Research library

Improved Monitoring of Honey bee Colony Strength via Audio IoT Sensors, Modulation Tensorgrams and Recurrent Neural Networks

This work estimates honey bee colony strength from remotely recorded hive audio while preserving temporal information that conventional modulation-spectrum features discard. The authors introduce modulation tensorgrams and compare two-dimensional and three-dimensional CNN, attention, and CRDNN architectures using more than 3,000 hours of public UrBAN audio; the colony-strength experiment uses recordings from nine hives. Under hive-independent cross-validation, the best CRDNN-3D model achieved a mean absolute error of 3.31 frames of bees and a correlation of 0.78, improving on the prior random-forest benchmark. Saliency and Grad-CAM analyses indicate that temporal modulation dynamics contribute to the predictions.

Download PDF

Publication details

Authors
Mahsa Abdollahi, Yi Zhu, Heitor R. Guimarães, Nico Coallier, Ségolène Maucourt, Pierre Giovenazzo, Tiago H. Falk
Organizations
🇨🇦 Institut national de la recherche scientifique (INRS)🇨🇦 Université Laval🇨🇦 Nectar Technologies Inc.
Year
2026
Type
Preprint

Relevancy to Gratheon

This paper provides a direct acoustic route to a product-relevant colony metric rather than only classifying isolated events. Estimating frames of bees continuously could help prioritize feeding, treatment, pollination-contract, and inspection decisions while reducing disruptive manual strength assessments.

The hive-independent evaluation is particularly useful for Gratheon because it tests whether a model transfers to colonies absent from training. The modulation-tensorgram representation, reported CRDNN baseline, and explainability maps provide concrete components for benchmarking acoustic models on Gratheon's own microphones and for deciding whether the added temporal processing cost produces enough field-level accuracy.