Bee Together: Joining Bee Audio Datasets for Hive Extrapolation in AI-Based Monitoring

Area of Science:

• Apiology
• Ecology
• Agriculture
• Bioacoustics

Background:

• Beehive health monitoring is crucial for biology, ecology, and agriculture.
• Audio sensors offer a non-intrusive method for hive monitoring.
• Existing audio datasets for bee classification often lack generalization across different hives.

Purpose of the Study:

• To address the generalization limitations in beehive audio classification.
• To create a standardized, open-access dataset for bee health research.
• To develop novel classification methods for improved hive extrapolation.

Main Methods:

• Reviewed and merged open audio datasets into the "BeeTogether" dataset on Kaggle.
• Implemented data augmentation and a methodology for measuring hive extrapolation.
• Benchmarked a classical classifier and introduced contrastive learning-based classifiers.
• Prototyped a process for obtaining absolute labels on unsupervised data.

Main Results:

• Classical classifiers achieved good accuracy but poor generalization to unseen hives, consistent with prior research.
• Classification performance was significantly influenced by colony-specific audio characteristics.
• Contrastive learning classifiers demonstrated improved accuracy and hive extrapolation abilities.
• The "BeeTogether" dataset provides a standardized framework for evaluating generalization.

Conclusions:

• A unified dataset and advanced classification techniques are essential for robust beehive monitoring.
• Contrastive learning offers a promising approach to overcome colony-specific biases in audio classification.
• This work facilitates more effective applications of bee health monitoring in agriculture.