15%≤S < 25%

  • stable trends: 25%≤S < 50%

  • strongly stable trends: 50%≤S < 75%

  • very strongly stable trends: S ≥ 75%.

3. Results

3.1. Interannual and regional variability of the indices

This section focuses on the region-specific features of the variability and trends of the precipitation indices. The values of median and interquartile range throughout the three different time periods are assembled in Tables 3 and 4. The region Foothills of the Alps in summer (Sa-s) and the regions Southern Germany (S-w) and Rhineland/Sauerland (RS-w) in winter exhibit for most indices much larger median values than the other regions. As an example for the whole interannual distribution, the 90th percentile (Q90) value is shown in Figure 3 for the period 1901–2000. This figure closely resembles the interannual variability (not shown, cf the values in Tables 3 and 4) of the mean wet day intensity (INT), the maximum 5-day precipitation (RX5day), and the seasonal maximum (MAX). In winter, also the interannual variability is considerably larger in these regions. In summer, the largest variability is found in the area of Northeastern Central Germany (NeC-s). The wet day frequency (FREQ) in the regions Foothills of the Alps (Sa-s) in summer and Rhineland/Sauerland (RS-w) in winter is also slightly enhanced compared to the other regions, although to a lesser extent. In summer, this is also due to an overall higher interannual variability. In contrast, the maximum number of consecutive dry days is slightly lower in these two regions with lower interannual variability as well. The R90pTOT index shows quite similar mean values in all regions, but with considerable differences in the variability for the regions Southwestern Central Germany (SwC-s) and Northeastern Central Germany (NeC-s). Most distributions in both seasons are skewed with longer tails towards larger values.

Figure 3.

Interannual variability of the index Q90 (mm/d) in the different regions for (a) JJA and (b) DJF

Table 3. Median values of the precipitation indices for summer (JJA) in the periods 1901–1950, 1951–2000 and 1901–2000 (units of the indices cf Table 1). In addition to the regions, an average over all 118 stations is shown
Summer (JJA)
118 stations1901–1950236.837.
Winter (DJF)
118 stations1901–1950161.635.613.04.910.236.628.117.2
Table 4. Interquartile range (IQR) of the precipitation indices for summer (JJA) in the periods 1901–1950, 1951–2000 and 1901–2000 (units of the indices cf Table 1). In addition to the regions, an average over all 118 stations is shown
Summer (JJA)
118 stations1901–1950120.812.
Winter (DJF)
118 stations1901–195095.

3.2. Linear trends

3.2.1. Centennial trends

We consider first the whole 20th century and the differences of the results when splitting the century into two sub-periods 1901–1950 and 1951–2000, before discussing the robustness of the linear tendencies throughout the century.

In Figure 4, the findings on the significance of the linear trends are summarized for all indices, regions and seasons; results are shown separately for the whole century as well as for the first and second half of the past century. We discuss first the trends of the whole century (upper panel). In the summer season, few century-long trends are statistically significant. These include increasing trends at a significance level of 90% and above for mean intensity in the Southeastern Central Germany (SeC-s), Foothills of the Alps (Sa-s) and the Rhineland/Sauerland (RS-s) region, for the Q90 also in Southeastern Central Germany (SeC-s) and in the Rhineland/Sauerland (RS-s), and the maximum 5-day precipitation (RX5day) only in the Foothills of the Alps (Sa-s).

Figure 4.

Significance of linear trends for the eight indices in the different regions for JJA (left) and DJF (right) and the three different periods: (a) whole century 1901–2000, (b) first half 1901–1950 and (c) second half 1951–2000. Trends significant at the 95% level are indicated by a double trend sign (++ or − −), those which are significant at the 90% level by a single sign (+ or −). Blank fields indicate non-significant trends

In winter (Figure 4, right side of top panel), the number of significant trends for the entire century is three times higher than in summer, especially in the whole of Southern Germany (SC-w and S-w) but also—to a lesser degree-for Northern Germany (N-w). Apart from increased total precipitation (SUM) in these regions, many of the intensity-related indices such as wet-day intensity (INT), 90th percentile value (Q90), maximum 5-day precipitation (RX5day) and daily precipitation maximum (MAX) show upwards trends. The Rhineland/Sauerland region (RS-w) also exhibits a significant trend at the 95% level for Q90. Indices describing the occurrence of precipitation, such as wet-day frequency (FREQ) and maximum number of consecutive dry days, do not show any significant trends on the 95% level in both seasons for the 20th century as a whole. In winter, however, the wet day frequency (FREQ) is increasing at the 90% level in the regions Northern Germany (N-w) and Western Central Germany (WC-w). In the latter region, CDD is decreasing. In Western Central Germany these changes are accompanied by significant enhanced total precipitation (SUM).

Trend significance patterns are quite different when evaluated separately for the first and second half of the past century (Figure 4(b) and (c)). In summer of the first half of the century, a strong increase in the intensity-related indices is observed in the regions Sa-s, WC-s, SwC-s and NeC-s (Foothills of the Alps, Western Central Germany, Southwestern Central Germany and Northeastern Central Germany). These changes are in line with significant increased total precipitation in Western Central Germany. In winter, there is only one weak negative trend for the index Rx5day in Western Central Germany (WC-w); otherwise no significant changes in the precipitation characteristics can be detected for these first 50 years. The second half of the century is characterized by drying in summer in 7 of the 9 regions and significant increase in intensity-related indices in winter, although compared to the trend pattern for the entire century there is a remarkable shift from southern to northern regions concerning the trend strengths. In summer some—partially weak—negative trends related to a precipitation decrease for the northern half of Germany, especially for the wet-day frequency are noted. Accordingly, the length of dry periods (index CDD) has increased in 4 of the 9 regions, covering most of the western and south-western parts of Germany. These tendencies in opposite direction for the wet-day frequency and maximum number of consecutive dry days are also seen in the non-significant trends in the other periods. They are consistent with moderate negative correlations (between − 0.5 and − 0.7) of both index time series in general.

The different trends for the different sub-periods are well discernible in the time series (see Figure 5 for Northern Germany in winter): several indices exhibit large interannual (see, e.g. CDD) or also interdecadal (e.g. MAX) variability. The interannual variability prevents trends from being significant: when interannual variability changes with time, the trend significance is also affected. A large interdecadal variability results in a strong dependence of the trend on the considered period. Interdecadal variability might also explain a part of the trend differences between the two half-centuries, i.e. several cases exist where the tendencies for the first and second half of the century are in different direction. Thus, the centennial long trends in these cases are weak or have low significance.

Figure 5.

Time series of the eight indices in winter (DJF) in Northern Germany (region N-w) including the linear trend lines for the whole 20th century (solid line) as well as for the first and second half (dash-dotted and dashed line, respectively)

Trend magnitudes are generally stronger and regionally more diverse in the individual 50-year periods compared to the century trends especially in summer (Tables 5 and 6). The trend values are presented in % of the respective median of each time series in order to better compare the trend strength between the different indices. The absolute values of the trends can be assessed through Table 3. The strongest trends are found for the maximum number of consecutive dry days with an average of 5.7% of the median per decade or 0.7 d increase in summer in the period 1951–2000 (average over all regions). A maximum change of 12% or 1.3 d is noted in this period in the Rhineland/Sauerland region (RS-s). The maximum 5-day precipitation total also exhibits large trend magnitudes, with up to 8% or 4 mm in the regions SwC-s and NeC-s (Southwestern Central Germany and Northeastern Central Germany), however in the first half of the century. The strongest trends in summer, when comparing the different regions, are found in Southwestern Central Germany for 1901–1950 (3.9% average over all indices). In winter, the differences in the trend magnitudes are not that large between regions and indices, although somewhat stronger changes can be detected for the indices INT, Q90 and RX5day in the region N-w (Northern Germany). During the whole century in winter total precipitation (SUM), wettest 5 days (RX5day) and the seasonal maximum precipitation (MAX) in Southern Germany (S-w) are increased by more than 3% per decade. Similar magnitudes are found also in region SC-w (South Central Germany) for the wet-day intensity (INT) and Q90.

Table 5. Trend magnitude in % of median per decade in JJA for the periods 1901–1950, 1951–2000 and 1901–2000. Bold underlined: 95% significance level, bold: 90% significance level
N-s1901–1950+ 0.70− 0.49− 0.74+ 1.28+ 2.24+ 4.04+ 0.99+ 2.93
 1951–20004.223.89+ 6.87− 0.63− 0.78− 1.99− 0.06− 2.02
 1901–2000− 0.10− 0.18+ 1.16− 0.25− 0.13+ 0.07+ 0.15− 0.07
NwC-s1901–1950− 1.69− 2.28+ 3.69+ 0.88+ 1.60− 1.06− 0.04− 0.27
 1951–20004.784.02+ 8.60− 1.11− 1.57− 2.992.025.14
 1901–2000− 1.02− 0.84+ 0.93− 0.16− 0.02− 0.31+ 0.06− 0.72
NeC-s1901–1950+ 1.54− 1.54+ 0.83+ 2.96+ 1.90 + 7.93+ 1.45+ 6.28
 1951–20004.983.51+ 7.28− 2.465.14− 2.72− 0.99− 2.64
 1901–2000− 0.68− 0.88+ 1.64+ 0.18− 0.21+ 1.43− 0.12+ 0.61
RS-s1901–1950+ 0.89− 0.84+ 0.36+ 2.20+ 0.66+ 1.53− 1.12− 0.07
 1951–20005.515.17 + 12.08+ 0.77+ 1.66− 2.69− 0.30+ 0.03
 1901–2000+ 0.28− 0.57+ 0.80 + 1.22 + 1.71+ 1.40+ 0.36+ 1.16
WC-s1901–1950 + 3.96+ 0.07+ 0.38 + 2.75 + 3.20 + 6.27+ 1.12 + 4.22
 1951–2000− 4.53− 3.26 + 5.74− 0.99− 1.61− 3.93− 1.21− 3.84
 1901–2000+ 0.06− 0.80+ 1.34+ 0.41+ 0.44+ 0.37− 0.04+ 0.15
SwC-s1901–1950 + 7.94− 0.12+ 2.13 + 4.30 + 4.43+ 1.74 + 6.37+ 4.44
 1951–20004.83− 2.48+ 2.19− 1.11− 1.70− 3.90− 1.13− 3.06
 1901–2000− 0.12− 0.44+ 1.02+ 0.30+ 0.15+ 0.28− 0.21+ 0.39
SeC-s1901–1950+ 1.28+ 0.75+ 1.02+ 1.08+ 1.14+ 1.13− 0.16+ 1.71
 1951–2000− 2.50− 1.85+ 1.75− 0.25+ 1.22− 3.29− 0.78− 2.35
 1901–2000+ 0.06− 0.55+ 0.78 + 0.90 + 1.15+ 0.94− 0.13+ 0.80
Sw-s1901–1950+ 1.71− 0.97+ 4.98+ 0.90− 0.07+ 0.81− 0.49+ 2.04
 1951–20004.02− 1.92 + 4.20− 1.14− 2.47−
 1901–2000− 0.42− 0.43+ 0.61+ 0.03− 0.06+ 0.06+ 0.31+ 0.58
Sa-s1901–1950+ 0.31− 0.90 + 7.25+ 0.89+ 0.97+ 3.53− 0.84− 0.02
 1951–2000− 1.87− 0.75+ 2.51− 0.22+ 0.22− 0.68+ 0.27− 1.57
 1901–2000+ 0.51− 0.18+ 1.09 + 0.79+ 0.60 + 1.51− 0.09+ 0.76
Table 6. Trend magnitude in % of median per decade in DJF for the periods 1901–1950, 1951–2000 and 1901–2000. Bold underlined: 95% significance level, bold: 90% significance level
N-w1901–1950+ 0.75− 0.39− 0.58− 0.75− 0.87− 0.79+ 0.14− 1.62
 1951–2000+ 4.59+ 1.23+ 4.86 + 4.17 + 5.31 + 5.76+ 0.81+ 4.05
 1901–2000 + 2.10 + 1.27−0.63+ 0.34+ 0.81 + 1.74 + 0.48+ 1.41
NC-w1901–1950− 0.94− 0.99− 1.14− 1.24− 0.47−3.64−0.22− 1.02
 1951–2000+ 4.83+ 1.94+ 0.97 + 3.16+ 2.21+ 4.10+ 1.11 + 5.32
 1901–2000+ 1.34+ 0.72− 1.09+ 0.38+ 0.61+ 0.14+ 0.14+ 0.88
RS-w1901–1950+ 1.25+ 1.54− 5.35− 1.04+ 0.83− 1.01+ 1.99+ 0.44
 1951–2000+ 1.75+ 0.31− 0.35+ 2.48+ 2.90+ 3.44+ 0.15+ 2.37
 1901–2000+ 1.48+ 0.57− 1.21+ 0.71 + 1.39+ 0.74+ 0.38+ 0.85
WC-w1901–1950+ 0.68+ 1.78− 4.86− 2.45− 2.414.80−0.05−3.11
 1951–2000+ 3.39+ 1.65+ 0.94+ 1.85+ 2.04+ 3.13+ 0.64+ 2.72
 1901–2000 + 2.69 + 1.192.03+ 0.79+ 0.91+ 1.58+ 0.08+ 1.39
SC-w1901–1950+ 1.16+ 0.56−1.34−0.15+ 0.68− 0.61+ 1.48+ 1.53
 1951–2000+ 3.42+ 1.23+ 2.30 + 2.72+ 2.15+ 4.26+ 1.38 + 5.80
 1901–2000 + 2.60+ 0.57+ 0.04 + 1.73 + 1.60 + 2.80+ 0.63 + 2.99
S-w1901–1950+ 1.92+ 1.06− 3.85+ 0.98+ 1.42+ 0.19− 0.77− 0.72
 1951–2000+ 2.67+ 0.03+ 5.49 + 3.87 + 3.93+ 3.07− 0.76+ 2.45
 1901–2000 + 3.65+ 1.01+ 0.34 + 2.62 + 2.61 + 3.54+ 0.13 + 3.29

Averaged over all regions and indices, trends during the first half of the century are about 2.0% in summer and 1.4% in winter. In the second half, however, averaged trends are equally strong in both seasons and much stronger than in the first half century (2.6%). Averaged over the whole century trends are twice as strong in magnitude in the winter (1.2%) than in the summer season (0.6%). Thus, conclusions about the seasonality of trend magnitudes are very much dependent on the considered time period.

3.2.2. Robustness of trends

The results of the trend analysis clearly indicate that the sign, magnitude and significance of the trends strongly depend on the time period for which the trends are estimated. To examine this further, 30 year moving trends as described in section '2.2. Methods' are investigated.

In total, 5112/3360 (summer/winter) trend values are estimated, corresponding to the eight precipitation indices, 9/6 regions and 71/70 30-year sub-periods. Both seasons are dominated by cases with upward trends, although with a slightly higher percentage in winter (57.6%) than in summer (51.7%). However, only a very small part of the 30-year moving sub-periods show significant negative/positive trends at 90% and higher levels (Table 7). In summer, the percentage of significant trends with 5% of negative and 5.6% of positive cases is a little higher than in winter (2.4% negative and 6.3% positive trends). Accumulating the significant trends over all regions, most of the positive trend cases in summer are noted in the indices MAX, R90pTOT, CDD, INT and RX5day and negative cases in the indices FREQ and CDD. In winter, the precipitation intensity (INT) shows a maximum number of significant positive and negative trends. In summary over all indices, the regions SwC-s and WC-s in summer and the regions N-w, S-w and NC-w in winter are characterized by highest numbers of significant upward trends. The most of the significant downward trend cases are observed in the regions N-s, NwC-s and NeC-s in summer and in the regions NC-w and N-w in winter. The frequency of the significant trends in the different stability classes according to Lupikasza (2010) indicates that most of those trends in summer and winter can be classified as unstable (Table 7). In summer, 16 cases (6 negative and 10 positive) are found, which fall in the category ‘poor stable trends’. The classes of ‘stable’ trends are unfilled. Most of the ‘poor stable positive trends’ are concentrated in the regions SwC-s and WC-s. In winter, two cases of significant upward trends can be identified as ‘poor stable trends’ and only one case passed the criterion of ‘stable’ trends (the wet-day intensity in Northern Germany). All significant downward trends in winter are ‘unstable’. Consequently, it can be concluded that the robustness of trend estimates throughout the century is low and spatially variable.

Table 7. Percentage of significant negative and positive trends (at the 90% level) in the 30-year moving sub-periods during the 20th century (1901–2000) in different precipitation indices and in different regions for summer and winter. The total number of sub-periods is 71 for summer and 70 for winter. Bold numbers indicate sub-periods with ‘poor stable trends’ (15.0% ≤ S < 25.0%) and bold underlined numbers with ‘stable trends’ (25.0% ≤ S < 50.0%) according to Lupikasza, 2010; cf also section '2.2. Methods'
Summer (JJA)
% of sub-periods with significant negative trends
% of sub-periods with significant positive trends
Winter (DJF)
% of sub-periods with significant negative trends
% of sub-periods with significant positive trends

This finding is also evident in the time series of the moving trends, which are shown exemplarily for the index RX5day in Figure 6. The results are similar for the other intensity-related indices due to their strong inter-correlation. As can be seen, the 30-year trends vary strongly throughout the century with several changes from positive to negative values and vice-versa. Summarizing the results for all indices, some significant trends are clustered around the 1920s and 1930s for the summer and the regions WC-s and SwC-s (Western Central Germany and Southwestern Central Germany) and around the 1980s in all regions in winter (except for SC-w). While the period with significant trends around the 1920s and 1930s in summer is obvious in all the intensity-related indices, the trends in FREQ and CDD are mostly significant around the 1950s and 1970s. In winter, the clustering of significant 30-year trends around the 1980s is valid for 4 of the 5 intensity-related indices.

Figure 6.

Time series of 30-year moving trend values (in % of the respective median per decade, drawn at the centre of each 30 year period) for the index RX5day in (a) JJA and (b) DJF. The grey-shaded background plot is the shape of Germany and the inset plots are drawn approximately at the position of the respective regions. Axes limits in the inset plots are all [1900; 2000] for the x- and [−20; 20] for the y-axis, exemplarily shown only for one region in each season. Trends significant at the 90% level are plotted in green, those significant at the 95% level in red and non-significant trends in blue

The spatial variability, i.e. by comparing the time series for the different regions in Figure 6, is also very large: the regional time series have a very different structure, especially in summer. In winter, the trend time series in the different regions are more similar. In general, the curves suggest that the trends are superimposed by strong inter-decadal variability of the intensity-related indices.

3.3. Return levels of maximum precipitation

The estimation of return levels for extreme precipitation heavily depends on the representation of the tails of the PDFs. We fitted a GEV Distribution to the time series of daily maximum precipitation (MAX). From the parameters of the GEV, return levels related to a specified return time can then be estimated directly. The fit of a GEV was in all cases, i.e. all seasons and regions, 50- and 100-year time series, successful according to the Kolmogorov–Smirnov test. The confidence interval for the shape parameter of the fitted distribution includes zero in nearly all cases; thus, the GEV type is mostly neither distinctly Weibull nor Fréchet, and consequently the tail of the distribution is approximately exponential. Since few of the time series of maximum precipitation show significant trends in the different periods, it is worth to verify these by a non-stationary GEV model. We tested the non-stationarity by introducing a covariate in the location parameter, representing a linear trend. This covariate is not significant in most cases. In the other cases, the resulting return levels vary much less than the extent of the confidence intervals of the stationary ones. For our discussion of the differences in the spatial distribution of return level estimation, the simpler stationary model is therefore sufficient, since we concentrate more on the size of the confidence intervals and on the differences in regions and time periods than on absolute values of the return levels. This is more suitable because we analyse regional averaged time series and not the MAX index at individual stations with generally considerably larger magnitude of the return levels. Confidence intervals are large in all cases due to the relatively small sample sizes; nonetheless their size gives important information on the reliability of return level estimates.

For return periods from 2 to 100 years, return levels are shown in Figure 7 for two selected regions. In Northwestern Central Germany in summer (NwC-s), the regionally averaged 100-year return levels are around 50 mm/d with an uncertainty (95% confidence interval) of ± 10 mm, i.e. a range of 35% compared to the value itself. In roughly the same region in winter (NC-w), values are somewhat lower with return levels ranging from 18 mm for a 2-year return period to ∼30 mm for 100 years. Generally, return levels are lower in winter than in summer for the other regions as well. The confidence intervals are also slightly smaller in absolute values in winter. For NwC-s and NC-w, the shape of the curve is more Fréchet-type in summer (heavier tail) and more of the Weibull type in winter in this region. In most cases (regions and periods), however, the shape of the curve tends more towards the Weibull type in both seasons (although mostly not significantly, see above).

Figure 7.

Return level plots for (a) the region Northwestern Central Germany (NwC-s) in summer and (b) Northern Central Germany (NC-w) in winter. The entire 20th century (solid grey line and grey shading) is compared with the second half of the century (black dashed lines)

According to Figure 7(b), the shape of the return level curves for Northern Central Germany in winter differs considerably when data from different periods are used: the increase of return levels with return period is much flatter when we account for only the last 50 years of data instead of the whole century.

The 100-year return levels calculated from three time periods (1901–2000; 1901–1950 and 1951–2000) are summarized in Figure 8 for all regions and both seasons. We note that, considering the confidence intervals, most return level estimates cannot be distinguished statistically between the three time periods, despite considerable absolute differences. On average, the 100 year return levels differ by 4.7–9.6 mm in summer and by 1.7–4.4 mm in winter; in both seasons the deviations are largest between the two 50-year periods as expected. Maximum differences in summer occur in the region Foothills of the Alps in the far south with 19.6 mm, whereas in winter Northern Germany exhibits the largest differences between two time periods (3.3 mm).

Figure 8.

100-Year return levels (‘RL100’, mm/d) for all regions and both seasons estimated from the whole time series (blue) as well as from the first (red) and second (black) half of the century

A closer analysis of both half centuries reveals that in 5 of 9 regions in summer (which are spread throughout the country) and in 2 of 6 regions in winter (only in the northern part), the return level estimated from one period is not included in the confidence interval of the other period. This indicates differences between the two parts of the century and hints at a non-stationarity of the climate system with respect to extreme precipitation. In most cases, the 100-year return level estimate obtained from the whole century time series is included in the confidence interval of the two half century estimates, except for two regions in summer (NwC-s and WC-s) and for one region in winter (N-w) where the first 50 years estimate does not include the whole century estimate. In the latter cases we would considerably underestimate the 100-year return levels in the north-western parts of Germany if only the data from the first 50 years were considered.

It can be also noted that the confidence intervals of the return levels estimated from the shorter time periods are, as expected, considerably larger than those from the long period (cf Figures 7(a) and 8). On average, the range of the confidence interval is 22.6 mm/d (or 34% in relation to the value itself) in summer for the whole century, compared to 32.2 mm/d (48%) and 36.7 mm/d (54%) for the first and second half, respectively. Similarly, the values for the winter season are 12.6 mm/d (31%) for the long and 18.9 mm/d (47%) and 17.7 mm/d (43%) for the two short periods. Since the size of the confidence interval can be important, e.g. in planning purposes, such differences are non-negligible.

4. Summary and conclusions

We have investigated precipitation characteristics of century-long daily observation records over Germany extended recently from 65 to 118 stations, which cover homogeneously the western part of Germany. This data extension allows now a more detailed analysis of the spatial variability of daily precipitation over the whole century, which was not possible before. Our experience during digitizing and quality checking of historical precipitation records suggests that the observations are highly contaminated by errors and inhomogeneities. In addition to measurement errors (e.g. non-suitable exposition, wind influence, etc.), the most pronounced uncertainties result from incorrectly labelled days with ‘no precipitation’ or ‘no measurement’ by the observers, and from unspecified accumulated precipitation amounts over several days attributed to a single day. The accumulated values are very difficult to detect, in particular if the values cannot be identified as outliers. Such uncertainties influence strongly the apparent statistical characteristics of precipitation. In spite of our careful control of the data and elimination of detected artefacts by close examination of observer remarks in the original documents and by comparison with neighbouring stations, we cannot exclude remaining inhomogeneities in the data set. Due to these difficulties, we analysed the precipitation characteristics not for individual stations but rather for clusters of stations.

A regionalization by means of a rotated PCA leads to a satisfying delineation of spatially coherent and consistent regions of stations with similar variability in daily precipitation and hopefully reduces the influence of inhomogeneities. This hope is corroborated by several other studies, which have used PCA for classification/regionalization of monthly or daily precipitation records (see introduction). We obtained a clustering into nine regions in summer and six in winter. Because of the very inhomogeneous spatial coverage of stations with long records such a regionalization based on data of the whole 20th century was not possible before the data extension.

A trend analysis of the regionally averaged time series for the 20th century reveals some increasing tendencies over the century in the winter season, which are statistically significant only in few regions and for some of the indices. Indices describing rainfall intensities exhibit trends more often than those describing rainfall occurrence and duration. The comparison of linear trends from the whole century with those estimated from only the first or second half of the century reveals different trend significance patterns. This is due partly to decadal variability in the indices and also to trends being restricted to certain shorter periods.

The main results of the trend analysis are:

  1. Taking the whole century, several significant trends are observed in total precipitation and in intensity-related indices for some of the regions and for both seasons. Significant trends are, however, much more frequently observed (in terms of region and index) for the winter season than for summer.

  2. Trends—if significant—differ, however, considerably between both halves of the century:

    1. The first half of the century is dominated by insignificant trends in the winter season in any region and any index. In the summer season, however, the central regions of Western Germany show a marked increasing trend in most of the intensity-related indices while the maximum number of consecutive dry days increased significantly in the Foothills of the Alps in the very South of Germany.

    2. In the second half of the century, the summer season is characterized by reduced total precipitation and frequency of wet days accompanied by an increase of the maximum number of consecutive dry days for all regions. The winter season reveals increasing trends in many of the intensity-related indices for the whole of Germany except a small region in Western Central Germany.

Other studies on centennial-long daily precipitation records from European countries also suggest rather weak changes in precipitation characteristics in summer and mainly positive but regionally different trend magnitudes and significances in winter (e.g. Moberg and Jones, 2005; Schmidli and Frei, 2005). Concerning changes of precipitation characteristics during the first part of the 20th century no comparable recent studies are available. Despite the different investigation period and analysis method, the detected trends of the second half of the 20th century are qualitatively in line with the findings of Hundecha and Bárdossy (2005) and Zolina et al. (2008). They suggest increasing tendency in heavy precipitation indices especially in winter in some parts of Germany.

An analysis of 30-year moving trends indicates a low robustness of the trend estimates throughout the century with only one case of an upward trend classified as ‘stable’. In the summer season, compared to winter, we find a particularly large variability of trends (sign, significance and magnitudes) in time and space and among indices. In conclusion, tendencies in daily precipitation in Germany over the 20th century are variable and climate change statements are very sensitive to the considered time period and therefore need to be treated with care. Similar conclusions can be found in Brunetti et al. (2004, 2006a, 2006b) and Lupikasza (2010).

A GEV Distribution has been successfully fitted to the regionally averaged seasonal maxima of rainfall. The shape parameter is generally not significantly different from zero; thus the assumption of a Gumbel-type distribution with approximately exponential tails (Grieser et al., 2007) cannot be ruled out in general. However, regional and temporal variations exist and the shape parameter and thus the form (type) of the distribution could change in the future in accordance with DeGaetano (2009), who detected a weak tendency of an increasing shape parameter during the last 50 years in the United States.

Our estimates of return levels vary also between the three considered time periods, in accordance with the results of Fowler and Kilsby (2003) and DeGaetano (2009). Due to broad confidence intervals, most differences are not statistically significant. The uncertainty of the estimates, indicated by the size of the confidence intervals, however, can be considerably reduced when the fit is based on the longer data period. Comparable studies on the effect of record length on return level confidence intervals are, to our knowledge, not available at this time. Furthermore, we find that the magnitude of return levels can in some cases be underestimated by the short records. Both can be important factors in risk analysis and risk management, e.g. for the determination of flood prevention walls.

We conclude that with the newly extended data set of daily precipitation records in Germany we achieved a much better and more detailed insight into the long-term variability of daily precipitation characteristics, both temporally and spatially. Our results suggest changes in statistical estimates throughout the century, and consequently it is worth continuing and increasing the efforts for data archaeology, since there is still a large amount of data available for digitization. Once the knowledge and confidence in the homogenization of daily precipitation has improved—which hopefully further increases the quality of our data set—it will be also very interesting to expand this analysis to the individual stations in order to learn more about the spatial variability of daily precipitation in Germany.


This research was funded by the North Rhine-Westphalian Academy of Science and Arts in the project ‘Large Scale Climate Changes and their Environmental Relevance’. We thank the reviewer whose comments helped to improve the manuscript.