Profiling of pollen in honey using multispectral imaging flow cytometry

Pollen identification from honey (melissopalynology) is of great economic and consumer relevance. Traditional microscopy is time-consuming and dependent on expert knowledge, thus limiting high-throughput analyses. We tested whether pollen images generated with Multispectral Imaging Flow Cytometry (MIFC) classified by machine learning (ML) (convolution neural networks (CNNs) and random forest (RF)) can serve as a fast and reproducible method for identifying pollen in honey. We examined how the classification of honey pollen is affected by images from different matrices (anther and honey samples), how results compare with traditional microscopic analysis and used a set of human-interpretable image features to better explain ML results. The study focused on the six common target pollen in honey Brassica, Castanea, Centaurea, Helianthus, Phacelia and Tilia type. A classifier trained with a combination of anther and honey brightfield and fluorescence images provided the best predictions for 13 monofloral and 10 polyfloral honeys. Yielding pollen identification results comparable to microscopy, three times more pollen can be evaluated, and a measurement time of a maximum of 10 min, the method demonstrates its high potential as a powerful screening tool in melissopalynology.