Open Collections

Ma, Zhenyu

University of British Columbia

Agriculture has been instrumental in shaping human civilization and the social-economic conditions of society. However, the food supply chain has been severely disrupted by the rapid growth of the world's population, increased urbanization, and the unforeseen impact of the pandemic. A recent report suggests that the global population could reach 9.3 billion by 2060. This impending crisis highlights the urgent need for innovative techniques and efforts to increase food production and ensure adequate food supply for the world's population. Crop monitoring is an important part of the solution for the global issue of food scarcity by providing farmers with real-time crop information to support agricultural decision-making. Besides, crop monitoring is also considered a tool to support phenomics research. High-throughput plant phenotyping is considered a vital aspect of crop monitoring, as it enables the acquisition of large-scale crop characteristics data. This thesis reviews state-of-the-art, high-throughput plant phenotyping development, including corresponding image sensors and platforms. It also analyzes the current state of crop monitoring and identifies upcoming challenges and potential future trends for researchers in this field. Based on the review result, a case study proposes an innovative automated biomass estimation system using depth image sensors. Various point cloud data analysis methods are validated using conventional volume method and machine learning approaches. Notably, the case study pioneer applied a deep learning-based 3D point cloud analysis model with RGB-D data for biomass estimation. The system employs a 3D reconstruction algorithm to convert color and depth information into a 3D point cloud. Then, a biomass estimation model named Bio-DGCNN is developed based on Dynamic Graph CNN (DGCNN). Additionally, transfer learning is incorporated into the proposed system to improve the accuracy of biomass estimation further and make the system adaptable to tasks with small datasets. The experimental results demonstrate that the proposed system improved biomass estimation accuracy, outperforming other competitive approaches with small datasets. As a result, the proposed system provides a viable solution for biomass predictions even with limited datasets.

Text

2023-09-26

Attribution-NonCommercial-NoDerivatives 4.0 International

http://hdl.handle.net/2429/86017

Electrical Engineering

University of British Columbia

UBCO

http://creativecommons.org/licenses/by-nc-nd/4.0/