An automated system for identifying dragonflies (Odonata: Anisoptera) from wings

Dragonflies are excellent ecological indicators for aquatic habitats, but are not always easy to identify.  Here, we created and tested a computer program that uses neural networks (NNs) to identify dragonflies to species from scans of their wings.  The system was tested with 14 classes: 13 libellulid species classes, with 10-41 individuals each, and one class of “rare” species, containing 27 individuals from 11 species from Aeshnidae, Cordulegastridae, Corduliidae, Gomphidae, and Libellulidae.  After digitization and image processing, two types of information were extracted from the wings scans: 15 traditional morphometric features, which describe relative wing shape, and Gabor-wavelet filter features, which describe the arrangement of edges in an image; coefficients from the latter method were down-sampled using principal components analysis (PCA) and F-statistics.  NNs were then trained to positively classify each class and validated to determine classification success rate.  The system was optimized by testing different network architectures and 5 combinations of coefficients.  A one-hidden-layer neural network with 7 neurons gave the best success rate.  NNs trained using the Gabor features down-sampled with PCA in combination with the morphometric features had the best success rate (82.8%), while those trained with only the morphometric features had the least success (68.9%).  We hope to improve the accuracy of this tool, to incorporate more dragonfly species as well as damselflies, and to make it freely available for use through a website interface, with the goal of helping taxonomists, ecologists, government agencies, and others to rapidly and accurately identify odonates.