Data-driven prediction modeling of groundwater quality using integrated machine learning in Pinggu Basin, China

The Pinggu Basin of Beijing (Capital of China). Achieving high-accuracy (>95 %) groundwater quality prediction is key for sustainable groundwater management and protection. This study focused on data-driven prediction modeling — Support Vector Machine (SVM), Random Forest (RF), Back Propagation (BP) Neural Network, and Convolutional Neural Network (CNN) — to predict groundwater quality based on 1019 groundwater samples from the study area. This study provided new insights into model selection and model building for groundwater quality prediction. The dissolution of carbonate rocks primarily controlled major hydrochemical ions. More than 90 % of groundwater samples were clean for drinking. Poor-quality samples were distributed in the northwest of the Pinggu Basin in recent years, mainly due to high nitrate levels (>50 mg/L). That nitrate concentration was an important factor controlling the groundwater quality was also concluded from the machine learning (ML) models. The ion ratio diagram revealed that most of the nitrate originated from agricultural nitrogen fertilizer use, with some contribution from urban sewage sources. The BP Neural Network was the most accurate model for predicting nitrate concentration and groundwater quality in the Pinggu Basin (R 2 =0.99, accuracy=0.99). • Nitrate is the key sensitive indicator for groundwater quality and machine learning models. • Agricultural activities and urban sewage contributed to high nitrate levels. • Back propagation neural network outperformed in regression and classification. • Data-driven model selection should consider objectives, data size, and variable influence.