Graduate Studies
First Advisor
Hongfeng Yu
Committee Members
Ashok Samal, Yufeng Ge
Date of this Version
7-2024
Document Type
Thesis
Citation
A thesis presented to the faculty of the Graduate College at the University of Nebraska in partial fulfillment of requirements for the degree of Master of Science
Major: Computer Science
Under the supervision of Professor Hongfeng Yu
Lincoln, Nebraska, August 2024
Abstract
Enhancing maize breeding programs is crucial for increasing yield, stress resistance, and nutrient use efficiency, especially in Nebraska, a leading corn-producing state in the United States. Tassel traits are vital for understanding reproductive development and optimizing pollination success, which directly impacts yield. Existing publications offer public datasets and open-source models for developing customized algorithms to realize tassel detection and localization in the target environment. This study presents the development of a customized TasselLFANet model, a light-weight deep learning model designed for real-time quantification of maize tassel location and size, for the UNL Field Plant Phenotyping Facility (NU-Spidercam). Utilizing frequent imaging and transfer learning techniques, the model successfully outputs tassel count and total area of bounding boxes with limited labeled images. The model was deployed to predict tassel traits for a field experiment with 23 genotypes and 2 water treatments. The model showed significant performance improvements when the number of labeled images increased from 100 to over 600. Correlation analysis indicated medium or weak correlations between tassel traits and other phenotypic characteristics. The correlation between tassel traits and yield indicates its potential contribution to yield prediction as unique features in machine learning algorithms. However, the model struggled to detect small and occluded tassels in modern genotypes, indicating areas for future improvement. Overall, this study demonstrates that effective phenotyping tools (tassel traits detection in this case) can be developed with limited labeling effort through public datasets, open-source models, and transfer learning techniques. Most importantly, our model successfully tracked tassel count and size with high temporal resolution, comparable to canopy height and NDVI in detecting the VT stage. Additionally, the model shows promise in informing the earliest timing for relatively accurate yield prediction using common phenotypic traits.
Advisor: Hongfeng Yu
Comments
Copyright 2024, Geng (Frank) Bai. Used by permission