Effectiveness of Transfer Learning, Convolutional Neural Network and Standard Machine Learning in Computer Vision Assisted Bee Health Assessment

Abstract

Honeybees are vital to society, as they pollinate over 80% of plants. Unfortunately, honeybee colonies have been losing at an average rate of 39.7% per year. Beehive monitoring depends on human visual examinations, which is time consuming and disruptive to colonies. It is critical to apply an effective and efficient technique to monitor bee health and save bee colonies. The paper provides a systematic study of applying transfer learning, classic convolutional neural network (CNN) and standard machine learning models on image-based bee health classification. Five models (SVM, CNN, VGG19, InceptionV3, MobileNet) have been evaluated on a real-world dataset with more than 5000 bee images and six health sub-classes. The accuracy rates from all five models were above 90% in the test dataset. In particular, VGG19 and CNN achieved 98.65% and 96.91% accuracy, respectively. These accuracy rates were higher than the best accuracy rate of 95% in other previous research. The promising model results demonstrate the potential of applying AI techniques to build an intelligent beehive inspection system.