Advanced identification of geological discontinuities with deep learning

Rock mass characterization is essential for various applications in geosciences. Traditional methods, such as manual mapping and interpretation, are labor-intensive and prone to inconsistencies. Although machine learning has advanced in many fields, its application in structural geology, especially for distinguishing different discontinuity types, remains limited. This study presents a deep learning-based approach for identifying geological discontinuities in borehole images, classifying features such as intact walls, induced cracks, and tectonic fault planes, among others. We evaluate deep learning architectures, including standard Convolutional Neural Networks and Transformer-based models, and optimize segmentation performance with multi-scale training, tiling strategies, and tailored loss functions. Our results demonstrate that the Transformer model, particularly SegFormer, outperforms U-Net in detecting complex geological features. The combined use of weighted cross-entropy and focal loss further improves model robustness, especially for underrepresented and challenging features. In addition, the choice of the tiling size significantly affects the classification performance of different geological features. This research establishes an efficient and accurate pipeline for automated geological interpretation, with significant implications for subsurface exploration and geotechnical engineering.