Deterministic, probabilistic, and machine learning approaches for water quality index prediction, source identification via non-negative matrix factorization, and health risk evaluation

Abstract

Water quality is a crucial index of human health and environmental sustainability. The present study aimed to apply deterministic, probabilistic, and ML techniques, such as RF, DT, KNN, XGB, SVR, and GB, to classify the water quality in the southern region of Iran. The levels of TDS and alkalinity exhibited the greatest deviation from the standards set by the EPA, WHO, and BIS. The WQI findings revealed that 88.30% of the data were classified as excellent or good quality when employing the deterministic method. Conversely, 97.77% fell within these categories when the Monte Carlo simulation approach was used. The models were meticulously assessed using a set of statistical metrics, including R2, MAE, RMSE, MSE, and PREI. The results show that the RF and XGB were highly effective in predicting WQI. The factors influencing the WQI, identified by RF and XGB methodologies and based on MLP, were TDS and SO42− within the study area. According to the Piper diagram, the predominant groundwater type in the study region was HCO₃−–Na⁺, influenced by seawater intrusion, geological properties, and human activities. The deterministic method showed that the HI values exceeded the threshold of 1 in 51%, 2.13%, and 2.13% of the samples for children, teenagers, and adults, respectively. In contrast, the Monte Carlo simulation approach indicated that the HI values exceeding 1 were 34.8% for children, 2.9% for teenagers, and 0.4% for adults. Moreover, the HI was significantly affected by fluoride concentration, ingestion rate, and their interaction. The study's findings emphasized sustainable water management in this area.