Identification of spatially heterogeneous drivers of lake eutrophication in China with explainable automated machine learning

Abstract Lake eutrophication poses a pressing global environmental challenge, severely degrading freshwater ecosystems. This study presents an explainable automated machine learning (XAutoML) framework to identify and quantify the spatially heterogeneous drivers of lake eutrophication across China. AutoGluon was employed for robust predictive modeling and KernelSHAP for explainability, particularly focusing on the ecological divide represented by the Hu Huanyong Line (Hu Line). AutoGluon demonstrated superior predictive performance (R² = 0.8407) compared to traditional machine learning models. The SHAP analysis revealed a stark spatial dichotomy in eutrophication drivers aligning with the Hu Line: in western China, natural factors like soil organic matter content (SOM) and mean normalized difference vegetation index are the primary drivers, with their protective effects susceptible to extreme precipitation. In contrast, eutrophication in eastern China is predominantly driven by synergistic anthropogenic pressures, including industrial nitrogen emissions (TN_ID) and agricultural phosphorus inputs. The study quantifies these regional differences, demonstrating that TN_ID are the most influential predictor in the east, while SOM is key in the west. We also elucidated complex interaction effects, such as the amplification of nutrient loss by extreme precipitation when SOM is high in the west, and the compounding effects of industrial and agricultural pollution in the east. This XAutoML framework provides a scientific basis for developing precise, region-specific watershed management strategies, advocating for ecosystem conservation in the west and stringent control of industrial and agricultural pollution in the east.