Integrated GIS–AHP and ensemble machine learning for groundwater and managed aquifer recharge suitability assessment in the upstream part of Awash River, Ethiopia

Abstract

Abstract Managed aquifer recharge (MAR) is a key strategy for enhancing groundwater storage, reducing overexploitation, and improving climate resilience in semi-arid regions. However, its effectiveness depends on the accurate identification of hydrogeologically suitable recharge zones. This study developed an integrated Geographic Information System–Analytical Hierarchy Process (GIS–AHP) and ensemble machine learning (ML) framework to assess groundwater potential and MAR suitability in the upstream part of the Awash River, Ethiopia. A total of 689 samples were used, with 551 for training and 138 for independent testing, and the model performance was evaluated using 10-fold cross-validation and threshold-independent metrics. Eleven thematic layers were weighted using the AHP to derive a Groundwater Potential Index (GPI). Geophysical parameters from vertical electrical soundings, transmissivity from pumping tests, and borehole yield data were used for the calibration and independent validation. Random Forest (RF), Support Vector Machine (SVM), and Extreme Gradient Boosting (XGBoost) models were trained using geophysical predictors and GPI and combined into an ensemble MAR probability model. The AHP-derived GPI showed good predictive performance (Receiver Operating Characteristic Area Under the Curve (ROC–AUC) = 0.731; Precision-Recall Area Under the Curve (PR–AUC) = 0.79), whereas the ensemble ML model achieved excellent performance (ROC–AUC = 0.913; PR–AUC = 0.946). High-to-very-high MAR suitability zones cover approximately 39% of the basin and are mainly associated with fractured volcanic formations, moderate slopes, and high lineament density. The integrated framework significantly improved MAR site identification and provided a robust decision-support tool for basin-scale groundwater management.