Wheat spikelet detection on RGB images using deep machine learning.

Area of Science:

• Agricultural Science
• Computer Vision
• Machine Learning

Background:

• Manual phenotyping of wheat spikelet number is time-consuming and not scalable for large breeding datasets.
• Accurate spikelet counting is crucial for estimating wheat plant productivity.
• Existing computer vision methods require extensive annotation efforts (segmentation masks or bounding boxes).

Purpose of the Study:

• To develop an automated, high-throughput phenotyping method for wheat spike characteristics.
• To evaluate the effectiveness of simplified point annotations for training machine learning models.
• To compare different computational strategies for spikelet detection using minimal annotations.

Main Methods:

• Proposed a spikelet detection strategy using simplified point annotations (spikelet centers).
• Explored three computational methods: U-Net for segmentation, density regression, and YOLOv8 for object detection.
• Evaluated models using Mean Absolute Error (MAE), Mean Absolute Percentage Error (MAPE), Precision, Recall, and F1 score.

Main Results:

• U-Net-based segmentation achieved high accuracy and robustness in spikelet localization and counting.
• Density regression and YOLOv8 showed limitations, with YOLOv8 underperforming due to geometric mismatches.
• Minimalistic point-level annotation combined with segmentation models proved effective for automated phenotyping.

Conclusions:

• Simplified point annotations significantly reduce annotation time and cost for machine learning datasets.
• U-Net architecture is highly effective for accurate spikelet counting in automated wheat phenotyping.
• This approach accelerates breeding programs and enhances large-scale phenotypic data collection efficiency.