IoT-Based Bee Colony Health Monitoring: A Focus on Energy Impact and Audio Feature Extraction

Beekeeping is essential for maintaining ecosystem stability, improving pollination, and enriching biodiversity. The presence of the queen bee is critical for evaluating the health of bee colonies, which is fundamental for ecological equilibrium. Tine Machine Learning (TinyML), designed to predict the queen bee's presence, can be a valuable support for beekeepers in proactively assessing the colony's health from sound processing. Nonetheless, preprocessing the raw audio signals to prepare features for the TinyML model demands significant computational resources. This paper analyzes the energy impact of audio feature extractions for Tiny Machine Learning (TinyML) applications tailored for edge devices. We consider the Short Time Fourier Transform (STFT) and Mel-Frequency Cepstral Coefficients (MFCC) as audio features. Here, we deploy several models on our custom IoT device to detect the queen bee's presence from audio recordings. The system's architecture employs simplified Machine Learning models and low-precision audio processing to optimize energy efficiency. Despite these modifications aimed at reducing power usage, the system maintains performance metrics close to more complex setups, demonstrating minimal compromise in accuracy. This study underscores the impact of efficient feature extraction and data precision on reducing energy demands in IoT devices, which is crucial for sustainable TinyML deployment.