Interpretable machine learning to forecast hypoxia in a lagoon

Abstract Dissolved oxygen is a key indicator in aquatic ecosystems, reflecting changes in water quality. Low levels of dissolved oxygen may lead to oxygen depletion (called hypoxia), putting at high risk the survival of aquatic organisms. Identifying the environmental conditions inducing hypoxia is therefore a topic of high ecological importance. In this study, we used four machine learning algorithms (Extreme Gradient Boosting (XGBoost), Extremely Randomized Trees (EXT), Random Forest, and Logistic Regression) to forecast hypoxia in a lagoon, considering different time lags (2,5,10 and 20-days). To do so, we used data on dissolved oxygen and a total of nine physicochemical and meteorological variables from Papas lagoon, Greece during 2015–2018. Key drivers and synergies that increase the risk of hypoxia were identified using the Shapley Additive exPlanations (SHAP) methodology. EXT was slightly superior to the other algorithms in forecasting hypoxia, achieving a success between 89% and 94% with pH, water temperature, chlorophyll and salinity as top explanatory variables. SHAP showed that the synergistic effect of low pH and chlorophyll, and elevated water temperature and salinity tended to favor conditions leading to hypoxia. SHAP also illustrated that diverse synergies of the explanatory variables can induce hypoxia, indicating the complex and nonlinear relationships between environmental factors and hypoxia. Overall, the present approach may be proved useful for the development of a reliable forecasting tool for alarming hypoxia and, ultimately, the effective monitoring of the lagoon.