A Framework for Detecting and Extracting Discontinuities Based on Machine Learning

Rock mass discontinuities play a significant role in evaluating the stability of rock slopes. It is necessary to collect and analyze these discontinuities to fully understand the mechanical and deformational behaviors of rock masses. Based on the digital images of some slope faces, this paper proposes a framework based on machine learning for detecting and extracting discontinuities. First, digital images and photo poses are combined, the original scene is reconstructed, and 3D point cloud data are obtained. All captured point clouds can be mapped with each image by the direct linear transformation algorithm. Next, the detection machine learning algorithm is used to segment rock mass discontinuities on 2D digital images. Finally, based on the mapping relationship, the 2D digital images that were segmented are transformed into 3D point cloud data. The geometric parameters of the corresponding discontinuities can be extracted by the DBSCAN and the Skmeans algorithm and the characterization of specific rules. The proposed framework was evaluated utilizing a quantity of slope digital images in the Yebatan hydropower under construction, located in Sichuan Province, China, as a case study. The results illustrate that most deviations are less than 10° for the dip direction and dip angle, which verifies the reliability of the proposed framework and maintains acceptable measurement accuracy.