A Bayesian Neural Network approach to study Dissolved Oxygen  in Southern Ocean water masses 

Oxygen plays a critical role in the health of marine ecosystems. As oceanic O2 concentration decreases to hypoxic levels, marine organisms’ habitability decreases rapidly. However, identifying the physical patterns driving this reduction in dissolved oxygen remains challenging. This study employs a Bayesian Neural Network (BNN) to analyze the uncertainty in dissolved oxygen forecasts. The method’s significance lies in its ability to assess oxygen forecasts’ uncertainty with evolving physical dynamics. The BNN model outperforms traditional linear regression and persistence methods, particularly under changing climate conditions. Our approach leverages three Explainable AI (XAI) techniques—Integrated Gradients, Gradient SHAP, and DeepLIFT—to provide meaningful interpretations of 2- and 8-year forecasts. The XAI analysis reveals that buoyancy frequency and eddy kinetic energy is a critical predictor for short-term forecasts across the North Atlantic Deep Water (NADW), Upper Circumpolar Deep Water (UCDW), masses. While the LCDW variability emphasizes also a role played by advection processes, such as salinity, over short and long timescales.