Wheat spike and spikelet detection and counting from high-resolution digital imagery using YOLO with Oriented...

Area of Science:

• Agricultural Science
• Computer Vision
• Plant Breeding

Background:

• Accurate wheat yield potential estimation is vital for crop management and breeding.
• Manual spike and spikelet counting is labor-intensive and prone to errors.
• Deep learning with digital imagery has improved wheat spike detection, but spikelet-level analysis is underdeveloped.

Purpose of the Study:

• To evaluate YOLOv11 and YOLOv12 for wheat spike and spikelet detection and counting using oriented bounding boxes (OBB).
• To introduce a new large-scale benchmark dataset for wheat phenotyping with OBB annotations.
• To establish AI-driven approaches for robust and scalable wheat phenotyping.

Main Methods:

• Utilized high-resolution digital RGB imagery for wheat phenotyping.
• Applied two deep learning object detection models, YOLOv11 and YOLOv12, with oriented bounding boxes (OBB).
• Developed and annotated a large-scale dataset with 48,521 spike and 60,404 spikelet instances.

Main Results:

• Pre-trained YOLOv11 achieved high accuracy (mAP@0.5 = 95.8%) for spike detection with faster training/inference than YOLOv12.
• Non-pretrained YOLOv11 showed superior accuracy (mAP@0.5 = 99.0%) for spikelet detection.
• Counting performance was comparable between models, validating OBB-based YOLO detection.

Conclusions:

• Oriented bounding box-based YOLO detection offers a robust and scalable solution for AI-driven wheat phenotyping.
• This approach significantly enhances the precision of spikelet-level analysis, previously underexplored.
• The developed benchmark dataset supports further advancements in automated wheat crop assessment.