Synoptic‐climatological evaluation of the classifications of atmospheric circulation patterns over Europe

This study evaluates the classifications of atmospheric circulation patterns collected in the COST733 database (COST733cat) in terms of their ability to stratify daily surface temperature and precipitation in 12 domains covering the whole of Europe. The classifications differ in the classification methods used, in the number of types, the variable(s) classified, the number of days in a sequence that are classified and in whether the classification is based on year-round or seasonal data. Several classification methods that perform fairly well are identified; they include a simple k-means clustering, a k-means clustering preceded by hierarchical cluster analysis, Litynski's method, and a classification based on circulation prototypes. On the other hand, there are a couple of classification methods that do not provide a good stratification of temperature and precipitation: orthogonally and obliquely rotated principal component analysis in a T-mode, Lund's correlation method, Kirchhofer's sums-of-squares method, and Erpicum's method. Some methods tend to perform better on large domains, while others tend to perform better on smaller domains; however, the sensitivity of most classification methods to the domain size appears to be small. Several methods exhibit a geographical dependence of their performance, e.g. the method based on circulation prototypes tends to perform better in the northern domains, while Jenkinson–Collison and Erpicum's methods perform better in the southern domains. Classifications of 4-day sequences are usually better in stratifying surface temperature than ordinary instantaneous classifications; the opposite is true for precipitation. Adding a mid-tropospheric variable (500 hPa heights or 1000/500 hPa thickness) to sea level pressure as a classified variable improves the skill of classifications in stratifying temperature.