Intelligent computing of rock mass rock quality designation based on deep learning and borehole image analysis

Rock quality designation (RQD) is widely recognized as a fundamental index in geotechnical engineering for evaluating rock mass integrity. It is extensively applied in rock mass classification systems and serves as a key input parameter for engineering rating systems. Conventionally, RQD determination relies on manual logging of recovered drill cores. However, this approach is labor-intensive, time-consuming, and highly sensitive to drilling techniques and core quality, making it difficult to obtain objective and reliable RQD values.

To address these challenges, this study proposed a novel, non-destructive approach based on the deep learning algorithm YOLOv5 (You Only Look Once, version 5) to detect and localize discontinuities directly from borehole televiewer images. It eliminated the disturbances and bias introduced during physical core extraction, enabling intelligent RQD computing. First, raw televiewer images were preprocessed, annotated, and augmented to build a representative dataset that highlighted natural fractures, bedding planes, and other geological discontinuities. Then, a YOLOv5 detector was trained on this dataset to recognize and segment discontinuities with high spatial accuracy. Finally, the model output was post-processed to compute RQD automatically by quantifying the proportion of intact rock segments exceeding the standard 10 cm threshold.

To assess the method’s performance, a case study was conducted on borehole zk4, part of a tunnel project in Yongzhou City, Hunan Province, China. Intelligent RQD values derived from the televiewer images were compared with conventional RQD measurements obtained from core boxes in the field. The results indicated that the automated approach tended to overestimate RQD by around 20% relative to manual measurements, with a mean absolute error of 9.82%. Despite this systematic bias, the spatial trend of RQD variation identified by the intelligent method closely matched that of in-situ wave velocity profiles, suggesting that the technique accurately captured relative changes in rock mass properties along the borehole.

Overall, the proposed YOLOv5-based workflow effectively reduces the influence of drilling-induced biases and core extraction artifacts on RQD estimation. By enabling rapid, repeatable, and objective computation of RQD directly from borehole images, the method enhances both efficiency and reliability of rock quality assessment. Future work will explore calibration strategies to correct systematic deviations and integrate complementary geophysical datasets. This approach demonstrates significant potential to digitalize geotechnical investigation processes, streamline tunnel engineering workflows, and advance rock mass characterization in a more robust and data-driven manner.