Unsupervised classification of the northwestern European seas based on satellite altimetry data

Abstract

From generating metrics representative of a wide region to saving costs by reducing the density of an observational network, the reasons to split the ocean into distinct regions are many. Traditionally, this has been done somewhat arbitrarily using the bathymetry and potentially some artificial latitude–longitude boundaries. We use an ensemble of Gaussian mixture models (GMMs, unsupervised classification) to separate the complex northwestern European coastal region into classes based on sea level variability observed by satellite altimetry. To reduce the dimensionality of the data, we perform a principal component analysis on 27 years of observations and use the spatial components as input for the GMM. The number of classes or mixture components is determined by locating the maximum of the silhouette score and by testing several models. We use an ensemble approach to increase the robustness of the classification and to allow the separation into more regions than a single GMM can achieve. We also vary the number of empirical orthogonal function (EOF) maps and show that more EOFs result in a more detailed classification. With three EOFs, the area is classified into four distinct regions delimited mainly by bathymetry. Adding more EOFs results in further subdivisions that resemble oceanic fronts. To achieve a more detailed separation, we use a model focused on smaller regions, specifically the Baltic Sea, North Sea, and the Norwegian Sea.