What you sample is what you get: ecomorphological variation in Trithemis (Odonata, Libellulidae) dragonfly wings reconsidered

Abstract

Background The phylogenetic ecology of the Afro-Asian dragonfly genus Trithemis has been investigated previously by Damm et al. (in Mol Phylogenet Evol 54:870–882, 2010) and wing ecomorphology by Outomuro et al. (in J Evol Biol 26:1866–1874, 2013). However, the latter investigation employed a somewhat coarse sampling of forewing and hindwing outlines and reported results that were at odds in some ways with expectations given the mapping of landscape and water-body preference over the Trithemis cladogram produced by Damm et al. (in Mol Phylogenet Evol 54:870–882, 2010). To further explore the link between species-specific wing shape variation and habitat we studied a new sample of 27 Trithemis species employing a more robust statistical test for phylogenetic covariation, more comprehensive representations of Trithemis wing morphology and a wider range of morphometric data-analysis procedures. Results Contrary to the Outomuro et al. (in J Evol Biol 26:1866–1874, 2013) report, our results indicate that no statistically significant pattern of phylogenetic covariation exists in our Trithemis forewing and hindwing data and that both male and female wing datasets exhibit substantial shape differences between species that inhabit open and forested landscapes and species that hunt over temporary/standing or running water bodies. Among the morphometric analyses performed, landmark data and geometric morphometric data-analysis methods yielded the worst performance in identifying ecomorphometric shape distinctions between Trithemis habitat guilds. Direct analysis of wing images using an embedded convolution (deep learning) neural network delivered the best performance. Bootstrap and jackknife tests of group separations and discriminant-function stability confirm that our results are not artifacts of overtrained discriminant systems or the “curse of dimensionality” despite the modest size of our sample. Conclusion Our results suggest that Trithemis wing morphology reflects the environment’s “push” to a much greater extent than phylogeny’s “pull”. In addition, they indicate that close attention should be paid to the manner in which morphologies are sampled for morphometric analysis and, if no prior information is available to guide sampling strategy, the sample that most comprehensively represents the morphologies of interest should be obtained. In many cases this will be digital images (2D) or scans (3D) of the entire morphology or morphological feature rather than sparse sets of landmark/semilandmark point locations.

Challenges and opportunities in applying AI to evolutionary morphology

“…AI has been implemented in this field through the use of algorithms that infer present and past ecomorphologies by reducing the dimensionality of ecomorphological data through ML pipelines such as Random Forest analyses (Mahendiran et al, 2022;Rabinovich, 2021;Sosiak and Barden, 2021;Spradley et al, 2019). Similarly, ML procedures have been used to discriminate and sort phenotypes (especially morphology) based on their belonging to specific ecomorphs or ecological guilds (MacLeod et al, 2022). These studies have highlighted the advantages of AI-based approaches compared to standard procedures used to test the links between morphology and ecology, such as Canonical Variate Analysis (Albrecht, 1980).…”

Male and female contributions to diversity among birdwing butterfly images

“…Here, we use machine-learnt embeddings to quantify and characterise, relative to predictions of sexual versus natural selection in phenotypic diversification 1 – 6 , sexual and interspecific variation across 16,734 dorsal and ventral photographs of birdwing butterflies, covering the entire Natural History Museum (NHMUK) birdwing collection, the largest and most comprehensive known on this group, including the three genera, 35 species OTUs (operational taxonomic units) and 131 recognised subspecies. Until very recently, methods capable of quantitatively capturing phenotypic variation approaching this scale and complexity did not exist 30 , 35 – 42 . We use deep learning with a triplet-trained convolutional neural network (CNN), ButterflyNet version 1.2 (Supplementary Software 1 , modified from ButterflyNet version 1 35 ), to generate Euclidean spatial embeddings of uniformly scaled, dorsal and ventral photographs.…”