Explainable machine learning reveals spatially heterogeneous BOD–COD relationships in a large river basin to inform targeted monitoring strategies

Abstract

Accurate monitoring of biochemical oxygen demand (BOD) is essential for assessing river health, but its measurement is labor-intensive and time-consuming. In contrast, chemical oxygen demand (COD) is routinely and automatically monitored in China's national network, yet it does not fully capture ecological stress. To address this limitation, we developed a machine learning framework to predict hard-to-measure BOD dynamics using frequently measured COD together with other environmental variables. Monthly observations from 135 monitoring sites across the Yellow River Basin during 2016–2022 were integrated with water quality parameters, catchment attributes, and anthropogenic emission indicators to construct a stacking model. The ensemble combined five base learners, including random forest, svmRadial, xgbTree, Cubist, and gbm. The stacking model achieved satisfactory predictive performance, with an R² of 0.78, representing an improvement of approximately 5.4% to 15% compared with the individual base models. SHAP analysis identified COD as the dominant predictor of BOD, exerting a strong positive effect. The strength of the BOD–COD coupling varied spatially, being stronger in urbanized watersheds with intensive domestic emissions and weaker in relatively natural, less-disturbed basins. Structural equation modeling further indicated that the spatial heterogeneity of human activities amplifies regional differentiation in the BOD-COD relationship, a pattern likely mediated by basin-specific dissolved organic matter characteristics. These findings support a differentiated monitoring strategy, whereby model predictions can complement BOD measurements in human-impacted basins, while higher sampling density is warranted in natural systems. Overall, the proposed framework enables optimized resource allocation and targeted monitoring of organic pollution in large, heterogeneous river basins.