Deep learning model for honey bee health detection

Abstract

Honey bee pollination plays an important role in the world’s food production. There is there-fore a need to safeguard the ecological balance by protecting honeybees to ensure continued production. One step towards protecting bees is to identify health issues that a✏ict honey- bees during the course of their lives early so as to prevent the transmission of the illnesses to other healthy honey bees. Existing mechanisms have proven ine↵ective in making the early bee health problem detection a reality. The disciplines of machine learning and computer vision are rapidly expanding and have already demonstrated their ability to handle challeng- ing problems. Existing studies have explored image classification that is based on images of healthy and infected bees. In an attempt to improve on the existing studies this thesis employed a venture of utilizing specific features for bee health detection. It considers the abdomen coloration and the leg/wing orientation of the bees for detection if a hive is varroa- infected or not. An object detection task and a segmentation task are applied and their performance evaluated. The YOLOv8 object detection model gave the best performance results and is thus recommended. The FASTER RCNN model also performed well giving and can also be employed in predicting bee health. The findings of the models outperform the state-of-the-art models used in similar applications. The new model will be instrumental in future object detection related studies. In addition, the use of such a model will go a long way towards the attainment of sustainable development goal in Agriculture more specifically in bee keeping.