Related Experiment Video
Updated: Jun 16, 2026
05:56
Objectification of Tongue Diagnosis in Traditional Medicine, Data Analysis, and Study Application
Published on: April 14, 2023
3.1K
Segmentation of roots in soil with U-Net
Abraham George Smith1,2, Jens Petersen2, Raghavendra Selvan2
11Department of Plant and Environmental Sciences, University of Copenhagen, Højbakkegaard Allé 13, 2630 Taastrup, Denmark.
Plant Methods
|February 15, 2020
Summary
An automated image segmentation system using U-Net Convolutional Neural Networks (CNNs) effectively quantifies plant roots in soil. This deep learning approach offers a faster, more accurate alternative to manual methods for root phenotyping in crop research.
Area of Science:
- Agricultural Science
- Computer Vision
- Plant Biology
Background:
- Plant root research is crucial for developing stress-tolerant crops with improved yields.
- Traditional root phenotyping methods are labor-intensive and challenging.
- Rhizotrons enable visual root inspection, but analysis remains manual.
Purpose of the Study:
- To evaluate an automated image segmentation method using U-Net Convolutional Neural Network (CNN) for root quantification.
- To compare the CNN-based method against the manual line-intersect technique.
- To assess the feasibility of deep learning for root analysis in small research settings.
Main Methods:
- Developed a dataset of 50 annotated chicory root images.
- Trained, validated, and tested a U-Net CNN model.
- Compared the automated system against a Frangi vesselness filter baseline.
- Evaluated performance using manual annotations and line-intersect counts on 867 images.
Main Results:
- The automated segmentation system achieved a Spearman rank correlation of 0.9748 and an R-squared of 0.9217 against line-intersect counts.
- The system demonstrated a 0.7 IoU (Intersection over Union) compared to manual annotations.
- Automated segmentations were of higher quality than manual annotations in many image areas.
Conclusions:
- A U-Net based CNN system is feasible for segmenting soil-based root images.
- The automated system can effectively replace the manual line-intersect method.
- Deep learning is a viable tool for custom datasets in small research groups.
