Automatic Segmentation and Characterization of Structure Planes From Borehole Images Based on Deep Learning

Structural characteristics of rock masses are crucial in geotechnical engineering, yet manual identification of structural planes from borehole images is limited by efficiency and reliability. To address this, we developed an improved U-Net-based segmentation network, specifically tailored for structural planes in borehole images, enabling automatic identification and characterization. The model integrates deformable convolutions and channel attention mechanisms to improve structure plane detection, allowing for precise segmentation of geological structures under various challenging conditions. Furthermore, we introduce an automated workflow that couples segmentation results with curve fitting and parameter estimation techniques to precisely quantify critical structural attributes, including dip direction, dip angle, and aperture. The proposed method was evaluated on a borehole image dataset, achieving a mean IoU of 69.53%, with 93.59% for the background class and 45.47% for the structure plane class, demonstrating its effectiveness in segmenting both dominant and complex regions. The dataset is available upon request and will be made publicly available in future studies. Results also validates the effectiveness of estimating structural plane parameters. Compared to exist methods, our approach specifically addresses the unique challenges of borehole images. By providing a reliable and efficient tool for structure plane segmentation and parameter characterization, this study enhances the accuracy and efficiency of rock mass structure detection and analysis.