Automated Lithology Segmentation of 3D Point Clouds from Highwalls Using Deep Learning

Digital twins are increasingly being adopted to support efficient hazard assessment and predictive modelling. A key prerequisite for the development of reliable digital twins in rock slope engineering is the accurate identification and segmentation of the rock lithology, as material properties significantly influence rock mass behaviour and govern the occurrence and severity of geotechnical hazards. Traditional methods for characterising rock lithology, primarily based on visual interpretation and borehole data, can be labour-intensive, subjective, and unsuitable for large-scale, dynamic environments such as mining operations. To address this gap, this paper proposes the application of deep learning-based 3D point cloud segmentation models to automate the segmentation of rock lithology in open-pit mine highwalls. Four different segmentation models (SparseUNet, Point Transformer version 2, Point Transformer version 3, Sonata) are explored and their performance in efficient and accurate rock lithology segmentation is evaluated using high-resolution 3D point clouds. The models are trained using a mine highwall dataset consisting of 1498 point cloud segments generated using terrestrial and aerial photogrammetry of nine open-pit mine sites from Australia. The data is split into an 80:20 ratio for training and validation purposes. The results show that Point Transformer version 3 outperforms Sonata, Point Transformer version 2 and Saprse UNet by 21%, 26% and 55%, respectively, measured by mean intersection over union on the unseen validation dataset split. The outcomes of this work provide the basis for developing lithology-informed digital twins of rock slopes, enabling data-driven rockfall hazard assessment, predictive monitoring, and sustainable slope management.