Scholarlink

<p>Accurate water quality assessment is critical for environmental monitoring and public health. Conventional Water Quality Index (WQI) computation methods, however, often rely on numerous parameters and labor-intensive processes, thus limiting their practicality for rapid assessments. While Machine Learning (ML) offers promising alternatives, the development of high-performing models typically demands extensive expertise and computational resources. This study addresses the latter gap by leveraging Automated Machine Learning (AutoML), specifically the AutoGluon platform, to predict WQI from a reduced set of readily available water quality parameters. Our objectives were to (i) evaluate the predictive performance of AutoML with reduced inputs, (ii) assess model interpretability via feature importance, and (iii) propose an automated framework for efficient water quality monitoring. This study analyzed a 36-year dataset from Taiwan’s national river water quality monitoring network, focusing on four parameters: electrical conductivity (EC), suspended solids (SS), water temperature (WT), and pH. A fully automated pipeline handled model selection, hyperparameter tuning, and ensemble construction by systematically testing multiple algorithm families and stacking strategies to determine the optimal setup for each parameter. This eliminated manual intervention and delivered reproducible, data-driven results that matched the distinct spatiotemporal patterns in the long-term records. Among the evaluated algorithms, ensemble-based tree models (CatBoost, Random Forest, and XGBoost) demonstrated superior performance, achieving a mean <InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(R^2\)</EquationSource> </InlineEquation> of 0.76 and low predictive errors. Feature importance analysis identified EC as the most influential predictor. Such results highlight the feasibility of reducing the number of input parameters without compromising prediction accuracy, thereby enabling faster and cost-effective water quality assessments.</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

Automated machine learning achieves accurate water quality prediction with reduced parameter requirements

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年1月7日

查看全文
  • Deivid Campos,
  • Viviane Galvão,
  • Matheus Lopes de Rezende,
  • Allan Braga,
  • Matteo Bodini,
  • Uilson R. V. Aires,
  • Roland Yonaba,
  • Leonardo Goliatt

摘要

Accurate water quality assessment is critical for environmental monitoring and public health. Conventional Water Quality Index (WQI) computation methods, however, often rely on numerous parameters and labor-intensive processes, thus limiting their practicality for rapid assessments. While Machine Learning (ML) offers promising alternatives, the development of high-performing models typically demands extensive expertise and computational resources. This study addresses the latter gap by leveraging Automated Machine Learning (AutoML), specifically the AutoGluon platform, to predict WQI from a reduced set of readily available water quality parameters. Our objectives were to (i) evaluate the predictive performance of AutoML with reduced inputs, (ii) assess model interpretability via feature importance, and (iii) propose an automated framework for efficient water quality monitoring. This study analyzed a 36-year dataset from Taiwan’s national river water quality monitoring network, focusing on four parameters: electrical conductivity (EC), suspended solids (SS), water temperature (WT), and pH. A fully automated pipeline handled model selection, hyperparameter tuning, and ensemble construction by systematically testing multiple algorithm families and stacking strategies to determine the optimal setup for each parameter. This eliminated manual intervention and delivered reproducible, data-driven results that matched the distinct spatiotemporal patterns in the long-term records. Among the evaluated algorithms, ensemble-based tree models (CatBoost, Random Forest, and XGBoost) demonstrated superior performance, achieving a mean \(R^2\) of 0.76 and low predictive errors. Feature importance analysis identified EC as the most influential predictor. Such results highlight the feasibility of reducing the number of input parameters without compromising prediction accuracy, thereby enabling faster and cost-effective water quality assessments.