Objective Phenotyping of Root System Architecture Using Image Augmentation and Machine Learning in Alfalfa (Medicago...

Area of Science:

• Plant genetics and breeding
• Computational biology
• Agricultural science

Background:

• Root system architecture (RSA) phenotyping for crop breeding is challenging, with limited objective methods.
• Current RSA phenotyping relies on subjective visual scoring, hindering efficient breeding programs.
• Alfalfa breeding for specific root types is ongoing but lacks advanced phenotyping tools.

Purpose of the Study:

• To develop and compare image-based RSA phenotyping methods for alfalfa using machine and deep learning.
• To enable objective classification of root types to support ongoing alfalfa breeding efforts.
• To enhance genomic breeding for climate-resilient alfalfa through accurate RSA phenotyping.

Main Methods:

• Collected and analyzed 617 field-grown alfalfa root images.
• Applied unsupervised machine learning, random forest, and TensorFlow-based neural networks for root classification.
• Utilized image augmentation techniques to improve model accuracy.

Main Results:

• Unsupervised machine learning showed limitations, incorrectly classifying most roots into an intermediate type.
• Random forest and deep learning models achieved 86% accuracy in classifying branch-type, taproot-type, and intermediate root types.
• Image augmentation significantly improved prediction accuracy to 97%.

Conclusions:

• Machine and deep learning offer accurate and objective RSA phenotyping for alfalfa.
• High prediction accuracy supports informed decisions in breeding programs.
• This approach facilitates genomic breeding for climate-resilient alfalfa with improved root traits.