Surface water quality assessment and its evaluation of potential pollution risks for drinking purposes employing water quality indices and various machine learning techniques
Abhijeet Das
Abstract
The long-term effects of human activities on surface water quality in Lakhanpur Block, Jharsuguda District, Odisha, remain poorly understood, especially regarding the combined impacts of urbanization, industrialization, and agricultural practices. Between 2017 and 2024, during the pre-monsoon seasons, nine samples were collected from 20 sampling stations to assess hydro-chemical characteristics and evaluate long-term water suitability, with consideration given to both natural influences and anthropogenic factors, employing water quality indices (WQIs) and various Machine Learning (ML) Models such as multiple linear regression (MLR), logistic regression (LR), support vector machine (SVM), decision tree (DT), extra tree (Ex T), random forest (RF), K-nearest neighbor (KNN) and extreme gradient boosting (XG Boost). Slight alkalinity may be attributed to biological activity, industrial effluents, or carbonate buffering. Majority of the inventoried locations, approximately 55%, fell within the excellent to good category for drinking water quality based on WA (weighted arithmetic) water quality index (WQI) values, while the remaining 45% indicated poor quality, necessitating appropriate treatment and the implementation of an effective water management strategy. The E(entropy - weighted) WQI method is more effective in reducing eclipsing effects, making it the preferred choice for forecasting WQ. A smaller fraction of the samples, indicated as 10% and 15%, were classified as excellent and good, respectively, while the E (entropy-weighted) WQI, holding a corresponding mean value of 148.20, indicated that most samples (50%) fell into the poor to extremely poor categories. Decreasing EWQI trend visualizes a decline in water quality from west to east across the region, with notably poor surface water quality observed in the southern areas near the sampling sites. Among regression models, the multiple linear regression (MLR) demonstrated excellent predictive performance, attaining a high R 2 of 0.9992 and a minimum RMSE of 0.338. SVM model demonstrated strong predictive power and generalization ability, though it appeared to be slightly less consistent across different evaluation sets compared to the MLR model. For the non-ensemble models (DT and KNN), while DT showed balanced performance between training, testing, and cross-validation, KNN performed exceptionally well through the training dataset, moving to bad performance on the testing and cross-validation datasets. Ensemble models (RF, Ex T, and XG Boost) generally performed well in forecasting the WQI scores in the three stages. XG Boost performs exceptionally well than other models, with a training RMSE of 0.001 (R 2 = 0.9998), testing RMSE of 1.96 (R 2 = 0.9726), and cross-validation RMSE of 2.93 (R 2 = 0.9309), respectively. • This study assessed the comprehensive quality of drinking water. • Investigates the impact of industrialization on water quality parameters. • WAWQI is employed to detect the variability of the different parameters. • EWQI exhibited a west-to-east decreasing trend in surface water quality. • ML was used to identify the correlation and comparison of the experimental values with predicted values. • Recommendations for physical, biological, and chemical contamination management.