Virtual outcrop models of geological structures: problems and best practices related to extraction of 3D structural data

The advancement of computer vision–based photogrammetric image processing pipelines, particularly Structure from Motion–Multi-View Stereophotogrammetry (SfM-MVS), has rapidly evolved. This evolution, coupled with the accessibility of low-cost and portable acquisition tools such as DSLR and mirrorless cameras, Uncrewed Aerial Vehicles (UAVs) and smartphones, has transformed outcrop studies in structural geology, propelling traditional field geology into the digital era. Notably, this revolution has significantly impacted Virtual Outcrop Models (VOMs), elevating them from mere visualization media to fully interrogable quantitative objects. 

Among the various applications of VOMs in structural geology, the extraction of near-planar features, including fracture and bedding surfaces, stands out as crucial. Numerous procedures exist for this purpose, ranging from fully automated segmentation and best-fitting of point clouds to the manual picking of 3D traces on both point clouds and textured meshes.

In this work, we explore the advantages, disadvantages, best practices, and drawbacks associated with the principal procedures for extracting near-planar geological data from VOMs. While automated or supervised recognition and subsequent best-fitting of coplanar patches in point clouds have garnered significant attention, their application is generally limited to specific case studies. Geological outcrops commonly lack patches of sufficiently large near planar surfaces for robust best fitting, necessitating manual picking procedures based on visual and/or structural interpretation. In such cases, the use of textured meshes is preferred over point clouds, and consideration must be given to the accuracy of the textured mesh during digitization, as well as the intrinsic roughness of geological surfaces. 

The analysis of coplanarity and collinearity of picked point sets aids in identifying traces deviating from idealized configurations. However, commonly suggested threshold values often result in small datasets. Nevertheless, relying on the visual inspection of the best-fit plane and real-time computation of best-fit planes from picked point sets generally yields acceptable results, handling coplanarity and collinearity dynamically during the extraction process.