High-Throughput Phenotyping of Leaf Discs Infected with Grapevine Downy Mildew Using Shallow Convolutional Neural Networks

Abstract

Objective and standardized recording of disease severity in mapping crosses and breeding lines is a crucial step in characterizing resistance traits utilized in breeding programs and to conduct QTL or GWAS studies. Here we report a system for automated high-throughput scoring of disease severity on inoculated leaf discs. As proof of concept, we used leaf discs inoculated with Plasmopara viticola ((Berk. and Curt.) Berl. and de Toni) causing grapevine downy mildew (DM). This oomycete is one of the major grapevine pathogens and has the potential to reduce grape yield dramatically if environmental conditions are favorable. Breeding of DM resistant grapevine cultivars is an approach for a novel and more sustainable viticulture. This involves the evaluation of several thousand inoculated leaf discs from mapping crosses and breeding lines every year. Therefore, we trained a shallow convolutional neural-network (SCNN) for efficient detection of leaf disc segments showing P. viticola sporangiophores. We could illustrate a high and significant correlation with manually scored disease severity used as ground truth data for evaluation of the SCNN performance. Combined with an automated imaging system, this leaf disc-scoring pipeline has the potential to considerably reduce the amount of time during leaf disc phenotyping. The pipeline with all necessary documentation for adaptation to other pathogens is freely available.

Code and data availability

The authors explicitly state that the complete leaf-disc-scoring pipeline (SCNN training scripts, Jupyter notebook, microscope workflow, and the training/validation image datasets) is publicly available as an open-source GitHub repository. This directly reproduces the paper's phenotyping analysis and includes the plant