Image-based yield prediction for tall fescue using random forests and convolutional neural networks.

Area of Science:

• Plant breeding
• Agricultural technology
• Machine learning in agriculture

Background:

• Traditional plant breeding relies on subjective visual trait evaluation, which is time-consuming, labor-intensive, and difficult to standardize.
• Automated high-throughput phenotyping offers a solution to these limitations by leveraging technology for objective and efficient trait assessment.

Purpose of the Study:

• To evaluate the accuracy of automated phenotyping using drone-based RGB images and machine learning for assessing tall fescue dry matter yield.
• To compare the performance of machine learning models against traditional breeder evaluations.

Main Methods:

• RGB images of tall fescue were captured and processed using random forest and convolutional neural network (CNN) models.
• The models predicted dry matter yield, identified top-yielding plants, and estimated breeder scores, with field measurements serving as ground truth.

Main Results:

• The CNN model achieved an R² of 0.62 for dry matter yield prediction, outperforming the random forest model and exceeding breeder accuracy by 8 percentage points.
• The CNN demonstrated strong performance in identifying elite genotypes (balanced accuracy 0.81) and predicting breeder scores (balanced accuracy 0.74).

Conclusions:

• Automated phenotyping with RGB imagery and machine learning provides a cost-effective, objective, and accurate alternative to visual evaluation in tall fescue breeding.
• This approach enhances selection accuracy, accelerates genetic progress, and potentially shortens the time to market for improved cultivars.