A-Maize-ing Deep Vision : developing deep learning and computer vision algorithms for UAV-based automated digital phenotyping in agricultural fields

Global food demand is projected to increase by up to 62 percent by 2050, while climate change is expected to place an additional 170 million people at risk of hunger. Meeting these challenges requires developing crop varieties that can withstand environmental stressors such as drought, heat, and poor soil. Achieving this goal depends not only on plant science but also on innovations in scalable data acquisition and analysis. Digital phenotyping offers a promising solution by enabling high-throughput, high-resolution trait measurement across large plant populations. While progress has been made in controlled environments, field-based phenotyping remains difficult due to environmental variability and the need for efficient processing at scale. This dissertation presents computer vision, deep learning, and image processing algorithms using unmanned aerial vehicles (UAVs) for scalable, low-cost phenotyping of maize, particularly in mutant lines that exhibit quantifiable leaf lesions. These lesions vary in color, size, shape, morphology, and temporal development depending on both environmental conditions and genetic background. Individual lesions are quantified in the field, and the resulting data are leveraged for improved phenotype prediction in selective breeding and network inference. Together, the following chapters contribute to the development of an integrated, scalable framework for field-based phenotyping -- a system I refer to as A-Maize-ing Deep Vision. Chapter 1 introduces CorNet, an unsupervised deep learning model for planar homography estimation from freely flown UAV video. Unlike traditional waypointxx guided systems, CorNet enables faster, flexible, metadata-free image registration using consumer-grade drones, with accuracy comparable to ASIFT. Chapter 2 builds on this with MaiZaic, a robust end-to-end pipeline that automates video frame sampling, camera calibration, homography estimation with the improved CorNetv3 , shot detection, and mini-mosaicking. MaiZaic improves mosaic accuracy by 13.1 percent while computing 14.11 times faster compared to ASIFT, offering a practical solution for real-time aerial imaging. Chapter 3 presents SeeMaiDetect, a high-resolution, labeled dataset for maize seedling detection and stand counting. Annotated for single, double, and triple plant classes, the dataset supports model evaluation under realistic variations in growth stage, lighting, soil, and density. Detection of single seedlings achieves precision up to 0.984 and recall up to 0.873, while multi-seedling detection remains more challenging due to class imbalance and morphological overlap. A YOLO-based model provides fast inference, averaging 155 ms per image. Chapter 4 introduces DeepMaizeCounter (DMC), an end-to-end system for automated maize stand counting in nursery fields. Operating in two modes, using either mosaicked imagery or raw frames with homography matrices, DMC leverages a YOLOv9 backbone for seedling detection and spatial reasoning for row and range segmentation. In field trials, DMC achieved an R2value of 0.616 using mosaicked data and 0.906 using raw video frames, with full-frame processing completed in 60.63 seconds across 83 images. Chapter 5, PointillistMaize, explores 3D plant reconstruction using Structure from Motion (SfM), Multi-View Stereo (MVS), Neural Radiance Fields (NeRF), and Gaussian Splatting. These methods generate detailed 3D point clouds of individual plants and nursery rows from UAV video. While COLMAP yielded the highest point density and geometric accuracy, NeRF offered comparable results with reduced computation time. Gaussian Splatting delivered photorealistic outputs despite lower point density. Together, these contributions present an integrated approach to field-based digital phenotyping, from UAV-based image capture and registration to detection, counting, and 3D reconstruction. The presented methods lay the technical foundation for large-scale lesion quantification and phenotype prediction, supporting data-driven crop improvement and genotype-phenotype modeling in plant breeding.