The Southern Ocean connects the Indian, Pacific and Atlantic Oceans and provides a direct pathway to exchange mass, heat and salt across the Global Ocean, therefore playing an important role in the global climate system. Due to the complexity of its structure and the general inadequacy of its sampling, both in time and space, it remains a challenge to describe and visualize the three dimensional pattern of its circulation and the associated tracer distribution (temperature, salinity, oxygen or nutrients). This thesis contributes to the understanding of the thermohaline structure of the ocean and especially of the remote Southern Ocean by introducing a novel decomposition method, the Functional Principal Component Analysis applied on vertical profiles of temperature and salinity. To this end, we first normalize hydrographic profiles by using a functional spline representation. Then the statistical method of dimension reduction and feature extraction reveals the main spatial patterns of the temperature and salinity variations. The first two vertical modes contribute to 90% of the combined variance and are related to very robust structures of the Global Ocean. The first mode is mainly controlled by temperature and the second by salinity. In the Southern Ocean, the vertical modes present circumpolar patterns that can be closely related to the stratification regimes that define the circumpolar fronts. Notably the Polar Front is located at the natural boundary between the region controlled by the first (thermal) mode to the north and the second (haline) mode to the south. A mapping of the fundamental zonation is provided with an estimate of the width of the water mass boundaries. As a validation of this method, the Antarctic Polar Front is investigated further in the Indian sector using the same statistical framework. We show that the Polar Front latitudinal position varies seasonally upstream of the Kerguelen Plateau. This meandering is confirmed by hydrographic data gathered by elephant seals equipped with miniaturized sensors. The proposed statistical method provides an objective way to define water mass boundaries and their spatial variability. It offers a useful framework for representing the density structure of the ocean in a reduced-dimension space while maximizing the variance explained. The functional approach also provides a robust way to validate model outputs against observations from any platforms.