Enhancing tree-based machine learning for chlorophyll-a prediction in coastal seawater through spatiotemporal feature...

The excessive growth of phytoplankton in water can deplete oxygen, release toxins, harm aquatic life, cause economic losses, and threaten coastal residents. Accurately predicting phytoplankton levels is crucial for safeguarding marine life and coastal communities. Taking Hong Kong as an example, this study aims to identify the best machine learning approach among 5 tree-based models in predicting chlorophyll-a levels in coastal waters; improve model performance by considering spatiotemporal features; advance model interpretation through applying the SHapley Additive exPlanations (SHAP) approach. The findings show that LightGBM performs the best in prediction. Adding spatiotemporal features improves the model accuracy from R²: 0.492 to R²: 0.702, while temporal features have a more significant contribution than spatial features. Five-day biochemical oxygen demand (BOD5), total phosphate, and wind (both speed and direction) are highlighted as the top three factors that influence chlorophyll-a concentration. This underscores the importance of careful management in controlling anthropogenic organic matter and nutrient input, while wind-driven water mixing may redistribute phytoplankton. These findings provide a scientific basis to enhance environmental management measures and ensure the sustainability of coastal ecosystems.