GeoCrack: A High-Resolution Dataset For Segmentation of Fracture Edges in Geological Outcrops

Area of Science:

• Geology
• Computer Science
• Artificial Intelligence

Background:

• Characterizing natural fractures in geological outcrops is crucial for subsurface analysis.
• Existing datasets lack the scale and annotation detail for advanced deep learning applications.
• Challenges in fracture segmentation include variations in lithology, weathering, and occlusions.

Purpose of the Study:

• Introduce GeoCrack, the first large-scale, open-source annotated dataset for fracture trace segmentation.
• Enable and standardize deep learning-based fracture characterization.
• Address real-world complexities in annotating geological fracture data.

Main Methods:

• Collected photogrammetric images from 11 diverse geological sites.
• Manually segmented fracture traces with a rigorous annotation vetting process.
• Created 12,158 image-mask pairs (224x224 pixels) accounting for geological and environmental factors.

Main Results:

• GeoCrack dataset comprises 12,158 annotated image pairs of fracture traces.
• The dataset captures diverse geological conditions and annotation challenges.
• A U-Net model achieved an 85% Intersection over Union (IoU) for fracture segmentation.

Conclusions:

• GeoCrack sets a new standard for natural fracture characterization datasets.
• The dataset facilitates robust deep learning models for geological fracture identification.
• It effectively addresses diverse real-world challenges in fracture segmentation.