Automated Specimen Triage for Dark Taxa: Deep Learning Enables Orientation, Sex Identification, and Anatomical Segmentation from Robotic Imaging

Abstract

Abstract Robotic specimen processing is transforming biodiversity discovery by replacing labor-intensive handling with scalable systems that can simultaneously generate high-quality specimen images. We demonstrate that these images can be leveraged by deep learning to efficiently extract key biological information and guide targeted specimen processing. Using a model dark taxon, the Phoridae (Diptera), the workflow performs three core tasks: sex identification, specimen orientation classification, and anatomical segmentation. Sex identification allows selective retention of diagnostically informative specimens, avoiding wasted effort on non-diagnostic individuals. Orientation classification enables specimens in the desired orientation to proceed immediately, while suboptimally oriented specimens can be repositioned for informative processing. Anatomical segmentation allows targeted processing of specimens displaying diagnostic characters or targeted analysis of specific anatomical regions in subsequent workflow steps. Comparative analysis of model architectures shows task-specific selection is crucial: a Convolutional Neural Network achieved an accuracy of 0.94 for orientation, a Vision Transformer achieved 0.88 for sex, and a U-Net precisely segmented nine anatomical regions with a mean IoU of 0.78. These results demonstrate that robotic imaging combined with deep learning provides a validated foundation for high-throughput, targeted specimen processing, maximising efficiency and utility for taxonomic and trait-based analyses, and supporting scalable, sustainable biodiversity workflows.

Ähnliche Arbeiten