Deep Learning Based Detection of Wild Bee Parasites under Natural Conditions

Svetlana Ionova · Henri Greil · Patrick Mäder · Marco Seeland

Data

The downloaded urls are available in data/filtered_obs.csv and data/filtered_inat.csv (cleaned). The list of added healthy bee images is available in data/healthy_bees.csv.

Composite Images

Parasites segmented from the images mentioned above comprised the foregrounds. The bee images which were used as backgrounds for composite image generation can be found in data/background_bees.csv.

Use Case Evaluation

We reserved 2500 images for the use case evaluation which are listed in data/use_case_bees.txt.

Training

Model can be trained using train.py. We started from OpenImages-v7 weights for YOLOv8m. Additionally, we modified the training script ultralytics/engine/trainer.py from the Ultralytics' YOLO repository (version 8.3.176) to progressively unfreeze the model. Our updated version can be found in yolo/trainer.py

Citation

@article{Ionova2026,
  title = {Deep learning based detection of wild bee parasites under natural conditions},
  volume = {95},
  ISSN = {1574-9541},
  url = {http://dx.doi.org/10.1016/j.ecoinf.2026.103754},
  DOI = {10.1016/j.ecoinf.2026.103754},
  journal = {Ecological Informatics},
  publisher = {Elsevier BV},
  author = {Ionova,  Svetlana and Greil,  Henri and M\"{a}der,  Patrick and Seeland,  Marco},
  year = {2026},
  month = May,
  pages = {103754}
}