- Semi-dry grasslands along a climatic gradient across Central Europe - 835 Journal of Vegetation Science 18: 835-846, 2007 © IAVS; Opulus Press Uppsala. Abstract Question: What is the variation in species composition of Central European semi-dry grasslands? Can we apply a training- and-test validation approach for identifying phytosociological associations which are loristically well deined in a broad geographic comparison; can we separate them from earlier described associations with only a local validity? Location: A 1200 km long transect running along a gradient of increasing continentality from central Germany via Czech Republic, Slovakia, NE Austria, Hungary to NW Romania. Methods: Relevés with > 25% cover of Brachypodium pin- natum and/or Bromus erectus were geographically selected from a larger database. They were randomly split into two data sets, TRAINING and TEST, each with 422 relevés. Cluster analysis was performed for each data set on scores from signiicant principal coordinates. Different partitions of the TRAINING data set were validated on the TEST data set, using a new method based on the comparison of % frequencies of species occurrence in clusters. Clusters were characterized by statistically deined groups of diagnostic species and values of climatic variables. Results: Species composition changed along the NW-SE gra- dient and valid clusters were geographically well separated. Optimal partition level was at 11 clusters, six being valid: two clusters Germany and the Czech Republic corresponded to the Bromion erecti; two clusters from the Czech Republic and Hungary to the Cirsio-Brachypodion, and two clusters were transitional between these two alliances. Conclusion: The training-and-test validation method used in this paper proved to be eficient for discriminating between robust clusters, which are appropriate candidates for inclusion in the national or regional syntaxonomic overviews, and weak clusters, which are speciic to the particular classiication of the given data set. Keywords: Austria; Bromion; Cirsio-Brachypodion; Czech Republic; Germany; Hungary; Phytosociology; Romania; Slovakia; Training and test data sets; Vegetation database. Nomenclature: Ehrendorfer (1973). Semi-dry grasslands along a climatic gradient across Central Europe: Vegetation classiication with validation Illyés, Eszter 1* ; Chytrý, Milan 2,6 ; Botta-Dukát, Zoltán 1,7, ; Jandt, Ute 3 ; Škodová, Iveta 4,8 ; Janišová, Monika 4,9 ; Willner, Wolfgang 5 & Hájek, Ondřej 2,10 1 Institute of Ecology and Botany, Hungarian Academy of Sciences, H-2163 Vácrátót, Hungary; 2 Department of Botany and Zoology, Masaryk University, Kotláská 2, CZ-611 37 Brno, Czech Republic; 3 Institute of Geobotany and Botanical Garden, Am Kirchtor 1, D-06108 Halle, Germany; E-mail ute.jandt@botanik.uni-halle.de; 4 Institute of Botany, Slovak Academy of Sciences, Dúbravská cesta 14, SK-845 23 Bratislava, Slovakia; 5 VINCA – Vienna Institute for Nature Conser- vation and Analyses, Giessergasse 6/7, A-1090 Vienna, Austria; E-mail wolfgang.willner@vinca.at; 6 E-mail chytry@sci. muni.cz; 7 E-mail bdz@botanika.hu; 8 E-mail iveta.skodova@savba.sk; E-mail 9 monika.janisova@savba.sk; 10 E-mail ohajek@sci.muni.cz; * Corresponding author; Fax +36 28360110; E-mail illyese@botanika.hu Introduction The past decade has witnessed a rapid development of electronic phytosociological databases (Ewald 2001; Hennekens & Schaminée 2001), which can be used to create vegetation classiication schemes valid over large areas and across national boundaries. In Europe, this of- fers a unique opportunity for international harmonization of vegetation classiication, habitat typologies and the subsequent planning of conservation strategies. However, vegetation units based on numerical classiications of data from selected areas or selected vegetation types are often not appropriate for direct inclu- sion in large-scale vegetation overviews, because such classiications are highly idiosyncratic. They accurately relect the structure of the input data set but do not use any external information; therefore some clusters are often speciic to the particular classiication but are rarely found in the classiications of other data sets from the same vegetation type. National or international systems of vegetation classiication, however, should be more robust and include only those vegetation units which have been recognized in several independent classiications. Large-scale vegetation classiication projects would be greatly improved if the case studies involving numeri- cal vegetation classiication clearly separated clusters with a more general validity from clusters speciic to the particular data set. One approach to achieve this is simple validation with a training and test data set (Duda et al. 2001), i.e. making a classiication on a training data set, then applying the same classiication method to a different (test) data set, comparing the classiications of the training and test data sets, and inally identifying the corresponding clusters (vegetation types) revealed in both data sets. So far, validation has been rarely used in studies describing vegetation patterns across landscapes, mainly due to the limited amounts of available data, which rarely