Amodal Segmentation and Trait Extraction of On-Branch Soybean Pods with a Synthetic Dual-Mask Dataset

Area of Science:

• Agricultural Science
• Computer Vision
• Plant Biology

Background:

• Occlusions in on-branch soybean images hinder accurate pod-level phenotyping.
• Existing methods struggle with incomplete visual data, limiting trait extraction precision.

Purpose of the Study:

• To develop a lab on-branch pipeline for accurate soybean phenotyping despite occlusions.
• To enable non-destructive, high-precision extraction of key soybean pod traits.

Main Methods:

• A prior-guided synthetic data generator producing visible and amodal labels.
• An amodal instance segmentation framework with an improved Swin Transformer backbone and dual heads.
• Three-stage transfer learning (synthetic excised → synthetic on-branch → few-shot real) and a morphology-driven module for trait extraction.

Main Results:

• Achieved high Visible Average Precision (AP) of 91.6/77.6 and amodal AP of 90.1/74.7 on real on-branch data.
• Demonstrated consistent performance gains by incorporating synthetic data, indicating effective occlusion reasoning.
• Excised pod tests showed low error for seed per pod (MAE 0.07) and geometric traits (R² up to 0.94).

Conclusions:

• The co-designed data-model-task pipeline effectively recovers complete pod geometry under heavy occlusion.
• Enables non-destructive, high-precision, and low-annotation-cost extraction of key soybean traits.
• Provides a practical foundation for standardized laboratory phenotyping and downstream breeding applications.