Stratified Monte Carlo Sampled Weights-of-Evidence for Gold Prospectivity Mapping in the Yilgarn Craton, Western Australia

Declining gold discovery rates require prospectivity workflows that are statistically transparent at the regional scale, while remaining adaptable to project-scale geological interpretation. We present a robust sampled weights-of-evidence (sampled-WoE) workflow for gold prospectivity mapping in the Yilgarn Craton, Western Australia. Mine and deposit records were cleaned using 400 m spatial deduplication, yielding 7203 representative mineralized points from 12,036 strict training records. Unlabeled background points were repeatedly sampled within the Yilgarn Craton and stratified by lithology, greenstone-belt membership, and structural-density class. Ten evidence variables were evaluated diagnostically, and an independence audit based on Spearman correlation and Cramér’s V defined a six-layer regional stack comprising lithological setting, fault density, magnetic anomaly, gravity anomaly, K, and Th. Stress tests showed limited sensitivity to random seeds and background-sample sizes, whereas larger exclusion buffers systematically inflated several weights. The conservative no-buffer scenario was therefore selected as the primary model. The highest-ranked 5% of the study area captured 49.23% of the valid representative mineralized points, corresponding to a descriptive 9.85-fold enrichment over random spatial selection; spatial-block out-of-sample validation retained a 45.4% Top-5% capture and 9.1-fold enrichment. The reproducible regional baseline provides a consistent basis for separate project-scale geological refinement and validation. Under spatial-block cross-validation, the out-of-sample Top-5% capture was 45.4% (9.1-fold enrichment), and a non-spatial random cross-validation (49.3%) confirmed that spatial blocking removes the optimism introduced by spatial autocorrelation.