Project: Mapping Dissolved Oxygen using Observations and Machine Learning

NSF Award Abstract:
Oxygen is produced by algae in the sunlit surface waters and is released into the atmosphere. This process contributes to about the half of atmospheric oxygen. However, there is a growing consensus in the scientific community that the global ocean oxygen inventory has declined in recent decades. Ocean heat uptake causes the reduction of solubility, and changes in circulation and biogeochemical processes associated with the ocean warming can further change ocean oxygen content. The reduction of dissolved oxygen can have far-reaching impacts on the marine habitats. Recent estimates of the global oxygen decline are in the range of 0.5-3.3% over the period of 1970- 2010. Distribution of the historical O2 measurements is irregular and sparse, causing significant uncertainty in these estimates. The objective of this project is to determine changes in the dissolved oxygen content of the oceans based on observational data and machine learning techniques. The overarching hypothesis of this project is that there are significant, regional relationships between O2 and other observed quantities. Dissolved oxygen is ultimately controlled by the combination of ocean circulation, air-sea gas transfer and biological processes. These processes can be linked with other observed quantities such as temperature (T) and salinity (S), but such relationships can be complex and non-linear. Therefore, it is difficult to determine a universal relationship that governs the distribution of O2 based on the first principle. However, machine learning algorithms can extract empirical relationships between O2 and other variables from existing observations, allowing us to estimate O2 where direct observation is not available. The work will also support one graduate and one undergraduate student research and outreach activities at local events.

In this project, machine learning will be used to fill data gaps in the historical O2 dataset and to generate an improved, gridded estimates of O2 from 1960 to present. This approach takes advantage of the large amount of accumulated in-situ observations over multiple decades including not only O2 itself but also other related variables such as T and S. The proposed work revolves around three hypotheses. First, the current estimates of global O2 trend and variability are strongly influenced by relatively data-rich regions such as North Atlantic and North Pacific. Machine-learning based O2 dataset with an improved gap-fill approaches is hypothesized to better represent relatively data-poor regions such as tropics and southern hemisphere oceans. Secondly, the current estimates indicate that less than half of O2 decline is explained by the solubility effect. The global O2-heat relationship measures the reinforcing effects of ocean ventilation and biogeochemistry. Machine learning can estimate empirical relationships between O2, T and other physical variables, which can be manipulated to perform sensitivity experiments. The empirical model of O2 can constrain the regional and global O2-heat relationship. Thirdly, it is hypothesized that observed O2 decline in the tropical thermocline are driven by the combination of natural climate variability and long-term trends. In the proposed work, sensitivity experiments are performed with the empirical model of O2 to evaluate the influences of long-term trends and decadal-scale changes associated with the modes of natural climate variability.

This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.