Odonates (order Odonata: dragonflies and damselflies) are important ecosystem service providers, preying on and controlling populations of mosquitos and other vectors of human disease, while also acting as indicators sensitive to changes in habitats, environment and climate.
Recognizing the complexities in accurately identifying odonate species, this study proposed a Convolutional Neural Network approach to fast and effective identification of dragonflies and damselflies.
Focussing on India, the authors sourced a collection of ~55,000 images of 256 native species from iNaturalist (via GBIF), Flickr and other sources. They then trained nine customized CNN architectures at four different image resolutions, dividing the dataset into 80 per cent for training and 20 per cent for validation.
The best performing model correctly identified the species in question in 94 per cent of cases. When prompted to select the top three candidates, the model included the correct species 98 per cent of the time. Model performance increased relatively with image resolution, although only to a certain degree after which additional irrelevant features led to overfitting.
While the best performing models had the most depth, size and system requirements, even the most lightweight model, SqueezeNet, which uses just five megabytes of storage, had a 94 per cent top-three accuracy rate, making it ideal for small computing platforms, like web browsers and mobile phones.