Rock Mass Characterization of Discontinuities by Unsupervised Machine Learning

Rock slope landslide potential, tunnel design, and the engineering of underground spaces critically require an analysis of rock joint orientation and spacing. Joint set measurements are typically determined by hand in the field using a compass–clinometer to measure the orientation of the geologic planes with respect to north and the dip of the plane with respect to the horizontal. Unmanned aerial vehicles (UAVs) can be utilized to capture hundreds of photos to create high-resolution 3D models of a rock outcrop, capturing visible joint set discontinuities on a faceted surface of a triangular irregular network (TIN). Computing methods of facets and facet normals from a point cloud allow for the characterization of discontinuity orientations without the need for manual measurements in the field. However, these methods used to calculate facets and facet normals result in the addition of noise in the dataset, which increases the difficulty of analysis. A two-stage filtering process employing density-based spatial clustering of applications with noise (DBSCAN), and the second derivative of the remaining clusters, removes data that are not representative of a discontinuity within the intact rock mass. Finally, an unsupervised clustering algorithm, K-Means Clustering, is applied to the dataset to extract the dip and dip direction of discontinuities. The methodology used to identify the orthogonal joint sets on a dam spillway demonstrated good performance across all identified joint sets, with the estimated variability in dip and dip direction broadly comparable to that observed in the manual field dataset. This indicates that this newly developed proof-of-concept approach can reliably capture the orientation and variability of orthogonal joint sets from large datasets.