Automated Identification of Parasitoid Wasps for Biological Control Monitoring Using an Optimized Single-Stage Detector

Parasitoid wasps are important biological control agents, but their small size and subtle morphological differences hinder rapid identification and require taxonomic expertise. This study developed a computer vision framework for recognizing parasitoid wasps from single-plane microscopy images. A dataset of 3420 images representing 36 parasitoid taxa and 38 terminal biological classes was assembled. Five YOLO detector generations and six CNN- and ViT-based classifiers were evaluated across three dataset sizes. YOLO11m was selected for further optimization. The optimized model integrated multiscale feature processing, attention mechanisms, a modified bounding-box regression loss, and tuned training settings. On an independent set, the baseline YOLO11m achieved an mAP@0.5:0.95 of 0.867 at 640 × 640 pixels. At 960 × 960 pixels under untuned settings, YOLO11m and YOLO11m-opt achieved 0.873 and 0.917, respectively. After applying the selected training configuration, YOLO11m-opt achieved a precision of 0.985, recall of 0.989, mAP@0.5 of 0.995, and mAP@0.5:0.95 of 0.933 across all 105 detection labels. Eigen-CAM visualizations showed spatial feature activations overlapping with potentially informative morphological structures. A Windows application supported batch analysis, offering a basis for computer-assisted specimen screening under laboratory conditions.