Scholarlink

<p>Missing groundwater level (GWL) records can significantly affect hydrogeological assessment and long-term water resource management. This study presents a structured comparative framework to evaluate five statistical imputation techniques-Expectation Maximization (EM), NIPALS, MICE, MCMC, and Nearest Neighbor-alongside a hybrid machine learning approach based on Adaptive Neuro-Fuzzy Inference System (ANFIS) for reconstructing incomplete GWL datasets. Artificial missingness scenarios of 5%, 10%, and 15% were systematically introduced into monitoring well data to simulate realistic data loss conditions. Model performance was evaluated using RMSE, MAE, MAPE, Nash–Sutcliffe Efficiency (NSE), and correlation coefficient (CC). Results indicate that covariance-based statistical methods, particularly EM and NIPALS, provide stable recovery for wells influenced by seasonal and trend-dominated groundwater behavior. The ANFIS framework was further examined using different membership functions to assess nonlinear modelling capability under increasing data sparsity. Sensitivity analysis demonstrated that membership function selection plays a critical role in maintaining predictive stability. The proposed evaluation framework offers practical guidance for groundwater monitoring agencies and data managers in selecting appropriate imputation strategies based on hydrogeological characteristics and data availability, thereby supporting reliable aquifer analysis and sustainable groundwater 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

Groundwater level data recovery: a comparative analysis of statistical and AI-driven imputation strategies

  • 期刊论文

  • 发表日期：

    2026年5月28日

查看全文
  • G. S. Sushmitha,
  • V. Amruthavarshini,
  • Siddesha Hanumanthappa

摘要

Missing groundwater level (GWL) records can significantly affect hydrogeological assessment and long-term water resource management. This study presents a structured comparative framework to evaluate five statistical imputation techniques-Expectation Maximization (EM), NIPALS, MICE, MCMC, and Nearest Neighbor-alongside a hybrid machine learning approach based on Adaptive Neuro-Fuzzy Inference System (ANFIS) for reconstructing incomplete GWL datasets. Artificial missingness scenarios of 5%, 10%, and 15% were systematically introduced into monitoring well data to simulate realistic data loss conditions. Model performance was evaluated using RMSE, MAE, MAPE, Nash–Sutcliffe Efficiency (NSE), and correlation coefficient (CC). Results indicate that covariance-based statistical methods, particularly EM and NIPALS, provide stable recovery for wells influenced by seasonal and trend-dominated groundwater behavior. The ANFIS framework was further examined using different membership functions to assess nonlinear modelling capability under increasing data sparsity. Sensitivity analysis demonstrated that membership function selection plays a critical role in maintaining predictive stability. The proposed evaluation framework offers practical guidance for groundwater monitoring agencies and data managers in selecting appropriate imputation strategies based on hydrogeological characteristics and data availability, thereby supporting reliable aquifer analysis and sustainable groundwater management.