BugNet: a rapid and scalable pipeline for automated insect monitoring using hierarchical data

Abstract

Despite the importance of monitoring insect diversity to ecological and conservation questions, we lack sufficient technologies to monitor insects at scale. While research into automated systems for monitoring biodiversity through camera traps has led to the development of a number of machine learning approaches for insect monitoring, these tools suffer from a lack of training data and face challenges in classifying insects in highly diverse systems where the majority of species are unknown to science. To address these challenges, we developed BugNet, an automated pipeline for aggregating insect image data from online databases and training hierarchical classification models, and test a large-scale insect detection model on GBIF and field images. We show that this system can be used to rapidly create and validate classification models with high accuracy on internet and field images. Furthermore, we show that incorporating hierarchical data into classification models improves their ability of models to handle unknown taxa. These systems are an important step towards a generalized and scalable insect detection platform. While not capable of monitoring every dimension of insect diversity, BugNet can be used to accurately classify insects from camera trap images, and is can be scaled to meet the data needs of larger ecological and conservation questions.