Acoustic Honeybee Queen-State Detection Under Unseen Conditions

This preprint benchmarks acoustic representations for detecting queen absence using 5,129 recordings from 3,285 hives across 47 apiaries, collected between October 2024 and August 2026. Task-specific convolutional networks using modulation spectrograms are compared with conventional acoustic features and pretrained audio transformers. The 2D modulation-spectrogram model obtains AUROC 0.81 and queenless-class AUPRC 0.37 on unseen hives, but only AUROC 0.63 and AUPRC 0.26 on unseen apiaries. The evaluation demonstrates both the promise of non-invasive acoustic screening and the substantial difficulty of transferring models across apiaries.

Download PDF

Publication details

Authors
Mahsa Abdollahi, Nico Coallier, Maxime Fraser Franco, Tiago H. Falk
Organizations
INRS-EMT, Université du Québec, Montréal, CanadaNectar Technologies, Montréal, Canada
Year
2026
Type
Preprint

Publication and access

  • arXiv, version 1, posted 29 September 2026 (eess.AS). Preprint; not peer reviewed.
  • DOI · Version-specific source · Original PDF
  • Affiliations above are taken from the PDF, rather than inferred from author-name matching.
  • License: Creative Commons Attribution 4.0, as stated on the arXiv record. The local PDF is an unmodified copy, attributed to the four authors above.

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

Relevancy to Gratheon

This is an actionable benchmark for acoustic colony-health alerts. Splitting recordings by hive is not enough: a separate apiary-held-out test exposes a much larger generalization gap. Evaluation should include precision-recall metrics for the uncommon queenless class, rather than presenting AUROC as detection accuracy.

For the 2D modulation model, queenless precision/recall changes from 0.30/0.67 on unseen hives to 0.42/0.27 on unseen apiaries at the reported operating points. These results support screening followed by beekeeper confirmation, not an autonomous diagnosis. The paper compares representations on a large operational dataset, but no public dataset or implementation release was verified in this run.