ITErRoot: High Throughput Segmentation of 2-Dimensional Root System Architecture

Abstract

Root system architecture (RSA) analysis is a form of high-throughput plant phenotyping which has recently benefited from the application of various deep learning techniques. A typical RSA pipeline includes a segmentation step, where the root system is extracted from 2D images. The segmented image is then passed to subsequent steps for processing, which result in some representation of the architectural properties of the root system. This representation is then used for trait computation, which can be used to identify various desirable properties of a plant’s RSA. Errors which arise at the segmentation stage can propagate themselves throughout the remainder of the pipeline and impact results of trait analysis.

This work aims to design an iterative neural network architecture, called ITErRoot, which is particularly well suited to the segmentation of root structure from 2D images in the presence of non-root objects. A novel 2D root image dataset is created along with a ground truth annotation tool designed to facilitate consistent manual annotation of RSA. The proposed architecture is able to take advantage of the root structure to obtain a high quality segmentation and is generalizable to root systems with thin roots, showing improved quality over recent approaches to RSA segmentation. We provide rigorous analysis designed to identify the strengths and weaknesses of the proposed model as well as to validate the effectiveness of the approach for producing high-quality segmentations.