Exploration
Marco Antonio Cotrina-Teatino; Jairo Jhonatan Marquina-Araujo
Abstract
Integrating entropy-based uncertainty analysis with machine learning offers a novel approach to improving lithological classification in mineral exploration. This study applies supervised algorithms to predict lithology from spatial and geochemical data collected at a gold deposit in northern Peru. The ...
Read More
Integrating entropy-based uncertainty analysis with machine learning offers a novel approach to improving lithological classification in mineral exploration. This study applies supervised algorithms to predict lithology from spatial and geochemical data collected at a gold deposit in northern Peru. The dataset includes 2,129 composited samples from 140 drillholes, containing spatial coordinates (East, North, Elevation) and gold content (Au). Six classifiers were tested: Random Forest, XGBoost, Support Vector Machine, K-Nearest Neighbors, Naive Bayes, and Multilayer Perceptron. Stratified five-fold cross-validation was applied to a 70/30 train-test split. The best performance was achieved by ANN-MLP (94.5% accuracy) and XGBoost (93.9%), with F1-scores above 94%. In zones of low uncertainty, models reached up to 100% precision, while accuracy dropped to 71.9% in highly uncertain regions. Entropy-based uncertainty mapping highlighted areas of geological ambiguity, such as lithological boundaries or sparsely sampled zones. The Friedman test confirmed statistically significant differences among classifiers (p < 0.001). These findings demonstrate that combining machine learning with spatial uncertainty quantification enhances both predictive reliability and geological interpretability, offering a practical tool for guiding exploration and reducing risk in complex mineral systems.
