ASSESSMENT OF WATER QUALITY USING PHYSICOCHEMICAL PARAMETERS AND MACHINE LEARNING TECHNIQUES

Abstract

Water quality degradation represents a big threat to environmental sustainability, the stability of ecosystems and public health. Water quality parameters, such as selected physicochemical parameters and machine-learning techniques, were used for assessing water quality in this study. The temperature, pH, electrical conductivity, total dissolved solids, chloride, chemical oxygen demand, and biological oxygen demand were measured. Descriptive statistics, correlation and principal component analysis were first used to analyse data distribution, relations and major pollution patterns. Then, repeated stratified cross-validation was used to compare multiple machine learning algorithms: logistic regression, decision tree, random forest, support vector machine, k-nearest neighbours, and gradient boosting. values found to be more than the selected reference limits for all observed samples. Electrical conductivity was strongly correlated with total dissolved solids (TDS), and the total variance in the data was primarily explained by dissolved solids, ionic concentration and chemical pollution as determined by principal component analysis. Gradient boosting gave the highest predictive accuracy of 0.880, F1 score of 0.837, and the ROC AUC of 0.926, followed by random forest. Total dissolved solids, electrical conductivity, and chloride were determined as the most important variables by feature-importance analysis for the estimation of the relative pollution burden.