Tracking Growth and Decay of Plant Roots in Minirhizotron Images

Alexander Gillert, Bo Peters, Uwe Freiherr von Lukas, Jürgen Kreyling, Gesche Blume-WerryWACV 2023

TL;DR

This paper introduces a novel unsupervised deep learning method for tracking plant root growth and decay over time using minirhizotron images. By comparing sequential images, the method moves beyond simple root length estimation to quantify root turnover, providing a more nuanced understanding of root dynamics crucial for ecological modeling.

What Problem Does This Paper Address?

Understanding plant root systems is vital for many scientific fields, as roots constitute a significant portion of plant biomass and play a critical role in nutrient cycling and carbon sequestration. However, studying roots in situ is challenging. Traditional methods are destructive, offering only snapshots. Minirhizotrons, transparent tubes inserted into the soil, allow for non-destructive monitoring. While automated analysis of minirhizotron images exists, current methods primarily focus on estimating total root length per image. This metric fails to capture the dynamic processes of root growth and decay (turnover), which are essential for understanding plant health and ecosystem function, especially for fine roots with short lifespans. The paper addresses this gap by developing a method to quantify root turnover by comparing sequential minirhizotron images.

Methods

The core of the proposed method is a two-stage unsupervised deep learning approach.

  1. Root Segmentation Network (f): A pre-trained semantic segmentation network (based on U-Net) is used to identify root pixels in an image. This network is assumed to be available or can be trained with existing annotations.
  2. Feature Embedding Network (g): A second neural network, similar in architecture to the segmentation network but without the final classification layer, is trained to extract dense feature embeddings for image patches. This network learns to represent root appearance in a way that captures similarity.
  3. Training Stage 1: Contrastive Learning: Network g is trained using contrastive learning on augmented views of the same image. Positive pairs are created using pixel-wise image transformations (color, contrast, brightness jitter) to simulate varying environmental conditions. Negative pairs are generated using rotations and flips, encouraging the network to learn shape-based similarity rather than appearance.
  4. Training Stage 2: Self-Supervision: The Stage 1 model is used to find initial correspondences between root pixels in different images. These correspondences then serve as positive examples for training network g on pairs of images. This stage helps the network generalize to significant environmental changes between image acquisitions.
  5. Inference:
    • Root Matching: The segmentation maps from two images are used to identify keypoints (root centers). A cross-checking and ratio test are applied to ensure robust matching of feature descriptors extracted by network g.
  6. Alignment and Turnover Map: The segmentation map of the earlier image is warped to align with the later image using the matched keypoints. This warped map is then combined with the segmentation map of the later image to create a "turnover map" classifying pixels as "same" (present in both), "decay" (present only in the first), or "growth" (present only in the second).
  7. Length Estimation: A skeletonization method is applied to the turnover map to estimate the total length of newly grown or decayed roots.

Datasets

  • Main Dataset: 2550x2273px minirhizotron images acquired with a CI-600 In-Situ Root Imager. This dataset comprises images from mesocosm and field experiments, containing roots of Carex rostrata, Mentha aquatica, and Equisetum fluviatile. 854 unannotated images were used for training.
  • Evaluation Dataset: 62 image pairs from the main dataset were annotated with matched root keypoints and turnover maps. Annotations were focused on correcting mistakes from the proposed method.
  • Secondary Dataset: The Sunflower subset of the PRMI dataset was used for additional evaluation.

Key Insights and Ideas

  • Beyond Total Length: The paper convincingly argues that simply measuring total root length is insufficient for understanding root dynamics. Quantifying growth and decay separately provides a much richer picture of root turnover.
  • Unsupervised Learning is Key: The ability to train the core matching network in an unsupervised manner is crucial. Obtaining ground truth for root correspondences across time is extremely labor-intensive, making unsupervised methods essential for practical adoption.
  • Two-Stage Training for Robustness: The two-stage training process (contrastive on same image, then self-supervised on different images) is critical. The first stage learns robust shape-based features, while the second stage adapts these features to handle significant environmental changes between image acquisitions.
  • Challenges of Real-World Data: The paper highlights practical challenges like soil movement (swelling/shrinking), root appearance changes (color), and roots becoming obscured, which make simple geometric transformations (like homography) inadequate.
  • Annotation Bias: The authors demonstrate a significant bias in evaluation metrics (specifically IoU) when annotations are created by correcting the output of a specific model. This highlights the importance of careful annotation strategies and choosing metrics less susceptible to such biases.

Key Results

  • The proposed method (Ours, Stage 2) significantly outperforms baseline methods like SIFT and SIFT+VoxelMorph across various metrics, including relative error in root counts and lengths, and success rate.
  • For the main dataset, the method achieves the best relative error for "same" and "growth" classes (0.08/0.10 for counts, 0.09/0.12 for lengths) and the highest success rate (93.5%).
  • Performance is best for shorter time spans between image acquisitions (optimal around 2 weeks), but the method remains robust for longer periods compared to baselines.
  • The PRMI dataset results show similar trends, with the proposed method achieving the best mean IoU (0.86) and lowest relative errors.

Limitations and Future Work

  • Root Obstruction: Roots hidden by soil are counted as "decay," which is an inherent limitation of minirhizotron imagery.
  • Large Root Movements: The current descriptor-based matching struggles with very large root movements that violate the assumption of local shape invariance.
  • VoxelMorph Limitations: While VoxelMorph shows promise, it can hallucinate new roots or shrink decayed ones, leading to errors. The authors suggest a potential combination of their method with VoxelMorph.
  • Future Work: Improving matching performance for longer time periods and exploring combinations with methods like VoxelMorph are suggested avenues for future research. The authors believe their method is already in a usable state for ecological research.

Summary updated

BibTeX
@inproceedings{gillert2023tracking,
  author = {Alexander Gillert and Bo Peters and Uwe Freiherr von Lukas and Jürgen Kreyling and Gesche Blume-Werry},
  title = {{Tracking Growth and Decay of Plant Roots in Minirhizotron Images}},
  booktitle = {IEEE/CVF Winter Conference on Applications of Computer Vision},
  year = {2023},
}
Figure