Evaluating the performance of machine learning approaches in predicting Albanian Shkumbini River's waters using water quality index model

Abstract

A common technique for assessing the overall water quality state of surface water and groundwater systems globally is the water quality index (WQI) method. The aim of the research is to use four machine learning classifier algorithms: Gradient boosting, Naive Bayes, Random Forest, and K-Nearest Neighbour to determine which model was most effective at forecasting the various water quality index and classes of the Albanian Shkumbini River. The analysis was performed on the data collected during a 4-year period, in six monitoring points, for nine parameters. The predictive accuracy of the models, XGBoost, Random Forest, K-Nearest Neighbour, and Naive Bayes, was determined to be 98.61%, 94.44%, 91.22%, and 94.45%, respectively. Notably, the XGBoost algorithm demonstrated superior performance in terms of F1 score, sensitivity, and prediction accuracy, the lowest errors during both learning (RMSE = 2.1, MSE = 9.8, MAE = 1.13) and evaluating (RMSE = 0.0, MSE = 0.01, MAE = 0.01) stages. The findings highlighted that Biochemical oxygen demand (BOD), Bicarbonate (HCO3), and Total Phosphor had the most positive impact on the Shkumbini River’s water quality. Additionally, a statistically significant, strong positive correlation (r = 0.85) was identified between BOD and WQI, emphasizing its crucial role in influencing water quality in the Shkumbini River.

Keywords

• Water Quality Index model
• Shkumbini River
• machine learning classifier
• model accuracy