Spatially honest validation of machine-learning prospectivity maps for karst bauxite in Kazakhstan

Abstract

Abstract Machine-learning prospectivity maps often report near-perfect apparent discrimination, but such estimates may not survive validation designs that reflect deployment in spatially clustered mineral systems. We map karst-bauxite prospectivity across covered terrain in Kazakhstan using freely available Sentinel-2 surface reflectance, nine spectral indices and three terrain derivatives, while withholding absolute coordinates for sensitive inventory protection. Instead of reporting a single cross-validation score, we evaluate the model through a graded validation ladder: random k-fold cross-validation, spatial-block cross-validation, k-fold nearest-neighbour distance matching and leave-one-deposit-cluster-out validation. For the operational random-forest model, random cross-validation gives ROC-AUC = 0.984, whereas spatial-block validation lowers performance to approximately 0.84 and the stricter prediction-oriented schemes converge to ROC-AUC $\approx$ 0.77--0.78, with performance stabilising near this honest level as block size approaches the residual autocorrelation range. Deposit-level bootstrapping gives mean held-out deposit-cluster skill of ROC-AUC = 0.72 with a 95\% confidence interval of 0.64--0.80. Conformal prediction shows asymmetric uncertainty: background-like terrain is rejected reliably, whereas bauxite-like terrain remains heterogeneous across unseen deposit clusters, and an area-of-applicability mask flags 20 of 69 delineated high-score targets as extrapolative. The resulting map should therefore be interpreted as a calibrated, uncertainty-aware and applicability-screened exploration triage product, not as a probability-of-discovery surface.

More by Marat Nurtas

• Two-stage synthetic-to-real transfer learning for automated mammography report generation using vision-language modelsScientific Reports · 2026
• Macro-scale earthquake forecasting under class imbalance in Central AsiaJournal Of Big Data · 2026
• Enhanced Fine-Tuning Optimization for Real-Time Fire and Smoke Detection Using a YOLOv9 Architecture with a Comprehensive Ablation AnalysisEngineering Technology & Applied Science Research · 2026 · Citations: 1
• Satellite-Based Deep Learning Models for Mineral ExplorationIEEE International Conference on Smart Information Systems and Technologies · 2026