Comparative GIS-Based Ensemble Machine Learning for Groundwater Potential Mapping and Hydrogeological Controls in the Arid Stampriet Transboundary Aquifer System, Namibia

Abstract Groundwater potential mapping is increasingly used to support well siting and resource planning in arid regions, particularly where hydrogeological data are scarce. This study presents the first GIS based assessment of groundwater potential for the Namibian section of the Stampriet Transboundary Aquifer System (STAS). It applies comparative ensemble machine learning models, namely Boosted Regression Trees (BRT) and Stacked Ensemble Random Forest (SERF). These were selected for their ability to capture nonlinear hydrogeological interactions and maintain predictive robustness under data-limited and heterogeneous conditions. The workflow, including model parameterisation and input datasets, is fully documented in the methods section to support reproducibility in principle. Predictor variables representing hydrogeomorphological, geological, and climatic controls, including drainage and lineament proximity and density, depth to bedrock, terrain attributes, rainfall, bulk density, and surface wetness, were integrated within BRT and SERF. Model performance was evaluated using five-fold spatial block cross-validation to reduce bias arising from spatial autocorrelation and assessed using receiver operating curve–area under curve (ROC–AUC) and frequency ratio analyses. Although both models achieved high predictive accuracy (AUC > 0.90), results remain influenced by data scarcity and hydrogeological heterogeneity and should therefore be interpreted as probabilistic indicators rather than deterministic predictions. Both models produced coherent groundwater potential patterns, consistently delineating a north-central to north-western corridor of elevated groundwater potential extending from Stampriet towards Gross Ums and Onderombapa. BRT demonstrated higher predictive skill (AUC = 0.958) than SERF (AUC = 0.915), indicating stronger representation of nonlinear hydrogeological interactions. Drainage proximity, net recharge, rainfall, and depth to bedrock emerged as the dominant predictors. These demonstrate the value of focused recharge along ephemeral drainage pathways, combined with subsurface storage constraints in this recharge-limited, multi-layered aquifer system. High to very high groundwater potential zones occupy a limited proportion of the basin (7.08% for BRT; 12.04% for SERF). This provides a practical basis for prioritising exploration, siting efficiency of new production boreholes, and supporting groundwater licensing, drought planning, and transboundary aquifer governance in the STAS. Graphical Abstract This graphical abstract presents a concise visual representation of groundwater potential mapping within the Namibian sector of the Stampriet Transboundary Aquifer System (STAS). It depicts the study area, along with the compiled geospatial and hydrogeological datasets, and outlines the machine-learning workflow employed to delineate groundwater potential zones. The framework assesses Boosted Regression Trees (BRT) and Stacked Ensemble Random Forest (SERF) models, emphasising the superior predictive performance of BRT, as evidenced by high area under curve (AUC) values. The findings highlighted drainage networks, net recharge, rainfall, and depth to bedrock as the critical factors influencing groundwater occurrence. High potential zones were predominantly mapped in the north-central to north-western portions of the STAS, accounting for approximately 7.08–12.04% of the total area. These spatially explicit outputs offer actionable insights for targeted groundwater development, aquifer protection, and sustainable resource management in both arid and semi-arid environments. The study also addressed methodological limitations and suggests future improvements, including the integration of subsurface data, robust spatial validation and reproducible workflows. It also included a thorough consideration of the multi-layered aquifer system, strengthening the study’s contribution to data-driven, sustainable groundwater management.