Prospectivity mapping and exploration targeting for sediment-hosted Pb–Zn deposits in NW Guizhou of SW China using an integrated machine learning framework

Abstract

Northwestern (NW) Guizhou in SW China hosts more than 100 Pb-Zn deposits which are primarily small in scale. In recent years, the discovery of a few large to super-large Pb-Zn deposits has revealed significant mineralization potential. It is thus necessary to reevaluate the mineral prospectivity using advanced predictive methods. This study presents an integrated machine learning framework that combines deep learning (DL), traditional machine learning (TML), and unsupervised deep learning via a variational autoencoder (VAE) for high-precision Pb-Zn prospectivity mapping in NW Guizhou. A total of 135,385 high-density soil geochemical samples (17 elements) with structural datasets are used to establish 18 evidence layers for hierarchical mineral potential assessment. Using four-fold cross-validation, this study shows that both the DL model trained on 5 × 5 km multi-channel image patches and the TML models utilizing 1–2 km buffer-processed numerical data achieved accuracies of 90 % or above on the test set. In practical application, the DL prospectivity map delineated high-confidence prospect areas more precisely than the TML models, suppressed background noise more effectively and captured nearly 90 % of known deposits. Although the TML model shows slightly lower prediction success rates (∼80 %), it provides smoother anomaly transition and identifies local anomalies in the target area. The integration of VAE with mineralization-related elements further enhances the system’s capability, enabling anomaly detection, improving the contrast of target areas, and refining them to kilometer-scale precision. Through the integration of the machine learning methods, this study successfully generates high-precision prospectivity maps and optimizes target areas at a kilometer scale, offering significant insights for sediment-hosted Pb-Zn deposit exploration.