Segmentation of Roots in Soil with U-Net

Abraham George Smith, Jens Petersen, Raghavendra Selvan, Camilla Ruø Rasmussen

Abstract

Plant root research is essential for developing stress-tolerant crops, yet phenotyping roots in soil remains a significant challenge. Rhizotrons facilitate visual inspection through transparent surfaces, but manual analysis of these images is labor-intensive. This study investigates an automated image segmentation method using the U-Net Convolutional Neural Network (CNN) architecture. Trained on a dataset of 50 annotated Chicory (Cichorium intybus L.) root images, the system achieved a Spearman rank correlation of 0.9748 and an $r^2$ of 0.9217 against manual line-intersect counts. The automated system demonstrated high-quality segmentations, occasionally outperforming manual annotations by identifying roots missed by human experts.

Introduction: The Hidden Half

Understanding root architecture is critical for developing stress-tolerant, high-yield crops. However, roots are the "hidden half" of the plant. To observe them without destruction, agronomists utilize rhizotrons—transparent panels that allow for non-invasive photography of root growth in situ.

The rhizotron facility used for non-invasive root imaging.

The primary bottleneck in root research is not image acquisition, but analysis. Researchers currently rely on manual "line-intersect" counting to estimate root density, a grueling process that can take over an hour for a small area of imagery. This paper presents a fully automated, deep-learning-based solution designed to generalize across varied conditions and eliminate manual labor.

Methodology

The researchers employed the U-Net architecture, a standard in biomedical image segmentation, adapted for agricultural imagery.

Technical Implementation

  • Architecture & Normalization: The U-Net was modified for 3-channel RGB input. Standard Batch Normalization was replaced with Group Normalization to maintain performance with the small batch sizes necessitated by high-resolution image tiles.
  • Loss Function: A hybrid of Dice Loss and Cross-Entropy was used to address extreme class imbalance, as roots typically occupy only ~0.5% of the total pixels.
  • Data Augmentation: To prevent overfitting on the limited dataset, aggressive augmentation—including color jittering and elastic grid deformations—was applied to simulate organic root variations.

Dataset and Training

The study utilized 50 high-resolution images of Chicory roots grown in a 4-meter deep rhizotron facility. These images were meticulously annotated at a per-pixel level, involving over 7 million labeled pixels. An additional 867 images with existing manual line-intersect counts were used for large-scale validation of the system's real-world utility.

Results and Discussion

The U-Net system significantly outperformed traditional computer vision baselines, such as the Frangi vesselness filter.

Side-by-side comparison of an original root image and the corresponding U-Net segmentation mask.

Quantitative Performance

  • Correlation: The system achieved a Spearman rank correlation of 0.9748 compared to manual counts, proving it a viable replacement for human labor.
  • Accuracy: The model reached an $F_1$ score of 0.701, compared to 0.462 for the Frangi baseline.

The "Super-Human" Effect

A notable finding was the model's ability to identify faint roots overlooked by human annotators. Upon re-evaluation, experts confirmed the model's detections were correct, suggesting that automated systems can provide a more consistent and objective "ground truth" than fatigued human labelers.

Conclusion

This research demonstrates that a robust agricultural segmenter can be trained with a relatively small but high-quality dataset. By automating root quantification, researchers can significantly accelerate the pace of plant phenotyping. Future work will focus on improving generalization across different crop species and soil types through transfer learning.

Summary updated

Abstract

Plant root research can provide a way to attain stress-tolerant crops that produce greater yield in a diverse array of conditions. Phenotyping roots in soil is often challenging due to the roots being difficult to access and the use of time consuming manual methods. Rhizotrons allow visual inspection of root growth through transparent surfaces. Agronomists currently manually label photographs of roots obtained from rhizotrons using a line-intersect method to obtain root length density and rooting depth measurements which are essential for their experiments. We investigate the effectiveness of an automated image segmentation method based on the U-Net Convolutional Neural Network (CNN) architecture to enable such measurements. We design a data-set of 50 annotated Chicory (Cichorium intybus L.) root images which we use to train, validate and test the system and compare against a baseline built using the Frangi vesselness filter. We obtain metrics using manual annotations and line-intersect counts. Our results on the held out data show our proposed automated segmentation system to be a viable solution for detecting and quantifying roots. We evaluate our system using 867 images for which we have obtained line-intersect counts, attaining a Spearman rank correlation of 0.9748 and an $r^2$ of 0.9217. We also achieve an $F_1$ of 0.7 when comparing the automated segmentation to the manual annotations, with our automated segmentation system producing segmentations with higher quality than the manual annotations for large portions of the image.
BibTeX
@inproceedings{smith2020segmentation,
  author = {Abraham George Smith and Jens Petersen and Raghavendra Selvan and Camilla Ruø Rasmussen},
  title = {{Segmentation of Roots in Soil with U-Net}},
  booktitle = {Plant Methods},
  year = {2020},
  url = {https://openreview.net/forum?id=uDr2XHw8vd},
}
Figure