Investigation into the use of image processing and machine learning techniques to monitor varroa mite infestation in honey bee colonies

According to research published in the Royal Society B: Biological Sciences, approximately a third of the food consumed around the world is reliant on animal pollinators to produce. Of all these pollinators, the Western Honey Bee (Apis mellifera) is the most significant human-managed insect pollinator, present on all continents except Antarctica. This management includes the movement of colonies by humans around countries to provide pollination services where and when needed.

Several stressors of Honey Bee Colonies exist – monoculture, diseases, climate change and pests. However, the greatest threat to the Western Honey Bee is the ectoparasite Varroa destructor mite, which can, in the absence of any mitigation e.g. medicaments or honey bee colonies developing resistant traits against varroa, cause the death of a honey bee colony with three years of first infestation. Varroa mites attack the colony via different actions, firstly by weakening individual bees both as pupa and adults by feeding on their vital organs, secondly by acting as a vector for harmful viruses between adult honey bees. Varroa destructor is present in most countries across the world with the Australian government executing a fruitless 15-month exercise to try to prevent the mite becoming endemic, accepting its arrival in 2023. To understand whether the colony is in danger of failure, it is important to monitor the level of varroa infestation in the honey bee colony and there are several methods to achieve this.

In this project, a survey has been conducted which shows that a sample of UK Beekeepers regard varroa as a serious problem to which honey bees need beekeeper assistance. The method of monitoring the level of varroa infestation with the highest reported usage in the survey was the debris board placed underneath the bee hive onto which varroa mites fall and can be counted.

The main work of the project was to investigate the ability of certain image processing and machine learning techniques to distinguish varroa mites from ordinary debris on the boards which could then be used by a beekeeper in the field, with a view to supporting the evaluation of the infestation level. To this end the project captured 380 magnified images of varroa mites using a smartphone camera in the field without any special lighting requirements. Using standard augmentation methods this varroa data set was then expanded to 4000 images. To match this, a 4000 data set of images not containing varroa mites was also created. The project then used a variety of deep learning machine learning algorithms to build classification models to operate on original images via upscaling.

The resulting models using the ResNet50 and Vision Transformers algorithms resulted in a sensitivity/recall score of 89% and 87% respectively in identifying varroa mites from other debris objects on the varroa board via upscaling. This could be extended to an application appropriate for either hobby beekeepers or even industrial scale beekeepers.