Japanese mayfly family classification with a vision transformer model

Benthic macroinvertebrates are a frequently used indicator group for biomonitoring and biological assessment of river ecosystems. However, their taxonomic identification is laborious and requires special expertise. In this study, we aimed to assess the capability of a vision transformer (ViT) model for family-level identification of mayflies (order Ephemeroptera). Specifically, we focused on evaluating the model’s capacity to classify three commonly found mayfly families (Baetidae, Ephemerellidae, and Heptageniidae) as well as other families that were grouped together. For the modeling, we originally constructed two different image datasets containing a total of 1,110 images of mayflies, which were split into training and validation datasets, and a test dataset was prepared from two different online photo galleries. The developed ViT model achieved reasonable accuracy, reaching 94.2% and 82.9% for the validation and test datasets, respectively. Given the use of a relatively small number of images in the training process, as well as some variations in the visual styles of the test dataset compared to the training dataset, we consider the level of accuracy to be high. Our results are encouraging toward the use of computer vision for taxonomic identification of macroinvertebrates, although there is still a need to develop specific designs and plans for this purpose, which can vary depending on regional differences in biodiversity as well as sampling and survey methods.