Instance Segmentation for Geotechnical Core Analysis in Mining

Geotechnical core logging lies at the foundation of mine design, safety, and planning, yet it remains laborious, subjective, and prone to methodological lock-in. Meanwhile, core box images represent a prevalent and largely untapped primary data source. Recent studies have shown that these images can be used to derive geotechnical metrics such as RQD by training an instance segmentation algorithm, a deep learning method that detects objects in an image and distinguishes individual instances so that each core piece can be identified and measured. This thesis presents the first zero-shot approach to instance segmentation by using Meta¿s Segment Anything Model (SAM), which requires no retraining and offers the potential for rapid deployment across sites. It evaluates the technical performance and the practical realities of integrating this novel approach when applied to legacy datasets. This demo was created as a QA/QC tool to verify manual RQD logs. The methodology applies image preprocessing, SAM inference, and post-processing to a pre-existing picture database at a major European base metals mine. Using this mine as a case study allows us to contextualize this methodology by considering the complex operational constraints of an active mining environment, including the need for compatibility with existing workflows. It also allows assessment of the data quality in legacy datasets. The methodology accurately segmented intact core pieces, achieving a MAE of 7.7 percentage points, thus demonstrating alignment between automated and human-derived results. Outliers demonstrated shortcomings in the input data and errors during the manual logging process, rather than errors with the application. Actual errors appeared due to failure to detect core pieces (false negatives), failure to separate multiple core pieces (undersegmentation) and the inability to filter out certain artefacts (false positives). This work demonstrates that SAM-based segmentation can form a practical foundation for extracting fracture data from core box photographs, when combined with filtering steps based on individual and relational mask geometry and targeted removal of unwanted masks. An important limitation to computer vision workflows such as this one is the inability to distinguish natural from artificial fractures, which must be marked by a human operator. Improving the quality and consistency of core box photographs would not only enhance current segmentation accuracy but also increase their long-term value as AI advances. Some obvious improvements to images include consistent lighting, consistent marking of core loss, and ensuring clear and complete visibility of depth tags. This thesis represents a significant advance in applying instance segmentation to geotechnical core box images, providing key insights into practical use cases, limitations, and requirements. Unlike approaches that require annotating over a thousand images and training a site-specific model, the zero-shot SAM method works without retraining, yet still delivers accurate segmentation from standard core photographs. Future research could investigate automated natural fracture detection, box detection and row detection and investigate improved segmentation performance. In its current form, SAM-based instance segmentation modules can already be applied in operational use cases to increase data availability, support more consistent logging, and reduce some of the uncertainties inherent in current core logging practices.