TLDR: Researchers have developed a low-power, edge computing system that accurately detects the presence of a queen bee in a hive using only environmental sensors like temperature, humidity, and pressure differentials. This system achieves over 99% accuracy, significantly outperforms audio-based methods in energy efficiency, and provides a scalable, non-invasive solution for precision beekeeping.
The health and stability of honeybee colonies are critically dependent on the presence of a queen bee. Traditionally, monitoring a queen’s presence involves manual inspections, which are not only labor-intensive and disruptive to the colony but also impractical for large-scale beekeeping operations. While audio-based methods have emerged as a promising alternative, they often come with high power consumption, complex data processing, and susceptibility to ambient noise.
A new research paper, titled “Queen Detection in Beehives via Environmental Sensor Fusion for Low-Power Edge Computing,” by Chiara De Luca and Elisa Donati, introduces an innovative solution to these challenges. The researchers propose a lightweight, multimodal system designed for queen detection that relies on environmental sensor fusion. This system specifically uses temperature, humidity, and pressure differentials measured both inside and outside the beehive.
The core of this approach involves quantized decision tree inference, which is executed on a commercial STM32 microcontroller. This setup enables real-time, low-power edge computing, ensuring high accuracy without demanding excessive energy. The study demonstrates that this system can achieve over 99% queen detection accuracy using only environmental inputs. Interestingly, the inclusion of audio features did not provide any significant performance gain, suggesting that environmental data alone is sufficient and more energy-efficient for this task.
How the System Works
The proposed system consists of two main parts: a Python-based data processing pipeline for training and validation, and a real-time prediction system implemented on an STM32 microcontroller. The Python component handles data preprocessing, model training, and performance evaluation. Meanwhile, the STM32 system manages real-time operations, including acquiring environmental data, computing features, and making predictions.
The researchers utilized the Bee Audio Dataset, a publicly available resource that includes environmental measurements (temperature, humidity, and pressure) collected both inside and outside beehives, alongside audio signals. Crucially, each data point is annotated with the presence or absence of the queen bee, serving as the ground truth for training the classification model.
Instead of using raw sensor readings, the system computes differential features, which represent the relative changes between internal hive conditions and the ambient environment. These differentials are believed to be more indicative of the hive’s status and internal activity, reflecting the colony’s ability to regulate its internal environment in response to external changes.
Machine Learning and Performance
To detect the queen bee, a Light Gradient Boosting Machine (LightGBM) classifier was employed. The model was trained using environmental differential features, with careful attention to class balance and preventing overfitting through stratified splitting and 5-fold cross-validation. The best-performing model was then converted into a C-compatible format for deployment on the STM32 microcontroller, ensuring compatibility with the embedded environment.
The system’s performance was rigorously evaluated. Models trained on individual environmental features (temperature, humidity, or pressure differentials) achieved accuracies around 86.7% to 86.9%. However, combining these features significantly boosted performance. For instance, combining temperature and humidity differentials resulted in a 97.2% test accuracy. The highest accuracy, 99.4%, was achieved when all three environmental features were used together.
A key finding was the minimal impact of audio features. When a basic audio feature (Root Mean Square energy) was included, it resulted in only a marginal improvement in test accuracy (up to 98.7%), which was still lower than the 99.4% achieved with environmental features alone. Given the minimal benefit and higher resource demands, the researchers concluded that audio sensing is not justified for energy-constrained embedded applications in this context.
Also Read:
- Automated Pig Behavior Monitoring: A New Computer Vision System for Farm Welfare
- Forecasting Cattle Weight Gain: Machine Learning Models Offer New Insights for Livestock Management
Energy Efficiency and Scalability
The STM32 implementation closely matched the Python baseline, achieving an accuracy of 99.2% compared to 99.3%. More importantly, the system demonstrated remarkable energy efficiency. With an estimated average power consumption of approximately 44.5 mW and an energy cost of about 98 mJ per inference, it represents a significant improvement over audio-based TinyML systems, which typically exceed 600 mJ per inference. This makes the system highly suitable for long-term, low-power deployment in field conditions.
This work lays the foundation for scalable, sustainable hive monitoring systems that are practical, non-invasive, and compatible with edge devices. It paves the way for autonomous precision beekeeping and embedded ecological sensing, offering a more efficient and less disruptive way to ensure the health of honeybee colonies. For more details, you can read the full research paper here.
