Overview
This research investigates automated methods for extracting defect-spacing information from borehole core archives. The approach leverages existing core tray photographs and corresponding digital log reports, addressing the common absence of native pixel-level crack annotations. Two complementary strategies are employed: one utilizing structured spacing categories from report text for weak interval-level classification, and another implementing fully supervised crack segmentation on core-row images.
Research Context
Borehole archives typically contain visual data in the form of core tray photographs alongside textual digital log reports. A significant challenge in utilizing these archives for detailed analysis is the lack of explicit pixel-level annotations for defects like cracks. This absence necessitates methods that can infer or identify such features from available data, particularly for deriving quantitative information such as defect spacing.
Approach
The research integrates two primary approaches to extract defect-spacing information.
Weak Labeling and Classification
- Structured spacing categories are extracted from the text layer of digital log reports. These categories serve as weak interval-level labels for classification tasks.
- A DINO encoder, trained on unlabeled core crops, is used to generate domain-specific representations.
- A subset of data is manually verified to identify and address label inconsistencies arising from the report-derived weak labels.
Supervised Crack Segmentation
- A dataset of 5,087 extracted core-row images is manually annotated to create ground truth for crack segmentation.
- Multiple fully supervised crack-segmentation models are evaluated.
- A specific architecture, termed a gated U-Net, is developed. This model integrates PiDiNet edge maps with Mask R-CNN masks through a learned spatial gating mechanism.
- Post-processing: Predicted crack locations from the segmentation models are converted into defect-spacing categories using deterministic rules.
Ancillary Rule-Based Estimators
Separate rule-based branches are developed to estimate additional geological parameters:
- Core-relative bedding angles: These estimations show agreement with log-report references on 75.4% of 1,200 evaluated images.
- Lithological color descriptors: Predictions for these descriptors agree with log-report references on 84.7% of 1,200 evaluated images.
The reported agreement values for these ancillary estimators measure concordance with recorded geological observations in existing reports, rather than independent physical accuracy.
Findings
Crack Segmentation Performance
The gated U-Net configuration demonstrated the highest performance among the evaluated segmentation configurations:
- F1 score: 0.860
- Crack-class Intersection over Union (IoU): 0.754
Ancillary Estimator Agreement
- Core-relative bedding angle estimations agreed with log-report references for 75.4% of images.
- Lithological color descriptor predictions agreed with log-report references for 84.7% of images.
These agreement figures are based on comparison with existing report data, not independent physical measurements.
Why This Matters
The developed framework offers a method to derive quantitative defect-spacing information from widely available borehole core photographs and digital log reports without requiring native pixel-level crack annotations. By combining report-derived weak supervision with image-based crack localization, the system addresses a critical gap in automating the analysis of existing geological archives. The ability to estimate parameters such as bedding angles and lithological color descriptors further augments the utility of these archived visual and textual datasets for geological interpretation.