Performance of tree-based ensemble techniques in predicting groundwater quality for irrigation purposes

Abstract

This study evaluates the performance of eight different machine learning (ML) methods to predict the Irrigation Water Quality Index (IWQI), an important metric for assessing groundwater quality for agricultural purposes. The study domain was selected as the Saïss Plain in northern Morocco as the region stands out as an area with intense agricultural activities where groundwater quality is of critical importance for irrigation. Groundwater quality is affected by natural factors such as salinity and ion concentrations, as well as anthropogenic activities such as agricultural and industrial practices. Among eight ML approaches, the XGBoost model outperformed its counterparts, including other tree-based ML algorithms and benchmarking models, and yielded the highest prediction accuracy with Nash Sutcliffe Efficiency (NSE) index of 0.963 and 0.892 for training and testing sets, respectively. Other tree-based models such as Random Forest, AdaBoost, and Extra Trees also showed strong performance, while benchmarking models such as ANN, KNN, and SVR were less effective due to the size and non-linear nature of the dataset. The analysis revealed that chloride (Cl⁻) and sodium (Na+) ions are the most critical factors in IWQI estimations. This study highlights the importance of robust ML models in groundwater quality management and provides insights to guide future research for sustainable irrigation practices.