Discovering Flash Pattern Motifs in Fireflies through Unsupervised Learning

Bioluminescent firefly species have evolved distinguishable flash patterns that are crucial for species

recognition and mating. Each species has a unique flash signal that helps conspecifics identify each

other in the darkness. This is especially critical when two or more species live in sympatry with

one another, as they must somehow share the space to communicate and reproduce successfully.

Prior work with flash pattern datasets has applied human-observer species labels broadly to entire

swarms without the capacity for fine-tuning required to parse apart sympatric species. Using

unsupervised learning techniques, we analyze spatiotemporal trajectories from video recording data

known to contain multiple unknown firefly species and flash motifs. By incorporating spatial

features, such as the velocity and shape of firefly flash streaks, alongside temporal features like

flash duration and flash interval, we extract distinguishing characteristics of each species’ flash

patterns. Subsequent clustering of the principal components enables identification and downstream

classification of the myriad flash motifs that coexist within the same swarm. Overall, our method

identifies the flash pattern composition of filmed mixed-species swarms from across the country

to uncover how sympatric light-communicating species adapt to shared presence and ensure signal

separation and interference reduction