Can we use artificial intelligence to identify insects? Opportunities, pitfalls, and hurdles in adoption

Arthropod identification poses a significant challenge in biodiversity science due to the complexity of sorting and identifying these diverse creatures. With arthropods comprising 84% of multicellular animal species, the task is formidable, exacerbated by limited taxonomists, unsorted collections, and time-intensive identification processes.



The bottleneck stems from a shortage of taxonomists with specialized knowledge, as no individual can cover all arthropod groups comprehensively. This necessitates collaboration and specimen exchange among experts, further impeding progress. Additionally, declining arthropod populations worldwide highlight the urgency of addressing this issue.



We leveraged deep learning models to arthropod identification from images, obtaining 80% accuracy on 7,524 genera of North American arthropods. Through a funded project, we are applying this technology to identification of pests and beneficials in cereal cropping systems. Such tools, alongside digitization and robotic sorting methods, promise to accelerate biodiversity research, facilitate ecosystem monitoring, and aid in pest management decisions. At the same time, there are significant limitations to deep learning approaches and barriers exist to widespread adoption in agriculture.