Open Collections

Zaji, Amirhossein

University of British Columbia

Wheat plant phenotyping is crucial for plant breeding and crop management. Traditional methods, however, are laborious and error prone. Computer vision and Deep Learning (DL) can address these issues by quickly and accurately analyzing wheat plant images, but data collection, annotation, and preprocessing for DL models can be costly and time-consuming. This thesis aims to reduce data preparation time for high-throughput wheat plant phenotyping using DL models. It explores the use of readily available sensors like cellphone RGB cameras for image acquisition and develops more efficient and automated annotation and preprocessing methods. The proposed aug- mentation algorithm reduces the number of training samples needed for accurate DL models, while transfer learning and other techniques are investigated to de- crease the amount of labeled data required. A state-of-the-art framework for localizing and counting wheat spikes using dotted annotation datasets and generation algorithms is proposed and evaluated. This approach is faster than bounding box and polygon annotation methods for wheat spike localization and counting. The computational component employs hybrid UNet architectures, and the proposed models demonstrate improved spike localization and counting compared to previous research studies. This research introduces Automatic Object Level Augmentation (AutoOLA), a novel augmentation algorithm that reduces DL model training sample requirements. AutoOLA augments spikes, leaves, stems, and backgrounds in wheat images separately using their optimized augmentation policies. Results indicate that AutoOLA can improve wheat spike counting by 60%. Stereovision cameras, which are more user-friendly, are employed for plant height measurement. An active learning model is used to annotate images from one camera and transfer them to the other camera’s images to expedite the process. Mask R-CNN is utilized for spike localization, and spike heights are estimated using depth layers. In conclusion, this thesis develops methods to reduce time and effort for apply- ing DL models in plant phenotyping. It investigates image acquisition with accessible sensors and develops more efficient annotation and preprocessing methods, including the proposed augmentation algorithm. By making DL models more accessible and scalable, this research advances plant breeding, genetics, and crop production.

Text

2023-05-05

Attribution-NonCommercial-NoDerivatives 4.0 International

http://hdl.handle.net/2429/84580

Civil Engineering

University of British Columbia

UBCO

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