Identification of Melipona species using geometric morphometrics and neural networks
Juliana Stephanie Galaschi Teixeira, Juliana Stephanie Galaschi Teixeira , Ana Carolina Quintão Siravenha , Schubert Ribeiro de Carvalho , Vera Lucia Imperatriz Fonseca
Instituto Tecnológico Vale - Desenvolvimento Sustentável, Belém, Brazil; Instituto Tecnológico Vale - Desenvolvimento Sustentável, Belém, Brazil ; Instituto Tecnológico Vale - Desenvolvimento Sustentável, Belém, Brazil ; Instituto Tecnológico Vale - Desenvolvimento Sustentável, Belém, Brazil ; Instituto Tecnológico Vale - Desenvolvimento Sustentável, Belém, Brazil
The genus Melipona is one of the most diverse of the tribe Meliponini, with approximately seventy species with an exclusive Neotropical distribution, divided into four subgenera. They are eusocial bees, which produce their queens in undifferentiated brood cells, being species of interest for honey production and pollination. Due to their high diversity and recent divergence, it’s difficult to identify their species, which makes it necessary to use alternative methods of identification. This study aims to identify species of Melipona through wing shape using two approaches: geometric morphometrics and artificial neural networks. For the analysis we used the right wing of 1588 workers of eight species: M. (Melikerria) fasciculata, M. (Melikerria) interrupta, M. (Michmelia) fuliginosa, M. (Michmelia) melanoventer, M. (Michmelia) nebulosa, M (Michmelia) flavolineata, M. (Michmelia) paraensis and M. (Michmelia) seminigra. For the morphometric analysis, we plotted 12 anatomical landmarks at the wing vein junctions and Procrustes optimization, Canonical Variate Analysis with Mahalanobis distances and Procrustes distances, and Discriminant Function Analysis were calculated. The correct proportiton of classifications of discriminant analysis ranged between 82% and 100%, with an average of 97.5%. The species with the highest score of correct classifications were M. fuliginosa and M. melanoventer, with 100% of accuracy in all comparisons. By using artificial intelligence approaches, in particular, the convolutional neural networks, the percentage of species identification accuracy was approximately 98%. One of the advantages of convolutional neural networks is the ability of automating the species identification from photographs of bee wings. Both methodologies were effective in the classification of similar species, proving the efficiency of image-based methods for the taxa identification.