Decoding mimicry: Quantifying variations in butterfly coloration patterns

Animal coloration is a complex morphological trait that exhibits fascinating patterns of variation. Butterflies from different species may resemble one another due to mimicry or convergent evolution, while individuals from the same species can appear quite distinct because of phenotypic plasticity and developmental noise. Such inter- and intraspecific variation is important for many ecological and evolutionary processes, including natural and sexual selection. To study such variation, we apply machine learning methods to quantify intricate coloration patterns. We collected a large dataset of butterfly images from museum collections, with many specimens from each population. We used a denoising autoencoder to embed these images into a low-dimensional space that captures the main features of their wing patterns. This unsupervised approach allows us to measure morphological similarity between mimetic species by comparing the degrees of variation within and across species. Patterns of variation in this morphospace may bring new insights into biological processes such as predator interactions and conspecific signaling.