Deep learning-based methods for phenotypic trait extraction in rice panicles

Area of Science:

• Agricultural Science
• Plant Breeding
• Computer Vision

Background:

• Accurate measurement of rice panicle traits is crucial for improving crop yield and quality.
• Traditional phenotyping methods are often labor-intensive and lack high-throughput capabilities.

Purpose of the Study:

• To develop and validate a deep learning pipeline for automated, high-precision rice panicle phenotyping.
• To address challenges in measuring traits like grain number, dimensions, and maturity stages, especially under occlusion.

Main Methods:

• A dataset of 5300 rice panicle images was curated, encompassing diverse types and maturity stages.
• A deep learning pipeline, OPG-YOLOv8, was implemented for image analysis and trait extraction.
• The model was trained and validated on subsets of the image data.

Main Results:

• High accuracy was achieved for panicle length extraction (R²=0.9583) and grain counting across different panicle types (R² up to 0.9799).
• Accurate measurements for grain length (R²=0.8823) and grain width (MAPE=6.64%) were obtained.
• The OPG-YOLOv8 model demonstrated robust performance in phenotyping.

Conclusions:

• The developed automated tool offers a comprehensive solution for rice panicle phenotyping.
• This technology effectively overcomes occlusion issues and bridges the gap between advanced AI models and practical breeding applications.