Technology identifies the absence of a queen through short audio recordings from the hive

Researchers evaluated the use of artificial intelligence to detect the absence of the queen in bee hives using short audio recordings. The study indicates viability for non-invasive, continuous, and low-cost monitoring. The best strategy combined MFCC-type acoustic attributes at the Mel scale with machine learning and deep learning models.

The absence of the queen compromises the viability of the colony. Her loss can alter reproduction, favor colonies with laying worker bees, and harm the stability of the nest. Traditional inspection methods require time, trained personnel, and opening the hive. This study proposes an alternative based on the sound produced by the bees.

The research analyzed recordings of beehives with and without a queen. The scientists used one-, two-, and three-second segments. The approach sought to identify sufficient acoustic signals for a rapid response in the field. The goal included reducing computational costs and enabling use in embedded systems.

Three sources

The data came from three sources. The Hiveeyes project provided twenty recordings from eleven urban beehives in Berlin, containing bees Apis mellifera carnica or Carnica-Buckfast hybrids. The LongHive project provided thirty-five recordings of the set “To bee or not to bee”. The USM Bee Lab, from the Universidad Técnica Federico Santa María, in Valparaíso, Chile, provided recordings of two beehives in an urban environment, about one hundred to two hundred meters from a busy avenue.

The Chilean recordings took place on two sunny days, between 3 p.m. and 8 p.m. One hive maintained its queen throughout the entire evaluation period. The other hive had no queen before December 2nd, presented queen larvae, and received a new queen by December 9th. This arrangement allowed for the evaluation of colony state transitions under natural conditions.

The researchers installed microphones inside the beehives. The equipment was placed between frames close to the population center. The approximate distance to the area between the brood chamber and the honey super was fifteen centimeters. The recordings used a sampling rate of forty-four thousand one hundred samples per second. The files came in mp3, wav, and m4a formats.

The processing divided the audio into non-overlapping windows. Then, each segment underwent acoustic attribute extraction. The study compared spectrograms, Mel spectrograms, and Mel cepstral frequency coefficients, known as MFCC. This last representation compacts sound information on a scale related to auditory perception.

Models evaluated

The models evaluated included SVM, XGBoost, convolutional neural networks, and multilayer perceptron. The comparison considered accuracy, precision, recall, F1-score, and Cohen's Kappa. The test set presented an imbalance, with approximately five hundred samples without a queen and two hundred and fifty with a queen. Therefore, the study paid attention to agreement beyond chance.

The best results appeared with MFCC attributes on the Mel scale. In one-second segments, the convolutional neural network achieved an average accuracy of 0,726. In two-second segments, XGBoost achieved 0,732. In three-second segments, SVM obtained 0,697. Confidence intervals indicated greater stability for the MFCC Mel attributes compared to linear spectrograms.

The study also showed limitations. The sample involved a limited number of beehives, urban noise sources, different recording equipment, and compression formats. The scientists did not isolate the effect of each noise source, such as wind, traffic, or human activity. They also did not perform formal significance tests between models due to the limited number of repetitions.

Even so, the results support the use of short audio recordings for hive screening. The technology can reduce invasive inspections and generate alerts after a possible queen loss. In commercial apiaries, this type of tool can integrate sensor networks and support management decisions with less direct intervention in the colonies.

Further information at doi.org/10.3390/insects17060547