Introduction: Cracks are a key feature that determines the structural integrity of rocks, and their angular distribution can be used to determine the local or regional stress patterns. The temporal growth of cracks can be monitored in order to predict impending failures of materials or structures such as a weakened dam. Thus, cracks and their spatial-temporal distributions should be automatically monitored for assessing their structural integrity, the associated stress patterns and their potential for failure. Method: We show that the U-Net convolutional neural network, semantic segmentation and transfer learning can be used to accurately detect cracks in drone photos of sedimentary massifs. In this case, the crack distributions are used to assess the safest areas for tunnel excavation. Compared to the coarse performance of ridge detection, the U-Net accuracy in identifying cracks in images can be as high as 98% when evaluated against human identification, which is sufficient for assessing the general crack properties of the rock faces for the engineering project. Result: Based on approximately 100 h of manual cracks labeling in 127 drone photos and 20 h of network training, the U-Net was able to successfully detect cracks in 23,845 high-resolution photographs in less than 22 h using two Nvidia V100 GPUs. Meanwhile, the network was able to detect more than 80% of the observable cracks of a volcanic outcrop in Idaho without additional training. With a modest amount of extra labeling on photos of the volcanic outcrop and transfer training, we found that the accuracy significantly improved. The surprising outcome of this research is that the U-Net crack detector laboriously trained on photos of sedimentary rocks can also be effectively applied to photos of volcanic rock faces. This can be important for real-time assessment of geological hazards and lithology information for dam inspection and planetary exploration by autonomous vehicles. For another application, we accurately detected fractures and faults with a scale of tens of kilometers from Martian photographs. Conclusions: In summary, our methodology of using CNN with transfer training suggests that it can be used as a semi-universal detector of cracks in across a range of diverse geological settings.
Bibliographical noteFunding Information:
The research reported in this publication was supported by the King Abdullah University of Science and Technology (KAUST) in Thuwal, Saudi Arabia. For computer time, this research used the resources of the Supercomputing Laboratory at KAUST and the IT Research Computing Group. We thank them for providing the computational resources required for carrying out this work.
Copyright © 2023 Shi, Ballesio, Johansen, Trentman, Huang, McCabe, Bruhn and Schuster.
- convolutional neural network
- geo-crack detection
- machine learning
- rock cracks
- transfer learning
ASJC Scopus subject areas
- Earth and Planetary Sciences(all)