BEE-YOND BUZZ: Exploring Deep Learning Techniques for Beehive Audio Classification

Monitoring beehive stability is critical amidst declining honeybee populations. This paper investigates the classification of beehive audio signals using transfer learning models. The performance of YAMNET and VGGish was compared in a binary classification task to distinguish bee from non-bee sounds, and a multiclass classification framework was developed to identify four distinct queen statuses within the hive. The models were evaluated using key benchmarks such as precision, recall, and F1-score. YAMNET achieved a precision of 74.5% and an area under the curve (AUC) of 0.87 in the binary classification task, outperforming VGGish. For the multiclass classification, deep learning models like Convolutional Neural Networks (CNN) and feedforward neural networks (FNN) achieved favorable performance metrics, CNN models exhibiting an average precision of 91.4%. These results demonstrate the potential for audio-based monitoring in non-invasive hive health assessment, offering real-time insights into queen presence, foraging behavior, and swarming, while promising future applications in detecting environmental stressors such as pesticide exposure.