Groundwater estimation and determination of its probable recharge source in the Lower Swat District, Khyber Pakhtunkhwa, Pakistan, using analytical data and multiple machine learning models

This study addresses the lack of integrated hydrogeochemical and machine learning approaches in groundwater assessment, particularly in complex mountainous terrains like the Lower Swat District, Pakistan. It aims to identify recharge sources using a combination of analytical data and advanced machine learning (ML) algorithms. Groundwater recharge sources and demarcation of feasible exploration sites via actual field data and machine learning-based approaches in the Lower Swat District were carried out. Based on variations in subsurface lithological composition (e.g., relative proportions of gravel, clay, silt, and bedrock) and the varying distances of selected well sites from the Swat River, the study area was divided into seven zones. Water samples were collected from surface runoff (river and canals) and groundwater (wells and springs) and analyzed for various physicochemical parameters, including major and trace elements, to identify the probable recharge source in the floodplain area of the Swat River. X-ray fluorescence (XRF) analysis of rock samples collected from the spring hosts was also performed to compare their mineral constituents with the dissolved load of the analyzed groundwater samples. Analytical data interpretation reveals that the recharge source for groundwater in the floodplain regime is the Swat River, while infiltration and percolation of rainwater act as probable recharge sources in the mountainous and elevated areas. Acceptable similarities were observed in the geochemical composition of the rock samples, spring water samples, and representative wells in their immediate neighborhood. A linear relationship was observed between the water table and distance from the Swat River, illustrating that water depth in wells increases with increasing distance from the main recharge source. The study applied six ML models, including random forest, support vector machine (SVM), and ridge Regression, to predict groundwater zones, with random forest achieving the highest accuracy ( R 2 = 0.95, root mean square error (RMSE) = 8.49, and mean absolute error (MAE) = 4.03), followed by decision tree ( R 2 = 0.93). These metrics validate the precision of our groundwater mapping and recharge zone predictions. This integrated approach improves groundwater exploration strategies and supports sustainable water resource management. Furthermore, predicted zones for potential water wells were marked in model wells using artificial intelligence (AI) and machine learning techniques.