Predicting groundwater electrical conductivity with ensemble learning and clustering optimization

Groundwater quality is a vital concern in arid regions, where rising salinity levels threaten drinking water and agricultural sustainability. Electrical conductivity (EC), a key indicator of salinity, requires accurate prediction for effective groundwater management. For this purpose, this study developed and evaluated four machine learning (ML) models; gradient boosting regressor (GBR), histogram gradient boosting regressor (HGBR), random forest regressor (RFR), and extreme gradient boosting regressor (XGBR) to predict EC in Herat province, Afghanistan. A dataset comprising monthly observations from 30 groundwater monitoring wells over an eight-year period (2016–2023) was used. The modeling process included data randomization, 70/30 training-testing splits, and performance evaluation using coefficient of determination (R²), root mean squared error (RMSE), mean absolute error (MAE), and median absolute error (MedAE). K-means clustering was applied to improve model precision. Results showed that GBR achieved the most balanced performance, with training and testing R² values of 0.94 and 0.90, and RMSE values of 0.23 and 0.33, respectively. While RFR and XGBR models achieved high training accuracy (R² = 0.98), they exhibited signs of overfitting. The findings suggest that GBR offers a robust, generalizable solution for EC prediction. This ML-based approach provides a scalable and data-driven tool for groundwater quality monitoring, contributing to sustainable water resource management in arid regions. Future work should incorporate additional parameters and expanded monitoring networks.