Efficient uranium deposit exploration necessitates advanced predictive methodologies capable of effectively managing complex and imbalanced geoscientific big data. This paper introduces a novel prospectivity mapping framework that integrates geological, remote sensing, aeromagnetic, and geophysical datasets through advanced machine learning techniques optimized for severe data imbalance. Data from the Husab uranium mine, located in the Erongo region of Namibia within the central zone of the Damara orogenic belt, ~ 7 km south of the Rössing uranium mine and ~ 50 km east of Swakopmund, were standardized and spatially sampled, with clearly delineated mineralized and non-mineralized samples. Several machine learning models were evaluated, including balanced random forest, class-weighted LightGBM, class-weighted CatBoost, and a stacking ensemble method. The stacking ensemble exhibited superior performance, achieving an accuracy of 99.2%, balanced accuracy of 97.3%, recall of 95.0%, and F1 score of 95.3%. Notably, the class-weighted CatBoost and class-weighted LightGBM models demonstrated exceptional recall capabilities, with values reaching 97.5% and 97.3%, respectively, highlighting their strength in identifying mineralized samples. Evaluation using comprehensive metrics including receiver operating characteristic (ROC) and Precision–Recall (P–R) curves (ROC-AUC = 0.998, P–R-AUC = 0.989 for the stacking ensemble) confirmed the robustness of the proposed models. Moreover, model interpretability was addressed using the SHapley Additive exPlanations (SHAP) framework, a widely used model-agnostic explainable artificial intelligence technique that assigns each feature an importance value based on cooperative game theory, which quantitatively elucidated the contribution of each input feature to the prediction results. The SHAP analysis identified magnetic anomalies, lithology, and airborne radiometric anomalies (Th/K, Th, K, U) as major predictors of uranium mineralization occurrence, in agreement with established metallogenic mechanisms. The integration of multi-source geoscience data, advanced machine learning, and explainable artificial intelligence significantly enhances exploration efficiency and provides transparent, practical insights for identifying high-potential uranium targets.
Li et al. (Mon,) studied this question.