Predicting Summer River Hypoxia from DOM Fluorescence Signatures Using Interpretable Machine Learning and Kinetic Validation

Summer hypoxia increasingly occurs in subtropical urban rivers even where conventional pollutants are effectively controlled, suggesting overlooked drivers of oxygen loss. We integrate approximately 10 years of monitoring data from rivers in southeastern China (666 EEM fluorescence records) with interpretable machine learning (XGBoost-SHAP), 432 incubation assays, and Streeter–Phelps-based kinetic analysis to identify controls on dissolved oxygen (DO) decline. The models show strong predictive performance and consistently identify temperature and a tryptophan-like DOM component (C1) as dominant predictors of DO variability. SHAP interaction analysis indicates that the influence of C1 on DO becomes stronger under warm conditions. Incubation experiments further demonstrate rapid oxygen consumption associated with protein-like DOM, supporting its role as a highly bioavailable microbial substrate. Incorporating C1 into the Streeter–Phelps framework substantially improves model performance ( R 2 ≈ 0.85–0.98), linking fluorescence-resolved DOM to DO dynamics. These results suggest that bioavailable protein-like DOM, likely associated with diffuse sewage inputs, plays a key role in summer oxygen depletion. Monitoring and controlling bioavailable DOM may therefore improve hypoxia mitigation and urban river management.