A computer vision pipeline for automated phenotyping of maize brace roots
Date
2025
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
University of Delaware
Abstract
This thesis demonstrates a novel application of machine learning in imaging for automated phenotyping in plant science, specifically focusing on maize brace roots to enhance crop yield and resilience. It addresses the limitations of manual and semi-automated phenotyping methods for brace roots, which are labor-intensive. It highlights the challenges of accurately segmenting complex field images due to inconsistent lighting and obstructions. The work proposes a new machine learning-based approach to efficiently create datasets and develop robust models for comprehensive phenotypic analysis in diverse field conditions. ☐ A survey of related works provides foundational context by reviewing modern image segmentation techniques, including training processes, various loss functions, and optimization algorithms. The discussion differentiates between Convolutional Neural Networks (CNNs) and Vision Transformers (ViTs), emphasizing their respective strengths in image processing and the benefits of encoder-decoder architectures like U-net for segmentation. Crucially, the role of large foundational models such as SAM2 for accelerating annotation tasks in plant science is detailed. The limitations of these models on out-of-distribution data are also acknowledged, underscoring the necessity of data augmentation techniques to enhance model robustness and generalization from limited and diverse datasets. ☐ The proposed approach outlines a comprehensive data collection strategy utilizing robotic platforms (Brobot 2021, 2023) and a handheld device (POGO stick) across multiple field seasons (2021-2024), yielding a large and varied image dataset. A novel dataset creation pipeline leverages state-of-the-art promptable segmentation models (SAM2) to efficiently annotate critical features like roots, stalks, and scale markers, significantly reducing human annotation time. To bolster model robustness against field variability, a rigorous preprocessing and data augmentation regimen, including custom padding and geometric transformations, is applied. The core segmentation architecture is a Unet with a pre-trained ResNet-34 encoder, optimized with weighted cross-entropy loss to manage class imbalance, and a classifier head is integrated to filter usable samples by leveraging latent space embeddings. Additionally, the thesis details methods for geometric analysis of segmented masks to quantify specific brace root phenotypic traits. ☐ The results section demonstrates that variability in data quality across field seasons significantly affects model performance, highlighting the critical contributions of data from 2021 and 2024 to generalization. The combination of transfer learning and data augmentation proved instrumental in achieving fast and effective convergence, with the best-performing segmentation model exhibiting strong class-wise metrics (IoU, F1 score, precision, recall) across background, markers, roots, and stalks. For classification of usable samples, a CLIP-based model achieved superior accuracy (96.95\%) compared to the custom CNN model, which was subsequently used for large-scale data filtering. Furthermore, significant positive correlations were observed between automated measurements and human-driven RootTaggingGUI measurements for key traits such as root count, spread width, stalk width, and root height, validating the efficacy of the automated phenotyping pipeline for these traits. ☐ In conclusion, this thesis successfully establishes an automated machine learning pipeline for robust and efficient brace root phenotyping in maize, effectively addressing the challenges posed by variable field imaging conditions and the need for multi-class segmentation. The methodology, encompassing advanced annotation techniques, comprehensive data augmentation, and a high-performing segmentation and classification model, offers a significant stride towards accelerating plant science research and crop improvement efforts by enabling high-throughput measurement of crucial phenotypic traits. Future work will aim to refine approximations for traits with weaker correlations and ways that segmentation can continue to improve.
Description
Keywords
Brace roots, Computer vision, Image analysis, Machine learning, Semantic segmentation