Dragonfly-patterner: A Python pipeline for extracting color patterns from iNaturalist images of dragonflies for broad-scale analysis of pattern evolution

Dragonflies and damselfies are a diverse and colorful insect group, and color is known to serve a critical role in visual communication, thermoregulation and environmental adaptation. However, evolutionary and ecological correlates of color and pattern variation in dragonflies are still underexplored at broad phylogenetic, spatial and temporal scales, in part because colors quickly fade from museum specimens. Here, we leverage citizen science field photographs from iNaturalist to generate spatially and temporally localized color information of live dragonflies. A key challenge is that images from these resources vary among multiple dimensions, including lighting, camera setup, and camera post processing. We developed a pipeline that utilizes machine learning and image processing approaches to filter out pattern signals using large quantities of iNaturalist data. We first filter out bad images while classifying angle views (e.g. dorsal versus lateral) using a trained ResNet. We then segment the dragonfly body from the background using a trained DeepLabV3 segmentation model. Next we apply color clustering using gaussian mixtures to the body segments to reduce noisy and continuous pixel colors to discrete and comparable color patterns. Our pipeline represents, to our knowledge, the first attempt at using mass citizen science images for organismal color pattern data. It will also allow the first assessment of color pattern diversity and its correlates at scale for the Odonate clade.