A new supervised learning framework for mineral prospectivity mapping: The three-class labeling (3CL) approach

Abstract

Mineral exploration depends on the ability to accurately distinguish geological characteristics and predict mineralization zones. This study introduces a novel three-class labeling (3CL) strategy, which categorizes regions as mineralized, non-mineralized, and neutral. Compared to traditional two-class labeling (2CL), this approach reduces false positives and enhances prediction precision by capturing intermediate areas that do not fit clearly into mineralized or non-mineralized categories. To implement this approach, we employed the Random Forest method to model mineralization potential, training it with optimized hyperparameters using GridSearchCV. The optimized trained model evaluation metrics demonstrated that the 3CL model outperformed the 2CL one, with higher accuracy (98 % vs. 96 %), F1-score (0.98 vs. 0.96), and Kappa coefficient (0.97 vs. 0.92), confirming its superior capability in distinguishing prospective zones while reducing misclassification. The 3CL approach enhances the spatial precision of prediction maps and offers a more geologically realistic interpretation of mineralization potential. These findings highlight the practical advantages of 3CL in improving exploration efficiency and guiding decision-making for resource allocation in mineral prospectivity studies.