Colour Extraction Pipeline for Odonates using Computer Vision

A research team from Leiden University and Naturalis Biodiversity Center has developed a novel computer vision pipeline specifically designed for ecological analysis of Odonates (dragonflies and damselflies). The system uses deep neural networks to perform two critical tasks: first, it identifies and segments individual insects in images into four key body parts—head, thorax, abdomen, and wings. Second, it extracts detailed color palettes from each segmented body part. This addresses a major bottleneck in ecological studies, where manual annotation of morphological traits like coloration is extremely time-consuming and limits the scale of research.

The pipeline was trained on a limited dataset of manually annotated images and then refined using pseudo-supervised learning on a larger set of unlabeled data. Crucially, it's designed to work with images from open-source citizen science platforms, allowing researchers to tap into vast repositories of biodiversity photography. This automation enables scientists to conduct statistical analyses on a previously impossible scale, examining correlations between insect coloration and environmental factors like climate change, geolocation, or habitat loss. The 18-page paper, submitted to NCCV 2026, demonstrates how AI can transform field biology by turning casual photographs into quantitative ecological data.

The technical approach is noteworthy for its efficiency. By leveraging pseudo-supervised learning, the researchers minimized the need for expensive, labor-intensive manual annotations. The model's ability to process standard citizen science images—which often vary in quality, angle, and background—makes it particularly practical for real-world deployment. This work represents a significant step toward automated, large-scale monitoring of ecosystem biodiversity, using AI to detect subtle phenotypic changes in insect populations that may serve as early warning signs of environmental stress.