Automating bee wing analysis: Computer vision for biodiversity monitoring and evolutionary ecology research

The global diversity of bees (over 20,000 species) presents practical challenges for taxonomic identification, hindering biodiversity research and monitoring. Patterns in the venation of insect wings are strongly conserved, and therefore are useful indicators of species and population identity. However, the spatial complexity of these patterns makes comparisons between individuals challenging and labor-intensive. We apply traditional (geometric morphometrics) and AI (computer vision) methods to quantify wing venation differences among and within bee species. We trained a convolutional neural network on a large image dataset, resulting in a model that can classify local bees to species from images of their wings, with >95% accuracy. We outline a standardized wing imaging protocol that can be deployed on live bees in the field, using computer vision to correct for variation in perspective and focal distance. Finally, we explore manual and automated methods for quantifying differences in wing vein patterns, which can be applied to monitor bee health (via fluctuating asymmetry) or explore patterns of phenotypic divergence among species and populations. Together, these results demonstrate the potential of high-throughput machine learning tools for accelerating basic evolutionary and applied conservation research.