Point Clouds and Machine Learning Applications in Rock Slope Modelling

Thumbnail Image
Farmakis, Ioannis
Point cloud , Rock slope , Machine learning , 3D Modelling , Rockfall , Monitoring , Susceptibility , LiDAR , Photogrammetry
The analysis of 3D point cloud data has been a valuable practice in rock slope modelling, with increasing development and acceptance starting in the early 2000’s. We use point clouds as a reference for geometric rock mass characterization, as terrain models for physical simulations and monitoring, and as a digital asset for remote visual inspections, amongst other applications. However, their analysis is often limited to daunting manual processing for the extraction of meaningful information. If we consider point clouds as digital reality assets, a whole new field of intelligent applications becomes available with the integration of Machine Learning (ML), following the lead of industries such as automotive and robotics engineering. Given the ever-growing multi-temporal digitization of rock slopes, rock slope management can also benefit from the digital transformation. This thesis explores the potential of 3D Computer Vision (CV) applications in the monitoring and modelling of rock slope hazards along transportation corridors. CV has recently found applications in rock slope research for – mainly – satellite image analysis but these applications do not fully consider the 3D complexity of the natural rock slopes, nor facilitate data-driven modelling at site-specific scales as 3D point clouds do. We address questions related to data structures and representation, prove the superiority of fully 3D processing of digital rock slope models for the extraction of valuable information related to rock slope management and propose a flexible object-oriented methodology for tailored knowledge extraction. We also tackle the integration of supervised learning in automating rockfall monitoring with point cloud classification neural networks and present a novel approach to slope-scale rockfall susceptibility modelling (RSM) as a point cloud semantic segmentation (PCSS) problem. The findings highlight the potential of multi-temporal digital monitoring data by providing evidence that AI with engineering geological perception can be developed with supervised ML and point clouds in rock slope modelling.
External DOI