Acoustic Signatures of Hive: Detecting Queen Bee Absence Through Machine Learning of Short Audio Segments

Area of Science:

• Agricultural Science
• Bioacoustics
• Machine Learning

Background:

• Honeybee decline threatens biodiversity and agriculture, necessitating non-invasive monitoring.
• Early detection of queen absence is vital for colony survival.
• Acoustic monitoring offers a promising solution for continuous hive assessment.

Purpose of the Study:

• To evaluate machine learning and deep learning models for acoustic queen-presence detection.
• To determine the effectiveness of different audio features (spectrogram, Mel-spectrogram, MFCC) for this task.
• To assess the feasibility of using short audio segments for real-time monitoring.

Main Methods:

• Collected hive audio data from various sources.
• Extracted audio features including spectrogram, Mel-spectrogram, and Mel-frequency cepstral coefficients (MFCC).
• Evaluated classical machine learning classifiers and convolutional neural networks (CNNs) for classification.

Main Results:

• Mel-frequency cepstral coefficient (MFCC) features consistently outperformed spectrogram-based features.
• Convolutional neural networks (CNNs) with Mel features achieved high accuracy on short audio segments.
• Gradient-boosted models performed well on longer audio windows.
• Brief acoustic segments proved sufficient for reliable queen-presence classification.

Conclusions:

• Acoustic analysis using machine learning is effective for non-invasive, real-time queen bee detection.
• The developed framework supports scalable, low-cost precision beekeeping.
• This approach contributes to sustainable apiculture through automated anomaly detection.