A Hybrid Physics-Machine Learning Modeling Framework Enhances Nonpoint Source Pollution Forecasting

Timely and accurate predictions of nonpoint source (NPS) pollution are critical for the protection and management of the water environment. However, the inability to obtain real-time runoff measurements and the scarcity of fertilizer application data often limit the model accuracy and timeliness. In this study, we developed a hybrid modeling framework that integrates a process-based HYPE model with a support vector machine (SVM) algorithm to improve the prediction of total nitrogen (TN) concentrations at a daily scale in a typical agricultural subwatershed in Guangdong Province, China. A precipitation-dependent hypothetical fertilizer application scenario was introduced to enhance the simulation realism. The results indicated that including runoff as an input factor notably improved model performance, with the best results achieved when using a 3 day cumulative flow as input. Compared with the original HYPE model, the proposed HYPE-SVM model achieved substantially better performance (NSE = 0.6501, KGE = 0.6931, RMSE = 5.7664 mg/L, WI = 0.8795, and PBIAS = – 8.54%), indicating both improved accuracy and reduced systematic bias. Overall, the HYPE-SVM model outperforms the physical model and offers a promising approach for the prediction of NPS pollution in data-scarce areas.