Groundwater quality degradation through natural and anthropogenic processes necessitates continuous monitoring and advanced assessment techniques. Limited groundwater studies in Ghana’s easternmost regions create uncertainty regarding pollution status, despite the critical need for comprehensive quality evaluation. This investigation employed multiple indexical proxies, groundwater pollution index (GPI), water quality index (WQI), water pollution index (WPI), percentage pollution index (PPI), and overall pollution index (OPI), alongside machine learning techniques to characterize groundwater pollution and evaluate predictive models performance in parts of the Dahomeyides. Hydrochemical analysis of 80 groundwater samples revealed that physicochemical parameters generally remained within WHO acceptable limits, with exceptions for bicarbonate (10 communities), nitrate (2 communities), and fluoride (1 community). The ionic concentration hierarchy (HCO₃⁻ >Na⁺ >Ca²⁺ >SO₄²⁻ >Mg²⁺ >Cl⁻ >NO₃⁻ >K⁺ >CO₃²⁻ >F⁻) indicated significant carbonate mineral dissolution influence. Multi-index assessment revealed 4.1%, 8.2%, and 95.9% pollution levels based on OPI, PPI, and WPI, respectively. Random Forest (RF) and Artificial Neural Network (ANN) models were systematically evaluated using Mean Absolute Error (MAE), Root Mean Squared Error (RMSE), and R² metrics. The ANN model demonstrated superior performance consistency across training and testing datasets, achieving R² = 1.00 for GPI with RMSE = 0.003359, compared to RF performance, showing potential overfitting with R² = 0.96 for PPI and RMSE = 2.524694 on testing data. Sensitivity analysis identified TDS, SO42-, and NO3- as dominant predictive features, reflecting anthropogenic activity impacts on groundwater quality. Spatial analysis identified pollution hotspots in Ofosu No. 1, Pare Pare, Old Duflumkpa, Obanda, and Kukurantumi communities, requiring targeted intervention strategies. The findings support ANN model adoption for operational groundwater monitoring systems and emphasize the effectiveness of integrated indexical assessment approaches for comprehensive pollution characterization in crystalline aquifer environments. Graphical Abstract  This is an indication of a step-wise procedure and presentation of the study. The basic statistics of the hydrochemistry of the samples were compared in the first step, and then the calculated indices were obtained from the hydrochemistry. The various stages of the models outline and the performance matrices of the prediction models of the generated indices are compared, and then the sensitivity of the hydrochemical parameters on the indices is elucidated. Each step is necessary for the next step; the statistical summary gives an overview of the hydrochemistry relative to world standards while giving a first-hand understanding of the groundwater quality. The pollution indices are important as they give the indexical pollution status of the groundwater. The stages of the prediction modes are outlined to give the reading audience the steps in the models generated. The comparison of the models is necessary to ascertain the acceptable model based on the performance and accuracy, for decision-making purposes. Lastly, the sensitivity of the hydrochemical parameters in the prediction of the pollution is necessary and a key guide to groundwater monitoring. All these stages and steps were necessary to give a comprehensive understanding of the prediction studies.

Indexical Proxies and Machine Learning Applications in Groundwater Pollution Prediction in a Typical Crystalline Aquifer Terrain of Ghana

期刊论文
发表日期：

2025年8月28日

查看全文

Mahamuda Abu,
Sidique Gawusu

摘要

Groundwater quality degradation through natural and anthropogenic processes necessitates continuous monitoring and advanced assessment techniques. Limited groundwater studies in Ghana’s easternmost regions create uncertainty regarding pollution status, despite the critical need for comprehensive quality evaluation. This investigation employed multiple indexical proxies, groundwater pollution index (GPI), water quality index (WQI), water pollution index (WPI), percentage pollution index (PPI), and overall pollution index (OPI), alongside machine learning techniques to characterize groundwater pollution and evaluate predictive models performance in parts of the Dahomeyides. Hydrochemical analysis of 80 groundwater samples revealed that physicochemical parameters generally remained within WHO acceptable limits, with exceptions for bicarbonate (10 communities), nitrate (2 communities), and fluoride (1 community). The ionic concentration hierarchy (HCO₃⁻ >Na⁺ >Ca²⁺ >SO₄²⁻ >Mg²⁺ >Cl⁻ >NO₃⁻ >K⁺ >CO₃²⁻ >F⁻) indicated significant carbonate mineral dissolution influence. Multi-index assessment revealed 4.1%, 8.2%, and 95.9% pollution levels based on OPI, PPI, and WPI, respectively. Random Forest (RF) and Artificial Neural Network (ANN) models were systematically evaluated using Mean Absolute Error (MAE), Root Mean Squared Error (RMSE), and R² metrics. The ANN model demonstrated superior performance consistency across training and testing datasets, achieving R² = 1.00 for GPI with RMSE = 0.003359, compared to RF performance, showing potential overfitting with R² = 0.96 for PPI and RMSE = 2.524694 on testing data. Sensitivity analysis identified TDS, SO₄^2-, and NO₃^- as dominant predictive features, reflecting anthropogenic activity impacts on groundwater quality. Spatial analysis identified pollution hotspots in Ofosu No. 1, Pare Pare, Old Duflumkpa, Obanda, and Kukurantumi communities, requiring targeted intervention strategies. The findings support ANN model adoption for operational groundwater monitoring systems and emphasize the effectiveness of integrated indexical assessment approaches for comprehensive pollution characterization in crystalline aquifer environments.

Graphical Abstract

This is an indication of a step-wise procedure and presentation of the study. The basic statistics of the hydrochemistry of the samples were compared in the first step, and then the calculated indices were obtained from the hydrochemistry. The various stages of the models outline and the performance matrices of the prediction models of the generated indices are compared, and then the sensitivity of the hydrochemical parameters on the indices is elucidated. Each step is necessary for the next step; the statistical summary gives an overview of the hydrochemistry relative to world standards while giving a first-hand understanding of the groundwater quality. The pollution indices are important as they give the indexical pollution status of the groundwater. The stages of the prediction modes are outlined to give the reading audience the steps in the models generated. The comparison of the models is necessary to ascertain the acceptable model based on the performance and accuracy, for decision-making purposes. Lastly, the sensitivity of the hydrochemical parameters in the prediction of the pollution is necessary and a key guide to groundwater monitoring. All these stages and steps were necessary to give a comprehensive understanding of the prediction studies.