Benchmarking YOLO Variants for Detecting Varroa Mites in Beehive Images

Youcef Naas; Slim Rouabah; Rabah Abdelkader; Meziane Kaci; Salem Merabti; Hassane Ezziane

doi:10.5935/jetia.v12i60.3873

Autores/as

Youcef Naas Department of Electronic and communication, Faculty of Sciences and Technology, University of Djilali Bounaama, Khemis Miliana, Ain Defla, Algeria.
Slim Rouabah Djilali Bounaama University of Khemis Miliana
Rabah Abdelkader Department of Electronic and communication, Faculty of Sciences and Technology, University of Djilali Bounaama, Khemis Miliana, Ain Defla, Algeria.
Meziane Kaci Department of Electrical Engineering, Faculty of Sciences and Technology, Djilali Bounaama University of Khemis Miliana , Ain Defla, Algeria.
Salem Merabti Department of Acoustics and Civil Engineering, Faculty of Sciences and Technology, Djilali Bounaama University of Khemis Miliana , Ain Defla, Algeria
Hassane Ezziane Research Laboratory of Electrical Engineering and Automatics LREA, Yahia Fares University, Medea, Algeria

DOI:

https://doi.org/10.5935/jetia.v12i60.3873

Resumen

The Varroa destructor mite drives major honey bee colony losses and economic damage in apiculture. Early detection supports timely intervention, yet manual inspection is slow and prone to error due to the mite’s small size and dense hive conditions. Automated vision systems offer a scalable solution for continuous monitoring.

This study benchmarks three YOLO variants, YOLOv7, YOLOv8, and YOLOv11, for automated Varroa detection. YOLOv11 achieves the highest performance, with mAP at 0.5 of 0.941 and recall of 87.9 percent, outperforming earlier versions. The shift from anchor-based detection in YOLOv7 to anchor free designs in YOLOv8 and YOLOv11, along with improved feature extraction through C2f modules and attention mechanisms, enhances small object detection critical for identifying mites. YOLOv8 provides the fastest training speed, operating twice as fast as YOLOv11 and 4.5 times faster than YOLOv7, which supports rapid deployment.

All modern variants exceed 93 percent mAP at 0.5 and maintain real time inference at 30 frames per second or higher. Results position YOLOv11 as the preferred model for maximum accuracy and YOLOv8 as an efficient option for resource constrained beekeeping applications.