High-throughput soybean seeds phenotyping with convolutional neural networks and transfer learning

Area of Science:

• Agricultural Science
• Computer Vision
• Plant Breeding

Background:

• Accurate soybean seed phenotyping is crucial for crop improvement but is hindered by manual data collection's inefficiency and errors.
• Existing image-based methods for high-throughput phenotyping lack robustness, and deep learning approaches require extensive ground truth data.

Purpose of the Study:

• To develop an automated method for soybean seed segmentation and morphological parameter calculation.
• To overcome the limitations of manual annotation and data scarcity in deep learning for agricultural applications.

Main Methods:

• A novel synthetic image generation and augmentation technique based on domain randomization was employed.
• A Mask R-CNN instance segmentation network was trained using the synthesized dataset, incorporating transfer learning to reduce computational costs.
• The method's performance was validated on both synthetic and real-world soybean seed datasets.

Main Results:

• The developed method successfully synthesized a large, labeled dataset for training, significantly reducing manual annotation efforts.
• A convolutional neural network trained solely on synthetic data achieved good performance in soybean seed segmentation.
• The approach demonstrated robustness and generalization capabilities across datasets of varying resolutions and real-world conditions.

Conclusions:

• The proposed method enables effective segmentation of individual soybean seeds and efficient calculation of morphological parameters.
• This approach is practical for high-throughput object instance segmentation and large-scale soybean seed phenotyping, facilitating advancements in crop breeding.