Resumen
Honey bee classification by wing geometric morphometrics entails the first step of manual annotation of 19 landmarks in the forewing vein junctions. This is a time-consuming and error-prone endeavor, with implications for classification accuracy. Herein, we developed a software called DeepWings© that overcomes this constraint in wing geometric morphometrics classification by automatically detecting the 19 landmarks on digital images of the right forewing. We used a database containing 7634 forewing images, including 1864 analyzed by F. Ruttner in the original delineation of 26 honey bee subspecies, to tune a convolutional neural network as a wing detector, a deep learning U-Net as a landmarks segmenter, and a support vector machine as a subspecies classifier. The implemented MobileNet wing detector was able to achieve a mAP of 0.975 and the landmarks segmenter was able to detect the 19 landmarks with 91.8% accuracy, with an average positional precision of 0.943 resemblance to manually annotated landmarks. The subspecies classifier, in turn, presented an average accuracy of 86.6% for 26 subspecies and 95.8% for a subset of five important subspecies. The final implementation of the system showed good speed performance, requiring only 14 s to process 10 images. DeepWings© is very user-friendly and is the first fully automated software, offered as a free Web service, for honey bee classification from wing geometric morphometrics. DeepWings© can be used for honey bee breeding, conservation, and even scientific purposes as it provides the coordinates of the landmarks in excel format, facilitating the work of research teams using classical identification approaches and alternative analytical tools.