Towards automated joint detection and RQD estimation in acoustic televiewer imaging using deep learning (instance segmentation)

A thorough understanding of rock mass structural complexity is essential for geotechnical design and analysis of surface and underground excavations in rock. Borehole imaging is commonly used to rapidly and accurately characterize fractures without handling core specimens. Acoustic televiewer (ATV) imaging is an effective tool for detecting structural fractures and determining Rock Quality Designation (RQD) along a borehole. As part of interpreting the ATV data, the logger typically detects and identifies joints manually. This is a time-consuming, subjective, and inconsistent process. This study introduces a method that can automate joint detection, joint orientation (alpha and beta angles), and RQD estimation. For this study, a total of 1390 m of ATV data, including 1847 joints, were collected from 24 boreholes. In the first step, several filtering techniques were used, including Canny, Laplacian of Gaussian, K-Means, Multiple thresholding, Hough transform, and watershed segmentation for automated joint segmentation. In comparison, watershed segmentation outperforms other techniques, but it is sensitive to noise and outbreaks present in some of the ATV images. As a result, a deep learning algorithm called Mask R-CNN was used. This approach is an instance segmentation method that showed promising results with an F1-score of 0.82 in automated joint detection on an unseen test dataset. Based on the model, the mean absolute errors of alpha and beta angles and the RQD calculated by the model are 1.4^o, 20.1^o, and 1%, respectively.