Award: OCE-2123546

The ocean is essential for sustaining biodiversity and human livelihood, health, and security. Key ecosystem services provided by the global ocean include climate regulation and biogeochemical cycling of carbon, nutrients, and dissolved gases such as oxygen. On the global scale, the observed oxygen decline is negatively correlated with ocean heat content changes, where the ratio of oxygen to heat changes is greater than that expected from the temperature-solubility relationship for oxygen, indicative of additional effects from non-solubility processes. It is difficult to reconstruct long term trends of dissolved oxygen accurately due to sparse and irregular sampling patterns in existing observations. This project aims to reconstruct dissolved oxygen concentration maps by application of machine learning. First, we train algorithms with oxygen measurements from ship and float platforms since 1965. A small fraction of the data is withheld from training and are used for validation. Secondly, the trained algorithms are tested against the validation data to measure its uncertainties. Third, the algorithm is run in the inference mode to produce time-varying, 3-dimensional maps, which are further validated against non-machine learning methods (e.g. optimal interpolation). Finally, the validated dataset is released in the public domain for the scientific community and any other interested parties. We also release the source code such that our work is reproducible by third parties. Since the initial publication of the optimal interpolation and machine learning based dataset, our data products have been downloaded over 1500 times (about 400 downloads of optimal interpolation data and 1100 downloads of machine-learning based maps). This project produced 4 peer-review publications. This project also provided financial support for the training of early career scientists. Two PhD students are trained under this project at Georgia Institute of Technology. Furthermore, several undergraduate students are involved and provided research experiences in ocean sciences and application of machine learning in large oceanographic dataset.

Last Modified: 08/16/2026
Modified by: Takamitsu Ito