Soybean Seedling Root Segmentation Using Improved U-Net Network.

Area of Science:

• Agricultural Science
• Computer Vision
• Biotechnology

Background:

• Accurate segmentation of soybean seedling roots is crucial for genetic breeding and understanding root morphology.
• Existing methods struggle with background interference (water stains, noise), leading to over-segmentation, unsmooth edges, and root disconnection.

Purpose of the Study:

• To develop an improved U-Net based semantic segmentation model for precise soybean seedling root image segmentation.
• To enhance root region identification and suppress background noise and interference.

Main Methods:

• Utilized an improved U-Net network incorporating a double attention mechanism during downsampling and an Attention Gate in skip connections.
• Employed connected component analysis for post-processing to remove residual background noise.
• Visual interpretation of model predictions using feature maps and class activation mapping maps.

Main Results:

• Achieved high segmentation performance with Accuracy (0.9962), Precision (0.9883), Recall (0.9794), F1-Score (0.9837), and Intersection over Union (0.9683).
• Demonstrated rapid image processing with an average time of 0.153 seconds per image.
• Showcased strong generalization ability with accurate segmentation in soil-culturing environments.

Conclusions:

• The proposed model accurately segments soybean seedling roots, overcoming limitations of previous methods.
• It provides a robust technical foundation for quantitative evaluation of soybean root morphology, supporting genetic improvement efforts.