ARADEEPOPSIS, an Automated Workflow for Top-View Plant Phenomics using Semantic Segmentation of Leaf States

Abstract

Linking plant phenotype to genotype is a common goal to both plant breeders and geneticists. However, collecting phenotypic data for large numbers of plants remain a bottleneck. Plant phenotyping is mostly image based and therefore requires rapid and robust extraction of phenotypic measurements from image data. However, because segmentation tools usually rely on color information, they are sensitive to background or plant color deviations. We have developed a versatile, fully open-source pipeline to extract phenotypic measurements from plant images in an unsupervised manner. ARADEEPOPSIS (https://github.com/Gregor-Mendel-Institute/aradeepopsis) uses semantic segmentation of top-view images to classify leaf tissue into three categories: healthy, anthocyanin rich, and senescent. This makes it particularly powerful at quantitative phenotyping of different developmental stages, mutants with aberrant leaf color and/or phenotype, and plants growing in stressful conditions. On a panel of 210 natural Arabidopsis (Arabidopsis thaliana) accessions, we were able to not only accurately segment images of phenotypically diverse genotypes but also to identify known loci related to anthocyanin production and early necrosis in genome-wide association analyses. Our pipeline accurately processed images of diverse origin, quality, and background composition, and of a distantly related Brassicaceae. ARADEEPOPSIS is deployable on most operating systems and high-performance computing environments and can be used independently of bioinformatics expertise and resources.

Plant phenotyping relevance

植物画像から葉の状態を分割・定量するオープンソースの表現型解析パイプライン開発が中心であり、植物フェノタイピング手法に該当する。

Code and data availability

The paper deposits its ARADEEPOPSIS Nextflow pipeline on GitHub, its manually annotated ground-truth training images on Zenodo (3946393), its trained DeepLabV3+ model checkpoints on Zenodo (3946618), and its model training code on GitHub. All are paper-specific, public, and actionable.