Prediction of Irrigation Water Quality Indices Using Random Committee, Discretization Regression, REPTree, and Additive Regression

This study aims to evaluate the performance of four ensemble machine learning methods, i.e., Random Committee, Discretization Regression, Reduced Error Pruning Tree, and Additive Regression, to estimate water quality parameters of Biochemical Oxygen Demand BOD and Dissolved Oxygen DO. Data from Anbar City on the Euphrates River in western Iraq was employed for the model's training and validation. The best subset regression analysis and correlation analysis were used to determine the best input combinations and to ascertain variable correlation, respectively. Besides, sensitivity analysis was employed to determine the standardized coefficient for BOD and DO predictions, hence knowing the significance of the relevant physical and chemical parameters. Results revealed that temperature, turbidity, electrical conductivity, Ca++, and chemical oxygen demand were identified as the best input combinations for BOD prediction. In contrast, the variable combination of temperature, turbidity, chemical oxygen demand, SO4−1, and total suspended solids was identified as the best input combination for DO prediction. It was also demonstrated that the random committee model was superior for predictions of BOD and DO, followed by the discretization regression model. For predicting BOD (DO), the correlation coefficient and root mean square error were 0.8176 (0.7833) and 0.3291 (0.3544), respectively, during the testing stage. The present investigation provided approaches for addressing difficulties in irrigation water quality prediction through artificial intelligence techniques and thence serve as a tool to overcome the obstacles towards better water management.