Speaker
Description
Fossil pollen analysis is a key tool for reconstructing past vegetation and climate, but conventional microscopic identification and counting are time-consuming, require substantial taxonomic expertise, and can be affected by observer bias. We develop an automated workflow combining multispectral imaging flow cytometry (MIFC) with convolutional neural networks (CNNs) for high-throughput and reproducible identification of fossil pollen from sediment samples. MIFC rapidly acquires multispectral images of individual particles, providing the basis for CNN-based image recognition and taxonomic classification.
Our workflow follows a hierarchical classification strategy. A first CNN distinguishes pollen grains from non-target sediment particles, including charcoal, debris, and Lycopodium marker spores. A subsequent CNN assigns detected pollen grains to taxonomic classes. As a palaeoecological test case, we apply this approach to fossil pollen samples from Lake Emanda, eastern Siberia, which provide a Late Quaternary record with conventional pollen counts for validation of automated classification.
Currently, we compare CNN-classified pollen abundances with conventional microscopic pollen counts at family level to assess quantitative agreement between approaches. In parallel, image-feature analyses investigate factors limiting classification performance, including particle orientation, focal plane, and image quality. The resulting benchmark image datasets will be curated and stored using OMERO-based image data management, supporting standardized metadata, accessibility, and future model development.
Ultimately, we aim to develop a CNN-based system that supports or partially replaces microscopic pollen identification and counting, enabling faster, scalable, and more reproducible reconstructions of past vegetation and climate.