RootShape: Automated Analysis of In Situ Fine Root Images
Authors
Kenneth Ball* (kenneth.ball@geomdata.com), Nirav Patel, Andrew Bartels, Anastasia Deckard, Jay Hineman
Institutions
Geometric Data Analytics, Inc., Durham, NC
URLs
Abstract
The ability to detect and measure plant features is fundamental to a host of problems in horticultural science and plant phenotyping. Observations of roots, how they grow, and how they interact with their soil environment have important scientific and commercial implications.
Though largely obscured, root features are essential for discovering how plants respond to short- and long-term environmental changes. Measuring the growth and distribution of fine roots is important for understanding how plants derive nutrients and water from soil and relationships between plants and mycorrhizal fungi.
Images of plant root systems can be collected nondestructively without damaging or killing the plant using rhizotrons—transparent interfaces that enable imaging of soil and roots. In situ imaging is an area of active research. Rhizotrons and minirhizotrons—transparent tubes and associated camera and scanning systems—are examples of imaging modalities that enable in situ collection.
The standard approach to segmenting roots in rhizotron images is manual tracing. Manual tracing and annotation are time consuming, tedious, and are the primary bottleneck in rhizotron image analysis. Supervised learning approaches can reduce the need for manual tracing but do not eliminate it because they are unlikely to generalize across diverse experimental designs and environments.
The approach to automating segmentation and annotation utilizes differential geometry applied to digital rhizotron and minirhizotron images to recognize and filter root-like regions from background. Then a low-dimensional vectorization of these segmented regions is derived, which can be used to train a very simple classifier with limited user intervention.