AN INTERCOMPARISON OF COMPUTER-ASSISTED CIRCULATION CLASSIFICATION METHODS

Five different methods that have been used for classification of circu lation patterns (correlation method, sums-of-squares method, average l inkage, K-means, and rotated principal component analysis) are examine d as to their ability to detect dominant circulation types. The perfor mance of the methods is evaluated according to the degree of meeting t he following demands made on the groups formed: The groups should (i) be consistent when preset parameters are changed (ii) be well separate d both from each other and from the entire data set, (iii) be stable i n space and time, and (iv) reproduce the predefined types. All the met hods proved to be capable of yielding meaningful classifications. None of them can be thought of as the best in all aspects. Which method to use will depend mainly on the aim of the classification. Nevertheless , the principal component analysis is most successful in reproducing t he predefined types and is therefore considered as the most promising method among those examined.