A high-throughput ResNet CNN approach for automated grapevine leaf hair quantification.

Area of Science:

• Plant Science
• Genetics
• Agricultural Technology

Background:

• Leaf hairiness in Vitis is a key trait for physical defense against pathogens like downy mildew and anthracnose.
• Leaf hairs influence wettability and provide habitats for biological control agents, impacting grapevine health.
• Quantifying leaf hair density objectively is challenging due to the lack of efficient tools.

Purpose of the Study:

• To develop and validate a high-throughput phenotyping tool for accurate and objective quantification of grapevine leaf hair density.
• To leverage convolution neural networks (CNNs) for image-based analysis of leaf hair coverage.
• To compare the performance of the developed tool against traditional manual evaluations by experts and non-experts.

Main Methods:

• Development of a phenotyping tool using modified ResNet CNNs trained on grapevine leaf disc images.
• Classification of leaf hair area using a minimalistic image dataset.
• Validation using 10,120 images from an F1 biparental population and comparison with expert ground truth data.

Main Results:

• The CNN model achieved 95.41% prediction accuracy.
• Phenotypic results showed strong correlation (R=0.98, 0.92) and low RMSE (8.20%, 14.18%) compared to expert data.
• The AI tool demonstrated superior consistency and accuracy over manual expert and novice evaluations, highlighting significant bias in non-expert assessments.

Conclusions:

• The developed CNN-based tool provides an objective, accurate, and efficient method for quantifying grapevine leaf hairiness.
• This high-throughput phenotyping approach is essential for advancing research on the role of leaf morphology in plant defense and breeding.
• The study underscores the limitations of subjective manual assessments and the need for automated tools in plant trait evaluation.