Tung, Chiao-JoChiao-JoTungWu, Yi-HuiYi-HuiWuLee, Shih-YangShih-YangLeeKuo, Yan-FuYan-FuKuo2025-10-162025-10-162025https://www.scopus.com/record/display.uri?eid=2-s2.0-105015540968&origin=resultslisthttps://scholars.lib.ntu.edu.tw/handle/123456789/7326942025 American Society of Agricultural and Biological Engineers Annual International Meeting, ASABE 2025, 13 July 2025 - 16 July 2025, TorontoThe use of parasitoid wasps for pest control is an essential strategy in sustainable agriculture. Accurate counting and gender identification are critical for establishing mass production techniques and ensuring product quality, as only female wasps possess parasitism capabilities. Parasitoid wasps are typically tiny, and their sex is often determined by subtle morphological traits. In this study, antennae with distinctly large clava was used to distinguish male and female Trissolcus sp., the target species. Conventional microscopic examination for gender identification is labor-intensive and time-consuming. To address these challenges, this study develops an automated system for counting and sex identification of Trissoclus sp., a parasitoid wasp approximately 1 mm in length that parasitizes Rhynchocoris humeralis, a citrus pest. High-resolution images (6400 dpi) of the wasps were captured using a flatbed scanner. Subsequently, a two-stage deep learning-based approach was implemented for the tasks. In the first stage, the original high-resolution images were downsampled to accelerate the wasp counting. The wasps in the downsampled images were localized and counted using YOLOv7. In the second stage, the bounding boxes of the wasps from YOLOv7 were used to localize and crop the wasps in the original high-resolution images. The cropped wasp images were then used to identify the genders of the wasps using ResNet18. The trained YOLOv7 and ResNet18, respectively, achieved a mean average precision of 98.91% and a validation accuracy of 96.16%. The proposed approach provides a scalable and automated solution for parasitoid wasp identification, streamlining workflows and reducing reliance on manual processes. Future work will focus on improving classification accuracy and enhancing system usability for broader implementation.falseBiological ControlDeep LearningParasitoid WaspResidual Neural Network (ResNet)You Only Look Once (YOLO)[SDGs]SDG2[SDGs]SDG5[SDGs]SDG8[SDGs]SDG12conference paper not in proceedings10.13031/aim.2025001462-s2.0-105015540968