WEATHER REGIMES - RECURRENCE AND QUASI STATIONARITY

Two different definitions of midlatitude weather regimes are compared. The first seeks recurrent atmospheric patterns. The second seeks quas i-stationary patterns, whose average tendency vanishes. Recurrent patt erns are identified by cluster analysis, and quasi-stationary patterns are identified by solving a nonlinear equilibration equation. Both me thods are applied on the same dataset: the NMC final analyses of 700-h Pa geopotential heights covering 44 winters. The analysis is performed separately over the Atlantic and Pacific sectors. The two methods giv e the same number of weather regimes-four over the Atlantic sector and three over the Pacific sector. However, the patterns differ significa ntly. The investigation of the tendency, or drift, of the clusters sho ws that recurrent Bows have a systematic slow evolution, explaining th is difference. The patterns are in agreement with the ones obtained fr om previous studies, but their number differs. The cluster analysis al gorithm used here is a partitioning algorithm, which agglomerates data around randomly chosen seeds and iteratively finds the partition that minimizes the variance within clusters, given a prescribed number of clusters. The authors develop a classifiability index, based on the co rrelation between the cluster centroids obtained from different initia l pullings. By comparing the classifiability index of observations wit h that obtained from a multivariate noise model, an objective definiti on of the; number of clusters present in the data is given. Although t he classifiability index is maximal by prescribing two clusters in bot h sectors, it only differs significantly from that obtained with the n oise model using four Atlantic clusters and three Pacific clusters. Th e partitioning clustering method turns out to give more statistically stable clusters than hierarchical clustering schemes.