Extracting Quantitative Information from Images Taken in the Wild: A Case Study of Two Vicariants of the Ophrys aveyronensis Species Complex

Abstract

Characterising phenotypic differentiation is crucial to understand which traits are involved in population divergence and establish the evolutionary scenario underlying the speciation process. Species harbouring a disjunct spatial distribution or cryptic taxa suggest that scientists often fail to detect subtle phenotypic differentiation at first sight. We used image-based analyses coupled with a simple machine learning algorithm to test whether we could distinguish two vicariant population groups of an orchid species complex known to be difficult to tease apart based on morphological criteria. To assess whether these groups can be distinguished on the basis of their phenotypes, and to highlight the traits likely to be the most informative in supporting a putative differentiation, we (i) photographed and measured a set of 109 individuals in the field, (ii) extracted morphometric, colour, and colour pattern information from pictures, and (iii) used random forest algorithms for classification. When combined, field- and image-based information provided identification accuracy of 95%. Interestingly, the variables used by random forests to discriminate the groups were different from those suggested in the literature. Our results demonstrate the interest of field-captured pictures coupled with machine learning classification approaches to improve taxon identification and highlight candidate traits for further eco-evolutionary studies.