A U-Net–Based Segmentation Approach for Honeybee Pollen Detection Using Deep Learning

This research article applies U-Net to segment pollen carried by honeybees in the public Pollen Dataset. The 714 images include 368 pollen-bearing bees and 346 without pollen. The authors manually annotate pollen regions, generate binary masks and use a random 80/20 split with 571 training images and 143 held-out test images. They report an F1 score of 99.4%. The study demonstrates a compact image-analysis approach for pollen-load monitoring, but does not establish transfer to unseen apiaries, camera systems or continuous entrance video.

Publication details

Authors
Esra Yılmaz, Ali Sarıkaş
Organizations
Marmara University, İstanbul, Türkiye
Year
2026
Type
Journal-Article

Publication and access

  • Published 30 September 2026 in Dokuz Eylül Üniversitesi Mühendislik Fakültesi Fen ve Mühendislik Dergisi, volume 28, issue 84, pages 443-448; publisher category: Research Article.
  • DOI · Publisher record · Open publisher PDF
  • Author names, affiliations and publication date agree between the publisher record, Crossref and the PDF.
  • The publisher states Creative Commons Attribution-NonCommercial 4.0. The six-page PDF was downloaded and parsed for verification, but is not mirrored on Gratheon's product website because commercial redistribution permission was not established.

The abstract above is an editorial summary, not a verbatim author abstract.

Relevancy to Gratheon

Pollen-load segmentation is directly relevant to entrance-camera estimates of incoming pollen-bearing foragers. The manual-mask workflow also illustrates how an existing image-classification collection can be extended into a segmentation task.

The reported F1 is an in-dataset result, not an estimate of field performance. The paper uses an image-random split rather than a demonstrated apiary-, camera- or video-disjoint evaluation. Segmentation scores and whole-bee pollen-presence classification scores should not be treated as interchangeable when comparing methods. No release of the new masks or trained checkpoint was verified. A Gratheon evaluation should use independent entrance video and separately measure missed pollen carriers, false positives and mask quality.