Multispectral Imaging Flow Cytometry for Spatio-Temporal Pollen Trait Variation Measurements of Insect-Pollinated...

Artificial intelligence (AI) surpasses human accuracy in identifying ordinary objects, but it is still challenging for AI to be competitive in pollen grain identification. One reason for this gap is the extensive trait variation in pollen grains. In classical textbooks, pollen size relies on only 25-50 pollen grains, mostly for one plant and site. Lack of variation in pollen databases can cause limited application of machine learning approaches to real-world samples. Therefore, our study aims to investigate sources of spatial and temporal pollen trait variation for pollen morphology and fluorescence. For this purpose, 64,001 pollen grains from the four herbaceous and insect-pollinated plant species Achillea millefolium L., Lamium album L., Lathyrus vernus (L.) Bernh., and Lotus corniculatus L. sampled across four years and seven locations across Central Germany were measured using multispectral imaging flow cytometry. Observed trait variations were very species-specific; however, for most species, significant differences in spatial as well as temporal variation were found for at least one pollen trait. We could also show that this variability and the identity of a particular sample influence the accuracy of AI classifications and that multiple measurements of different origins provide the most robust AI-based identifications.