Automatic Detection and Counting of Wheat Spikelet Using Semi-Automatic Labeling and Deep Learning.

Area of Science:

• Agricultural Science
• Plant Phenotyping
• Computational Biology

Background:

• Wheat spikelet number is a key trait for grain yield prediction.
• Manual spikelet counting is laborious and time-consuming.
• Efficient phenotyping methods are crucial for crop improvement.

Purpose of the Study:

• To develop an automated, cost-effective system for counting wheat spikelets.
• To utilize imaging processing and deep learning for accurate spikelet detection.
• To enhance the efficiency of wheat phenotyping.

Main Methods:

• Developed an unsupervised learning method for automatic spikelet detection and dataset creation.
• Employed transfer learning to retrain a deep convolutional neural network (CNN) for spikelet counting.
• Captured color images of wheat spikes during the grain-filling stage.

Main Results:

• The automated system demonstrated high accuracy in counting spikelets across four wheat lines.
• Quantitative metrics (RMSE, RRMSE, R²) indicate strong correlation between automated and manual counts.
• The developed methods effectively estimate spikelet numbers, improving counting efficiency.

Conclusions:

• The proposed imaging and deep learning approach provides an efficient solution for wheat spikelet counting.
• This automated phenotyping system aids in analyzing wheat spike developmental characteristics.
• The technology contributes to more accurate wheat yield assessment and breeding programs.