A deep learning classification pipeline for identifying economically important tephritid fruit flies

Computer vision approaches utilising deep learning offer significant potential benefits for entomological applications, particularly for image-based taxonomic identification. Fruit flies (Tephritidae) represent economically damaging pests where species-level identification is critical for effective pest control, management, surveillance, and eradication programs. We assessed the capability of a deep learning convolutional neural network (CNN) pipeline to identify tephritid species from wing images, which serve as key diagnostic features. Our dataset comprised 1380 tephritid wing images spanning 34 tephritid species and 12 genera, with additional images from two 'other' classes (other Diptera and Hymenoptera). We employed a two-stage approach: (1) object detection using an Ultralytics YOLOv11n model to detect wing objects in images, treating all wings as a single class, followed by (2) species classification using an Ultralytics YOLOv11-cls model applied to cropped and augmented wing images generated from the object detection stage. The models demonstrated high accuracy in both wing detection (mAP50-95 value of 0.99 on a novel test set) and species classification (overall accuracy of 0.98 on a novel test set). Class-wise accuracy for different species varied (0.67-1) but showed general correlation with the number of original images available per class (10–285). Our results provide a potentially valuable tool for detecting pest tephritid species in biosecurity contexts. While deep learning technology remains in early development stages for entomological applications, such approaches hold promise to transform diagnostic and surveillance capabilities for biosecurity and pest management.