A machine learning-based rapid water quality estimation method for stormwater outfalls: A case study in Xiamen, China

Urban stormwater outfall pollution is characterized by abrupt onset and drastic fluctuations, making it difficult for conventional monitoring methods to achieve rapid response. Taking the typical stormwater outfall in Xiamen as a case study, this research developed a machine learning-based rapid water quality estimation method. Two ensemble learning models, Random Forest (RF) and eXtreme Gradient Boosting (XGBoost), were compared, and the Harris Hawks Optimization (HHO) algorithm was introduced for automatic hyperparameter tuning. The results showed that both models performed poorly under default hyperparameters, while HHO optimization significantly improved model accuracy. Specifically, the HHO-XGBoost model achieved R² values of 0.905 and 0.866 for NH₃-N and TN, respectively. Feature importance and SHapley Additive exPlanations (SHAP) analyses revealed that HHO optimization enhances model interpretability by suppressing redundant feature interference and strengthening the weights of dominant variables. Furthermore, by comparing HHO with Random Search and Bayesian Optimization, this study confirmed that HHO possesses stronger global search capability in nonlinear hyperparameter spaces. The findings demonstrate that a machine learning framework driven by easily measurable and real-time monitoring indicators can serve as an effective supplement to existing monitoring systems, providing technical support for trend diagnosis and rapid screening of urban stormwater outfall pollution.