Integrating landscape "source-transport-sink" mechanisms into GeoAI to enhance surrogate modeling of watershed nutrient loads.

Accurate prediction and effective management of watershed nitrogen (N) and phosphorus (P) loads remain critical challenges in controlling non-point source (NPS) pollution. Most surrogate models rely primarily on land-use variables, overlooking the influence of landscape structural and functional transformations triggered by land-use change on nutrient load dynamics, limiting the interpretability and generalization capacity of surrogate models. To address this gap, we propose a novel geospatial artificial intelligence (GeoAI)-based surrogate model framework that incorporates the "source-transport-sink" mechanism. In this framework, source-sink landscape types, hydrological connectivity (IC), and landscape fragmentation index were introduced as proxy variables representing the "source-transport-sink" mechanism. Pearson correlation analysis, geographic detectors, and PLS-SEM confirmed their relevance, interaction effects, and mechanistic pathways in explaining watershed nutrient loads. Sub-basin TN and TP load data for 2000-2022, simulated by the coupled Export Coefficient Model (ECM) and Soil and Water Assessment Tool (SWAT) model (R² > 0.9), served as training samples to compare traditional machine learning models (XGBoost, CatBoost, ANN) with GeoAI models (GeoXGBoost, GNNWR, GTNNWR) for watershed NPS simulation. Among these models, the GTNNWR model demonstrated superior generalization and overfitting control, achieving R² values of 0.698 for TN and 0.755 for TP in the test set. Coupling the intPLUS model with GTNNWR enabled projections of TN and TP load changes in the Dongjiang River Basin (DRB) by 2030 under three scenarios: economic development, ecological protection, and natural development. The ecological protection scenario markedly slows the increase in TN and TP across the DRB. Risk maps indicate that TP hotspots are concentrated in the middle basin, whereas TN risk is more widespread across the middle and lower basins, with the largest increases occurring around Dongguan and Shenzhen. Overall, the GeoAI surrogate, grounded in the "source-transport-sink" mechanism, provides a well-founded and transferable tool with broad applicability for simulating and managing watershed NPS pollution risk.