Three-dimensional geological modeling of joint-dense zones using multi-scale feature learning under incomplete resistivity data conditions

Tunnel engineering frequently intersects joint-dense zones, increasing the risk of hazards such as water inrush and rock collapse. Accurate geological models are crucial for project safety and construction quality. However, harsh conditions and limited infrastructure often lead to incomplete resistivity-based geophysical data, posing challenges for fine-scale modeling. Existing voxel-based 3D geological modeling methods commonly neglect the spatial distribution of resistivity, resulting in insufficient voxel classification. To address these limitations, this study focuses on two aspects: completion of missing resistivity data and voxel classification based on spatial features of resistivity data. We propose a resistivity prediction model that combines random forest techniques with multi-interpolation methods, alongside a voxel classification model based on a multi-scale 3D convolutional neural network. By accurately predicting subsurface resistivity, voxel classification is performed for joint-dense zones, establishing a framework for refined 3D geological modeling despite incomplete data. Experimental results show that the proposed resistivity prediction model outperforms traditional standalone interpolation methods. Moreover, when integrated with completed resistivity data, the multi-scale 3D CNN achieves higher voxel classification accuracy than non-deep learning and other deep learning approaches. This framework enables precise, fine-scale 3D modeling of joint-dense zones and provides a flexible reference for geological modeling in similar contexts.