Advance Image Processing For Hydroponic Plant Root Phenotyping
Date
2025-02-14
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
ORCID
0009-0004-2127-8615
Type
Thesis
Degree Level
Doctoral
Abstract
Plant root system architecture (RSA) plays a critical role in biology, plant science, and agriculture due to its responsibility to support plant growth, interact with the environment, be the primary organ absorbing water and essential nutrients from the soil, and releasing low molecular weight compounds into the soil to promote root health and function. Optical 3D root phenotyping systems employed on hydroponically-grown plants offer advantages such as high throughput, large resolution, and exceptional visual clarity. However, they also face challenges such as intensive user labour, problematic segmentation, and gaps caused by the occlusion of the root supporting system. This project proposes a series of upgrades to the baseline system, RootReader3D, including hardware upgrades, an auto-cropping module, a segmentation module, an inpainting module, and an integration testing network to evaluate module performance. The hardware upgrades include enhancements to the root-supporting mesh colour and the axis of rotation calibration mechanism. The auto-cropping and segmentation modules eliminate the need for user input by reviewing images from different angles to ensure the root's completeness. The new segmentation module uses random-walker segmentation, which performs better while removing the need for user-selected parameters. The inpainting module leverages an Adversarial Generative Network to restore missing root segments, improving the reconstructed 3D models and estimated root traits. A soybean dataset with multiple cultivars with significant variation in root architecture and replicates for each cultivar is processed using the proposed integration testing network. The test results indicate that the new auto-cropping and segmentation modules demonstrate comparable overall performance, with significantly higher IoU and DSC scores and slightly lower Recall scores than the baseline system. The inpainting module restores most of the missing root segments and improves the accuracy of overall root traits estimation. During the integration test, the Mean Squared Error (MSE) of estimated root volume traits are reduced from 0.234 to 0.081 cm^3. The modules proposed by this project benefits the software users by minimizing the operation time and avoiding parameter selection bias. Furthermore, these modules improve the quality of 2D root images and 3D root models, which also optimizes the accuracy of the estimated root traits.
Description
Keywords
image processing, deep learning, plant root phenotyping, generative adversarial network
Citation
Degree
Doctor of Philosophy (Ph.D.)
Department
Computer Science
Program
Computer Science