Mineral prospectivity mapping based on wavelet neural network and Monte Carlo simulations in the Nanling W-Sn metallogenic province

The Nanling Range in South China is endowed with abundant W-Sn and other important rare metal resources associated with granitic intrusions, but the rate of new major mineral discoveries has been drastically diminished over the last decades. Data-driven mineral prospectivity mapping (MPM) using machine learning is emerging as a powerful tool in support of mineral exploration targeting at regional or camp scale. However, MPM based on supervised learning faces some general problems, notably including subtle information obscured in exploration data and imbalanced training samples. In this paper, we proposed a novel ensemble scheme for MPM using wavelet neural network (WNN) and Monte Carlo simulation (MCs) to address the forementioned issues. Specifically, the purpose of using WNN based on multiscale wavelet analysis is to capture weak or hidden (geophysical and geochemical) information caused by concealed ore deposits, while an undersampling MCs (UMCs) algorithm is employed to assess uncertainties caused by data imbalance and lessen its impact on MPM. Subtle (weak and mixing) information related to concealed mineralization, including geophysical and geochemical anomalies, were extracted using wavelet analysis at first, and then a dozen of predictive evidences was integrated based on WNN using UMCs for producing a W-Sn perspectivity map. The results suggest that WNN has a higher prediction accuracy (-90%) in comparison to the artificial neural network (ANN) (-85%). Moreover, the improved AUC value (-0.9) of WNN based on UMCs compared to the classic ANN (-0.82) and WNN (-0.85) suggests that the proposed ensemble algorithm improves the geological generalization of machine learning. Revisiting WNN predictions by uncertainties produced more sharply focused exploration targets of WSn deposits (covering -5% of the study area). The resulting predictive map provides important clues of W-Sn deposit occurrences which could guide and stimulate future mineral exploration in the Nanling Range.

Automated summary

This paper proposes a novel ensemble scheme using wavelet neural network (WNN) and Monte Carlo simulation (MCs) for data-driven mineral prospectivity mapping (MPM). The results show that the proposed algorithm improves the geological generalization of machine learning and provides important clues for mineral exploration.