JULY 2014AMERICAN METEOROLOGICAL SOCIETY |JULY 2014| 995PB

AFFILIATIONS : Zolina—LGGE, Grenoble, France, and P.P. Shirshov Institute of Oceanology, Moscow, Russia; Simmer and Kapala—Meteorologisches Institut, Universität Bonn, Bonn, Germany; Shabanov and Gulev—P. P. Shirshov Institute of Oceanology, Moscow, Russia; becKer and mächel—Deutsche Wetterdienst (German Weather Service, DWD), Offenbach, Germany; GroiSman—National Climate Data Center (NCDC), Asheville, North CarolinaCORRESPONDING AUTHOR : Olga Zolina, LGGE, CNRS/UJF-Grenoble 1, 54 rue Molière, BP 96, 38402 Saint Martin d’Hères Cedex, FranceE-mail: ozolina@lgge.obs.ujf-grenoble.fr

DOI:10.1175/BAMS-D-12-00134.1

A supplement to this article is available online (10.1175/BAMS-D-12-00134.2).

©2014 American Meteorological Society

records—cannot effectively capture such strongly localized precipitation extremes, especially in regions with complex orography where precipitation patterns are extremely inhomogeneous in space and highly variable in time. Multiannual gridded products based on these sparse network rain gauge collections, although extremely valuable for the assessments of long-term continental-scale changes in precipitation and their further intercomparison with climate model simulations, are hardly useful for the evaluation of experiments with regional high-resolution models targeting case studies of extreme events.

Here we provide a brief overview of the newly developed set of high-resolution daily precipitation grids over Germany, which are derived from the daily-observing precipitation network of the Deutsche Wet-terdienst (DWD). High-resolution daily grids covering in different data streams the period from 1931 onward were developed under the STAMMEX (Spatial and Temporal Scales and Mechanisms of Extreme Pre-cipitation Events over Central Europe) project funded by the DFG (Deutsche Forschungsgemeinschaft) and supported by DWD, the Meteorological Institute of the University of Bonn, and the North Rhine Westphalian Academy for Science and Arts.

NATIONAL DWD RAIN GAUGE NET-WORK. The DWD rain gauge network is one of the densest and most properly maintained regional precipitation networks. It consists of 11,617 stations, of which about 7,500 have been digitized, quality-controlled, and included in a digital database. In the last four decades, when sampling was the densest, the typical station-to-station distance was 3–10 km in the southern and central parts of Germany and 5–20 km in the northern and eastern parts (Fig. 1a–c). The number of continuously reporting gauges increased since the late 1940s from about 2,000 to more than

H igh-quality quantitative information about precipitation characteristics in Europe, includ-ing extremes, is crucial for estimating and

modeling observed climate variability and trends in European hydroclimate and f loods. The recent flood in early June 2013 over Southern and South-eastern Germany exceeded the water levels of the already record-breaking flood of August 2002. This event was followed by very localized and extremely heavy precipitation events in the latter half of June in Western Germany. These two events give a flavor of the strongly localized character of European pre-cipitation, with spatial patterns ranging from a few to several tens of kilometers. The strongest clustering of precipitation patterns is observed in summer when its structure is dominated by convective processes. This poses strong requirements for the spatial resolution of rain gauge networks used for quantification of spatial and temporal variability of rainfall patterns. The sparse rain gauge networks usually used for climate assessments—despite their long and continuous daily

Precipitation Variability and Extremes in Central Europe

New View from STAMMEX Resultsby olGa Zolina, clemenS Simmer, alice Kapala, pavel Shabanov,

paul becKer, hermann mächel, SerGey Gulev, and pavel GroiSman

JULY 2014AMERICAN METEOROLOGICAL SOCIETY |JULY 2014| 997996

4,000 stations until about the year 2000, when the collection experienced a decline of about 10%–15% (Fig. 1d). All DWD stations are equipped with Hellmann rain gauges with a 200-cm2 collecting sur-face and seasonally activated snow crosses and lids. Since the 1950s, the gauges have been equipped with antifreezing agents. In the 1990s, the Hellmann gaug-es were replaced at most stations by tipping buckets combined with drop-counting (for low rain rates) rain gauges with the same orifice of 200 cm2 and a heated funnel controlled by thermostats. In the early 2000s, DWD started to implement the second generation of automatic weighing gauges (the so-called Pluvio devices). Generally, the heated tipping-bucket gauges may imply an uncertainty resulting from evaporation and/or sublimation from the gauge. The new Pluvio systems measure precipitation continuously; thus, this effect is negligible compared to the earlier gauges collecting precipitation over 24 h. The uncertainty

associated with the use of heating devices built in different types of instruments is, however, estimated to be much smaller than those related to wind effects.

The measurement precision of the DWD gauges is 0.1 mm. With the installation of the Pluvio devices in the early 2000s, automatic readings started to replace the manual readings. However, this upgrade did not affect the reading time for daily precipitation sums, which remained at 07:30 local time. Every month, the daily records are transmitted to the regional prepro-cessing offices for quality checks and thereafter trans-ferred to DWD headquarters at Offenbach, where they are repeatedly quality-checked, undergo correc-tion procedures, and are finally archived. Correction procedures include adjustments of the reading time and accounting for instrument exposition (when possible). The main observational error source is the impact of wind on rain gauge measurements. Although most of the DWD rain gauges are still run

Fig. 1. Maps with the locations of stations used for the production of gridded data for (a) 1951–2008; (b) 1951–2000; and (c) 1971–2000; and (d) time series of changes in the total number of stations (black line, left axis) and the number of gaps (red line, right axis) for different STAMMEX datasets. The total number of stations shown in panel (d) includes only gauges reporting for at least 30 years. The actual total number of gauges (including those reporting for only a few years and which were used for the development of STAMMEX 0.25° dataset based on all available stations) is larger and amounts to 7,561. Horizontal lines mark the numbers of stations for which the number of gaps was less than 10% and which were used for the development of STAMMEX products based on simultaneously reporting stations for different periods.

JULY 2014AMERICAN METEOROLOGICAL SOCIETY |JULY 2014| 997996

by volunteers with gauges being placed in their gar-dens and, thus, to some extent being protected against the wind impact by the surrounding vegetation and buildings, the wind-associated error is present in the daily DWD rain gauge data. Appropriate corrections, which should be based on collocated wind measure-ments, are hardly applicable due to missing wind information for most of the gauge locations. The total error of the Hellmann gauges, due to evaporation loss and wind effects, is estimated to be in the range of ±10% for the DWD network.

STAMMEX GRIDDED PRODUCTS. Data selec-tion. The STAMMEX strategy was to derive different gridded datasets optimized for different purposes. Specifications may range from merging all available observations to using subsampled datasets, in order to guarantee temporal homogeneity of sampling. Some resulting gridded datasets should be exclusively based on stations continuously reporting during the time period for which the respective dataset is built. Products based on the full observational dataset will be suitable for case studies, while products with homogenized sampling will be more applicable to climate variability analyses. Thus, our approach dif-fers from, for example, the ECA (European Climate Assessment) strategy: their E-OBS dataset based on the European collection of rain gauges contains about 500 DWD stations in 1950 and increases to about 800 in the 1990s, with many stations provid-ing data for just a few years. Using all available (and

thus, a time-varying number of) stations may result in time-dependent biases and, consequently, impact estimates of trends and variability. To avoid such inhomogeneity, STAMMEX products were made available for four time periods with spatially varying extent (Table 1, Fig. 1):

1) The Western Germany dataset from 1931 onward has a 0.5° spatial resolution and is based on 471 stations;

2) 1951 onward, with 0.25° and 0.5° spatial resolu-tions (based on all available 7,561 stations) and with 0.5° resolution (based on 888 simultaneously reporting stations);

3) 1951–2000, with 0.25° and 0.5° spatial resolutions based on 2,157 simultaneously reporting stations; and

4) datasets for the best-sampled 30-yr period, 1971–2000, with 0.1°, 0.25°, and 0.5° resolution (3,088 simultaneously reporting stations).

During these periods, the number of gaps in se-lected subsets of reporting rain gauges was less than 3% except for a 2-yr period (1945–46). Supplement A provides information on data coverage for selected periods and the number of stations per grid cell for different spatial resolutions.

Gr idd ing procedures . Spatia l gridding for a l l STAMMEX gridded products has been provided by several gridding procedures. As a reference,

Table 1. Overview of the STAMMEX gridded daily precipitation products: setting of methodologies, periods of availability, and spatial resolution.

GRIDDING METHOD

KrigingGrid-cell averaging

Distance weightingBackground window:

30 daysBackground window:

20 daysBackground window:

10 days

Time period Resolution

1951–2008 0.25° × 0.25°, 0.5° × 0.5°, 7,561 stations, daily means and uncertainties

1931–2008 0.5° × 0.5°, 471 stations (West Germany only), daily means and uncertainties

1951–2008 0.5° × 0.5°, 888 stations, daily means and uncertainties

1951–2000 0.25° × 0.25°, 0.5° × 0.5°, 2,157 stations, daily means and uncertainties

1971–2000 0.1° × 0.1°, 0.25° × 0.25°, 0. 5° × 0.5°, 3,088 stations, daily means and uncertainties

JULY 2014AMERICAN METEOROLOGICAL SOCIETY |JULY 2014| 999998

we used kriging, in which semivariogram parameters, influence radii, and weight-ing coefficients were derived for each individual grid cell and day, accounting for the local arrangement of the available reporting gauges. Kriging was applied to the fractional anomalies around long-term background val-ues. Fractional anomalies were computed by trans-forming actual anomalies to the anomalies of con-tributions of daily values to the totals accumulated during the background pe-riod. In order to derive the background values we first estimated shape and scale parameters of the gamma distribution for all stations in the grid cell. Averages of the grid-cell parameters of the Gamma distribution were used to derive back-ground values for a given time window (e.g., month). Computations were per-formed for the background window varying from 10 to 30 days, with a 30-day window applied for the pro-duction of the reference datasets and the others used for supplementary gridded datasets. Besides kriging, we also used several alternative gridding procedures (e.g., conventional grid-cell aver-aging and distance weight-ing supplemented with a modified method of local procedures for interpolation in fully unsampled cells). This allowed for the development of supplementary gridded products that are equally available for the users and can be used for studying the sensitivity of the method of gridding of long-term tendencies and short term signals. Further details on the gridding procedures are presented in Supplement B.

Uncertainty estimates. All STAMMEX daily precipi-tation grid-cell values are accompanied with uncer-tainty estimates. Besides standard estimates of the kriging uncertainty, we developed error estimates by a bootstrapping algorithm. For each grid cell (0.5°, 0.25°, and 0.1°), we excluded up to half of the randomly selected stations from those falling within

Fig. 2. (a) Long-term annual average of the intensity of very heavy precipita-tion (99th percentile of daily precipitation) for the period 1950–2000 at 0.25° resolution derived from STAMMEX daily climatology; (b) differences in the 99th percentiles of daily precipitation between STAMMEX and E-OBS datas-ets, and smoothed 5-yr running mean time series of the 99th percentile of daily precipitation derived from STAMMEX grids using kriging for 1951–2008 (red), 1951–2000 (brown), and using grid-cell averaging for the period 1951–2008 (blue) as well as from E-OBS grids (green) for (c) the winter season over Brandenburg, Saxony-Anhalt, and Berlin and (d) for the summer period over Baden-Württemberg. Regions for which the time series were derived are shown in inlay maps by blue boxes.

JULY 2014AMERICAN METEOROLOGICAL SOCIETY |JULY 2014| 999998

the impact radius, and then repeated the gridding procedure without these stations. This procedure was applied several times to every grid cell. Statistics of the differences between values derived from all stations and from the randomly generated subsets were used to characterize uncertainties (see Supple-ment C for details). These uncertainty estimates were thereafter used to select an optimal setting of the gridding algorithm, including the choice of semivariogram, background window, and interpola-tion methodology.

Gridded products. Finally, we developed several sets of gridded products for the four periods of data streams with different spatial resolutions (Table 1). Each stream includes daily precipitation grids and uncertainty estimates (derived using kriging with a 30-day background window) as a reference dataset. Furthermore, each stream includes similar daily grids derived using different background windows (grids are provided for 10- and 20-day windows) and two alternative gridding procedures, all of which are accompanied with their uncertainty estimates. For all streams, we also provide files with the number of stations per grid cell. Furthermore, for the grids we also provide supplementary files with blanking fully unsampled grid cells. These additional files can guide users on the actual data coverage in STAM-MEX products. The complete STAMMEX database consists of 27 gridded arrays for the periods; all are publically available.

CLIMATOLOGICAL AND SYNOPTIC APPLICATIONS. Long-term climatology and extremes. New high-resolution STAMMEX daily precipitation grids allow the development of a large set of climatological characteristics of precipitation relevant for regional climate studies, extreme-event statistics, and the validation of regional climate models. The supplementary datasets provided with the STAMMEX daily grid data include monthly and seasonal precipitation totals, intensities, number of wet days, characteristics of heavy (95th percentile) and very heavy (99th percentile) precipitation, as well as characteristics of relative extremeness derived from standard indices, the distribution of fractional con-tribution (DFC), and statistics of wet- and dry-spell durations. As an example, Fig. 2a shows the annual climatology of the intensity of very heavy precipita-tion (99th percentile) for the 1951–2000 period as derived from the STAMMEX 0.25°-resolution prod-

uct. This figure reveals areas with locally high and very heavy precipitation in central western Germany and in the mountain regions of southern Germany with a climatological intensity reaching 50 mm day-1. Over most of the country, the STAMMEX products reveal very heavy precipitation 10%–25% higher than the E-OBS 0.25° grids based on a limited col-lection of stations (Fig. 2b), while climatologies of the mean precipitation are generally consistent in the two products, with somewhat better regionaliza-tion in the STAMMEX product (See Supplement D). Figure 2b systematically identifies higher estimates of very heavy precipitation over Eastern Germany by 4–7 mm day-1 in STAMMEX and the regions in southern Germany where STAMMEX-based esti-mates indicate an annual very heavy precipitation more than 10 mm day-1 higher compared to the E-OBS product (i.e., up to a 25% increase). Although the considerably higher STAMMEX precipitation extremes are persistent in all seasons, the largest dif-ferences (up to 40%) are observed in summer, where only very dense observational networks can capture extremes associated with strong convective processes.

JULY 2014AMERICAN METEOROLOGICAL SOCIETY |JULY 2014| 10011000

Climate variabil i ty in precipitat ion. STAMMEX products allow for detailed regional analyses of the observed climate variability in mean precipitation intensity and in precipitation extremes. Estima-tion of linear trends for the last six decades from STAMMEX daily grids reveals upward trends in cold seasons with the maxima of more than 6% per decade in northern Ger-many, and mostly negative trends in summer, with the strongest downward ten-dencies in central Germany and in Baden-Württemberg, amounting to 5.5%–6% per decade. This seasonality in heavy and extreme pre-cipitation is also evident in the mean intensity and confirms earlier analyses based on individual station data. Spatial structures of linear trends in precipita-tion characteristics derived from STAMMEX grids are more detailed compared to those revealed by E-OBS. In particular, important local features associated with orography are ev i-dent that are not seen in the coarser resolution grids based on a limited collec-tion of rain gauges as well as in reanalyses and satel-lite (e.g., GPCP, TRMM, CMOR PH, PERSIAN N) data. Importantly, the ho-mogenized sampling in the STAMMEX daily products results in both quantita-tive and qualitative differ-ences in trend estimates compared to the E-OBS

grids for the same large regions. Figures 2c and 2d show seasonal time series of the 99th percentile of daily precipitation for central eastern (winter) and southwestern (summer) Germany derived from different STAMMEX products and E-OBS data. There is a very close agreement between the time series and linear trend estimates across different

Fig. 3. (a–d) Accumulated daily precipitation over Germany for the period 9–12 Aug 2002 from STAMMEX daily grids, and daily precipitation for 12 Aug 2002 derived from (e) STAMMEX and (f) E-OBS grids.

JULY 2014AMERICAN METEOROLOGICAL SOCIETY |JULY 2014| 10011000

STAMMEX products based on kriging and grid-cell averages as well as on different subsets of stations (Figs. 2c,d). However, they all reveal both quantita-tive and qualitative differences in the estimates of linear trends compared to E-OBS data. Thus, in southern Germany—which is known for a strong seasonality in trends of extreme precipitation—STAMMEX data show a negative trend of about -0.54 mm day-1 decade-1 (about 12% decrease dur-ing the observational period), while E-OBS reports alternatively a positive trend of about 8% during the last six decades. Importantly, the major discrepancy in the time series results from the first part of the record, where the differences between STAMMEX and E-OBS estimates are approximately twice as large compared to the latest decades. Similarly, in eastern Germany in winter (Fig. 2c), STAMMEX reports trends twice as small compared to E-OBS, with the difference between the area-averaged time series progressively declining from 2 to 0.3 mm day-

1. The most probable cause of the differences in mean precipitation between the two products for the first 20 years of the assessed period [i.e., in the 1951–70 period (Figs. 2c and 2d)] is a changing number of rain gauges used in the E-OBS product.

Synoptic applications and case studies of extreme pre-cipitation. However, the major potential of STAMMEX data is the diagnostics of extreme precipitation events, particularly those leading to disastrous f loods in Central Europe. Figure 3 shows the development of precipitation patterns during the disastrous European floods in August 2002 that resulted in unprecedented economic losses exceeding 16 billion Euros and more than 110 fatalities in Austria, the Czech Republic, and Germany. Accumulated precipitation during the 5-day period amounted to more than 250 mm in many regions of southeastern Germany (especially over the Free State of Saxony) and peaked at more than 300 mm south of the city of Dresden. The city was heavily flooded, with the water level in the Elbe River 9.4 m above normal. STAMMEX grids ac-curately replicate the maximum daily precipitation on 8 December 2002, which amounted to more than 265 mm day-1 in the area of Zinnwald-Georgenfösld (Fig. 3e), where the local rain gauge reported just slightly higher values. Based on a coarser resolution network, the E-OBS precipitation grids show for this date only half of the daily precipitation, which peaked at 133 mm day-1 (Fig. 3f) and was located about 50 km northwest. Here STAMMEX reports nearly 100

mm day-1 higher values. Importantly, the STAMMEX grids also capture the tongues of extreme precipita-tion higher than 80 mm day-1 aligning to the Polish border, which have heavily affected western Poland, while the E-OBS grids report 10–25 mm day-1 less precipitation in this region (Fig. 3f). This example shows that STAMMEX products are extremely use-ful for the validation and further improvement of the regional and local climate estimates and NWP models simulating precipitation and flooding events over other densely populated river valleys, during severe thunderstorms over city areas, and at heads of mountain rivers. They are thus indispensable for accurately monitoring extreme events and impact assessments. Another important realization from analyzing the STAMMEX results for some other regions where dense rain gauge networks similar to those available for Central Europe are absent or are not shared/accessible is that one has to be careful when using coarse resolution grids for assessments of the consequences of extreme rain events. These coarse resolution products may not necessarily ac-curately reproduce individual rainfall clusters typical for the region and event analyzed. In addition, a spa-tial resolution of 3–20 km in a terrain as varied as in Germany seems appropriate to quantify small-scale precipitation events.

AVAILABILITY AND UPDATES. Daily grid-ded data for different time periods, along with the readme files and access codes, are available through the STAMMEX website (www.olgazolina .com/?f ile=kop10.php). In addition, long-term monthly and seasonal precipitation statistics for different periods as well as a dynamic map environ-ment is now under construction at the same site. STAMMEX products currently cover the period until 2008 and will be regularly updated in the future.

ACKNOWLEDGMENTS. We appreciate comments and suggestions of the two anonymous reviewers, which helped to improve the previous version of this manu-script. Discussions with Klaus Peter Koltermann of MSU (Moscow) are appreciated. We thank the Deutsche Forsc-hungsgemeinschaft for funding the STAMMEX Project under the contract PN-50160119 and the Deutsche Wet-terdienst for continuous support of the project and data supply. We also benefited from the support of the Russian Ministry of Education and Science under the contracts 14.B25.31.0026 and 14.577.21.0048, as well as from the support of Université Joseph Fourier and LGGE (Grenoble).

JULY 2014AMERICAN METEOROLOGICAL SOCIETY |JULY 2014| PB1002

FOR FURTHER READINGAbatzoglou, J. T., 2013: Development of gridded surface

meteorological data for ecological applications and modeling. Int. J. Climatol., 33, 121–131, doi:10.1002/joc.3413.

Brienen, S., A. Kapala, H. Maechel, and C. Simmer, 2013: Regional centennial precipitation variability over Germany from extended observation records. Int. J. Climatol., 33, doi:10.1002/joc.3581.

Di Luzio, M., G. L. Johnson, C. Daly, J. K. Eischeid, and J. G. Arnold, 2008: Constructing retrospective gridded daily precipitation and temperature datasets for the conterminous United States. J. Appl. Meteor. Climatol., 47, 475–497.

Giesecke, J., and H. Meyer, 1987: On the automatic as-sessment of rainfall and its evaluation—experience in the Federal Republic of Germany. Water for the Future: Hydrology in Perspective (Proceedings of the Rome Symposium, April 1987). IAHS publ. no. 164.

Groisman, P. Ya., and D. R. Legates, 1994: The accu-racy of United States precipitation data. Bull. Amer. Meteor. Soc., 75, 215–227.

——, R. W. Knight, and T. R. Karl, 2012: Changes in intense precipitation over the Central U.S. J. Hydrometeor., 13, 47–66.

Gutowski, W. J., and Coauthors, 2010: Regional, extreme monthly precipitation simulated by NARCCAP RCMs. J. Hydrometeor., 11, 1373–1379, doi:10.1175/2010JHM1297.1.

Haylock, M., N. Hofstra, A. Klein Tank, E. J. Klok, P. Jones, and M. New, 2008: A European daily high-resolution gridded data set of surface temperature and precipitation for 1950–2006. J. Geophys. Res., 113, D20, doi:10.1029/2008JD010201.

Herrera, S., J. M. Gutierrez, R. Ancell, M. R. Pons, M. D. Frias, and J. Fernandez, 2012: Development and analysis of a 50-year high-resolution daily gridded precipitation dataset over Spain (Spain02). Int. J. Climatol., 32, 74–85.

Higgins, R. W., J. E. Janowiak, and Y.-P. Yao, 1996: A gridded hourly precipitation data base for the United States (1963–1993). NCEP/Climate Prediction Center Atlas 1, National Centers for Environmental Predic-tion, 46 pp.

Hofstra N., M. Haylock, M. New, P. Jones, and C. Frei, 2008: The comparison of six methods for the inter-polation of daily European climate data. J. Geophys. Res., 113, D21 110, doi:10.1029/2008JD010100.

Klok, E. J., and A. M. G. Klein Tank, 2009: Undated and extended European data set of daily climate observa-tions. Int. J. Climatol., 29, 1182–1191.

Maurer, E. P., A. W. Wood, J. C. Adam, D. P. Lettenmaier, and B. Nijssen, 2002: A long-term hydrologically based dataset of land surface f luxes and states for the conterminous United States. J. Climate, 15, 3237–3251.

Menne, M. J., I. Durre, R. S. Vose, B. E. Gleason, and T. G. Houston, 2012: An overview of the Global Historical Climatology Network-Daily database. J. Atmos. Oceanic Technol., 29, 897–910.

Nespor, V., and B. Sevruk, 1999: Estimation of wind-induced error of rainfall gauge measurements using a numerical simulation. J. Atmos. Oceanic Technol., 16, 450–464.

Pryor, S. C., J. A. Howea, and K. E. Kunkel, 2009: How spatially coherent and statistically robust are tempo-ral changes in extreme precipitation in the contigu-ous USA? Int. J. Climatol., 29, 31–45.

Rauscher, S. A., E. Coppola, C. Piani, and F. Giorgi, 2010: Resolution effects on regional climate model simula-tions of seasonal precipitation over Europe. Climate Dyn., 35, 685–711, doi:10.1007/s00382-009-0607-7.

Villarini, G., P. V. Mandapaka, W. F. Krajewski, and R. J. Moore, 2008: Rainfall and sampling uncertainties: A rain gauge perspective. J. Geophys. Res., 113, D11102, doi:10.1029/2007JD009214.

Zolina, O., C. Simmer, A. Kapala, S. Bachner, S. K. Gulev, and H. Maechel, 2008: Seasonality of pre-cipitation extremes over Central Europe during the last 50 years. J. Geophys. Res., 113, D06110, doi:10.1029/2007JD008393.

——, ——, S. K. Gulev, and S. Kollet, 2010: Changing structure of European precipitation: Longer wet periods leasing to stronger extremes. Geophys. Res. Lett., 37, L06704, doi:10.1029/2010GL042468.

——, ——, K. P. Belyaev, S. K. Gulev, and K. P. Koltermann, 2013: Changes in European wet and dry spells over the last decades. J. Climate, 26, 2022–2047, doi:10.1175/JCLI-D-11-00498.1.

U P P L E M E N T

This document is a supplement to “Precipitation Variability and Extremes in Central Europe: New View from STAMMEX Results,” by Olga Zolina, Clemens Simmer, Alice Kapala, Pavel Shabanov, Paul Becker, Hermann Mächel, Sergey Gulev, and Pavel Groisman (Bull. Amer. Meteor. Soc., 95, 995–1002) • ©2014 American Meteorological Society • Corresponding author: Olga Zolina, LGGE, CNRS/UJF-Grenoble 1, 54 rue Molière, BP 96, 38402 Saint Martin d’Hères Cedex, France • E-mail: ozolina@lgge.obs.ujf-grenoble.fr • DOI:10.1175/BAMS-D-12-00134.2

PRECIPITATION VARIABILITY AND EXTREMES

IN CENTRAL EUROPENew View from STAMMEX Results

by Olga ZOlina, Clemens simmer, aliCe Kapala, pavel shabanOv, paul beCKer, hermann mäChel, sergey gulev, and pavel grOisman

Even for the high-density DWD rain gauge network some grid cells remain fully unsampled during some years. Figure S1 shows the locations of fully unsampled grid cells for different periods. The number of these grid cells amounts to 6.8% for the 0.5° spatial resolution (1951–2008), to 7.8% for the 0.25° spatial resolution (1951–2000) and to 30% for the 0.1° spatial resolution (1971–2000) when only simultaneously reporting stations are considered. Most of the fully unsampled grid cells are located in eastern Germany (Fig. S1), for instance 62 of 88 unsampled grid cells for the 0.25° resolution. For the products based on all available stations (0.25° and 0.5°

SUPPLEMENT A: DATA COVERAGE AND THE NUMBER OF UNSAMPLED GRID CELLS FOR DIFFERENT PERIODS.

Editor’s Note: This supplement is composed of four sections. To jump to a specific section, click the appropriate supplement title below.

Supplement A: Data coverage and the number of unsampled grid cells for different periods. Supplement B: Gridding procedures and production of daily precipitation grids

Supplement C: Estimation of uncertainties in gridded daily precipitation products. Supplement D: Comparison with E-OBS grids

spatial resolution) the number of fully unsampled boxes is practically negligible during the period prior to 2000. When the gridded products were produced by grid cell averaging, interpolation into these boxes has been performed using the modified method of local procedures by Akima (1970) (See Supplement B). Figure S2 shows the number of stations per grid cell for 0.25° and 0.5° spatial resolution used for STAMMEX grids produced by grid cell averaging and distance weighting. For 0.5° resolution in most locations the number of stations per grid cell exceeds 10 with only few cells containing less than 5 stations. For 0.25° resolution most boxes have at least 4 stations

ES109JULY 2014AMERICAN METEOROLOGICAL SOCIETY |

(for the product based on all available rain gauges) and at least over western Germany all grid cells contain more than 2 stations (for the product based on simultaneously reporting gauges). Uncertainties of kriging and of grid cell averaging for these grid cells are discussed in Supplement C. For the period

from 1931 to 2008 the STAMMEX gridded products (0.5° resolution) cover only western Germany and are based on 471 simultaneously reporting stations with the number of unsampled grid cells being 13 (Fig. S3). These cells are primarily located along the eastern border of western Germany.

number

Fig. S1. Fully unsampled grid cells (orange boxes) in STAMMEX grids of different resolutions: (a) 0.5°, (b) 0.25°, and (c) 0.1°.

ES110 JULY 2014|

Fig. S2. Number of reporting rain gauges per (a),(b) 0.25° × 0.25° grid cell and (c),(d) 0.5° × 0.5° grid cell for (left) all available stations during 1951–2008 and (right) simultaneously existing stations during 1951–2008.

Fig. S3. Simultaneously reporting stations for the period 1931–2008 (green points) and locations of the fully unsampled cells (orange boxes) for this period.

ES111JULY 2014AMERICAN METEOROLOGICAL SOCIETY |

Several gridding procedures were used for the production of the STAMMEX grids. As a reference we used kriging, which was a lso applied for the production of E-OBS (e.g., Haylock et al. 2008). Kriging was applied to the fractional daily precipitation anomalies computed with respect to the background values equal to precipitation totals during the background window. In contrast to Haylock et al. (2008), where background values were produced with thin-plate splines from climatological monthly totals computed from station values, we first approximated the empirical distribution of daily precipitation values at each station for the background window by a gamma distribution (see, e.g., Wilks 2006; Zolina et al. 2004) and then provided separate grid cell averaging of shape and scale parameters of gamma PDF (Groisman et al. 1999; NOAA 2003). This procedure shows a better performance compared to averaging of precipitation totals, which are frequently computed from incomplete data records adding another source of uncertainty (see, e.g., Haylock et al. 2008). Furthermore, the background windows in STAMMEX products were centered on the analyzed daily field (e.g., for 25 Mar 1971 background was computed for the period from 10 Mar 1971 to 10 Apr 1971). This is different from the use of a background computed for calendar months that may result in uncertainties associated with strong seasonal changes in precipitation regimes, especially in spring and autumn.

Kriging was performed using semivariagrams built for every grid cell and for every day on the ba-sis of varying number of stations. This number was determined from estimation of the decorrelation radius. First all stations within 50-km radius from the center of grid cell were used to estimate the spatial correlation. The latter was estimated on the scale of the background window (e.g., 1 month). Impact of the seasonal cycle onto estimation of the correlation radius was not removed; however, our analysis shows

that even on monthly scales this impact is minor and results in the changes of the number of stations falling into the radius, ranging within a few percent. Then only stations falling into the decorrelation radius corresponding to a correlation of 0.8 were chosen for building of the semivariagram for this grid cell and this particular day. Figure S4 shows the aver-aged number of stations used for the computation for different grids and different periods. The number of stations falling within the decorrelation radius varies from less than 10 to more than 70 with a 10%–15% larger number of stations in winter compared to sum-mer, consistent with a higher decorrelation radius in the winter season. Interannual standard deviations of the number of stations used for computations (Fig. S5) show that the number of stations varies from year to year within 10%–35%.

An alternative procedure of grid cell averaging was applied for daily precipitation values for all grid cells containing at least one rain gauge value. For un-sampled grid cells daily gridded values were provided by spatial interpolation using the modified method of local procedures (Akima 1970), which is based on a piecewise function with slopes at the junction points determined locally by a set of polynomials. This method yields better accuracy and does not result in unrealistic local extrema compared with the alternative procedures (e.g., spline functions). Figure S6 shows long-term differences between climatological precipitation estimates derived using grid cell averaging and kriging for different seasons and for the annual mean as well as RMS differences computed from individual daily values derived from the two methods. Grid cell averaging tends to show slightly smaller values of precipitation with long-term differences within 0.2 mm day–1 in most locations. RMS differences vary from 0.2 to 0.4 mm day–1 in winter to about 1 mm day–1 in summer. Comparisons of the STAMMEX 0.25° grids with the E-OBS dataset are presented in Supplement D.

SUPPLEMENT B: GRIDDING PROCEDURES AND PRODUCTION OF DAILY PRECIPITATION GRIDS

ES112 JULY 2014|

Fig. S4. Number of stations used for computations in kriging products (stations per grid cell) for (a)–(d) the an-nual mean, (e)–(h) January, and (i)–(l) July, and, in columns from left to right, for periods 1951–2008, 1951–2000, 1970–2000 for 0.25° resolution, and 1970–2000 for 0.1° resolution.

Fig. S5. Interannual standard deviations of the number of stations used for computations in kriging products (stations per grid cell) for periods (a) 1951–2008, (b) 1951–2000, (c) 1970–2000 for 0.25° resolution, and (d) 1970–2000 for 0.1° resolution.

ES113JULY 2014AMERICAN METEOROLOGICAL SOCIETY |

SUPPLEMENT C: ESTIMATION OF UNCERTAINTIES IN GRIDDED DAILY PRECIPITATION PRODUCTS.

There are different metrics for estimating the uncertainty of gridded fields (see, e.g., Hofstra et al. 2008). Analysis of skills of interpolation by comparison with single station measurements (e.g., when the location for which the interpolation is performed is excluded from the procedure) is hardly applicable. For precipitation fields derived with either method, grid cell values cannot be directly compared with point measurements since they represent area-averaged values. For the STAMMEX daily grids derived with kriging, we supply two standard metrics of the uncertainty of interpolation, namely the kriging variance (Fig. S7) and the Yamamoto (2000) interpolation variance, which accounts also for the local data dispersion (Fig. S7). Besides these measures, we also estimated the uncertainty based on the random elimination of stations from the kriging procedure. For this purpose for every grid cell we repeated the kriging procedure for the subsets formed after random elimination of up to 50% of stations

Fig. S6. (a)–(c) Long-term differences between climatological precipitation derived using grid cell averaging and kriging as well as (d)–(f) RMS differences computed from individual daily values derived from the two methods for (left) annual mean, (middle) winter, and (right) summer seasons.

falling within impact radius. This bootstrapping approach provided several estimates of daily gridded values for each grid cell based on a reduced subset of stations. Then we estimated the RMS of differences between the reference gridded value and solutions derived from the reduced subsets. These RMS of differences were further used as uncertainty estimates termed variational interpolation uncertainty. Figure S7 shows maps of the uncertainty estimates derived from the bootstrapping approach for 0.25° resolution. Estimates using the Yamamoto (2000) interpolation variance and the variational interpolation uncertainty for 0.1° grids are presented in Fig. S8. Figures S7 and S8 show relative uncertainties in computation of fractional anomalies. The relative kriging variance (RKV) shows generally a uniform distribution with relative uncertainties varying from 2 to more than 5%. The relative Yamomoto interpolation variance (RYIV) varies from 1%–4% in winter to 3%–9% in summer and shows maxima over Rheinland-

ES114 JULY 2014|

Pfalz, Hessen, Thüringen, Sachsen, and southern Brandenburg. The structure of the relative variational interpolation uncertainty (EVIU) is similar to that for RYIV, with slightly higher values of uncertainties, and show maxima of about 10% in summer in Rheinland-Pfalz and southern Brandenburg. Of interest is the analysis of the uncertainties of gridding for the unsampled grid cells compared to the grid cells where observations were present. One can argue that in summer when precipitation is largely provided by convective small scale events uncertainties should be larger for unsampled boxes.

Table S1 compares estimates of three types of errors (mapped in Fig. S7) built for the unsampled cells and the neighboring grid cells containing reporting rain gauges. The differences are quite small and in most cases statistically insignificant. The only statistically significant (at 80% significance level) differences were found for RYIV and RVUV in the summer season; however, for RYIV the difference is negative, implying an even slightly higher level of uncertainty in sampled boxes compared to the unsampled ones.

Using the variational approach we also esti-mated the actual uncertainty of our computations

Fig. S7. Estimates of the (a)–(c) relative kriging variance (RKV), (d)–(f) relative Yamomoto interpolation vari-ance (RYIV), and (g)–(i) relative variational interpolation uncertainty (RVIU) for 0.25° resolution STAMMEX grids for (left) January, (middle) July, and (right) the full year. All estimates are presented in percent of the fractional anomaly.

ES115JULY 2014AMERICAN METEOROLOGICAL SOCIETY |

in mm day–1 (Fig. S9). In winter the maximum uncertainty amounts to 2 mm day–1 in southern Baden-Württemberg. Maximum summer values of up to 3.5–4 mm day–1 are observed in the mountain regions of southern Germany. It is also important to know the extent to that gridding allows for the rep-lication of the structure of wet and dry days which is important for quantifying durations of wet and dry spells (Zolina et al. 2013). Figure S10 shows skill scores

of matching wet days between the gridded value and the neighboring stations contributing to the kriging procedure. In these estimates we assumed the match-ing to be successive if the interpolated grid cell value reported the same day (wet or dry) at more than 50% of stations participating in the interpolation for this grid cell. When the threshold for the identification of wet days is 1 mm day–1, the matching score exceeds 95% in winter and 90% in summer. If we choose the

threshold of 0.1 mm day–1, matching scores are typically higher than 90% in winter and 85% in summer. We also performed an alternative estimate of matching of wet days. In this computation we analyzed for every grid cell the matching scores for the identification of wet days be-tween the interpolated value and all stations contributing in the interpolation for this grid cell. If the number of stations participating in the

Table S1. Differences between the estimates of different uncertainties (RKV, RYIV, RVIU) computed for the unsampled cells and for the neighbor-ing grid cells containing reporting rain gauges. Statistical significance is given according to a Student’s t test.

Season RKVus–RKVs % RYIVus–RYIVs % RVIUus–RVIUs %

Winter (DJF) –0.16 –0.06 –0.11

Summer (JJA) 0.21 –0.76* 0.94*

Year 0.09 0.23 –0.36

* significant at 80% level

** significant at 90% level

*** significant at 95% level

Fig. S8. Estimates of the (a)–(c) relative Yamomoto interpolation variance (RYIV) and (d)–(f) relative varia-tional interpolation uncertainty (RVIU) for 0.1° resolution STAMMEX grids for (left) January, (middle) July, and (right) the full year. All estimates are presented in percent of the fractional anomaly.

ES116 JULY 2014|

interpolation for the grid cell (identifying wet day) equals 10 of which 5 stations report a wet day, the score will be 0.5; if all 10 stations report wet days, the score will be 1. Figure S11 demonstrates histograms of scores for 4 grid cells for the wet day thresholds

1 mm day–1 and 0.1 mm day–1. Clearly, in more than 60%–70% of cases these multiple matching scores exceed 0.85, implying consistency of the identification of wet days across most stations used for computa-tions. Finally, we estimated skill scores of matching

Fig. S9. Absolute uncertainties in daily precipitation (mm day–1) estimated using variational approach for (a) January, (b) July, and (c) the full year.

Fig. S10. Matching scores of the number of wet days in STAMMEX 0.25° grids estimated using a variational approach for (a),(d) January, (b),(e) July, and (c),(f) theannual mean for wet day thresholds of (top) 1 and (bot-tom) 0.1 mm day–1.

ES117JULY 2014AMERICAN METEOROLOGICAL SOCIETY |

wet days, which were suggested by Hofstra et al. (2008), namely percent correct (PC) index and critical success index (CSI). If a is the fraction of hits (both the reconstructed and observed time series report wet day), b is the fraction of false alarms (wet day in the reconstructed time series and dry day in the observed series), c is the fraction of misses (dry day in the re-constructed time series and wet day in the observed series), and d is the fraction of correct rejections (dry day in both reconstructed and observed time series), then the PC index and CSI can be computed as

(S1)

Maps of these two indices (Fig. S12) demonstrate that over most of Germany the PC index exceeds 0.9 (more than 85% of locations) and the CSI index is higher than 0.7 in 98% of locations and exceeds 0.8 at nearly 50% of stations. Note that Hofstra et al. (2008) analyzed the ECA dataset (Klok and Klein Tank 2009) containing just several tens of stations over Germany, and reported a CSI index from 0.6 to 0.7 in most German locations. Note also that in winter, largely associated with stratiform precipitation regimes, estimates of these indices are somewhat higher than those for summer when the occurrence of convective precipitation increases. This is well seen in Fig. S13 showing empirical cumulative distribution functions of the PC index for all German stations. Whereas during the winter season only few percent of stations demonstrate a PC index smaller than 0.85, in summer this fraction of stations amounts to more than 30%.

Fig. S11. Occurrence histograms of multiply matching scores across all stations used for interpolation for the four locations (color boxes in inlay map correspond to the colors of bars) computed for (a),(c) January and (b),(d), July for wet day thresholds of (top) 1 and (bottom) 0.1 mm day–1.

,

.

ES118 JULY 2014|

Fig. S12. Estimates of (a) PC index and (b) CSI computed for annual time series and estimates of PC index computed for (c) winter (DJF) and (d) summer (JJA) .

Fig. S13. Empirical cumulative distribution functions of PC index for all Ger-man stations computed for winter (blue, DJF) and summer (red, JJA) seasons.

ES119JULY 2014AMERICAN METEOROLOGICAL SOCIETY |

SUPPLEMENT D: COMPARISON WITH E-OBS GRIDS

these differences are in agreement with those shown for very heavy precipitation in Fig. 2. Estimates of RMS differences derived from comparison of daily time series in every point (Figs. S14 d–f) imply RMS differences from 0.6–1.2 mm day–1 in winter to more than 2 mm day–1 in summer with again higher values in eastern Germany.

An important question is the origin of these dif-ferences and their potential time-dependent nature implicitly suggested by Figs. 2c and 2d, which show an obvious shift in E-OBS data around 1970. Figure S15 shows differences between the E-OBS and STAM-MEX precipitation for the period before 1970 and for the last two decades. Differences for the period 1951–70 are 2 to 3 times larger compared to those for the last two decades. Remarkable time-dependent differences in extreme precipitation revealed by Fig. 2 can be also explained by the differences in the number of wet days presented in Fig. S16. It shows a considerably higher number of wet days in E-OBS over eastern Germany during 1951–70 and a uniform pattern of differences with a slightly higher number of wet days in STAMMEX for the last two decades.

E-OBS daily precipitation grids (Haylock et al. 2008) cover the period from 1951 onward over the whole of Europe with a 0.25° spatial resolution and use a limited collection of European rain gauges (Klein Tank et al. 2002; Klok and Klein Tank 2009). Regional comparison of STAMMEX products with E-OBS data is important because E-OBS grids are widely used for conventional climate assessments and validation of regional climate model outputs. To ensure comparability for direct comparisons, we used STAMMEX products at the same 0.25° resolution grid. Figure S14 compares (in the same manner as Fig. S6) daily precipitation values in STAMMEX with E-OBS grids (Haylock et al. 2008).

Comparison of long-term climatological precipi-tation shows that STAMMEX values are on average somewhat higher than those of E-OBS with differenc-es in most locations not exceeding 0.2–0.3 mm day–1. The largest differences of up to 1 mm day–1 (about 10% of the mean intensity) are observed in the summer season in southern Bavaria. Climatological differ-ences in eastern Germany are on average also some-what higher than over western Germany. Generally,

Fig. S14. (a)–(c) Long-term differences between climatological precipitation derived from E-OBS and STAM-MEX-KR product for 1951–2008 well as (d)–(f) RMS differences computed from individual daily values derived from the two methods for the (left) annual mean, (middle) winter, and (right) summer seasons.

ES120 JULY 2014|

Fig. S15. Long-term differences between climatological precipitation derived from E-OBS and STAMMEX-KR product for the (a),(b) 1951–70 and (c),(d) 1991–2008 (left) annual mean and (right) summer season.

ES121JULY 2014AMERICAN METEOROLOGICAL SOCIETY |

REFERENCESAkima, H., 1970: A new method of interpolation and

smooth curve fitting based on local procedures. J. Assoc. Comput. Math, 17, 589–600.

Groisman, P. Y., and Coauthors, 1999: Changes in the probability of heavy precipitation: Important indicators of climatic change. Climatic Change, 42, 243–285.

Haylock, M., N. Hofstra, A. Klein-Tank, E. J. Klok, P. Jones, and M. New, 2008: A European daily high-resolution gridded data set of surface temperature and precipitation for 1950–2006. J. Geophys. Res., 113, D20119, doi:10.1029/2008JD010201.

Hofstra, N., M. Haylock, M. New, P. Jones, C. Frei, 2008: The comparison of six methods for the interpolation of daily European climate data. J. Geophys. Res. 113, D21110, doi:10.1029/2008JD010100.

Klein Tank, A. M. G., and Coauthors, 2002: Daily da-taset of 20th-century surface air temperature and precipitation series for the European Climate As-sessment. Int. J. Climatol., 22, 1441–1453.

Fig. S16. Difference in the total annual number of wet days between E-OBS and STAMMEX- products for the periods (a) 1951–70 and (b) 1991–2008.

Klok, E. J., and A. M. G. Klein Tank, 2009: Undated and extended European data set of daily climate observa-tions. Int. J. Climatol., 29, 1182–1191.

NOAA, 2003: Data documentation for Data Set 3721 (DSI-3721) “Gridded US Daily Precipitation and Snowfall Time Series”, March 24, 2003. National Cli-matic Data Center, Asheville, 20 pp. [Available from NOAA National Climatic Data Center, Asheville, NC 28801.]

Wilks, D. S., 2006: Statistical Methods in the Atmospheric Sciences. 2nd ed., Academic Press, 676 pp.

Yamamoto, J. K., 2000: An alternative measure of the reliability of ordinary kriging estimates. Math. Geol., 32, 489–509, doi:10.1023/A:1007577916868.

Zolina, O., A. Kapala, C. Simmer, and S. K. Gulev, 2004: Analysis of extreme precipitation over Europe from different reanalyses: A comparative assessment. Global Planet. Change, 44, 129–161.

Zolina, O., C. Simmer, K. P. Belyaev, S. K. Gulev, and K. P. Koltermann, 2013: Changes in European wet an dry spells over the last decades. J. Climate, 26, 2022–2047, doi:10.1175/JCLI-D-11-00498.1.

ES122 JULY 2014|