An in situ image-based phenotyping system for hydroponic maize seedling roots based on DB-UNet and customized...

Area of Science:

• Agricultural Science
• Computer Vision
• Plant Biology

Background:

• Root phenotype is crucial for maize growth and development.
• Traditional phenotyping lacks in situ monitoring and accurate segmentation.
• Existing models struggle with root segmentation accuracy.

Purpose of the Study:

• To develop an integrated system for crop root phenotyping.
• To propose an accurate maize root segmentation model (DB-UNet).
• To enable precise multi-trait root phenotyping.

Main Methods:

• Developed a crop root phenotyping system for cultivation and data collection.
• Proposed DB-UNet, a CNN-ViT dual-branch model with attention fusion.
• Utilized a mixed loss function (Dice, Focal, KL loss) for segmentation.
• Developed a skeleton-based algorithm for multi-trait root phenotyping.

Main Results:

• DB-UNet achieved 91.02% mIoU, 82.78% FG IoU, and 97.72% Centerline-Dice on a custom dataset.
• DB-UNet outperformed classic UNet in key segmentation metrics.
• Root length measurement error reduced to 3.14%, an 8.42% improvement.
• Strong positive correlation found between plant height and total root length.

Conclusions:

• The developed system and DB-UNet model significantly enhance hydroponic maize root segmentation accuracy.
• The custom phenotyping algorithm enables reliable extraction of multiple root traits.
• Improved segmentation accuracy directly correlates with reduced phenotypic measurement errors.
• Accurate root phenotyping provides valuable insights into maize growth and development.