A High-Throughput Phenotyping System Using Machine Vision to Quantify Severity of Grapevine Powdery Mildew

Area of Science:

• Plant pathology
• Agricultural imaging
• Computational biology

Background:

• Powdery mildews pose significant challenges for imaging-based phenotyping.
• Previous low-throughput microscopy methods were used for Erysiphe necator resistance phenotyping in grape genetic studies.
• A need exists for automated, high-throughput phenotyping systems for plant disease resistance.

Purpose of the Study:

• To develop automated imaging and analysis methods for assessing Erysiphe necator severity on grapevine leaf disks.
• To create a system capable of high-throughput, nondestructive phenotyping of grapevine responses to powdery mildew.
• To demonstrate the system's utility for assessing host resistance and treatment efficacy.

Main Methods:

• A high-resolution imaging system was developed using a 46-megapixel CMOS camera and a 3.5× magnification lens.
• Automated X-Y positioning and Z-axis focusing captured detailed images of grape leaf disks.
• A convolutional neural network (CNN) based on GoogLeNet analyzed subimages to determine Erysiphe necator presence and severity.

Main Results:

• The system captured 78% of a 1-cm leaf disk area in 3-10 focus-stacked images within 13.5-26 seconds.
• The CNN achieved 94.3% training validation accuracy and agreed with human experts on 89.3%–91.7% of subimages.
• The nondestructive live-imaging approach successfully differentiated between susceptible, moderate, and resistant grapevine samples over time.

Conclusions:

• The developed automated system enables accurate, nondestructive, and high-throughput phenotyping of grapevine powdery mildew resistance.
• The system can process over one thousand samples daily, facilitating host resistance and treatment efficacy assessments.
• The CNN-based approach is adaptable for phenotyping diverse pathosystems and traits using leaf disk assays.