Uncertainty-Aware Mineral Prospectivity Mapping with Positive-Unlabeled Learning and Conformal Calibration: Ni Mineralization in the Yilgarn Craton

Abstract

Abstract Mineral prospectivity mapping (MPM) supports exploration targeting, yet most machine learning approaches emphasize point probabilities with limited uncertainty quantification or coverage guarantees. Regional MPM faces incomplete deposit labels from biased exploration, spatial autocorrelation invalidating random cross-validation, and limited decision-relevant guidance linking probability surfaces to exploration priorities. This study develops a framework that combines positive-unlabeled learning, nested spatial block cross-validation, bootstrap ensemble uncertainty, and cross-conformal calibration, and converts the resulting signals into five priority zones defined by lift-based percentile thresholds and a 30% relative interquartile range ( $$\rho _{\text {IQR}}$$ ρ IQR ) cut-off. On the Yilgarn Craton, Western Australia (70,200 grid cells at 5 km resolution, 818 cells containing at least one known Ni deposit location, a 1.17% positive rate), the framework uses nested fivefold spatial block cross-validation with 50 km blocks informed by an empirical variogram. XGBoost and a BaggingPU-XGBoost variant achieve nearly identical outer test PR-AUC of 0.195 and 0.194 and ROC-AUC of 0.911, and the Excluded zone removes 9% of the positive cells at empirical coverage exceeding the 85% nominal target. BaggingPU-XGBoost produces more stable high-probability predictions, with median $$\rho _{\text {IQR}}$$ ρ IQR of 18% in the top 1% of bootstrap mean probabilities compared with 48% for XGBoost. This stability advantage shifts deposits into a more populated Immediate zone, so the combined Immediate and Priority zones of BaggingPU-XGBoost capture 33% of the positive cells compared with 23% for XGBoost. The framework provides an uncertainty-informed prioritization for MPM workflows in which discovery efficiency must be balanced against costs and investment risks.