Abstract
Root system architectures are complex, multidimensional, and challenging to characterize effectively for agronomic and ecological discovery. We propose a new method, Spatial and Texture Analysis of Root System architEcture with Earth mover's Distance (STARSEED), for comparing root architectures that incorporate spatial information through a novel application of the Earth Mover's Distance (EMD). We illustrate that the approach captures the response of sesame root systems for different genotypes and soil moisture levels. STARSEED provides quantitative and visual insights into changes that occur in root architectures across experimental treatments. STARSEED can be easily 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.
Introduction
Quantifying Root System Architecture (RSA)—the spatial arrangement of roots in soil—is essential for developing crops resilient to environmental stressors like drought or flooding. Traditional methods for RSA analysis typically fall into two categories: reductionist metrics and deep learning models. Reductionist tools measure specific traits like total root length or branching angles but often fail to capture the holistic spatial distribution of the root system. Conversely, deep learning models can segment and classify roots with high accuracy but require extensive labeled datasets and function as "black boxes," offering little interpretability for biological discovery.
To bridge this gap, we introduce STARSEED (Spatial and Texture Analysis of Root System architEcture with Earth mover’s Distance). This framework provides a spatially explicit, holistic comparison of root systems by treating root images as distributions of texture features and utilizing the Earth Mover’s Distance (EMD) for quantitative analysis.
Methodology
The core of the STARSEED framework is the application of Earth Mover’s Distance (EMD), or the Wasserstein-1 distance, to root imagery. EMD calculates the minimum "work" required to transform one distribution into another, making it ideal for comparing spatial structures.
The STARSEED pipeline involves several key steps:
- Preprocessing: Manual root tracings are converted into binary masks, which are then downsampled and smoothed to mitigate noise.
- Signature Generation: The processed image is partitioned into a grid of spatial bins. Each bin is assigned a "signature" comprising its (x, y) coordinates and a weight derived from a local texture feature, such as the percentage of root pixels.
- EMD Calculation: The algorithm computes the distance between these signatures, identifying not just the magnitude of difference between root systems, but also the specific spatial locations of these differences.
- Visualization: The resulting pairwise distance matrices are projected into 2D space using Multi-Dimensional Scaling (MDS), allowing for the visualization of clusters and experimental gradients.
Experimental Datasets
The effectiveness of STARSEED was evaluated using a dataset of 220 images of sesame (Sesamum indicum) roots grown in specialized rhizoboxes. The experimental design included four cultivars (Sesaco32, 35, 38, and 40), four soil moisture levels (60%, 80%, 100%, and 120% of water-holding capacity), and two fertilizer conditions.
Results and Discussion
Texture and Spatial Sensitivity
Analysis revealed that simple texture features, specifically the percentage of root pixels within a local bin, were highly effective at discriminating between experimental treatments. The method proved robust across various grid resolutions (100 to 2000 bins), though the optimal scale depended on whether the analysis focused on genotype or environmental factors.
Biological Insights: Avoidance vs. Tolerance
STARSEED successfully captured distinct biological responses to waterlogging. In unfertilized conditions, sesame plants exhibited a "tolerance" strategy, maintaining root growth into saturated soil. However, with the addition of fertilizer, plants shifted to an "avoidance" strategy, characterized by lateral root proliferation above the flooded zone.
Genotype vs. Environment
The framework highlighted the shifting dominance of factors influencing RSA. In unfertilized runs, the cultivar (genotype) was the primary determinant of root architecture. In contrast, the addition of fertilizer made the soil moisture level (environment) the dominant driver of architectural changes.
Explainability and Discovery
Unlike opaque neural networks, STARSEED provides "flow matrices" that visually demonstrate how one root system would need to be spatially transformed to match another. This provides researchers with immediate, intuitive insights into how treatments affect root development. MDS plots further confirmed that root architectures arrange themselves along meaningful gradients corresponding to treatment intensity.
Conclusion
STARSEED offers a powerful, explainable alternative to existing RSA analysis tools. While currently demonstrated on 2D rhizobox images, the underlying EMD framework is extensible to 3D data such as CT scans. Future work will focus on automating the initial root tracing step and applying the method to time-series data to characterize the dynamic development of root systems.
Code and data are available at: https://github.com/GatorSense/STARSEED
Summary updated