Comparison of Cluster Analysis Methods for Identification of Weather Regimes in Euro-Atlantic Region for Winter and Summer Seasons

Abstract

Various methods of cluster analysis are used for identification of large-scale atmospheric circulation regimes or weather regimes (WRs). In this paper we compare four most commonly used clustering methods – k-means (KM), Ward’s hierarchical clustering (HW), Gaussian mixture model (GM) and self-organizing maps (SOM) to analyze WRs in Euro-Atlantic region. The data used for WRs identification are 500 hPa geopotential height fields (z500) from the ERA5 reanalysis for the 1940–2022 period. Four classical wintertime weather regimes are identified by the KM method – two regimes associated with positive and negative phases of the North Atlantic Oscillation (NAO+ and NAO–), a regime associated with the Scandinavian blocking (SB) and a regime characterized by elevated pressure over the Northern Atlantics. For summer months KM method gets WRs that are similar by their spatial structure to the classical winter ones. The SOM method yields results that are almost identical to the results of KM method. Unlike KM and SOM methods, HW and GM do not catch the spatial structure of all four classical winter Euro-Atlantic weather regimes and their summer analogues. Compared to WRs of the KM and SOM methods, WRs obtained by HW and GM methods explain less z500 variance, they have different occurrences, persistence and transition features. Summer and winter WRs obtained by HW and GM methods are less similar to each other compared to WRs provided by KM method. Average spatial correlation coefficients between mean z500 fields of WRs obtained by KM and HW methods are 0.76 in winter and 0.83 in summer, 0.70 in winter and 0.72 in summer for KM and GM methods and 0.41 in winter and 0.44 in summer for the regimes between HW and GM methods, respectively. There are statistically significant trends of seasonal occurrence of WRs found by some of the studied clustering methods – a positive trend for the occurrence of the NAO+ regime and a negative trend for the occurrence of the NAO– regime.