Identification of Indian Butterflies and Moths with Deep Convolutional Neural Networks x

This paper reports our efforts to use artificial intelligence based on deep convolutional neural network (CNN) as a tool to identify Indian butterflies and moths. We compiled a dataset of over 170,000 images for 800 Indian butterfly species and 500 Indian moth species from diverse sources. We adopted the Effi-cientNet-B6 architecture for our CNN model, with about 44 million learnable parameters. We trained an ensemble of 5 such models on different subsets of the images in our data, employing artificial image augmentation techniques and transfer learning. This ensemble achieved a balanced top-1 accuracy of 86.5%, top-3 accuracy of 94.7%, and top-5 accuracy of 96.4% on the 1300 species, and a mean F1score of 0.867. Thus, our efforts demonstrate artificial intelligence can be effectively used for identifying these biological species that would substantially enhance the work efficiency of field level biologists in several spheres of investigations.