Over 90% of the global warming in the course of the most recent 50 years has occurred in oceans. Hence, long time series of oceanographic observations are vital for understanding and monitoring changes in the ocean climate.

XBTs (expendable bathy-thermograph) provided the longest continuous observing system, with some XBT lines having now collected more than 30 years of ocean temperature profiles at relatively high spatial and temporal frequencies. However, due to their unstructured sampling in both space and time, along with gaps in data coverage, practical use of XBT data is difficult.

 Here, we use modern machine learning methods, in particular, Gaussian Process regression, to create a new ocean temperature product with regular space and time sampling. We illustrate our approach using a "classic" data set (the "Keeling curve" of atmospheric CO2 collected from the Mauna Loa Observatory) before showing examples of our methodology applied to XBT data in the South Indian Ocean. We will discuss both the mathematical and the practical implementation.

A comparison of this Bayesian regression approach to other sophisticated Machine Learning mechanism is also included in our research.