Spring 2025 Project Updates

Public Datasets

Shane Andres, Minyoung Kim

This semester we’ve continued to work towards training machine learning models to detect the status of beehives. We are actively deploying our recording hardware inside of beehives to collect data, but as this is in the early stages we only have a small quantity of training examples. In the meantime, we turned to publicly available datasets to help us develop preliminary models. We performed a literature review and compiled a list of public datasets.

A good portion of these datasets is reviewed in a recent publication [1]. Their summary highlights that some papers make only audio features public, but not the raw audio. This prohibits feature engineering, limiting the usefulness of the data. Many datasets have limited labels for their data (ex. which hive a piece of audio is from), and do not include any quantitative description of hive health. The paper also introduced a new dataset called UrBAN, which has raw audio modality and labels from hive inspections. We chose to begin exploring this dataset by training a binary classifier for queen bee presence.

Model Development

Aaryan Gusain

Our project began with a comprehensive data and labeling effort, where we unified over 20,000 audio clips from the UrBAN dataset with queen-right and queen-less labels derived from meticulous “as-of” merges with historical inspection data. This process involved cleaning and consolidating multiple inspection logs. From this curated audio, we engineered 282-dimensional feature vectors for each clip, primarily composed of 128-band Mel-spectral statistics combined with MFCCs and their deltas. A side study also explored the effectiveness of LFCC and FFT features, finding that LFCCs performed particularly well with simpler models.

With our feature set prepared, we evaluated a range of machine learning models, including KNN, Random Forest, Logistic Regression, MLP, and gradient boosting methods like XGBoost and LightGBM.

The LightGBM model proved to be the most effective, achieving an outstanding 98.4% test accuracy.

Critically, it also reached 94% recall on the queen-less class, demonstrating its reliability in identifying unhealthy hives. As a simple yet powerful baseline, a model using LFCC features with a KNN classifier also achieved a high 94.6% weighted F1-score, confirming the strength of our feature engineering.

Recording Hardware

Shane Andres

Two semesters ago, the audio team developed an initial prototype for our audio recording package to collect data from bee hives. As we’ve started deploying our hardware on beehives, we’ve also been reviewing literature to ensure our setup aligns with best available practices. Through our literature review, we found that recording methodologies vary from paper to paper. We found one paper summarizing the trends seen across studies to be very helpful [2].

Most use microphones placed either in the middle of brood frames, on top of brood frames, or next to the hive entrance. We also came across papers using accelerometers instead of microphones. Bees cannot actually hear sound, and instead pick up on physical vibrations of the surface they are on. We are wondering if listening in on just the vibrations of honeycomb would give us less noisy data with more rich information about the bees. We’re hoping to try this soon and are looking into possible hardware solutions.

[1] Mahsa Abdollahi et al., “UrBAN: Urban Beehive Acoustics and PheNotyping Dataset,” Scientific Data 12, no. 1 (2025): 536, https://doi.org/10.1038/s41597-025-04869-1.

[2] Cassandra Uthoff et al., “Acoustic and Vibration Monitoring of Honeybee Colonies for Beekeeping-Relevant Aspects of Presence of Queen Bee and Swarming,” Computers and Electronics in Agriculture 205 (February 2023): 107589, https://doi.org/10.1016/j.compag.2022.107589.