NeuralBee - A Beehive Health Monitoring System

Abstract

Bees are essential as they are responsible for the pollination of one-third of the world’s food. Without bees, the availability of fresh produce would be significantly less and could also lead to the collapse of several ecosystems. This study proposes a system that uses computer vision to detect Varroa mite infestation levels in a beehive using object detection techniques and a beehive audio analysis system using Mel spectrograms and Mel-frequency cepstral coefficients (MFCCs) as input features to a deep learning model to discriminate between a healthy hive and a weak hive. For this experiment the object detection algorithms YOLOv8, YOLOv7, YOLOv5 and SSD, are compared based on their accuracy, speed, and compute requirements. A dataset consisting of over 10,000 ground-truth images of bees infected with varroa mites and healthy bees was used and the models achieved the highest precision of 0.962 for Varroa mite detection. For audio analysis, a custom dataset with over 2 hours of audio recordings from ‘‘strong’’ and ‘‘weak’’ beehives was used to train and evaluate a neural network that reached a maximum accuracy of 0.998.