Tiller estimation method using deep neural networks.

Area of Science:

• Agricultural Science
• Computer Vision
• Plant Biology

Background:

• Tiller number is a key determinant of grass plant yield.
• Manual tiller counting is labor-intensive and unsuitable for high-throughput phenotyping.
• Conventional methods rely on heuristic features, limiting estimation accuracy.

Purpose of the Study:

• To develop an automated method for estimating plant tiller number using deep neural networks (DNNs).
• To address the challenge of limited training data for DNNs in plant phenotyping.
• To improve tiller number estimation accuracy compared to conventional methods.

Main Methods:

• Utilized deep neural networks (DNNs) for feature extraction from plant images.
• Employed strategies to overcome insufficient training data: using pretrained DNN models and pretext tasks for feature representation learning.
• Applied regression techniques to estimate tiller numbers from extracted DNN features.
• Conducted experiments using side-view whole plant images with a plain background.

Main Results:

• The proposed DNN-based methods significantly improved tiller number estimation accuracy.
• The strategies of using pretrained models and pretext tasks were effective in handling limited training data.
• Achieved better performance compared to conventional heuristic-feature-based methods.

Conclusions:

• Deep neural networks offer a promising approach for automated plant tiller counting.
• Pretrained models and pretext tasks are viable solutions for DNN application in phenotyping with limited data.
• The developed method enhances the efficiency and accuracy of high-throughput plant phenotyping.