Improved estimation of oceanographic climatology using empirical orthogonal functions and clustering

Vertical profiles of temperature, salinity, and sound speed velocity are used in numerous applications where accurate vertical profiles are crucial. Conventional climatological representations of vertical oceanographic profiles are based on mean or median profiles of historic data in a rectangular area containing the position in question. In areas containing oceanographic fronts mean profiles may not be representative for the profiles in the area and may even be unphysical. We propose a different approach to generate more realistic climatological estimates of the vertical profiles at a given time and position. The depth-dependent behaviours of all historic temperature and salinity profiles are classified by combining Empirical Orthogonal Function (EOF) analysis with K-means clustering. All profiles with similar EOF-coefficients are sorted into a single cluster and averaged to find a representative profile for that cluster. The geographical extent and temporal validity of the cluster are given by the positions and measurement times of the contained profiles. The method is here illustrated using ARGO temperature profiles from the North Atlantic from 2001 to 2012. The proposed method automatically allocates a high density of clusters in areas with large oceanographic variability, such as areas with oceanographic fronts. On the eastern coast of North America cold water from the Labrador Sea runs southwards between the coastline and the warmer Gulf Stream running northeast, resulting in strong fronts. The depth-dependent behaviour of an average profile from all profiles contained in a rectangular, geographic window may differ strongly from the present oceanographic profiles. The profiles representing the nearby clusters, on the other hand, better represent the general depth-dependent behaviour of the profiles in this region.