<p>Climate-driven changes in precipitation, temperature, and runoff are intensifying nutrient pollution hazards in watersheds, particularly in urban-rural transitional zones characterized by fragmented land use and sparse monitoring. Predicting these climate-sensitive nonpoint source pollution (NPSP) dynamics is difficult because multisource inputs and rapidly evolving spatial patterns challenge conventional models. This study developed an artificial intelligence (AI) framework that integrates machine learning with remote sensing data assimilation to assess nutrient pollution in rapidly urbanizing basins. A random forest regression model, trained on 15 GIS-based predictors, serves as the core predictor for total nitrogen (TN) and total phosphorus (TP) concentrations. An ensemble Kalman filter (EnKF) assimilates remote sensing inversion (RSI) products into the model for dynamic updates. Applied to the Wenruitang River watershed in eastern China, the framework reduces TP prediction errors by up to 52.9% compared with standalone models and better captures short-term concentration fluctuations relevant for hazard assessment. Shapley additive explanation (SHAP) analysis reveals that socioeconomic and meteorological variables dominate temporal variability, while land-use indicators, specifically construction land and water area, serve as stable spatial drivers of nutrient hazard. Receiver operating characteristic (ROC)-based land-use thresholds identify critical source zones with &gt; 71% construction land or &lt; 5% water area. Although these zones occupy only 26% of the basin, they contain 60% of high-TN hotspot grids and 48% of high-TP hotspot grids. The findings demonstrate that remote sensing-assisted AI can effectively identify nutrient pollution hotspots and offer a transferable, interpretable tool for climate-resilient water quality management and hazard-informed watershed governance.</p>