Comprehensive Evaluation of Segmentation Models for Bee Varroa Mite Detection

Accurate and detailed identification of Varroa destructor, a parasitic mite threatening honeybee colonies, is essential for effective beekeeping management. To mitigate its severe impact on bee health, developing precise and reliable detection techniques has become a critical focus in apiculture research. In this study, we present a comprehensive evaluation of several state-of-the-art deep learning-based segmentation architectures for pixel-level detection of Varroa mites on bees. Specifically, we assess encoder-decoder networks (UNet, UNet++, ResUNet, UNet3+), atrous convolution-based models (DeepLabV3 and DeepLabV3+ with ResNet101 and EfficientNetB5 backbones), and single-stage detection frameworks adapted for segmentation (YOLOv8, YOLOv11) based on a dedicated dataset of Varroa-infected bees. Experimental results show that the DeepLabV3 architecture with a ResNet101 backbone outperforms the other approaches, achieving an Intersection over Union (IoU) of 52.11% and a Dice coefficient of 67.21%, while maintaining an acceptable inference time. These findings highlight the effectiveness of DeepLabV3 in accurately localizing and segmenting Varroa destructor, and underscore the importance of pixel-level analysis in automated bee health monitoring systems.