Striking the Perfect Balance: Employing a Hybrid Approach to Rapidly Generate Phytolith Training Data

Summary

This is an abstract from the "From the Agricultural to the AI Revolution: Analytical Advances in Paleoethnobotany" session, at the 90th annual meeting of the Society for American Archaeology.

Manual identification of phytoliths is time consuming and prone to error. Deep Learning (DL) algorithms are transforming this methodology, significantly reducing analysis time and increasing accuracy. However, developing an effective DL workflow to identify phytoliths depends on the manual acquisition of hundreds to thousands of high-quality images for training data. We must strike a balance between obtaining lots of high-quality images (balancing file size associated with full slide images) and minimizing the human time needed to capture those images (balancing time to annotate and crop individual phytoliths within full slide images). Therefore, integrating automation with expert input is essential. Employing a Nikon Eclipse Ni-E microscope and a Digital Sight 10 camera, we have opted to use lower magnification to first obtain full slide scans of ashed comparative samples. We then manually pinpoint areas for automated high magnification Z-stacks or single plane images, which also minimizes cropping time. This hybrid approach, engineered with Nikon Imaging Specialists, is facilitating the rapid generation of critical high quality training data, and this system will be adapted to form part of the labs developing DL workflow system, which is set to revolutionize the scale and accuracy of archaeological phytolith analysis.

This Resource is Part of the Following Collections

• Society for American Archaeology 90th Annual Meeting, Denver, CO (2025)   •
• From the Agricultural to the AI Revolution: Analytical Advances in Paleoethnobotany

Keywords

Record Identifiers

Abstract Id(s): 51238