MSPB: a longitudinal multi-sensor dataset with phenotypic trait measurements from honey bees

Yi Zhu, Mahsa Abdollahi, Ségolène Maucourt, Nico Coallier, Heitor R. Guimarães, Pierre Giovenazzo, Tiago H. Falk, Yi Zhu 00112023 · DOI

Introduction

Honey bees (Apis mellifera) are essential to global food security, yet they face increasing pressure from climate change, pesticides, and parasites like Varroa destructor. Traditional hive management relies on manual inspections, which are labor-intensive and stressful for the colony. While automated "smart hive" technologies offer a non-invasive alternative, the development of effective machine learning models has been hindered by a lack of high-quality, long-term datasets that link raw sensor data with verified biological outcomes.

Methodology

To address this gap, researchers monitored 53 honey bee colonies across two apiaries in Québec, Canada, for a full year (May 2020 to April 2021). The study employed a dual-track data collection strategy:

  1. Continuous Sensing: In-hive sensors recorded temperature and humidity every 15 minutes. Audio was sampled every 5 minutes and processed on-the-fly into four acoustic features (e.g., Hive Power and Audio Band Density Ratios) to optimize battery life and storage.
  2. Expert Phenotypic Annotations: Every two weeks, apicultural experts performed manual inspections to record "ground truth" health metrics, including brood population, honey yield, defensive behavior, and hygienic behavior.
Figure 2: Schematic of the MSPB data collection system, illustrating the integration of continuous sensor streams with bi-weekly expert annotations.

As shown in the system overview, the data was transmitted to a cloud platform for processing, bridging the gap between high-frequency environmental data and longitudinal health assessments.

Dataset Characteristics

The resulting MSPB (Multi-modal Sensor dataset with Phenotypic trait measurements from honey Bees) dataset represents a significant advancement over existing benchmarks like NU-Hive or OSBH. Key advantages include:

  • Unprecedented Duration: 365 days of continuous monitoring.
  • Scale: Data from 53 distinct hives.
  • Richness: Inclusion of rare expert labels for winter mortality and hygienic behavior, which are typically absent in shorter-term datasets.
Figure 3: (a) Top-down view of a hive frame with the sensor node installed; (b) Exploded view of the hive components showing the vertical integration of sensors.

Key Insights and Results

The researchers validated the dataset through several proof-of-concept tasks:

  • Acoustic Indicators: Audio signals proved to be highly responsive to colony changes. "Hive Power" (energy in the 122–519 Hz range) exhibited clear circadian rhythms and spiked during human interventions, such as the addition of honey supers.
  • Hygienic Behavior and Productivity: A strong correlation ($r = 0.54$) was found between a colony's hygienic behavior—the ability to remove diseased brood—and its total honey yield, suggesting that "cleaner" hives are significantly more productive.
  • Predicting Winter Survival: Using an unsupervised anomaly detection model (Isolation Forest), the researchers achieved an AUC-ROC of 0.730 in predicting winter mortality using only summer sensor data. Hives that failed to survive the winter showed lower audio activity levels months before the colony collapsed.
  • Population Estimation: A linear SVM classifier distinguished between small and large colonies with a balanced accuracy of 65.8%.

Conclusion and Future Work

The MSPB dataset provides a robust foundation for training AI-driven hive monitoring systems. While the decision to store acoustic features rather than raw waveforms limits certain signal-processing retrospective analyses, the dataset's scale and phenotypic richness are unparalleled. Future research will focus on multi-modal fusion—combining audio, temperature, and humidity—to improve the detection of complex events like swarming or queen loss.

Summary updated

BibTeX
@misc{zhu2023mspb,
  author = {Yi Zhu and Mahsa Abdollahi and Ségolène Maucourt and Nico Coallier and Heitor R. Guimarães and Pierre Giovenazzo and Tiago H. Falk and Yi Zhu and Mahsa Abdollahi and Ségolène Maucourt and Nico Coallier and Heitor R. Guimarães and Pierre Giovenazzo},
  title = {{MSPB: a longitudinal multi-sensor dataset with phenotypic trait measurements from honey bees}},
  year = {2023},
  doi = {10.5281/zenodo.8371700},
}
Figure