Figure 1

(A) Our standardized data collection method (top) films fireflies in their natural habitat with two cameras arranged in a stereoscopic vision configuration²² (photo reproduced and modified with permission from²³). Flash streaks (center, colored by time) in the resulting videos are triangulated and concatenated into trajectories, based on proximity and velocity (bottom). Each trajectory is represented by a time series of individual flashes. (B) Four example five-second flash sequence time series for each of the seven species in our study, labeled with the location and date of the corresponding recording. Our recurrent neural network is used to classify flash sequence time series; time series shown were selected from the top 100 sequences per species with the highest classification probabilities following the filtration process outlined in Methods Section "Characterization". (C) Two-dimensional t-SNE embedding of the output of the last hidden layer of the network on the top 100 predictions, just before inference. Flash patterns are clustered by similarity, and the distinguishability of each species’ characteristic flash pattern can be detected by the colored clusters.