High-Precision Detection of Homona magnanima and Adoxophyes honmai in Mandarin Orchards Using Deep Learning

Abstract The effective monitoring and management of agricultural pests are critical for sustainable crop production and minimizing economic losses. This study presents the development of the Adoxophyes honmai and Homona magnanima Detection Phenomics Dataset (AHMDP Dataset) and the design of a robust deep learning model for automated insect detection using the YOLOv11 framework. Images were collected from eight mandarin orange orchards on Jeju Island, South Korea, using pheromone-baited unmanned monitoring traps equipped with high-resolution cameras. The dataset comprises 15,531 annotated images containing 36,129 instances of the two major tortrix moth species, Adoxophyes honmai and Homona magnanima, annotated with precise bounding boxes. A comprehensive preprocessing pipeline was implemented, including image slicing via the Slicing Aided Hyper Inference (SAHI) method and annotation conversion to the You Only Look Once (YOLO) format, to enhance the detection of small insect instances and maintain data consistency. The YOLOv11-based detection model was trained on this dataset and evaluated on a separate validation set, achieving a mean average precision (mAP@0.5) of 0.991 and an mAP@0.5:0.95 of 0.832, demonstrating high precision and recall for both insect species. The model’s inference speed supports real-time deployment, offering a practical solution for continuous pest monitoring in orchard environments. This integrated approach, combining a well-curated dataset with a state-of-the-art detection framework, lays the foundation for scalable, accurate phenomics analysis of insect populations, facilitating informed decision-making in integrated pest management.