AI-Powered Semi-Automated Trace Detection Techniques on 3D Digital Rock Mass Models

Estimating rock mass structure relies heavily on deciphering trace lines on exposed rock surfaces. Manual detection of traces on 3D digital models can be tedious and time-consuming. This study proposes a novel AI-driven approach to enhance 3D trace surveying, overcoming potential limitations. A custom pixel-wise texture mapper algorithm transforms 3D rock models into virtual 2D images, eliminating access restrictions. DeepLabV3+, a deep-learning network, efficiently detects trace lines within these virtual images. The detected traces are transformed into 3D data through a multi-step process involving point cloud integration, thinning and segmentation, and re-projection. This meticulous reconstruction culminates in a 3D point cloud representing a trace network. This AI-powered method promises several advantages over traditional methods. Automating trace detection and minimizing human subjectivity enhances accuracy and reduces surveying time. Additionally, virtual image generation allows for surveying difficult-to-access outcrops or overcoming camera positioning challenges. Beyond the proposed methodology, the study will compare the performance of AI and human experts in trace pattern detection. Furthermore, it will explore potential applications of this technology in various geological contexts, paving the way for future advancements in rock structure analysis. 1. INTRODUCTION Manual rock discontinuity surveys, typically relying on handheld devices, face inherent drawbacks: difficulty, time consumption, safety hazards, and surveyor subjectivity (Ferrero et al., 2009; Franklin et al., 1988; Priest & Hudson, 1976). In response, remote surveying techniques emerged, generating 3D models through dense point clouds or textured mesh models (Assali et al., 2014). Recent advancements in computational methods have empowered the analysis of these models, promising a more accurate and detailed understanding of rock mass geometry (Battulwar et al., 2021). Algorithms leveraging 3D rock mass models offer solutions to traditional survey limitations, including challenging access, lengthy procedures, and surveyor bias. They enable significantly increased measurement volumes on rock outcrops, covering inaccessible areas with minimal effort (Ferrero et al., 2009). Moreover, automated analysis of rock faces using 3D models presents cost-effectiveness in terms of operational expenses and time compared to traditional surveying (Slob et al., 2005).