Automated identification of sedimentary structures in core images using object detection algorithms.

Area of Science:

• Geology
• Sedimentology
• Artificial Intelligence

Background:

• Manual interpretation of sedimentary structures in core analyses is crucial for subsurface geology but is slow, requires expertise, and can be biased.
• Automating this process can significantly improve efficiency and consistency in geological studies.

Purpose of the Study:

• To investigate the application of convolutional neural networks (CNNs) for automated identification of sedimentary structures in core images.
• To compare the performance of two object detection models, YOLOv4 and Faster R-CNN, for this task.

Main Methods:

• Training YOLOv4 and Faster R-CNN models on annotated datasets of siliciclastic core images representing 15 sedimentary structure types.
• Evaluating model performance based on precision, recall, inference time, and mean average precision.
• Testing model generalization on previously unseen datasets.

Main Results:

• YOLOv4 demonstrated high precision (up to 95%) and faster processing times compared to Faster R-CNN.
• Faster R-CNN achieved a higher mean average precision (94.44%) but had lower recall for common structures.
• Both models struggled with morphologically similar structures and showed slightly reduced performance on unseen data.

Conclusions:

• Deep learning, particularly using CNNs like YOLOv4, offers a promising approach to automate core interpretation in sedimentology.
• This automation can reduce manual effort, enhance reproducibility, and streamline geoscientific applications.
• Further development is needed to improve generalization across diverse core imagery and distinguish subtle structural variations.