Ant Detection using YOLOv8: Evaluation of Dataset Transfer Impact

Abstract

In order to avoid having to fight with aphids and plant virus diseases caused by them in gardens, it is very important to notice ant colonies. As a result we decided to train artificial intelligence to detect ant colonies, then this artificial intelligence can be integrated into an autonomous orchard monitoring system using unmanned aerial vehicles. However, there is restricted availability of open datasets, which contain natural images and region specific species. In the scope of pilot study we decided to train convolutional neural network using ANTS dataset and to test it on small domain-specific dataset to identify the need to collect new dataset. The experiment was completed using the popular architecture YOLOv8. The YOLOv8n and YOLOv8m models trained on ANTS showed accuracy 98% and 99% mAP@0.5. Meanwhile, their accuracy was only 6% and 5% mAP@0.5 respectively testing on our dataset called “WildAnts”. Our pilot study experimentally proves that it is important to collect natural dataset of ant images to train robust artificial intelligence for orchard monitoring using unmanned aerial vehicles. This study will be interesting for all machine learning specialists, because it numerically shows accuracy decrease in the result of dataset transfer.