D3070: Can we automate bug identification? A case study using deep learning to identify photographs of North American arthropods

This project seeks to identify arthropod specimens quickly and accurately through taxonomic classifications such as class, family, and genus. Identification is an integral step to any downstream biological application such as designing experiments with species-species interactions. However, the way we approach identification usually requires expert knowledge, is often time-consuming, and may be biased or inaccurate. Finding ways to minimize these issues leads to more efficient and better-quality research. We address this by training a deep learning model that can identify North American arthropod genera from images using deep learning. Deep learning is a new class of machine learning that can solve complex problems that are too difficult to program by hand. This approach has increased in popularity due to advances in both data collection and hardware used for deep learning. We chose to collect our data from two sources, BugGuide.net and iNaturalist.org, because they provide us with a large amount of images and complement each other. BugGuide.net offers high-quality, carefully curated images while iNaturalist.org contains more images of common species captured at often lower quality. iNaturalist’s identification tool “Seek” attempts automated identification of all living organisms at the expense of accuracy, while our method maximizes accuracy because of our focus on specifically arthropods. In the future, we plan to make this technology available as a phone application and website as well as integrate support for live video processing and object detection. This will allow researchers and the public to use our application easily and integrate our technology into larger projects.