Evaluation and comparison of circulation type classifications for the European domain

Within the COST733 Action Harmonisation and Applications of Weather Types Classifications for European Regions systematic evaluation and comparison studies of circulation type classifications (CTCs) for the North Atlantic European region and 12 embedded smaller sub-domains are performed on the basis of a comprehensive set of commonly used classification methods. In this contribution the performance of CTCs is investigated in terms of the separability and within-type variability of daily circulation types as basic properties of CTCs. Several statistical metrics for assessing these features are calculated for the variable used for classification (mean sea level pressure - MSLP) and as well for associated surface climate variables (2 m temperature - 2mT and precipitation - PREC) using daily gridded ERA-40 reanalysis data for the period from September 1957 to August 2002. Thus the ability of each classification to produce homogeneous well defined classes concerning circulation and as well surface climate variables can be quantified, including respective spatial variations. Based on evaluation results a comparison of the different classification approaches and superordinate basic method groups is performed. As evaluation criteria exhibit distinct sensitivity to the number of circulation types, the presentation of results focuses on a selection of 16 different automatic CTCs, each of them comprising 18 circulation types. Evaluation results averaged over the whole CTC ensemble indicate generally higher performance of CTCs for winter months, for the smaller and more westerly spatial domains, and for MSLP compared to 2mT and PREC. The comparison of evaluation indices estimated for individual CTCs and as well for method groups (based on predefined thresholds, principal component analysis, leader algorithms and optimization algorithms respectively) indicate that there is no overall best CTC or basic method featuring superior overall performance characteristics. Instead distinct differences in relative performance rankings of CTCs become evident between and within different spatial domains, between different seasons and when focusing on different variables. CTCs utilizing optimization algorithms (variants of non-hierarchical cluster analysis) for classification reach highest performance for MSLP in most cases. However this superior performance for MSLP does not necessarily result in comparable high performance for associated surface climate variables 2mT and PREC. Although mostly featuring lower performance for MSLP, CTCs based on principal component analysis or predefined thresholds reach performance values for 2mT and PREC that are comparable to or even higher than those of cluster based CTCs. Thus, results presented in this contribution allow for providing decision support with respect to the application of existing CTCs and the development of custom-designed CTCs rather than for pointing out one universal best CTC. (C) 2010 Elsevier Ltd. All rights reserved.