AI audio sensors monitor bee colony strength with 3000+ hours of data

A team led by Mahsa Abdollahi from Université du Québec and Université Laval developed an improved acoustic monitoring system for honey bee hives. Their key innovation is a modulation tensorgram representation that retains temporal dynamics of the modulation spectrum—information discarded by previous handcrafted feature approaches. Using convolutional neural networks (CNN) and convolutional recurrent deep neural networks (CRDNN), the model was trained and tested on the public UrBAN dataset containing over 3,000 hours of real-world beehive audio recordings.

The results show significant improvements in both accuracy and cross-hive generalizability over prior benchmark methods, with enhanced robustness to noisy in-the-wild conditions. The researchers used saliency maps and gradient-weighted class activation maps for explainability, confirming that modulation spectral temporal dynamics are critical for the task. This work demonstrates that accurate, generalizable, and robust remote monitoring of honey bee colony strength is achievable with advanced audio AI, potentially helping beekeepers and agricultural stakeholders detect colony health issues early.