Three-dimensional mineral prospectivity mapping using a residual convolutional neural network with lightweight attention mechanisms

Abstract

The continuous exploration and mining of surface and shallow mineral resources have promoted subsurface exploration. Over the past decade, mineral prospectivity mapping (MPM) has progressively expanded from two-dimensional (2D) to three-dimensional (3D). The 3D convolutional neural networks (CNN) and attention mechanisms (AMs) are adept at processing 3D voxel data, offering significant advantages for 3D MPM. However, the larger computational cost in a 3D CNN–AMs model presents limitations, constraining its application to 3D MPM. This study presents a new 3D CNN architecture composed of residual blocks (ResBlocks) and a lightweight Attention Mechanism (LAM) for 3D MPM. ResBlocks incorporate skip connections to deepen the network structure, thereby enhancing its ability to model complex nonlinear patterns and mitigating the vanishing gradient problem. The LAM utilizes the dimensionality reduction fully connected layer for channel attention and depthwise separable convolution for spatial attention, thus reducing computational costs. A case study in the Wulong gold district demonstrates that the proposed architecture achieves performance improvements in 3D MPM without significant increases in numbers of parameter count and FLOPs, highlighting its efficiency and effectiveness. Furthermore, the 3D mineral targets obtained in this study are beneficial for subsurface gold exploration in the Wulong Au district, China.