Groundwater recharge mapping using Forest-based classification and regression and uncertainty analysis in a semi-arid watershed

Abstract

Groundwater recharge mapping is essential for sustainable water planning in semi-arid basins where monitoring is limited and groundwater is heavily relied upon. This study mapped potential groundwater recharge zones in the Al-Khazer watershed, northern Iraq, by integrating twelve conditioning factors within a machine-learning framework. A Forest-based Classification and Regression (FBCR) model, implemented in ArcGIS Pro 3.4.2, was trained on groundwater data from 88 wells; model performance was evaluated using error statistics and ROC analysis, and the robustness of the resulting map was then assessed with the Assess Sensitivity to Attribute Uncertainty tool under three input-perturbation scenarios (± 5%, ± 10%, and ± 20%), each with 100 simulations. The model showed consistent performance between training and validation (R² ≈ 0.88, RMSE ≈ 3.4, MAE ≈ 2.75, ~ 90.5% matched features), and ROC analysis confirmed strong discrimination (AUC = 0.912 for training; 0.907 for testing). The predicted suitability pattern was physically coherent, with higher suitability concentrated along the main valley corridors and tributary fans and lower suitability across the steep northern uplands where runoff energy is high. Under increasing perturbation, stability declined systematically but remained dominated by a stable core, 88.80% stable features at ± 5%, 80.27% at ± 10%, and 70.21% at ± 20%, with instability expanding mainly along class boundaries and transition zones. FBCR coupled with simulation-based uncertainty assessment thus provides a transparent basis for identifying high-confidence recharge zones; in the Al-Khazer watershed, the stable high-suitability valley corridors are recommended for groundwater protection and managed-aquifer-recharge screening, while unstable transition zones warrant field verification and enhanced monitoring.