Improving micromorphological analysis with CNN-based segmentation of flint/obsidian, bone and charcoal

Area of Science:

• Archaeological Science
• Geoarchaeology
• Digital Archaeology

Background:

• Archaeological micromorphology analysis is often subjective and challenging.
• Accurate quantification of materials like bone, charcoal, and lithics is crucial for understanding past environments and human activities.

Purpose of the Study:

• To develop and evaluate a deep learning tool for the automatic segmentation and quantification of key materials in archaeological thin sections.
• To improve objectivity, accuracy, and reproducibility in the analysis of archaeological micromorphology.

Main Methods:

• Utilized high-resolution photomicrographs of 57 thin sections.
• Trained and evaluated state-of-the-art convolutional neural networks (CNNs), specifically a U-Net with an InceptionV4 encoder.
• Applied the models to segment and classify the relative abundance of flint/obsidian, bone, and charcoal.

Main Results:

• The U-Net with InceptionV4 encoder achieved high mean intersection over union (IoU) scores: 0.96 for flint/obsidian, 0.80 for bone, and 0.82 for charcoal.
• Balanced accuracies for material abundance classification were excellent: 0.99 for flint/obsidian, 0.92 for bone, and 0.85 for charcoal.
• Demonstrated the effectiveness of deep learning in material segmentation and quantification.

Conclusions:

• Deep learning offers a powerful solution to enhance objectivity and accuracy in archaeological micromorphology.
• The developed tool provides a valuable resource for geoarchaeological research, enabling more reliable analysis of Palaeolithic contexts.
• Automated segmentation and quantification can significantly improve reproducibility in the field.