Leveraging behavioral variability for robust classification of firefly flash signals

Swarms of fireflies flash in patterns to advertise their presence to potential mates, resulting in dazzling light displays. However, light pollution, climate change, and other factors are threatening firefly populations worldwide. This reality has kickstarted conservation efforts rooted in a quantitative understanding of firefly signaling behavior. Here, we present a high-throughput analysis pipeline to extract trajectories and flash patterns of individual fireflies recorded in the field. We observe significant behavioral variability within species, which mandates a characterization of firefly flash behavior that extends beyond the prior practice of representing a species by a single discrete pattern. We then train a recurrent neural network using flash pattern data to accurately classify firefly species. Our computational methods enable an extensive characterization of how signaling behaviors vary spatially and temporally across firefly populations and species around the globe.