Comparative Analysis of Deep Learning Models for Honeybee and Threat Detection at Hive Entrances

Abstract

This thesis examines and compares the performance of four recent object recognition models YOLOv8, YOLOv12, RF-DETR and Roboflow 3.0 in recognizing bees and exterior attacks at the hive entrance. For this, a dataset with labeled images of bees of different classes (worker, Drone, queen, bee_pollen and attacker) was developed and used for training and testing the models. Precision, recall, F1-score, mean average precision (mAP) of processing and time taken in post-processing were the evaluation criteria. The results showed that RF-DETR outperformed the others overall, exhibiting superiority in accurate bee identification with an F1 score of 0.9595 and a mAP@0.5 of 0.980. Roboflow 3.0 showed strong multi-class performance, achieving the best mAP@0.5:0.95 (0.679). Further proof of RF-DETR's superiority in real-world detection scenarios came from qualitative testing. This study shows how automated hive monitoring can be greatly improved by deep learning-based vision systems. Precision beekeeping and early threat detection are made possible by these scalable, non-invasive alternatives to manual inspection.