Prediction of groundwater total nitrogen via an interpretable ensemble machine learning framework: Implications for groundwater diversion management in complex catchments.

Groundwater total nitrogen (TN) is a key indicator of groundwater ecological security, and accurate prediction and driver identification are essential for targeted nitrogen control and watershed management. To address the limitations of single models and better capture complex interactions, this study investigated the Hangbu River Basin, a major tributary of Lake Chaohu. Groundwater TN data, together with information on multiple environmental variables related to climate, topography, land use, vegetation, and soil properties, were collected from 59 monitoring wells during six campaigns (Nov 2024-Sep 2025). Multiscale buffer analysis (500-3500 m) was conducted to identify the optimal spatial scale, followed by developing an interpretable ensemble machine learning framework integrating weighted voting and stacking strategies. The results show that a 3000 m buffer is the optimal prediction scale. Ensemble learning improved performance: voting achieved the highest accuracy (R2 = 0.87), while stacking had the lowest errors (RMSE = 1.11, MAE = 0.76); F tests (p < 0.05) confirmed the robustness of these improvements. SHAP analysis indicated that groundwater TN spatial variability is primarily associated with nonlinear interactions among landscape configuration (CIRCLE_MN), topography (DEM), and precipitation. Recursive feature elimination based on Spearman correlation ranking demonstrated that the top nine predictors capture the essential information governing groundwater TN variability (R2 = 0.81). The proposed ensemble learning framework improves the understanding of groundwater nitrogen dynamics in complex watersheds and provides practical methodological support for precise groundwater nitrogen pollution control.