Scholarlink

<p>Seasonal hypoxia threatens urban river ecosystems. Although machine learning (ML) models exhibit strong capabilities in predicting dissolved oxygen (DO), utilizing ML models to uncover intrinsic data patterns and translate predictions into actionable management insights remains a major challenge. To address this gap, we developed an integrated framework that combines high-performance ML, an enhanced Structured SHAP Analysis Approach (SSAA), and causal inference. This framework was applied to an important urbanized tributary of the Yangtze River. It quantifies trade-offs in DO prediction accuracy across spatio-temporal data resolutions. It further utilizes SSAA and Fast Causal Inference to identify underlying patterns and causal relationships among the variables, transforming predictive insights into actionable management strategies. The CatBoost model achieved the highest accuracy (R²=0.9728, KGE = 0.9776) using local, high-frequency (hourly) data. Our SSAA method quantified critical thresholds, revealing that hypoxia risk increases significantly when water temperature exceeds 21.60&#xa0;°C or pH falls below 7.77. We also identified that high water temperatures (&gt; 25&#xa0;°C) combined with low sunshine duration (&lt; 5.8&#xa0;h/day) may together create conditions with extreme risk of severe hypoxia. Causal inference further demonstrated that the strong pH-DO correlation is a spurious relationship mediated by algal activity. The framework identified the primary influencing tributary and quantified differentiated management targets, proposing a stricter DO standard for the key tributary (LD &gt; 8.56&#xa0;mg/L) over the mainstem (QH &gt; 7.95&#xa0;mg/L). This research transforms the ML model from a predictive tool into a decision support system, enabling proactive, precise, and spatially differentiated water quality management.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Applying Interpretable Machine Learning for Hypoxia Mitigation: an Operational Framework Coupling Predictive Modeling and Causal Inference for Urban River Management

  • 期刊论文

  • 发表日期：

    2026年1月3日

查看全文
  • Jia-yun Chen,
  • Da-wei Wang,
  • Zu-lin Hua,
  • Ke-jian Chu,
  • Li Gu

摘要

Seasonal hypoxia threatens urban river ecosystems. Although machine learning (ML) models exhibit strong capabilities in predicting dissolved oxygen (DO), utilizing ML models to uncover intrinsic data patterns and translate predictions into actionable management insights remains a major challenge. To address this gap, we developed an integrated framework that combines high-performance ML, an enhanced Structured SHAP Analysis Approach (SSAA), and causal inference. This framework was applied to an important urbanized tributary of the Yangtze River. It quantifies trade-offs in DO prediction accuracy across spatio-temporal data resolutions. It further utilizes SSAA and Fast Causal Inference to identify underlying patterns and causal relationships among the variables, transforming predictive insights into actionable management strategies. The CatBoost model achieved the highest accuracy (R²=0.9728, KGE = 0.9776) using local, high-frequency (hourly) data. Our SSAA method quantified critical thresholds, revealing that hypoxia risk increases significantly when water temperature exceeds 21.60 °C or pH falls below 7.77. We also identified that high water temperatures (> 25 °C) combined with low sunshine duration (< 5.8 h/day) may together create conditions with extreme risk of severe hypoxia. Causal inference further demonstrated that the strong pH-DO correlation is a spurious relationship mediated by algal activity. The framework identified the primary influencing tributary and quantified differentiated management targets, proposing a stricter DO standard for the key tributary (LD > 8.56 mg/L) over the mainstem (QH > 7.95 mg/L). This research transforms the ML model from a predictive tool into a decision support system, enabling proactive, precise, and spatially differentiated water quality management.