Spatiotemporal-aware machine learning approaches for dissolved oxygen prediction in coastal waters

Area of Science:

• Environmental Science
• Marine Biology
• Data Science

Background:

• Coastal hypoxia poses a significant threat to marine ecosystems and fisheries.
• Accurate monitoring and prediction of hypoxia are crucial for effective management.

Purpose of the Study:

• To develop and evaluate machine learning models for hypoxia monitoring in coastal waters.
• To investigate the impact of spatiotemporal factors on dissolved oxygen (DO) prediction.
• To interpret model predictions using SHapley Additive exPlanations (SHAP) to understand hypoxia drivers.

Main Methods:

• Utilized a long-term climate and marine monitoring dataset from Tolo Harbour and Mirs Bay, Hong Kong.
• Compared four tree-based machine learning models, including LightBoost, for DO concentration prediction.
• Incorporated spatiotemporal effects and applied SHAP for model interpretability.

Main Results:

• The LightBoost model demonstrated the highest effectiveness in predicting DO concentrations.
• Considering spatiotemporal effects significantly improved prediction accuracy (R² increased by 0.30 in Zone 1 and 0.68 in Zone 2).
• Hydrodynamics were identified as a primary driver of hypoxia, with anthropogenic activities and silicate levels (potentially from groundwater discharge) also playing significant roles.

Conclusions:

• Machine learning, particularly with spatiotemporal considerations, offers a powerful approach for hypoxia prediction.
• Understanding the interplay of hydrodynamics, anthropogenic impacts, and nutrient sources like silicate is key to managing coastal hypoxia.
• This research lays the groundwork for a near-future forecasting tool for hypoxia events.