Mineral Prospectivity Mapping Using Deep Self-Attention Model

Multi-source data integration for mineral prospectivity mapping (MPM) is an effective approach for reducing uncertainty and improving MPM accuracy. Multi-source data (e.g., geological, geophysical, geochemical, remote sensing, and drilling) should first be identified as evidence layers that represent ore-prospecting-related features. Traditional methods for MPM often neglect the correlations between different evidence layers that vary with their spatial locations, which results in the loss of useful information when integrating them into a mineral potential map. In this study, a deep self-attention model was adopted to integrate multiple evidence layers supported by a self-attention mechanism that can capture the internal relationships between various evidence layers and consider the spatial heterogeneity simultaneously. The attention matrix of the self-attention mechanism was further visualized to improve the interpretability of the proposed deep neural network model. A case study was conducted to demonstrate the advantages of the deep self-attention model for producing a potential map linked to gold mineralization in the Suizao district, Hubei Province, China. The results show that the delineated high potential area for gold mineralization has a close spatial association with known mineral deposits and ore-controlling geological factors, suggesting a robust predictive model with an accuracy of 0.88. The comparative experiments demonstrated the effectiveness of the self-attention mechanism and the optimum depth of the deep self-attention model. The targeted areas delineated in this study can guide gold mineral exploration in the future.