Off-campus UNL users: To download campus access dissertations, please use the following link to log into our proxy server with your NU ID and password. When you are done browsing please remember to return to this page and log out.
Non-UNL users: Please talk to your librarian about requesting this dissertation through interlibrary loan.
Analysis of High-dimensional Data for Plant Phenotyping
Traditional plant phenotyping usually relies on manual measurement of selected traits from a small number of plants. This kind of approach has very limited throughput and makes it infeasible to conduct a comprehensive analysis of traits on a large scale. Recent advances in techniques have alleviated the problem of phenotyping bottleneck. High-throughput phenotyping systems are already available in different environments, such as greenhouses and fields. Many types of 2D images can be obtained, including RGB, fluorescent, and hyperspectral images. 3D representation of plants is also available with the development of new tools such as LiDAR (Light Detection and Ranging) and high-resolution multi-view cameras. The spatial and temporal resolutions of data generated from these new techniques have been significantly increased. In my dissertation work, I aim to develop new techniques to help scientists tackle these large-scale high-dimensional data in an effective and efficient manner. First, I design an interface that allows users to study hyperspectral images. My interface can help users determine important bands, identify regions of interest, and generate image fusion results for time-varying hyperspectral plant images. Second, I use neural networks to design an improved interface to study hyperspectral images interactively. The neural network projects hyperspectral features into a scatterplot. The predicted scatterplot can not only differentiate different substances in the hyperspectral images but also show the relationship among substances. Third, I present an end-to-end pipeline to reconstruct surfaces from LiDAR-based point clouds of maize plants. I propose a two-step clustering approach to accurately segment the points of each individual plant component according to maize properties. I further employ surface fitting and edge fitting to ensure the smoothness of plant surfaces. Fourth, I use neural networks to design an improved end-to-end pipeline to reconstruct surfaces from 3D plant point clouds. The segmentation network provides good segmentation of plant point clouds. Then, a fitting network is used to reconstruct surfaces that change smoothly from the stem to the leaves. The reconstructed plant is more realistic than previous work.
Zhu, Feiyu, "Analysis of High-dimensional Data for Plant Phenotyping" (2020). ETD collection for University of Nebraska - Lincoln. AAI28028389.