Abstract
This study investigates the impact of microphone placement on honey bee audio monitoring for time-of-day classification, a key step toward automated activity monitoring and anomaly detection. Recognizing the time-dependent nature of bee behavior, we aimed to establish a baseline diurnal pattern recognition method. A custom apparatus enabled simultaneous audio acquisition from internal (brood frame, protected from propolization) and external hive locations. Sound signals were preprocessed using Power Spectral Density (PSD). Extra Trees and Convolutional Neural Network (CNN) classifiers were trained to identify diurnal activity patterns. Analysis focused on feature importance, particularly spectral characteristics. Interestingly, Extra Trees performance varied significantly. While achieving near-perfect accuracy (98-99%) with internal recordings, its accuracy was considerably lower (61-72%) with external recordings, even lower than CNNs trained on the same data (76-87%). Further investigation using Extra Trees and feature selection methods using Mean Decrease Impurity (MDI) and Recursive Feature Elimination with Cross-Validation (RFECV) revealed the importance of the 100-600 Hz band, with peaks around 100 Hz and 300 Hz. These findings inform future monitoring setups, suggesting potential for reduced sampling frequencies and underlining the need for monitoring of sound inside the beehive in order to validate methods being tested.