Han, Z, Schwartzkopff, A, Mathew, TJ, Moridi, M & Akdag, S 2026, 'An unsupervised manifold learning framework for automated identification of 
rock discontinuities in underground mines using raw LiDAR point cloud', in D Cumming-Potvin & J Wesseloo (eds), Ground Support 2026: Proceedings of the 11th International Conference on Ground Support in Mining, Australian Centre for Geomechanics, Perth, pp. 1-11, https://doi.org/10.36487/ACG_repo/2625_27
(https://papers.acg.uwa.edu.au/p/2625_27_Han/)
Abstract: Accurate mapping of rock discontinuities is essential for rock mass characterisation, stability analysis and ground-support design. Traditional manual approaches are time-consuming, labour-intensive, and subjective in interpretation. Digital approaches based on photogrammetry and Light Detection and Ranging (LiDAR) point cloud methods provide efficient tools for solving the problem. However, photogrammetry-based methods are challenging due to the low-light and dusty underground environments, making LiDAR point cloud data a superior tool to the traditional approach, which can actively emit lasers and is therefore independent to the illumination conditions. Supervised deep learning-based methods are widely adopted for 3D point cloud analysis in underground constructions due to their robust performance. Although the performance of these models is enhanced by adopting advanced deep learning strategies, the application and generalisation ability of these methods are limited by the supervised manner, which requires extensive labelled datasets. Therefore, this study proposes a novel unsupervised deep neural network-based method (ManiRock) for automatic rock discontinuities identification from LiDAR point clouds. We integrate a state-of-the-art unsupervised implicit learning method and a widely used clustering method. ManiRock directly takes the raw point cloud as input and constructs a continuous implicit representation of the rock surface manifold and identify and isolate each discontinuity subsequently in a two-step manner. This study provides a case study as a geometric proofofconcept using limited unsupported underground tunnel rock mass point clouds. The method is tested using LiDAR point cloud datasets acquired from a metal mine located in Kalgoorlie, Western Australia. The results are compared with ground-truth results derived from manual mapping and demonstrate that ManiRock achieves high-fidelity discontinuity mapping while bypassing the need for data labelling, providing a potentially useful tool for rapid geotechnical assessment of underground rock mass.

Keywords: rock mass discontinuities, LiDAR, unsupervised manifold learning