Abstract
Automated beehive monitoring offers a non-invasive way to assess colony health, but environmental noise often pollutes audio data. This paper investigates machine learning techniques to distinguish between pure beehive sounds and external noise. By introducing a new 12-hour annotated dataset, the authors evaluate Support Vector Machines (SVM) and Convolutional Neural Networks (CNN). While both models achieve high accuracy on known environments, performance drops significantly when testing on unseen hives, highlighting a critical generalization gap in bioacoustic monitoring.
Introduction
Automated monitoring is a transformative tool for beekeeping, enabling the detection of swarming, queen loss, or disease without disturbing the hive. Sound is the primary diagnostic medium, as healthy hives produce a characteristic low-frequency hum. However, microphones also capture "noise"—rain, wind, traffic, and wildlife. To build reliable diagnostic tools, researchers must first filter out these non-bee events. This study addresses the fundamental classification task: identifying whether an audio segment contains pure beehive sound or is "polluted" by external noise.
Methods and Materials
Machine Learning Approaches
The study compared two distinct methodologies:
- Support Vector Machines (SVM): A traditional approach testing various kernels (RBF, Linear, Polynomial) and features such as Mel-frequency cepstral coefficients (MFCCs) and Mel spectra.
- Convolutional Neural Networks (CNN): The authors adapted the Bulbul architecture, originally designed for bird sound detection. This model utilizes four convolutional layers and three dense layers, employing data augmentation (time and pitch shifting) to enhance robustness.
Dataset and Annotation
A significant contribution of this work is the release of a specialized dataset aggregated from the Open Source Beehive (OSBH) and NU-Hive projects. It comprises 78 recordings totaling approximately 12 hours of audio. Non-expert volunteers labeled the data into "Bee" (pure hive sound) and "noBee" (external sounds superimposed on hive sounds).
Key Results and Insights
Performance Metrics
Both models performed optimally with longer audio segments (60 seconds), as the added temporal context helped distinguish continuous bee hums from transient noises. The SVM configuration using MFCCs and a 5-second noise threshold achieved the highest performance, with Area Under the Curve (AUC) scores often exceeding 0.90 on internal test sets.
The Generalization Gap
The most critical finding was the failure of models to generalize to new environments. In "hive-independent" tests—where models were trained on one set of hives and tested on a completely different one—performance plummeted, with AUC scores dropping to the 0.50–0.60 range. This suggests that models overfit to the specific acoustic signatures or microphone placements of individual hives rather than learning universal bee sound characteristics.
Discussion and Future Work
The study reveals that while automated noise filtering is feasible, the "new hive" problem remains a major obstacle. The authors identify several paths forward:
- Domain Adaptation: Future research should focus on making models trained on one hive effective on another without requiring extensive new labeled data.
- Expert Annotation: Refining datasets with specialist labels could reduce bias and improve model precision regarding overlapping sound types.
- Deployment Strategy: Currently, the most practical approach may be training site-specific models rather than seeking a universal solution.
Ultimately, bridging the generalization gap is essential for the scalable deployment of automated beehive health monitoring systems.
Summary updated