Audio-based identification of beehive states

Inês Nolasco, Alessandro Terenzi, Stefania Cecchi, Simone Orcioni, Helen L. Bear, Emmanouil Benetos

Abstract

The absence of the queen in a beehive is a very strong indicator of the need for beekeeper intervention. Manually searching for the queen is an arduous recurrent task for beekeepers that disrupts the normal life cycle of the beehive and can be a source of stress for bees. Sound is an indicator for signalling different states of the beehive, including the absence of the queen bee. In this work, we apply machine learning methods to automatically recognise different states in a beehive using audio as input. We investigate both support vector machines and convolutional neural networks for beehive state recognition, using audio data of beehives collected from the NU-Hive project. Results indicate the potential of machine learning methods as well as the challenges of generalizing the system to new hives.

Introduction

In beekeeping, the loss of a queen bee is a critical emergency that requires immediate intervention to save the colony. Traditional monitoring involves manually opening hives to search for the queen, a labor-intensive process that stresses the bees and disrupts their natural life cycle. Fortunately, bees communicate through complex acoustic signals, and the hive's "hum" changes significantly when the queen is missing. This research explores non-invasive, audio-based machine learning methods to detect queen-state changes, focusing on the challenge of "hive-independent" generalization—ensuring a monitoring system works on new, unseen hives without prior training on their specific acoustic signatures.

Methods and Feature Extraction

The study compares traditional Support Vector Machines (SVMs) with Convolutional Neural Networks (CNNs). Beyond standard features like Mel-frequency cepstral coefficients (MFCCs), the authors introduce the Hilbert-Huang Transform (HHT) to better capture the non-stationary and non-linear nature of bee sounds, which result from various high-frequency muscle contractions and wing movements.

The HHT feature extraction pipeline, showing the decomposition of signal frames into Intrinsic Mode Functions (IMFs) before applying the Hilbert Transform.

As shown in Figure 1, the HHT process involves Empirical Mode Decomposition (EMD) to break audio frames into Intrinsic Mode Functions (IMFs). The Hilbert transform is then applied to these IMFs to extract instantaneous frequency and amplitude, creating a precise time-frequency representation. For classification, 10-minute audio clips from the NU-Hive project (totaling roughly 96 hours) were used, providing a robust real-world dataset for training and evaluation.

Key Results and Generalization

A primary finding of this work is the significant performance gap between "random" data splits and "hive-independent" splits. While models often achieve near-perfect accuracy when training and testing data are randomly mixed from the same hives, this performance typically drops when the model is tested on a completely isolated, unseen hive.

SVM results showing Area Under the Curve (AUC) for various feature sets. The gap between random splits (green dots) and hive-independent tests (black circles) highlights the challenge of generalization.

As illustrated in Figure 4, the SVM achieved its best hive-independent performance (AUC ≈ 0.94) by fusing MFCCs with HHT features. This combination proved exceptionally strong for capturing frequency behaviors that standard Mel spectra missed. In contrast, CNNs struggled with generalization in this setup, often falling to near-random performance (AUC < 0.5) when using Mel spectra on unseen hives. This suggests that deep learning models may overfit to the specific environmental acoustics of the training hives when the number of distinct hives in the dataset is limited.

Conclusion and Future Work

While the combination of MFCCs and HHT features shows great promise for robust queen-state detection, the study highlights the need for greater dataset diversity to improve deep learning generalization. Future work will focus on expanding the dataset to include more hives and exploring CNN architectures specifically designed for HHT and MFCC fusion. Additionally, the researchers aim to extend this pipeline to detect other critical states, such as the onset of swarming or the presence of pests like the destructive Varroa mite.

Summary updated

Abstract

The absence of the queen in a beehive is a very strong indicator of the need for beekeeper intervention. Manually searching for the queen is an arduous recurrent task for beekeepers that disrupts the normal life cycle of the beehive and can be a source of stress for bees. Sound is an indicator for signalling different states of the beehive, including the absence of the queen bee. In this work, we apply machine learning methods to automatically recognise different states in a beehive using audio as input. % The system is built on top of a method for beehive sound recognition in order to detect bee sounds from other external sounds. We investigate both support vector machines and convolutional neural networks for beehive state recognition, using audio data of beehives collected from the NU-Hive project. Results indicate the potential of machine learning methods as well as the challenges of generalizing the system to new hives.
BibTeX
@inproceedings{nolasco2019audio,
  author = {Inês Nolasco and Alessandro Terenzi and Stefania Cecchi and Simone Orcioni and Helen L. Bear and Emmanouil Benetos},
  title = {{Audio-based identification of beehive states}},
  booktitle = {IEEE International Conference on Acoustics, Speech and Signal Processing},
  year = {2019},
  url = {https://openreview.net/forum?id=2KtLoZcRu0},
}
Figure