PaperlayerView exact revision on arXiv

Source-linked AI summary

Segmentation of Roots in Soil with U-Net

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

arXiv:1902.11050v2cs.CV

TL;DR

Manual extraction of root information from rhizotron images is a bottleneck for root phenotyping. This paper uses U-Net-based segmentation trained on a small annotated dataset and finds strong agreement with manual root measurements, while annotation and scratched-panel errors remain limitations.

  • Problem

    Manual annotation and line-intersect counting are bottlenecks for extracting root length and depth information from rhizotron images.

  • Method

    The study trains, validates, and tests a U-Net CNN for root segmentation using 50 annotated images and an ordinary RGB camera, comparing it with a Frangi filter baseline.

  • Results

    The automated system achieves a Spearman rank correlation of 0.9748 and an r^2 of 0.9217 against line-intersect measurements, and consistently outperforms Frangi by F1 score on unseen data.

  • Takeaways & Limitations

    The results support U-Net as a feasible approach for segmenting and quantifying roots in soil and potentially replacing manual line-intersect measurements.

  • Takeaways & Limitations

    Scratches in acrylic panels can create false positives that overestimate rooting depth, while annotation-quality errors make F1 underrepresent true performance.

Abstract

from arXiv · show

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.

Background

Root phenotyping in soil is hindered by inaccessible roots, laborious manual measurements, and low image contrast. The study proposes U-Net segmentation from ordinary RGB photographs as an alternative to feature-engineered and traditional filtering approaches.

  • Root phenotyping challenges: Manual line-intersect analysis is laborious, can take 20 minutes per metre of grid line, and omits several architectural root traits.It counts grid crossings rather than directly measuring root length or traits such as branching, diameter, tip count, growth speed, and growth angle.
  • Root phenotyping challenges: Low contrast between roots and soil has limited prior automated detection, motivating methods that work on ordinary RGB photographs without specialized hardware.Earlier approaches often detected only young unpigmented roots or roots in black peat, while spectroscopy required customized equipment.
  • Prior approaches: Feature-engineered root segmentation depends on specialist knowledge and may restrict performance or reuse across datasets.Deep learning instead allows representations to be learned from annotated data, reducing the need for manually programmed feature extraction.
  • Study aim: The study develops a U-Net CNN for segmenting roots in soil, using an architecture suited to settings where large annotated datasets are difficult to obtain.A Frangi vessel enhancement filter provides a simpler traditional baseline with lower typical hardware and training requirements.
  • Study aim: The proposed approach tests whether RGB-based U-Net segmentations can support root quantification and outperform a traditional Frangi-based method.The hypotheses concern correlation with manual line-intersect measurements and agreement with test annotations.

Data collection

The study uses Chicory images collected through a large rhizotron facility, retaining images suitable for analysis and sampling across panel depths. Existing line-intersect measurements provide a comparison basis.

  • Imaging setup: Images came from Chicory roots photographed in 2016 at a 4 m deep rhizotron facility using the existing photography setup.The facility contained 12 rhizotrons with transparent acrylic panels, enabling inspection at the soil-rhizotron interface.
  • Imaging setup: The facility produced 3300 images across nine dates and depths from 0.3 to 4 m, with four photographs covering each panel width.Images were indexed by rhizotron, direction, and panel depth.
  • Image selection: Images unsuitable for analysis were excluded when panels contained equipment, incomplete soil coverage, or other obstructions, and depth distributions guided representative annotation sampling.The sampling procedure accounted for the available depth distribution and selected images randomly after determining depth quotas.

Annotation

Fifty Chicory images were selected, cropped to match the region used for existing root-intensity measurements, and densely annotated pixel by pixel by an experienced agronomist.

  • Dataset construction: Fifty images were chosen for annotation based on annotator availability and the time required for annotation.The selected images formed the annotated dataset used for subsequent model development and evaluation.
  • Image preparation: Images were cropped from 4608 × 2592 to 3991 × 1842 pixels, representing approximately 300 × 170 mm of rhizotron surface.The crop removed overlap, framing, and metal structures to match the region used for existing root-intensity measurements.
  • Annotation procedure: A trained agronomist created per-pixel root-or-soil annotations in Photoshop, producing 7351422 labelled pixels per image.Annotation took approximately 30 minutes per image and labelled pixels perceived as root.

Data split

The annotated images were divided into training, validation, and held-out test sets, with the test set reserved for evaluating performance on unseen data.

  • Dataset partitioning: Training optimized network weights with stochastic gradient descent, whereas validation guided performance monitoring and hyperparameter tuning.The test set was kept for final assessment on unseen data.
  • Dataset partitioning: Ten images were reserved for testing, while the remaining 38 usable images were split into training and validation datasets.Two images were removed because one lacked roots and another contained a sticker on the acrylic glass.

Metrics

The study evaluates root segmentation using F1, accuracy, correlation with manual line-intersect measurements, and agreement with manual annotations. F1 gives equal weight to precision and recall, while skeletonization converts segmentations into root-length estimates.

  • Metrics: F1 evaluates segmentation while weighting precision and recall equally, matching the goal of balancing root overestimation and underestimation.Precision measures correctly predicted root pixels among predicted roots, whereas recall measures them among actual roots.
  • Metrics: Accuracy is reported as the ratio of correctly predicted pixels to total pixels, alongside test-set F1 evaluation on unseen data.Training and validation F1 were used during development, while the completed system was evaluated on the test set.
  • Metrics: Skeletonization thins each segmentation so remaining pixels can provide a root-length estimate comparable with line-intersect counts.The same conversion procedure was applied to both U-Net and Frangi outputs.
  • Metrics: Correlation and coefficient-of-determination analyses compare predicted root length with manual line-intersect intensity on held-out test images.The evaluation also compares U-Net output lengths with manual per-pixel annotation lengths.

Frangi Vesselness Implementation

The baseline uses Frangi vesselness filtering to identify root-like tubular structures, then thresholds and cleans the result to produce a segmentation.

  • Frangi Vesselness Implementation: The Frangi baseline thresholds vesselness responses and removes small connected components before producing root segmentations.CMA-ES optimizes the filter and threshold parameters using 1 − mean(F1) as the objective.
  • Frangi Vesselness Implementation: Frangi filtering offers a traditional, off-the-shelf alternative with typically lower hardware requirements and training time than U-Net.The method was selected because roots resemble the blood vessels for which the filter was designed.
  • Frangi Vesselness Implementation: The filter estimates pixel-level tubularity, using vessel-like structure as the basis for distinguishing roots from soil.

U-Net Implementation

The U-Net implementation segments root-containing image tiles using contextual receptive fields, augmentation, class-imbalance-aware loss, and monitored optimization. Training performance continued improving while validation performance plateaued around epoch 40.

  • U-Net Implementation: The network uses an RGB-adapted original U-Net architecture implemented in PyTorch with group normalization and He initialization.The input layer was modified from a single channel to receive RGB images.
  • U-Net Implementation: U-Net receives 572 × 572 tiles and predicts the central 388 × 388 region, using surrounding context to classify output pixels.Mirroring pads image edges before tile extraction, providing synthetic context at boundaries.
  • U-Net Implementation: Each epoch samples root-containing tiles and caps the number retained per image at 40 to reduce the effect of empty regions and uneven root abundance.Ninety random locations are initially extracted from each training image before filtering and capping.
  • U-Net Implementation: Color jitter and elastic-grid deformation augment tiles to improve performance on similar data absent from training.Elastic deformation was applied with probability 0.9, alongside specified brightness, contrast, saturation, and hue perturbations.
  • U-Net Implementation: Dice loss combined with cross-entropy addresses severe class imbalance, with roots comprising only 0.54% of training pixels.The combined loss performed better than either loss alone in preliminary experiments.
  • U-Net Implementation: Training F1 kept improving while validation F1 plateaued around epoch 40, indicating fitting to training-specific noise and anomalies.

Results

On held-out test data, U-Net segmented roots more effectively than the Frangi baseline and closely tracked manual root-intensity measurements across 867 images.

  • Evaluation protocol: The final model was selected using validation performance, but train and validation results were not treated as reliable estimates of performance on new data.The held-out test set was used for the reported generalization assessment.
  • Segmentation quality: U-Net segmentations often matched or exceeded manual annotations, including roots missed by the annotator and cleaner root-soil boundaries.Some errors remained at ambiguous boundaries, and faint or small root sections were sometimes missed.
  • Held-out test results: U-Net achieved higher F1 than Frangi on the held-out test images, with every root-containing image above 0.6 versus some Frangi results below 0.4.The highest U-Net F1 was just below 0.8, and U-Net predicted more than twice as many root pixels as Frangi.
  • Metric interpretation: U-Net had higher recall than precision because it slightly overestimated total root pixels, while accuracy exceeded 99% for both systems.The high accuracy was driven by the predominance of soil pixels, which both systems usually classified correctly.
  • Held-out test results: 0.9748 Spearman correlation and r^2 of 0.9217 linked U-Net segmented root length with line-intersect measurements across 867 images.The relationship showed increasing variance at higher root intensities, with some deviations where U-Net detected roots missed by the annotator.

Discussion

The U-Net system appears feasible for segmenting roots in soil and quantifying root length, while several implementation and evaluation limitations remain. Its accessibility is supported by training on a small annotated dataset and ordinary RGB images, but hardware, annotation, and length-estimation constraints warrant caution.

  • The U-Net system consistently outperformed the Frangi approach on unseen test data according to F1, supporting its ability to discriminate roots from soil.The authors also report a high correlation between automated and manual measurements, supporting root quantification.
  • The method was trained, validated, and tested with 50 images annotated by one agronomist, demonstrating feasibility despite limited labelled data.The images were annotated in a few days, and the system used an existing ordinary RGB photography setup.
  • The combined Dice and cross-entropy loss was more effective in preliminary experiments, but the precise benefits of combining the losses remain unclear.The authors call for more research into why combined loss functions improve performance.
  • The automated method approximated manual line-intersect root-length measurements, but converting segmentation pixels into length was not the study’s main focus.Skeletonization followed by pixel counting can produce length estimates that vary with root orientation.
  • F1 may not represent true segmentation performance because annotation-quality errors and ambiguous boundaries can produce errors even for high-quality segmentations.The authors suggest improving annotations and note disagreement between human raters in related segmentation work.
  • The implemented CNN requires a high-end GPU, whereas the Frangi filter has lower hardware requirements, limiting immediate use on typical laptops.The authors identify model compression as a possible way to improve accessibility.
Loading 1902.11050v2…