Monitoring honey bee colony strength remotely using Internet of Things (IoT) sensors is crucial for agriculture and ecosystem stability. Previous methods focused on handcrafted features from audio modulation spectra. This paper introduces a new approach using a modulation tensorgram that preserves temporal dynamics, which are often discarded.

This new representation is fed into a convolutional neural network (CNN) and a convolutional recurrent deep neural network (CRDNN). Using the public UrBAN dataset, comprising over 3,000 hours of beehive audio, the proposed method demonstrates improved accuracy and cross-hive generalizability compared to prior benchmarks. The results also indicate enhanced robustness in noisy, real-world recording conditions.

Explainability techniques like saliency maps and gradient-weighted class activation maps highlight the significance of modulation spectral temporal dynamics. The findings suggest that accurate, generalizable, and robust acoustic monitoring of honey bee colony strength is achievable using this novel method.