Machine learning reads pollen images to identify honey sources in minutes

A research team now reports that pollen grains can be identified with a faster, more automated approach: multispectral imaging flow cytometry combined with machine learning. Tested on real honeys, the method produced pollen identification results comparable to traditional microscopic analysis while allowing three times more pollen to be evaluated and requiring a maximum measurement time of ten minutes.

The study, published in npj Science of Food, addresses melissopalynology, the identification of pollen from honey. Pollen identification carries great economic and consumer relevance. Traditional microscopy, however, is time-consuming and dependent on expert knowledge, which limits high-throughput analyses. The researchers therefore tested whether pollen images generated with multispectral imaging flow cytometry, classified by machine learning, could serve as a fast and reproducible method for identifying pollen in honey.

The machine learning relied on convolutional neural networks and random forest classifiers. The team examined how classification is affected by images from different matrices, comparing anther samples with actual honey samples. They also compared the imaging flow cytometry results with traditional microscopic analysis and used human-interpretable image features to better explain what the models were doing.

Six common target pollen types in honey formed the focus: Brassica, Castanea, Centaurea, Helianthus, Phacelia and Tilia type.

The best performance came from a classifier trained with a combination of anther and honey images in both brightfield and fluorescence channels. This combined model provided the best predictions for a test set of 13 monofloral and 10 polyfloral honeys.

The automated approach yielded results comparable to microscopy, meaning the classifications agreed with the expert-based method on the honeys tested.

The authors conclude that the method demonstrates high potential as a powerful screening tool in melissopalynology. The study focused on six target pollen types and 23 honeys, so broader validation across more pollen types and honey origins remains a natural next step.

Subject of Research: Using multispectral imaging flow cytometry and machine learning to identify pollen in honey as a fast alternative to traditional microscopy.

Article Title: Profiling of pollen in honey using multispectral imaging flow cytometry

Article References: Walther, F., Hofmann, M., Janke, M., Erler, S., Friedrich, E., Treder, M., Rakosy, D., Harpole, W. S., Mäder, P., Berdún, A., Hornick, T., & Dunker, S. (2026). Profiling of pollen in honey using multispectral imaging flow cytometry. npj Science of Food, 10(1), Article 337. https://doi.org/10.1038/s41538-026-01188-6

Image Credits: AI Generated

DOI: 10.1038/s41538-026-01188-6

Keywords: honey, pollen, melissopalynology, multispectral imaging flow cytometry, machine learning, convolutional neural networks, random forest, monofloral honey, brightfield imaging, fluorescence, food authentication, npj Science of Food