Deep Learning Pipeline for the Identification of Ground Beetles (Carabidae) in New Zealand

The identification of invertebrates is often difficult and time‐consuming because of their high abundance and diversity. This limits the use of invertebrates in environmental monitoring and conservation. However, computer vision approaches combined with deep learning models offer a range of benefits to insect conservation, particularly for image‐based identification. In this study, we assessed the ability of deep learning convolution neural network models to identify genera of Carabidae (Coleoptera) from dorsal images of the thorax and elytra. A two‐stage detection and identification pipeline, using the Ultralytics Python library and YOLOv11 models, was developed based on 4715 specimens from the New Zealand Arthropod Collection. The detection model was very reliable at extracting the thorax and abdomen of beetles from the overall specimen images with a mean average precision of 0.983, and precision and recall of 1. The classification model was trained and evaluated using five‐fold cross‐validation on cropped images from the detection model. Classification performance was high with mean (SD) accuracy across folds of 0.92 (0.005), precision of 0.91 (0.01), recall of 0.88 (0.001) and F1 of 0.88 (0.02), indicating that dorsal views of the thorax and abdomen have high predictive value. We also developed a web application (CarabID) that allows researchers to quickly upload and identify carabids using the above model. This tool is aimed at identification from images captured in semi‐controlled conditions (e.g. sorted pitfall samples) to provide greater taxonomic resolution and reliability than from field‐based images. Although deep learning technology is still in its infancy for entomology, such approaches could transform the identification of invertebrates.