Photogrammetry and artificial intelligence for rock mass classification

Rock mass classification is largely based on manual mapping, which is slow, subjective, and requires working close to exposed wall that can be hazardous. This thesis asks whether photogrammetrical three-dimensional point clouds, processed with artificial intelligence methods, can produce rock mass classification inputs and results that are accurate enough to match field observations.
A 31.5-million-point cloud data of a 2.6 m x 4.6 m drift wall at Under-ground Research Laboratory of Aalto University was processed at multiple point densities to assess robustness. Surface normals were estimated and clustered using artificial intelligence-based methods (K-means and mean-shift) to identify discontinuity sets, and planar patches were fitted to sup-port spacing calculations. Rock quality designation (RQD) derived from the point cloud processing was 78.9% and closely matched manual scanline mapping (77.5%). Joint roughness was estimated using small planar surface patches and converted to the joint roughness number (Jr) via the joint roughness coefficient normalized to 20 cm (JRC20).
Validation against manual measurement showed that the three dominant joint sets were reproduced within a few degrees; mean-shift provided cleaner results of minor sets than K-means; while K-means is quicker and simpler but contains noise. Using parameters obtained from the 3D point cloud, the drift face was classified as Q≈16 (good), RMR ≈ 64 (Class II, Good), and GSI ≈ 59 - 64, consistent with field assessment, while RMi deviated more due to its sensitivity to spacing through block volume estimation. Limitations were point cloud coverage and the presence of blast-induced fractures that could not be identified automatically. Overall, photogrammetry combined with artificial intelligence methods is a feasible way to obtain rock mass classification inputs from tunnel face 3D point clouds.