Deep learning-based detection of seedling development.

Area of Science:

• Plant Science
• Computational Biology
• Agricultural Technology

Background:

• High-throughput phenotyping using computer vision is crucial for monitoring seedling emergence and early development.
• Current research often overlooks seed-to-seed kinetics of early developmental events, focusing instead on metrics like leaf area index or single event detection.

Purpose of the Study:

• To develop and validate a comprehensive image processing and machine learning pipeline for classifying early seedling growth stages.
• To accurately identify and track developmental events on a seed-to-seed basis using computer vision.

Main Methods:

• Utilized top-view imaging systems to capture the entire seedling growth process.
• Precisely annotated key developmental stages: emergence, cotyledon opening, and first leaf appearance.
• Trained deep neural networks, exploring strategies to integrate prior knowledge of developmental stage order, with a deep neural network and long short-term memory cell combination yielding the best performance.

Main Results:

• Achieved over 90% accuracy in correctly detecting and classifying three distinct early growth stages (emergence, cotyledon opening, first leaf appearance) plus soil.
• Demonstrated the pipeline's effectiveness on red clover and alfalfa accessions.

Conclusions:

• The developed pipeline offers a robust solution for automated monitoring of early plant development.
• The methodology is adaptable for various crops and can be extended to analyze additional developmental stages.