Skip to main content
ARS Home » Midwest Area » Madison, Wisconsin » U.S. Dairy Forage Research Center » Dairy Forage Research » Research » Publications at this Location » Publication #367009

Research Project: Improving Forage Genetics and Management in Integrated Dairy Systems for Enhanced Productivity, Efficiency and Resilience, and Decreased Environmental Impact

Location: Dairy Forage Research

Title: Using image object recognition to increase biomass in red clover (Trifolium pratense L.) breeding

item Sindic, Caleb
item Riday, Heathcliffe

Submitted to: Crop Science
Publication Type: Peer Reviewed Journal
Publication Acceptance Date: 12/16/2019
Publication Date: 12/31/2019
Citation: Sindic, C.M., Riday, H. 2019. Using image object recognition to increase biomass in red clover (Trifolium pratense L.) breeding. Crop Science. 60:1771-1781.

Interpretive Summary: Using artificial intelligence (AI) to automate plant breeding processes is becoming more common. We report here the implementation of AI in red clover plant breeding. The AI was used to identify red clover plants in digital images and then to assign the red clover plants a plant dry matter yield estimate. The AI was over 94% successful at recognizing red clover plants in digital images. The AI plant dry matter yield estimate was correlated with human plant dry matter yield estimates at r-square equal to 0.528. We conclude that using AI in red clover breeding is currently feasible. We further conclude that additional plant traits could be obtained using AI from AI identified red clover plants observed in digital images.

Technical Abstract: Phenotyping in forage legume breeding can be time consuming and resource intensive. With computation advances and computational power cost reductions, utilizing artificial intelligence in automated phenotyping is becoming feasible for even resource limited forage legume breeding programs. Here we report on the use of machine learning to train a neural network to identify and isolate red clover plants from digital images of space plant red clover nurseries. Challenges to red clover plant identification include: 1) plants were grown with a grass companion; and 2) plants were close enough to each other so that plants often overlapped each other in the digital images. To estimate biomass yield, a second neural network was trained using machine learning to count leaves per plant among identified red clover plants from the plant classification neural network. The two neural networks were validated on six red clover digital image sets taken on red clover space plant nurseries. Average neural network plant classification success rates were measured at 94.6% across the six digital image sets. Neural network red clover leaf counts were correlated with human visual biomass scores at r2 = 0.528. We conclude that automated phenotyping based on digital image analysis of red clover breeding nurseries is currently feasible. We further conclude that additional phenotypic traits could be obtained on identified red clover plants from such images tests.