Machine learning and ensemble learning models for groundwater potential mapping in a fractured basin: Case of the Azrou-Khenifra basin, central massif, Morocco

Groundwater potential mapping (GWP) is essential for sustainable resource managmment, particlarly in semi-arid and data-scarce regions. This study aims to delineate groundwater potential zones in the Azrou-Khenifra basin by integrating topographical, hydrological, geological, cilmatological and Land use/ Land cover factors using machine learning and ensemble learning models. A total of twenty groundwater influencing factors were considered, including elevation, slope, slope aspect, profile and plan curvature, TPI, TRI, SPI, TWI, drainage density, distance to drainage, lithology, distance to faults, fault density, distance to lineaments, lineament density, rainfall, LST, as well LULC, NDVI, most of which were derived from satellite data. Frequency Ratio (FR) and Weight of Evidence (WoE) were aplied to analyze the spatial relationship between water points and individual factor classes and to evaluate their contribution. Groundwater potential mpas were generated using three machine learning algorithms : Random Forest (RF), Support Vector Machine (SVM), and K-Nearest Neighbors, and four ensemble learning models (RF-SVM, RF-KNN, SVM-KNN, and RF-SVM-KNN). Models performance was evaluated using precision, recall, F1-score, overall accuracy, and ROC-AUC. The results indicate that the RF model achieved the highest predictive perfomance among individual models, with an accuracy of 0.74 and ROC-AUC of 0.8112, followed by KNN (AUC = 0.787) and SVM (AUC = 0.778). Ensemble learning models competitiveand stable performance, particularly RF-based ensembles and the final RF-SVM-KNN voting model, which achieved balanced metrics with and overall accuracy of about 0.68-0.72 and AUC values reaching 0.8112. Feature importance analysis highlights rainfall, Fault density, and distance to faults as the most influential factors controlling groundwater occurrence, followed by topograohic indices and hydrological factors. Additional validation based on the spatial distribution of water points productivity across groundwater potential classes confirmed the reliability of the generated maps, as higher flow-rate points were predominantly located within high and very high potential zones. The findings demonstrate that integrating machine learning and ensemble learning techniques provides reliable groundwater potential predictions and offers a useful framework for groundwater exploration and sustainable management in hetergeneous and data-limited regions.