Automated Borehole Core Analysis with Report-Derived Weak Labels and Supervised Crack Segmentation

Geotechnical engineering relies on Rock Mass Classification Systems (RMCS) to inform tunnel support and excavation planning. These systems depend on precise measurements of crack spacing, bedding angles, and lithology. Traditionally, geologists manually inspect core tray photographs and assign these values based on professional judgment, a process that is time-consuming, subjective, and prone to inter-observer variability. This paper addresses the need for an automated, reproducible pipeline to extract these metrics from existing borehole archives.

The authors propose a dual-pathway system that processes two primary artifacts: high-resolution core tray photographs and digital log reports.

1. Crack Segmentation: The team manually annotated 5,087 core-row images to train a gated U-Net. This architecture integrates PiDiNet edge maps with Mask R-CNN outputs through a spatial gating mechanism to isolate thin, elongated cracks from complex rock textures.
2. Auxiliary Analysis: Parallel rule-based branches process the same images to estimate core-relative bedding angles and lithological color descriptors.
3. Weak Supervision: For scenarios lacking pixel-level annotations, the authors demonstrate a classification approach using structured spacing categories extracted directly from the text layer of digital log reports as weak labels.

The gated U-Net achieved an F1 score of 0.860 and a crack-class IoU of 0.754, outperforming other evaluated segmentation configurations. The rule-based branches for bedding angles and lithological color showed 75.4% and 84.7% agreement with log-report references, respectively. This framework is significant because it transforms legacy, annotation-free archives into quantitative datasets, reducing the subjectivity inherent in manual logging and providing a scalable method for geotechnical data processing.