水生态学杂志2026，Vol.47Issue(4)：118-130,13.DOI:10.15928/j.1674-3075.202412160001
基于可解释机器学习的美国纽约州湖泊水华频率和风险驱动因素分析
Analysis of Harmful Algal Bloom Frequency and Risk Drivers in New York State Lakes,USA:A Study Based on Explainable Machine Learning
摘要
Abstract
Lake algal blooms have negative impacts on human health and aquatic ecosystems,making their control and management crucial.In this study,113 lakes across New York State,USA were selected for research,and we explored the occurrence frequency of algal blooms and the factors driving risk,aim-ing to provide a feasible method for controlling harmful algal blooms(HABs).The study was based on lake survey data from June to September of 2018 and 2019.Water quality indicators,nutrient types,lake morphology,and watershed land use were selected as influencing factors,and algal bloom frequency types and algal bloom risk types were used as prediction targets to train machine learning(ML)models.The Shapley Additive Explanation(SHAP)and Partial Dependence Plot(PDP)were employed to rank the importance of influencing factors and interpret the ML models.Results show that the ML model based on the Random Forest algorithm achieved the highest accuracy in classifying algal bloom frequency types and risk types,with accuracy scores of 0.859 and 0.923,and AUC(Area Under the Curve)scores of 0.945 and 0.988,respectively.The ML model demonstrated excellent performance in classification tasks(ROC AUC score close to 1),exhibiting high predictive accuracy and generalization ability.Land use types(for-estland,agricultural land,and urban/residential land)were identified as significant influencing factors for lakes with different algal bloom frequency types.Forestland was the most important factor for lakes with frequent algal blooms,with a SHAP mean value of 0.55;agricultural land was the most important factor for lakes with no algal blooms,with a SHAP mean value of 0.30;and urban/residential land had the larg-est impact on lakes with periodic algal blooms,with a SHAP mean value of 0.20.For different algal bloom risk types,nutrient status was the most important influencing factor for low,medium,and high-risk lakes,with SHAP mean values of 0.12,0.07,and 0.05,respectively.Additionally,the impact of nutrient status on algal bloom risk types exhibited a"threshold effect".关键词
机器学习/湖泊/水华频率/水华风险/可解释性Key words
machine learning/lake/algal bloom frequency/algal bloom risk/explainability分类
资源环境引用本文复制引用
刘冰,孙海杰,叶晓语,时凯歌,刘辉,卢鑫,韩帅军,古励..基于可解释机器学习的美国纽约州湖泊水华频率和风险驱动因素分析[J].水生态学杂志,2026,47(4):118-130,13.基金项目
国家自然科学基金项目(51208448) （51208448）
河南省高等学校重点科研项目(24B610016) （24B610016）
河南省科技厅科技攻关项目(262102320243). （262102320243）
