An assessment of climatological synoptic typing by principal component analysis and kmeans clustering

Abstract

A common method of automated synoptic typing for climatological investigations involves data reduction by principal component analysis followed by the application of a clustering method. The number of eigenvectors kept in the principal component analysis is usually determined by a threshold value of relative variance retained, typically 85% to 95%, under the implicit assumption that varying this relative variance will not affect the resultant synoptic catalogue. This assumption is tested using daily 500-mb geopotential heights over northwest Canada during the winter period (December to February) from 1948 to 2006. Results show that the synoptic catalogue and associated surface climatological characteristics undergo changes for values of relative variance retained over 99%, indicating the typical thresholds are too low and calling into question the validity of performing principal component analysis prior to objective clustering.