High-throughput and separating-free phenotyping method for on-panicle rice grains based on deep learning.

Area of Science:

• Agricultural Science
• Computer Vision
• Deep Learning
• Plant Phenotyping

Background:

• Rice is a critical global food source, necessitating advancements in breeding for high yield and quality.
• Accurate measurement of rice grain traits is essential for assessing yield potential and quality.
• Current phenotyping methods are often manual, time-consuming, and costly.

Purpose of the Study:

• To develop a high-throughput, non-destructive phenotyping method for rice panicles.
• To accurately extract critical on-panicle rice grain traits without separating or threshing.
• To leverage visible light scanning and deep learning for precise trait analysis.

Main Methods:

• Visible light scanning imaging of rice panicles.
• Deep learning models (Faster R-CNN and Pix2Pix) for grain detection, segmentation, and occlusion compensation.
• An image processing pipeline to calculate fifteen phenotypic traits.

Main Results:

• High correlation (R² values 0.83–0.99) and low error (MAPE 1.65%–9.13%) for key traits compared to manual measurements.
• Rapid imaging (approx. 60 seconds/panicle) and data processing (<10 seconds).
• Identified varietal differences in grain length, width, and length/width ratio distributions.

Conclusions:

• The developed method offers high-throughput, non-destructive, and precise extraction of on-panicle rice grain traits.
• Its low cost and robust performance facilitate widespread adoption in crop breeding.
• Provides novel approaches for phenotyping rice and other crop panicle traits.