YOLOv8 Enables Automated Dragonfly Species Classification Using Wing Images

Abstract

This study investigates the digitization of insect collections and the application of deep learning models to improve this process. During a one-hour filming session, 141 images of dragonfly specimens from the Cornell University Insect Collection were captured and preprocessed using five distinct methods: (1) adding box annotations around the wings, (2) adding polygon annotations to outline the forewings and hindwings, (3) removing vein system images, (4) retaining only the wing outline images, and (5) grouping by automatically measured wing size and classifying species within those groups. By comparing YOLOv8 models trained on datasets with these different preprocessing methods, the study revealed three key findings: (1) datasets with bounding box annotations result in shorter preprocessing times and superior model performance compared to polygon annotations; (2) although models trained with polygon annotations may have lower accuracy, they provide more detailed information on wing length and phenotypic traits; and (3) the wing vein system, rather than the wing outline, is the critical factor in classification accuracy.