Assessing the effectiveness of non-point source pollution models in data-limited urban areas

Non-point source (NPS) pollution from stormwater runoff has become a major threat to urban water bodies. Rapid and reliable pollution profiling is essential for effective mitigation, yet early-stage stormwater management often lacks detailed drainage data and long-term monitoring, complicating model selection. This study evaluates the performance and practical utility of three widely used NPS modeling approaches—statistical regression, machine learning, and physical process-based models—using a large-scale field monitoring dataset. Improved Export Coefficient Method models achieved high accuracy for TN and COD (R² > 0.7) but showed overfitting risks due to collinearity. Random Forest Regression predicted COD, TN, NH₃-N, and TP well (R² > 0.6) but struggled with predicting TSS loads. In contrast, SWMM models failed to deliver reliable predictions, even after auto-calibration, underscoring their limitations without prior user expertise. Factor contribution analysis highlighted antecedent dry period, rainfall depth, and land use as key predictors. Nitrogen-related pollutants were more influenced by dry deposition, while phosphorus was more affected by rainfall-triggered wash-off. Finally, a practical multi-criteria evaluation framework, considering accuracy, generalizability, robustness, and cost-efficiency, is proposed to guide model selection under data-limited conditions. This study is expected to promote the utility of machine learning models in practice and provide theoretical support for NPS pollution mitigation in urban areas.