Scholarlink

<p>Reconstructing incomplete salinity data is crucial for enabling subsequent simulation, forecasting, and related research activities that support effective water quality management, particularly in downstream areas such as the Sai Gon–Dong Nai River, Vietnam. Therefore, this study proposes a methodological framework employing machine learning algorithms (e.g., KNN, MLP, RF, and XGB) to impute bi-hourly salinity data at Nha Be station from 2015 to 2020, with up to 50% missingness and long consecutive gaps ranging from 12 to 48&#xa0;h. Imputation was based on observed salinity at the same station and on water level data from surrounding stations, namely Nha Be, Phu An, and Bien Hoa. An artificial dataset was generated to evaluate model performance, reflecting the original missingness patterns and mechanisms. The findings indicate that the MLP model outperformed the other methods under both full-period and year-by-year imputation strategies. Notably, the year-by-year approach achieved higher statistical agreement and lower prediction errors, with R<sup>2</sup> values ranging from 0.860 to 0.960; both RMSE and MAE remained consistently below 1.001&#xa0;g/L and 0.787&#xa0;g/L, respectively, across the six years analyzed. Subsequently, the optimized model and scenario were applied to impute the original salinity data, successfully reconstructing its statistical distribution and demonstrating its effectiveness in handling consecutive gaps with high missing rates. Overall, the proposed imputation framework enhances the continuity and reliability of the salinity monitoring dataset, thereby supporting water resources management and adaptation strategies against salinity intrusion.</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

Machine Learning-Based Multivariate Time Series Imputation for Long Continuous Gaps in Surface Water Salinity Monitoring Data

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年2月27日

查看全文
  • Nguyen Thi Diem Thuy,
  • Hiroshi Ishidaira,
  • Kazuyoshi Souma,
  • Jun Magome

摘要

Reconstructing incomplete salinity data is crucial for enabling subsequent simulation, forecasting, and related research activities that support effective water quality management, particularly in downstream areas such as the Sai Gon–Dong Nai River, Vietnam. Therefore, this study proposes a methodological framework employing machine learning algorithms (e.g., KNN, MLP, RF, and XGB) to impute bi-hourly salinity data at Nha Be station from 2015 to 2020, with up to 50% missingness and long consecutive gaps ranging from 12 to 48 h. Imputation was based on observed salinity at the same station and on water level data from surrounding stations, namely Nha Be, Phu An, and Bien Hoa. An artificial dataset was generated to evaluate model performance, reflecting the original missingness patterns and mechanisms. The findings indicate that the MLP model outperformed the other methods under both full-period and year-by-year imputation strategies. Notably, the year-by-year approach achieved higher statistical agreement and lower prediction errors, with R2 values ranging from 0.860 to 0.960; both RMSE and MAE remained consistently below 1.001 g/L and 0.787 g/L, respectively, across the six years analyzed. Subsequently, the optimized model and scenario were applied to impute the original salinity data, successfully reconstructing its statistical distribution and demonstrating its effectiveness in handling consecutive gaps with high missing rates. Overall, the proposed imputation framework enhances the continuity and reliability of the salinity monitoring dataset, thereby supporting water resources management and adaptation strategies against salinity intrusion.