DeepWings: a machine learning tool for identification of honey bee subspecies

DeepWings© is a software that uses Machine Learning for fully automated identification of Apis mellifera subspecies based on wing geometric morphometrics (WGM). Here, we examined the performance of DeepWings© under realistic conditions by processing 14,782 wing images with varying quality and produced by different operators. These images represented 2,593 colonies covering the native ranges of A. m. iberiensis (Portugal, Spain and historical introduction in the Azores), A. m. mellifera (Belgium, France, Ireland, Poland, Russia, Sweden, Switzerland, UK) and A. m. carnica (Croatia, Hungary, Romania). The classification probability obtained for the colonies was contrasted with the endemic subspecies distribution. Additionally, the association between WGM classification and that inferred from microsatellites and SNPs was evaluated for 1,214 colonies. As much as 94.4% of the wings were accepted and classified by DeepWings©. In the Iberian honey bee native range, 92,6% of the colonies were classified as A. m. iberiensis with a median probability of 91.88 (IQR = 22.52). In the Azores, 85.7% of colonies were classified as A. m. iberiensis, with a median probability of 84.16 (32.40). In the Dark honey bee native range, 41.1 % of the colonies were classified as A. m mellifera with a median probability of 99.36 (8.02). The low percentage of colonies matching the native subspecies was mainly due to the low values registered in Avignon (20.0%), Poland (32.9%), and Wales (41.2%). In contrast, most of the colonies analyzed in other locations of the native range of A. m. mellifera matched this subspecies: Belgium (100.0%), Groix (63.9%), Ouessant (72.7%), Ireland (78.0%), Russia (96.2%), Sweden (84.2%) and Switzerland (55.6%). In the colonies from Croatia, Hungary, and Romania, 88.0% of the samples were classified as A. m. carnica, with a median probability of 98.49 (6.76). The association between WGM and molecular data was highly significant but not very strong (Spearman r = 0.31, p < 0.0001). A good agreement between ...