Integrating CCME-WQI and Machine Learning for Groundwater Quality in the Sedrata Plain

Abstract

Abstract The quality of groundwater in the Sedrata plain in northeastern Algeria is under increasing pressure due to semi-arid climatic conditions, geological mineralisation and growing agricultural demand. This study provides a thorough physicochemical characterisation of phreatic aquifer waters collected from 25 open wells over four sampling periods (February, May and December 2023 and May 2024; n = 100), combining a traditional Water Quality Index (WQI) analysis with six machine learning (ML) prediction models. The CCME-WQI framework was applied using 12 physicochemical parameters. Six ML algorithms, Gradient Boosting (GB), Random Forest (RF), Support Vector Regression (SVR), Bayesian Regression (BR), AdaBoost (AdB) and K-Nearest Neighbours (K-NN) were trained and evaluated across 12 nested input-variable combinations, using the Recursive Feature Elimination with Linear Cross-Validation (RFE-L) methodology. Gradient Boosting achieved the highest predictive accuracy on the full variable set (R2 = 0.88 and RMSE = 4.89 in the test set), while SVR demonstrated the best generalisation ratio (R2 = 0.83 in the test set with only two variables removed). Best-subset regression identified a parsimonious nine-variable model (T, EC, Turb, HCO3–, Ca2+, Cl–, K+, Na+; Adj. R2 = 0.812) as the recommended input for operational WQI prediction. Chloride, calcium and electrical conductivity exhibited the strongest negative correlations with WQI, reflecting evaporite dissolution and ion exchange processes. These findings provide actionable guidance for the management of groundwater in semi-arid regions facing analogous hydrochemical and climatic constraints.