Deep Learning and Pollen Detection in the Open World

Summary

This is an abstract from the "Advances in Macrobotanical and Microbotanical Archaeobotany, Part II" session, at the 89th annual meeting of the Society for American Archaeology.

Pollen-based paleoecological reconstructions rely on visual identifications that can be automated using computer vision. To date, most automated approaches have focused on taxonomic classification of pollen in cropped images. There are fewer protocols for pollen detection (i.e., localization) in whole-slide images. New samples potentially introduce rare and novel taxa, making pollen detection in the open world—a world where we constantly encounter new diversity—challenging. We explored pollen detection in the open world by focusing on three significant challenges. We first addressed taxonomic bias—missed detections of smaller, rarer pollen types. We fused an expert model trained on this minority class with our general pollen detector. We next addressed domain gaps—differences in image magnification and resolution across microscopes—by fine-tuning our detector on images from a new imaging domain. Lastly, we developed continual learning workflows that integrated expert feedback and allowed detectors to improve over time. Effective pollen detectors enable higher and more reproducible pollen counts that could improve the accuracy of diversity estimates and accelerate the creation of long-term, high-resolution paleoecological records. Our methods can be applied to other visually diverse biological data, including algae, fungal spores, and phytoliths.

This Resource is Part of the Following Collections

• Advances in Macrobotanical and Microbotanical Archaeobotany, Part II   •
• Society for American Archaeology 89th Annual Meeting, New Orleans, LA (2024)

Keywords

Record Identifiers

Abstract Id(s): 41664.0