Abstract
Accurate mapping of structural features from exposed rock surfaces is a fundamental part of rock engineering work. Traditional methods are often labor-intensive and prone to subjectivity. With the increasing availability of high-resolution 3D data from terrestrial laser scanning (LiDAR), automated methods have become essential for characterizing discontinuities over large and complex rock surfaces. This study presents a scale-dependent and automated workflow for structural mapping from point cloud data, integrating normal vector orientation and surface roughness. Given that noise and edges typically exhibit higher roughness values, roughness was used as a filtering tool to extract meaningful geological features. An unsupervised k-means clustering algorithm was employed in three configurations: (1) using orientation individually as clustering features, (2) using both features together, and (3) applying a recursive approach where roughness-based filtering is followed by orientation-based clustering. A range of neighborhood sizes was tested to evaluate the effect of scale on roughness and normal estimation. Results show that the recursive approach provides the clearest cluster separation, improved stereonet concentration, and better alignment with manually mapped structures. The methodology offers a reliable and scalable solution for automated structural analysis in geotechnical and geological applications.