Transfer Learning-Based Multi-Class Pest Detection using a Real-Time Object Detection Framework with Control-Oriented Grouping for Smart Agriculture
Artryan Nero A. Logarta, Lysa V. Comia
Abstract
Pest infestations cause annual crop yield losses of 20–40% and over USD 220 billion in damages, with smallholder farmers most affected. Conventional monitoring methods are labor-intensive, subjective, and slow, motivating automated, real-time approaches. This study proposes a control-oriented pest detection system using YOLO11, trained on 14 agriculturally significant pest classes and extended with a post-hoc grouping mechanism aligned to Integrated Pest Management (IPM). The model achieved a mean precision of 0.775, recall of 0.601, AP50 of 0.700, and mAP50–95 of 0.553, with an overall [email protected] of 0.704. Strong results were observed for tomato_hornworms (0.838), armyworm (0.829), and brown_marmorated_stink_bug (0.849), while aphid_cluster (0.477) and spider_mite_webs (0.264) remained challenging due to their small size and background clutter. Confusion matrix analysis revealed misclassifications among morphologically similar pests and occasional sensitivity in pollinator detection. Dataset inspection confirmed class imbalance and prevalence of small, centrally located instances. Beyond detection, the grouping mechanism mapped pests into biological, chemical, cultural, mechanical, and pollinator-protection strategies, with cultural controls achieving the highest confidence (0.889). These results demonstrate the potential of YOLO11 coupled with control-oriented grouping to provide reliable, decision-ready outputs for smart farm systems, UAV-based treatments, and sustainable pest management.