Counting and identifying invertebrates from pitfall traps: a novel method using images and deep learning

The identification of invertebrates from standardised sampling methods such as pitfall traps is often time-consuming because of their high abundance and diversity. However, new technologies based on images and deep learning techniques are rapidly improving the efficiency of invertebrate identification, particularly for non-specialists. In this study, an image dataset of New Zealand ground invertebrates was used to train object detection and classification models to identify invertebrates. We captured a single image of the contents of a pitfall trap, annotated the specimens to a taxonomic group, and applied a two-stage YOLO deep learning model training for identification. Our model precision varied between taxa but had an overall mean average precision of 0.95 for the object detection stage. The classification stage achieved a top1 accuracy of 0.92. Groups with the highest observation counts were Coleoptera, Isopoda, and Formicidae, which also achieved the highest identification accuracy. Although deep learning technology is still in its infancy for entomology, such approaches could transform field sampling, laboratory counting, and identification of invertebrates.