Using the Software DeepWings© to Classify Honey Bees across Europe through Wing Geometric Morphometrics

SIMPLE SUMMARY: Wing venation traits are used to identify honey bee subspecies. While several wing-based tools are available, they suffer from weaknesses that were addressed by the recently developed software DeepWings©. This software allows fully automated identification of wing images in a friendly, free, and rapid manner. Here, we sought to test DeepWings© on 14,816 wing images representing 2601 colonies sampled in the native areas of three widespread subspecies in Europe: the Iberian honey bee (Apis mellifera iberiensis), the dark honey bee (Apis mellifera mellifera), both belonging to the M lineage, and the Carniolan honey bee (Apis mellifera carnica), belonging to the C lineage. DeepWings© classification of these colonies largely matched the endemic M and C lineages, with proportions of 71.4% and 97.6%, respectively. At the subspecies-level the matching proportions were 89.7% for the Iberian honey bee, 41.1% for the dark honey bee and 88.3% for the Carniolan honey bee, which can be explained by DeepWings© sometimes confounding closely related subspecies and, more importantly, by genetic pollution. A comparison between DeepWings© data and molecular data revealed that the agreement between the two is weaker when there is genetic pollution. Our results suggest that DeepWings© is a valuable tool for honey bee identification, which can be used not only for breeding and conservation but also for research purposes. ABSTRACT: DeepWings© is a software that uses machine learning to automatically classify honey bee subspecies by wing geometric morphometrics. Here, we tested the five subspecies classifier (A. m. carnica, Apis mellifera caucasia, A. m. iberiensis, Apis mellifera ligustica, and A. m. mellifera) of DeepWings© on 14,816 wing images with variable quality and acquired by different beekeepers and researchers. These images represented 2601 colonies from the native ranges of the M-lineage A. m. iberiensis and A. m. mellifera, and the C-lineage A. m. carnica. In the A. m. iberiensis range, 92.6% of the colonies matched this subspecies, with a high median probability (0.919). In the Azores, where the Iberian subspecies was historically introduced, a lower proportion (85.7%) and probability (0.842) were observed. In the A. m mellifera range, only 41.1 % of the colonies matched this subspecies, which is compatible with a history of C-derived introgression. Yet, these colonies were classified with the highest probability (0.994) of the three subspecies. In the A. m. carnica range, 88.3% of the colonies matched this subspecies, with a probability of 0.984. The association between wing and molecular markers, assessed for 1214 colonies from the M-lineage range, was highly significant but not strong (r = 0.31, p < 0.0001). The agreement between the markers was influenced by C-derived introgression, with the best results obtained for colonies with high genetic integrity. This study indicates the good performance of DeepWings© on a realistic wing image dataset.