Performing Department
(N/A)
Non Technical Summary
Understanding plant roots is critical to increasing plant and crop efficiency and resiliance. However, studying roots is extremely challenging. Our current tools and mechanisms are very limited in what root information they can collect and the scale in which this root information can be obtained. Recent advances in machine-learning combined with improved imaging technology has opened doors to collect and explore root characteristics of field-grown plants more easily and efficiently.This includes root imagery from various imaging sources including X-ray Computed Tomography (X-ray CT) scans of part of or whole root structures. However, a critical bottleneck for using this big data, specifically X-ray CT imagery of roots, is the inability of machine-learning algorithms to recognize and differentiate root features from those of organic matter and other noise embedded in the imagery. In this project, we propose to develop and apply an approach that couples root architecture modeling with Graph Convolutional Neural Networks (GCNNs) for improved detection and characterization of switchgrass roots from X-ray CT scans of soil cores. We will develop, implement and share new machine learning methods to automate the understanding of X-ray CT scans of roots. The outcomes of this research will include a novel root phenotyping framework with highly streamlined workflow that will transform our ability to detect and extract features of root traits from scanned soil core images from the field.
Animal Health Component
(N/A)
Research Effort Categories
Basic
50%
Applied
(N/A)
Developmental
50%
Goals / Objectives
To increase carbon deposition in the soil and enhance crop resource acquisition efficiency, characterizing root form and function is critical. This is because root traits are linked to the efficiency of water and nutrient uptake and the potential for root-to-soil carbon transfer and sequestration of carbon in the soil. Recent advances in machine-learning combined with improved imaging technology has opened doors to collect and explore root characteristics of field-grown plants more easily and efficiently.This includes root imagery from various imaging sources including X-ray Computed Tomography (X-ray CT) scans of part of or whole root structures. However, a critical bottleneck for using this big data, specifically X-ray CT imagery of roots, is the inability of machine-learning algorithms to recognize and differentiate root features from those of organic matter and other noise embedded in the imagery. In this project, we propose to develop and apply an approach that couples root architecture modeling with Graph Convolutional Neural Networks (GCNNs) for improved detection and feature extraction of switchgrass roots from X-ray CT scans of soil cores. We will leverage a large number of CT scan images we accumulated from our previous research projects. We aim to 1) improve the adaptability of GCNNs for increased root detection in X-ray CT data and incorporate the use of model-simulated data during the GCNN training process, 2) develop model-informed GCNN algorithms to integrate root architecture modeling with the GCNN framework, and 3) identify genetic markers underpinning switchgrass root architecture and morphology. We will compare genetic analyses from data obtained manually from washed root cores with those obtained from the adaptive GCNN and the model-driven GCNN. The outcomes of this research will include a novel root phenotyping framework with highly streamlined workflow that will transform our ability to detect and extract features of root traits from scanned soil core images from the field.
Project Methods
The methods in this effort are spread across three objectives (each with multiple tasks) as outlined below:Objective 1: Develop new GNN approaches to improve root trait extraction. Through this objective, we will be developing and implementing weakly supervised adaptive graph convolutional networks (Task 1.1) as well as investigating multiple training approaches (e.g., using model-simulated soil cores) for the GCNNs (Task 1.2). Theprimary methods in this objective are artificial intelligence algorithm development tasks.Objective 2: Develop hybrid GCNN-modeling approaches to leverage root architectural modeling within GCNN network architectures and training approaches to improve performance.Through this objective, we will be advancing a root architectural model (specifically, CropRootBox) (Task 2.1) and integrating this modeling within novel GCNN approaches (Task 2.2). Theprimary methods in this objective are artificial intelligence algorithm development and root modeling tasks.Objective 3: Identify the genetic basis of switchgrass root architectural and morphological characteristics from both washed root cores versus those extracted using the model-informed GCNN approach.Through this objective we will be simulatenously evaluating Objectives 1 and 2 (through comparison on traditional approaches to those developed in the first two objectives) as well as carrying out genetic analyses of the root characterstics extracted from the soil core data. The primary methods in this objective are algorithm and modeling evaluation as well as genetic mapping tasks.