Convolutional Neural Networks to Estimate Dry Matter Yield in a Guineagrass Breeding Program Using UAV Remote...

Area of Science:

• Agricultural Science
• Computer Vision
• Plant Breeding

Background:

• Forage dry matter yield is a key trait in ruminant nutrition and forage breeding programs.
• Traditional methods for evaluating forage dry matter yield are labor-intensive and limit breeding efficiency.
• High-throughput phenotyping (HTP) using novel technologies is needed to accelerate breeding progress.

Purpose of the Study:

• To develop and evaluate a convolutional neural network (CNN) approach using UAV-RGB imagery for estimating dry matter yield traits in guineagrass.
• To assess the potential of CNNs for high-throughput phenotyping (HTP) in forage breeding programs.
• To compare the performance of different CNN architectures for estimating leaf dry matter yield (LDMY) and total dry matter yield (TDMY).

Main Methods:

• Utilized UAV-RGB imagery from a Phantom 4 PRO drone over 330 guineagrass plots.
• Employed various CNN architectures (AlexNet, ResNeXt50, DarkNet53, MaCNN, LF-CNN), including pretrained models.
• Validated CNN-based estimates against ground-truth data (LDMY, TDMY) using ten-fold cross-validation.

Main Results:

• CNN models achieved significant correlations (r: 0.60-0.79) with real dry matter yield traits.
• Estimates showed acceptable accuracy with root square mean errors (RSME) for LDMY (286.24-366.93 kg·ha⁻¹) and TDMY (413.07-506.56 kg·ha⁻¹).
• The pretrained ResNeXt50 architecture demonstrated the best performance for indirect selection of dry matter yield traits.

Conclusions:

• CNNs coupled with UAV remote sensing data offer a highly promising HTP solution for estimating dry matter yield in forage breeding.
• The developed HTP traits derived from CNNs were heritable, supporting their use in genetic selection.
• This approach can significantly enhance the efficiency and accuracy of guineagrass breeding programs.