Butterfly specimens (1269) from the species H. erato and H. melpomene were photographed at the NHM London (specimen numbers and image filenames, table S2), using consistent photographic protocols and lighting. Photographs were screened for poor image quality giving a final dataset of 1234 butterflies and 2468 photographs, including a dorsal and ventral photograph of each butterfly. For deep learning, images were then cropped and resized to a height of 64 pixels (maintaining the original image aspect ratio and padded to 140 pixels wide). The photographic dataset used for deep learning is provided in the Dryad Data Repository with filenames corresponding to joined data in table S2.

Taxonomic and locality data were recorded from NHM Heliconius butterfly specimen labels (table S2). Subspecies taxonomy follows reference (30). The complete photographic dataset covers 37 named subspecies and one-labeled cross: 21 subspecies from H. erato and 17 from H. melpomene. Specimens of these subspecies were sampled exhaustively from the NHM collection, within the limits of the data collection period. The complete photographic dataset (fig. S4, Dryad Data Repository) covers both specimens closely representative of subspecies descriptions (30, 31) (including available holotypes, syntypes, and paratypes; table S2) as well as other, naturally varying, individuals. These variants include some likely hybrid specimens showing varying levels of phenotypic admixture from other subspecies (see additional taxonomic information, table S2). Inclusion of all available specimens in machine learning covered a very broad range of the phenotypic diversity within these species, providing the deep learning network with all available information from which to learn phenotypic features correlated with subspecies identification. Two sets of statistical analyses were then conducted, one set including all 1269 photographed butterfly specimens and the second set excluding potential hybrid specimens to give a reduced dataset of 815 specimens and 1630 photographs (fig. S4 and table S2).

Note: The content above has been extracted from a research article, so it may not display correctly.

Please log in to submit your questions online.
Your question will be posted on the Bio-101 website. We will send your questions to the authors of this protocol and Bio-protocol community members who are experienced with this method. you will be informed using the email address associated with your Bio-protocol account.

We use cookies on this site to enhance your user experience. By using our website, you are agreeing to allow the storage of cookies on your computer.