A Weakly Supervised Deep Learning Framework for Sorghum Head Detection and Counting.

Area of Science:

• Agricultural Science
• Computer Science
• Plant Breeding

Background:

• Sorghum yield is linked to head distribution, making head counting crucial for breeding.
• Manual head counting is labor-intensive, inefficient, and prone to errors, especially in large-scale studies.
• Deep learning offers potential for automated counting but requires extensive labeled data.

Purpose of the Study:

• To develop a weakly supervised deep learning framework for sorghum head detection and counting using UAV imagery.
• To reduce the significant human labeling effort typically required for training deep learning models.
• To improve the trustworthiness of deep learning models by visualizing learned features.

Main Methods:

• Proposed an active learning-inspired, weakly supervised deep learning framework.
• Utilized a semi-trained Convolutional Neural Network (CNN) for synthetic annotation of limited labeled data.
• Employed Unmanned Aerial Vehicle (UAV)-based imagery for sorghum head detection and counting.

Main Results:

• Achieved high correlation (R² = 0.88) between human and machine counts, demonstrating comparable performance to fully supervised methods.
• Significantly reduced the need for extensive manual data labeling.
• Visualized key features learned by the network, enhancing model interpretability.

Conclusions:

• The proposed framework effectively reduces human labeling effort for sorghum head counting.
• Weakly supervised deep learning with synthetic annotation is a viable approach for agricultural phenotyping.
• Model interpretability is enhanced through feature visualization, increasing user trust.