Interpretable machine learning reveals regional heterogeneity and key environmental drivers of BOD₅ in complex river...

Accurate prediction of Biochemical oxygen demand (BOD₅) can substantially reduce the labor-intensive monitoring required to assess organic pollution in surface water, yet it remains challenging in regions subject to strong anthropogenic disruptions. This study investigated typical river networks within the Taihu Lake Basin, China, by integrating six-year water quality monitoring data with multiple influencing factors, including meteorological parameters, socio-economic indicators and channel characteristics. Both unsupervised and supervised machine learning approaches were applied, and SHapley Additive exPlanations (SHAP) were utilized to quantify the contribution of each factor to BOD₅ variation. Regional classification using k-means clustering, coupled with the integration of multiple environmental factors integration substantially enhances model performance. The Categorical Boosting (CatBoost) model demonstrated the best performance within this dense river-network system characterized by weak hydrodynamics and widespread hydraulic infrastructure. Water quality parameters, mainly the permanganate index and nitrogen species, accounted for the largest share of explained variance (42.31-50.60%), while meteorological factors (especially wind speed and water temperature) exhibited pronounced nonlinear effects and together contributed more than 30% across all regional types. In urbanized areas, channel characteristics such as pumping stations moderated the influence of meteorological and other water quality indicators, with their contribution increasing to 13.55%, whereas in agricultural regions, diffuse inputs and elevated nitrate concentrations were strongly associated with higher BOD₅ levels. This study underscores the uncertainty inherent in water quality prediction under a changing climate and demonstrates that machine learning-based modeling provides a quantifiable and interpretable data-driven framework for understanding BOD₅ dynamics.