Convolutional Neural Networks for Real Time Classification of Beehive Acoustic Patterns on Constrained Devices.

Area of Science:

• Agricultural Technology
• Computational Biology
• Machine Learning

Background:

• Convolutional Neural Networks (CNNs) effectively classify bee colony acoustic patterns for health assessment.
• Implementing CNNs for bee monitoring can lead to high computational costs and energy demands compared to traditional methods.

Purpose of the Study:

• To investigate the feasibility of CNN architectures for developing a bee colony monitoring system on constrained hardware.
• To analyze the performance trade-offs of various CNN models and acoustic data durations on single-board computers.

Main Methods:

• Ten CNN architectures were tested on Nvidia Jetson Nano, Raspberry Pi 5, and Orange Pi 5 single-board computers.
• Models were trained on acoustic spectrograms of varying durations (1-30s), with hyperparameter optimization via Optuna and k-fold cross-validation.
• Inference time and power consumption were measured for performance comparison.

Main Results:

• Performance varied across CNN architectures and single-board computers, with shorter acoustic sample durations impacting results.
• Specific CNN models demonstrated potential for efficient operation within the power and computational constraints of the tested devices.

Conclusions:

• CNNs can be adapted for effective bee health monitoring systems on resource-constrained devices.
• This research provides a foundation for developing practical, low-power apiculture monitoring solutions using advanced AI.