Process-Driven Cross-Factor Integration of SWAT and Interpretable Machine Learning for Enhancing Riverine Total Nitrogen Prediction

Nitrogen runoff from industrial wastewater and agricultural fertilization has intensified eutrophication in aquatic ecosystems, yet accurate daily total nitrogen (TN) simulations remain challenging due to sparse and discontinuous monitoring data. This study develops a process-informed cross-factor SWAT-ML framework that integrates the Soil and Water Assessment Tool (SWAT) with machine learning (ML) to predict daily TN concentrations in the Yihe River Basin. To address low TN simulation accuracy and limitations in monitoring data, this framework uses SWAT-simulated nontarget process variables that are physically relevant and more robust than direct SWAT TN outputs as ML inputs. After hyperparameter optimization, the SWAT-ML framework exhibited a notable performance improvement. Among the cross-factor strategies, the SWAT(TP)-LSTM configuration consistently performed best (NSE = 0.94, RMSE = 0.60 mg/L) by more effectively transferring process-related information into the learning framework. SHAP interpretation indicates that hydrological transport and soil-water interactions dominate TN dynamics, reflecting the shared runoff- and erosion-controlled pathways linking phosphorus and nitrogen export. SWAT(TP)-LSTM captures both rapid meteorological responses and delayed hydrological memory, whereas TN-driven models are more vulnerable to biogeochemical uncertainty. Overall, this interpretable framework enhances predictive robustness and provides mechanistic insights into watershed-scale nitrogen dynamics.