Enhancing Precision Agriculture Pest Control: A Generalized Deep Learning Approach With YOLOv8-Based Insect Detection

• Published in: IEEE access Vol. 12; pp. 84420 - 84434
• Main Authors: Vilar-Andreu, Mario, Garcia, Laura, Garcia-Sanchez, Antonio-Javier, Asorey-Cacheda, Rafael, Garcia-Haro, Joan
• Format: Journal Article
• Language: English
• Published: Piscataway IEEE 2024; The Institute of Electrical and Electronics Engineers, Inc. (IEEE)
• Subjects: Agriculture; Artificial neural networks; Computer vision; Convolutional neural networks; Crops; Datasets; Deep learning; Detection algorithms; Feature extraction; Insect detection; Insects; Machine learning; Object detection; Object recognition; Performance evaluation; pest control; Pests; Real time; Real-time systems; YOLO; you only look once (YOLO)
• ISSN: 2169-3536; 2169-3536
• DOI: 10.1109/ACCESS.2024.3413979

Precision Agriculture (PA) is gaining new momentum due to its ability to accurately adjust the resources to a crop's needs while maintaining/enhancing quality levels. However, crop-damaging pests compromise yields, jeopardizing the benefits of PA. Computer vision-based pest detection techniques offer promising avenues to overcome potential losses for farmers. The recent object detection framework, YOLOv8 (You Only Look Once) applied to real-time insect monitoring is an open-source, cutting-edge PA approach based on Convolutional Neural Network (CNN) models that enables precise and quick decision making in agricultural crops. Under this umbrella, traditional pest studies using YOLO or other deep-learning solutions focused on only one or a few insects for specific crops provide an excessively narrow solution. In this paper, we propose a new form of using YOLO for pest detection with a generalist perspective by intensively testing a YOLOv8-based tool implementing a single insect category. The goal is to detect the presence of any type of insect in any type of crop in real time. A comprehensive performance evaluation is carried out using a well-known dataset. The results of the training, validation, and testing phases are then discussed, obtaining an <inline-formula> <tex-math notation="LaTeX">mAP_{50} </tex-math></inline-formula> value of 0.967 for the m model and an <inline-formula> <tex-math notation="LaTeX">mAP_{50-95} </tex-math></inline-formula> value of 0.632 for the l model. Finally, we also identify the premises to elaborate a complete and useful dataset able to unleash the full potential of YOLOv8.