Spatial and Texture Analysis of Root System distribution with Earth mover’s Distance (STARSEED)

Joshua Peeples, Weihuang Xu, Romain Gloaguen, Diane Rowland, Alina Zare, Zachary BrymPlant 2023 · DOI

Abstract

STARSEED can be generalized to other plants and provides insight into root system architecture development and response to varying growth conditions not captured by existing root architecture metrics and models. The code and data for our experiments are publicly available: https://github.com/GatorSense/STARSEED.

Introduction

Plant roots, often referred to as the "hidden half" of agriculture, are critical for crop survival under environmental stressors such as drought or flooding. Understanding Root System Architecture (RSA) is essential for breeding resilient crops. However, characterizing RSA is notoriously difficult. Traditional industry-standard software typically reduces a root system to individual parameters like total length, average diameter, or branching angles. While informative, these metrics fail to capture the spatial context—the "texture" of how a root system occupies soil space. Researchers have lacked a robust, automated method to compare entire root systems holistically and determine exactly where and how much they differ in response to varying soil conditions.

The STARSEED Framework

To address these limitations, we propose STARSEED (Spatial and Texture Analysis of Root SystEm distribution with Earth mover’s Distance), a computational framework that leverages computer vision and optimal transport theory to provide a quantitative, holistic comparison of plant root architectures.

The STARSEED computational pipeline, illustrating the flow from raw image acquisition through preprocessing, signature generation, and EMD calculation to final MDS projection.

The STARSEED pipeline consists of several key stages:

  1. Preprocessing: 2D images of roots grown in transparent "rhizoboxes" are segmented to isolate root pixels from the soil background.
  2. Signature Generation: The segmented image is divided into a grid of spatial bins. For each bin, a "texture feature" (e.g., percentage of root pixels, fractal dimension, or lacunarity) is calculated to create a "signature"—a set of coordinates and weights representing the root distribution.
  3. Earth Mover’s Distance (EMD): At the core of STARSEED is the EMD (also known as the Wasserstein metric). It calculates the minimum "work" required to transform one distribution into another, measuring the distance and direction root mass would need to "move" to transform one architecture into another.
  4. Visualization & Analysis: Pairwise distances are projected into 2D space using Multidimensional Scaling (MDS), allowing for the visualization of clusters of similar architectures. The Calinski-Harabasz (CH) index is then used to quantify the quality of these groupings based on experimental treatments.
Detailed view of the signature representation process, where preprocessed images are converted into a weighted set of spatial features.

Experimental Setup and Datasets

The framework was evaluated using a dataset of 220 images of sesame (Sesamum indicum) root systems. The study accounted for four different cultivars (S32, S35, S38, S40), four soil moisture levels (60%, 80%, 100%, and 120% of water-holding capacity), and two nutrient levels (with and without fertilizer). Plants were grown in custom rhizoboxes for 16–21 days and scanned at 400 DPI. Roots were hand-traced to ensure high-quality ground truth for algorithmic validation.

Key Insights and Biological Findings

  • Avoidance vs. Tolerance Strategies: A significant biological finding was the sesame plant's response to waterlogging (120% moisture). Without fertilizer, plants exhibited a "tolerance" strategy, growing roots deep into saturated soil. Conversely, with fertilizer, they switched to an "avoidance" strategy, restricting deep growth in favor of lateral proliferation near the surface.
  • Feature Efficacy: While complex features like fractal dimension were tested, the simple percentage of root pixels per bin proved to be the most effective feature for distinguishing between different root architectures.
  • Genotype vs. Environment: The study revealed that environmental factors can override genetics. Without fertilizer, the cultivar (genotype) was the primary driver of root shape. However, with added nutrients, moisture level became the dominant factor determining architecture.

Results and Robustness

STARSEED successfully grouped root systems by treatment according to the CH index. In fertilized runs, the 120% moisture treatment yielded a high CH score (54.72), indicating a highly distinct and consistent architecture. Furthermore, EMD "flow matrices" allowed researchers to pinpoint specific areas where growth shifted between cultivars. The method also demonstrated robustness across various spatial bin sizes, suggesting that researchers do not need to perfectly tune grid resolution to achieve meaningful results.

Limitations and Future Work

Currently, the framework relies on hand-traced images, which presents a manual bottleneck. Integrating automated machine-learning segmentation (e.g., U-Net) is a priority for future development. Additionally, while 2D rhizobox analysis is a standard proxy, the EMD component of STARSEED is theoretically scalable to 3D data from CT or MRI scans. Finally, applying the framework to time-series data could help characterize how root "texture" evolves dynamically over a plant's lifecycle.

Summary updated

Abstract

STARSEED can be generalized to other plants and provides insight into root system architecture development and response to varying growth conditions not captured by existing root architecture metrics and models. The code and data for our experiments are publicly available: https://github.com/GatorSense/STARSEED .
BibTeX
@inproceedings{peeples2023spatial,
  author = {Joshua Peeples and Weihuang Xu and Romain Gloaguen and Diane Rowland and Alina Zare and Zachary Brym},
  title = {{Spatial and Texture Analysis of Root System distribution with Earth mover’s Distance (STARSEED)}},
  booktitle = {Plant},
  year = {2023},
  doi = {10.1186/s13007-022-00974-z},
}
Figure