The application of machine learning and acoustic signal processing to detecting and predicting important events in the lifecycle of honeybee colonies

Animal pollinators are crucial for food production, environmental sustainability, and ecosystem health. Honeybees are the primary pollinators of flowering plants making them and their monitoring highly important. Traditional hive monitoring by manual inspections can harm and disturb bees, is laborious, time-consuming and disruptive to their social life (Bencsik et al., 2011). Automated detection and prediction of bee colony anomalies, with instant communication to beekeepers, would benefit this field. This would enable effective and efficient colony management by beekeepers, avoiding unnecessary inspections, resource waste, and financial losses.

This project aims to develop an automated system using bee audio to detect two critical hive anomalies, namely queen absence and swarming. A colony without a queen for an extended period will perish, as the queen is the sole fertile female (Butler, 1954). Swarming, resulting in a substantial colony loss, can significantly reduce productivity and cause economic losses for beekeepers if not detected and managed promptly. The work described in thesis employs signal processing and machine learning methodologies to discriminate between queen presence and queen absence in hives and predict imminent swarming. This thesis utilizes frequency domain features (FFT, MFCC, Spectrograms and Mel spectrograms) to build binary classification models for both queen and swarm detection. A novel multiclassification model for swarm prediction up to four weeks in advance is also presented. This is achieved by analysing the weekly frequency content of bee audio from the four weeks preceding swarming, using a 5-class classification model. The classifiers compared are Multi-Layer Perceptron (MLP), Convolutional Neural Network (CNN), Long Short-Term Memory (LSTM), and Logistic Regression for detecting normal and anomalous bee colony sound patterns. For swarm prediction, LSTM and the CNN classifiers are employed for both the binary classification and the multiclass classification problems, trained on sensor data from beehives. The classifiers were trained using data collected by sensors from beehives in several different locations. The CNN and LSTM achieve significantly higher discrimination accuracy than the MLP and Logistic regression models across all classification tasks. The impressive average 96% accuracy of these demonstrates that successful bee monitoring is possible using the proposed system.