Deep learning-based high-throughput phenotyping for tiller quantification in interspecific bentgrass hybrids using...

Area of Science:

• Plant Science
• Computational Biology
• Agricultural Technology

Background:

• Tiller production is vital for turfgrass density and performance.
• Manual tiller counting is a bottleneck for large-scale breeding programs.

Purpose of the Study:

• To develop and compare automated methods for tiller quantification in turfgrass.
• To address the limitations of manual tiller counting in breeding programs.

Main Methods:

• Evaluated 770 plants from an interspecific bentgrass hybrid population.
• Developed and compared three automated tiller quantification methods: edge-based segmentation, Faster R-CNN, and YOLOv8.
• Assessed accuracy, robustness under occlusion, and computational efficiency using an annotated image dataset.

Main Results:

• The YOLOv8 deep learning model achieved the highest accuracy (R² = 0.97).
• YOLOv8 processed images significantly faster than Faster R-CNN.
• Edge-based and Faster R-CNN methods showed reduced performance in dense canopies.

Conclusions:

• One-stage deep learning detection (YOLOv8) outperforms complex two-stage models for phenotyping fine, occluded structures.
• The developed workflow provides a reliable, high-throughput solution for tiller counts.
• This framework is transferable for integrating image-derived phenotypes into grass breeding pipelines.