Wild bees are threatened by numerous parasites that can significantly impair their health and even lead to death. Such parasites can weaken entire colonies, ultimately causing their eradication. However, existing studies focus on domesticated honey bees and apply methods under well-controlled conditions. Methods for automated detection of parasites in wild bees and under natural conditions are lacking. We focus on two types of parasites: endoparasites of the family Stylopidae and kleptoparasitic larvae of specific blister beetles of the tribe Meloidae. We followed an opportunistic data collection approach and sampled images of parasites present in the wild in Germany. We investigate the feasibility of using deep learning methods to detect these parasites in images of wild bees with diverse natural backgrounds. In detail, we gathered, analyzed, and annotated publicly available images of parasitized bees. Then we trained an object detection model YOLO to localize and classify parasites in images of wild bees. Because the number of suitable images is limited, we applied data augmentation techniques to increase the dataset size. Most notably, we created composite images by overlaying segmented parasite crops on images of healthy bees. Our trained model is a proof-of-concept to demonstrate automated parasite detection in images of wild bees under natural conditions. We note that detecting parasites poses a significant challenge, because they are often difficult to discern. Issues such as blurry images, poor illumination, occluded and overlapping parasites further complicate detection. The scarcity of available images exacerbates the problem. However, we demonstrate in a use case that the trained model can be used to analyze images of wild bees to find and identify unlabeled parasites in public image repositories. We provide a publicly available demonstrator to showcase the model’s capabilities and to encourage further research in this area. • The first deep learning-based approach for detecting wild bee parasites, namely Stylopidae and Meloidae larvae, in natural environments. • Novel augmentation pipeline tailored for rare and underrepresented parasite classes. • Use-case evaluation demonstrating the model’s potential to discover previously unannotated parasites in public image collections and reduce manual annotation effort.
Ionova et al. (Wed,) studied this question.