This study proposes a method for bathythermal analysis applied to the temperature profiles collected in the Bay of Biscay since 1950 and recorded at the French Hydrographic and Oceanographic Service of the Navy (SHOM). We consider the ocean variability, historically sampled with temperature measurements, from a geostatistical point of view. Thanks to an appropriate model of the thermal structures and using the "kriging" method, we proceed to monthly mean optimal estimations on a regular gridnet. For this purpose, we choose a random function model able to take several scales of phenomena into account. In practice, a structural analysis is first performed from the monthly files by computing the experimental variogram. Secondly, we exclude the incoherent values. Following this filtering process, a second variogram is established and, thus, the monthly mean situations are evaluated using the "kriging" method.