Root2Graph: A Graph-based Semantic Segmentation Architecture For Plant Roots
Date
2023-10-02
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
ORCID
0009-0009-4160-5202
Type
Thesis
Degree Level
Masters
Abstract
Plant growth is significantly dependent on roots because roots play a crucial role in water and nutrient uptake. Traditional convolutional neural network (CNN) models extract binary segmentation masks, distinguishing root pixels from background pixels. However, it is important to advance further and classify different levels of roots, such as primary and lateral roots. Extracting primary and lateral roots from plant root images enables the calculation of root traits like primary and lateral root length and their branching angles. Computation of such traits can aid in breeding more stress-tolerant plants and crops, resulting in better yields. Existing approaches for the extraction of primary and lateral roots from plant root images often employ pixel-based semantic segmentation, which may not consider the structural information of plant roots and could lead to disconnected root structures. A plant root system is essentially a tree/graph-like structure with branching points as nodes and the root segments as edges. Leveraging this graph-like structure of plant roots, this thesis proposes a graph-based semantic segmentation approach using graph neural networks (GNN), which is named as 'Root2Graph' architecture. The Root2Graph architecture represents a shift from pixel-wise to graph-based classification.
Despite the GNN model displaying slightly lower performance in terms of F1 score and AUC-ROC metric in comparison to baseline pixel-based CNN model, it effectively addresses the issue of disconnected root structures observed in the pixel-based baseline model. We conduct a comprehensive analysis to identify the strengths and weaknesses of Root2Graph architecture in comparison to the baseline pixel-based CNN model and validate it's effectiveness in distinguishing between primary and lateral roots. The findings provide valuable insights for future advancements in root system analysis.
Description
Keywords
GNN, Skeletons, Semantic Segmentation
Citation
Degree
Master of Science (M.Sc.)
Department
Computer Science
Program
Computer Science