Résumé
Insect identification is essential to taxonomy, ecology, and public-health. Traditional methodsrely on expert knowledge and intact specimens, making them time-consuming and limiting. Insectwings carry rich structural and optical information that can reveal taxonomic identity. While wingvenation is a known diagnostic trait, we propose using optical features encoded in Wing Interfer-ential Patterns (WIPs), arising from thin-film interference, as a novel and robust marker. Usinga dataset of 5,502 high-resolution images from seven Dipteran families, we trained convolutionalneural networks on a species-subset represented by at least 10 specimens, to classify species basedon WIPs, achieving over 90% accuracy. These highlight the potential of WIPs as a scalable andeffective tool for insect identification.Building on this, we introduce an open-set framework supporting classification in the context ofincomplete taxonomic knowledge. We demonstrate that the latent representations learned by ourmodels, when visualized, form a structured space in which the taxonomic hierarchy (family, genus,species) emerges naturally, like a WIP-based insect atlas. This atlas is not only interpretablebut also expandable, supporting both the discovery of unseen species and their integration withminimal retraining.Beyond taxonomy, WIPs can convey additional information related to ecological and physio-logical traits. Preliminary analyses show WIP variations correlated with sex and blood-feedingstatus, suggesting that WIPs act as multidimensional biological sensors encoding more than iden-tity alone.These contributions establish an interpretable open-set-compatible framework for insect iden-tification and trait inference. The “WIP-based insect atlas” offers an extensible platform forbiodiversity research, ecological monitoring, and vector surveillance in dynamic, data-limited en-vironments.