1010 / 2024-09-20 03:38:05
Predictability of Southern Ocean Dissolved Oxygen: Bayesian1 vs. Deterministic Approach to Forecasting2
bayesian approach, dissolved oxygen
Session 15 - Ocean deoxygenation: drivers, trends, and biogeochemical-ecosystem impacts
Abstract Accepted
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, identify
ing 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’ certainty with evolving physical dynamics. The BNN model outperforms traditional linear regression and
persistence methods, particularly under changing climate conditions, where it captures increased uncertainty, as quantified by Bayesian entropy. 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 is a critical predictor for short-term forecasts across the North Atlantic Deep Water (NADW), Upper Circumpolar Deep Water (UCDW), and Lower Circumpolar Deep Water (LCDW) masses while
mixing processes and salinity become more influential over longer timescales.
decreases to hypoxic levels, marine organisms’ habitability decreases rapidly. However, identify
ing 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’ certainty with evolving physical dynamics. The BNN model outperforms traditional linear regression and
persistence methods, particularly under changing climate conditions, where it captures increased uncertainty, as quantified by Bayesian entropy. 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 is a critical predictor for short-term forecasts across the North Atlantic Deep Water (NADW), Upper Circumpolar Deep Water (UCDW), and Lower Circumpolar Deep Water (LCDW) masses while
mixing processes and salinity become more influential over longer timescales.