Novel Statistical Methods on High-Throughput Plant Phenotyping: Learning, Detection and Efficiency

This dissertation focused on developing novel statistical methods to address two challenges raised in high-throughput plant phenotyping (HTPP).HTPP has become an emerging technique to study plant traits due to its fast, labor-saving, accurate, and non-destructive nature. It has wide applications in plant breeding and crop management. However, the resulting massive image data has raised a challenge associated with efficient plant traits prediction and anomaly detection. We proposed a two-step image-based online detection framework of monitoring and quick change detection of the individual plant leaf area via real-time imaging data. The efficiency of the proposed framework is validated by a real data analysis.To extract detailed plant traits from HTPP images, semantic segmentation is always needed. However, the annotation cost for semantic segmentation is particularly expensive as it requires pixel-wise labeling. The second part of this dissertation focused on how to wisely select the initial training images to label within a limited cost budget. We proposed a novel stratified sampling strategy that can construct a representative initial dataset for active learning (AL), which achieves higher prediction performance. In addition, we further proposed a combined AL algorithm to investigate whether the diversity factor would improve the performance of the classical uncertainty-based AL algorithm. All the proposed methods are validated by an image set acquired from unmanned aircraft systems