作者：Xu, Y., Ma, R., Feng, Y., Zou, B., Li, S., Huang, W., Zhou, S. & Peng, W.

Lake eutrophication is a globally pervasive environmental issue. While its driving mechanisms exhibit significant spatial heterogeneity, the relative contributions of climate change versus anthropogenic pressure remain underexplored at large spatial scales. Leveraging the Trophic State Index (TSI) of 2693 lakes from 2000 to 2020, this study innovatively integrates a geographically explainable artificial intelligence framework (XGBoost-GeoShapley/SHAP) with causal inference (CausalForestDML) to elucidate the sensitivity of TSI to environmental factors, building upon the identification of regional driving patterns. We categorized Chinese lakes into three distinct driving patterns: Natural-Sensitive, Anthropogenic-Sensitive, and Mixed-Sensitive. Based on this classification, SHAP analysis identified the nonlinear threshold characteristics of key factors and the interactive regulatory effects of natural backgrounds on anthropogenic pressures. Causal inference further revealed sensitivity variations that conventional correlation analysis failed to capture. Specifically, potential evaporation exerted a significant positive driving effect in humid regions but shifted to a prominent inhibitory role in the arid Northwest (Mixed-Sensitive). In Anthropogenic-Sensitive regions, the eutrophication potential of impervious surfaces was confirmed to significantly outweigh their mitigation capacity; furthermore, these ecosystems exhibited heightened sensitivity to increased temperature and solar radiation. Notably, the causal model captured early signals of anthropogenic pressure encroaching upon the natural-dominated southeastern edge of the Qinghai-Tibet Plateau. By integrating the identification of dominant patterns with the quantification of heterogeneous causal effects, this study systematically elucidates the driving mechanisms of lake eutrophication. These findings provide a scientific basis for differentiated zonal governance and underscore the need for future management to proactively account for the potential impacts of climate change on the stability of these driving mechanisms.

湖泊富营养化是一个全球普遍的环境问题。尽管其驱动机制存在显著的空间异质性，但在大空间尺度上，气候变化与人为压力的相对贡献仍鲜有探讨。本研究利用2000年至2020年间2693个湖泊的营养状态指数（TSI），在识别区域驱动模式的基础上，创新性地融合了地理可解释人工智能框架（XGBoost-GeoShapley/SHAP）与因果推断方法（CausalForestDML），以阐明TSI对环境因子的敏感性。我们将中国湖泊划分为三种截然不同的驱动模式：自然敏感型、人为敏感型和混合敏感型。基于此分类，SHAP分析识别出关键因子的非线性阈值特征以及自然背景对人为压力的交互调节效应。因果推断进一步揭示了常规相关分析未能捕获的敏感性差异。具体而言，潜在蒸发量在湿润地区表现出显著的正向驱动效应，而在干旱的西北地区（混合敏感型）则转变为明显的抑制作用。在人为敏感型区域，不透水表面的富营养化潜力被证实远超其缓解能力；此外，这些生态系统对气温升高和太阳辐射增强表现出更高的敏感性。值得注意的是，因果模型捕捉到了人为压力向以自然主导的青藏高原东南缘侵袭的早期信号。通过将主导模式的识别与异质性因果效应的量化相结合，本研究系统阐明了湖泊富营养化的驱动机制。这些发现为差异化分区治理提供了科学依据，并凸显未来管理需主动考量气候变化对这些驱动机制稳定性可能产生的影响。

（来源：Water Research 2026  DOI: 10.1016/j.watres.2026.125746）