Missouri Maize Computation and Vision Laboratory

Dewi is a very dedicated fifth-year Ph.D. student in Computer Science. Her research primarily focuses on digital phenotyping, leveraging consumer grade UAVsystems to make advanced agricultural technologies more accessible. She is an experienced UAV pilot with part 107 license and has over 200 hours of flying experience. Dewi has developed sophisticated algorithms for 3D reconstruction, image registration, object detection, and object segmentation. Her projects include:

• cornet v1: An unsupervised deep learning based homography estimator for a faster and improved mosaicking without metadata. CorNet is ten times faster than ASIFT while maintaining comparable accuracy. The current version is CorNetv3, trained on a more diverse and balanced dataset.
• MaiZaic: An improved and more robust end-to-end mosaicking pipeline that incorporates dynamic sampling with pixel optical flow, automatic lens and gimbal calibration, the enhanced Cornetv3, shot detection based on UAV movements, and mini-mosaicking to reduce error accumulation. MaiZaic achieves Average Percentage Error (APE) of 11% lower than ASIFT.
• Deep Maize Counter (DMC): an automated seedling maize stand counting pipeline using YOLOv9 for nursery field. DMC has two modes: (1) counting from mosaics that provides the whole view of the field or (2) counting from raw images utilizing homography matrices to stitch both detection and raw into mosaics. The mosaics mode achieved r² of 0.616, while the raw frames mode achieved 0.906.
• Pointillist Maize: 3D maize reconstructions from 360° aerial videos using Structure from Motion (SfM), Neural Radiance Fields (NeRF), and Gaussian Splatting. Comparative analysis demonstrated that NeRF produces 90.4% of points and computes 7.3 times faster than SfM, while Gaussian Splatting produces 8.1% of points and operates 3.0 times faster than SfM.
• Maize Seedling Detection Dataset: A dataset containing aerial images of seedling maize across different growth stages, soil colors, and camera views. To ensure accurate counting, the dataset is categorized into three classes: one plant, two plants, and three plants. The two- and three-plant groups were intentionally planted together.
• Maize 3D Model Dataset: A dataset featuring aerial video orbital trajectories at various camera angles and views, including single plants, rows, and entire fields. Single plants were grown in a tiling field, with each plant spaced 15 feet apart to enable very low-altitude orbits. Evie and Chris are computing the 3D models of these plants.

Chimdi is a third year PhD student majoring in Electrical and Computer Engineering. His research focuses on computer vision and image processing. His work integrates deep learning, image processing, mathematical topology and computer vision to develop solutions and address challenges in plant phenotyping. Ultimately he works on mapping high-dimensional tensors of maize lesion phenotypes — biological attributes of leaf spots — onto manifolds that are only locally metric and using those maps to predict the structure of the underlying physiological network and faster routes for crop improvement. His projects include:

• Pre-Processing Pipeline and Color Correction Algorithm: Pre-processing pipeline and color correction algorithm for the ex situ imaging
• Wavelet Application and Snake Algorithm: Application of discrete and continuous wavelet transforms on color corrected 16-bit images with the application of a Navier-Stokes algorithm.

Our current phenotyping method relies on still images of single leaves from field-grown plants: an experienced team can image about 100 leaves/day. We previously developed a cascade of MATLAB algorithms to segment lesions from healthy tissue on leaves, and other code that quantitates lesion phenotypes. The cascade performs well on many phenotypes, but is less accurate on phenotypes that have many tiny lesions crowded together. We are now working on an improved pipeline to segment lesions from all the phenotypes. This more diverse collection of phenotypes will let us populate the manifold and recognize similar phenotypes, forming the model's blocks.

Our goal is to develop 3D models of maize plants, segment the leaves from each plant, and use each plant's leaves as input to our improved lesion segmentation pipeline.

We are looking to improve the efficiency of identifying lesions on maize plants from videos taken by drones by segmenting leaves. Segmenting leaves, or seperating the leaves out as individuals from a plant, can be achieved using modern computer vision advancements. By taking video from drones as input, we hope to output a result that clearly and accurately segments the leaves from the plant to be used as input to a model that will then segment the lesions on the leaves. Currently we are leveraging tools like Nautilus to automate extracting three-dimensional data in the form of point clouds, and using COLMAP for reconstruction. We wll then use these data to train and experiment with various computer vision models to optimize precise and accurate leaf segmentation.

I'm working on segmenting plants from rows imaged by the drone using both 2D and 3D methods.

An important challenge is to track leaves as they develop, since new leaves appear at the top of the growing plant while older, juvenile leaves die at its bottom. Viewed from above, the angle a leaf subtends with the culm (the plant's 'stem') is determined very early in development and does not change. This provides a time-invariant signature for each leaf. My current work is developing algorithms to detect these angles as the crop grows.

My research focuses on the development of a robotic planter designed to improve the efficiency of planting genetically modified corn seed. Currently, the planting process is done by hand, requiring hours of labor in high temperatures, often exceeding 90° F. This not only makes the task physically demanding but also time-consuming. The robotic planter we are developing will streamline this process by automating the planting of all varieties of corn seeds used by our lab with the flexibility to accommodate future varieties. In addition to planting, the system will be capable of digging and closing holes, significantly reducing manual labor. A key future objective is to make the robotic planter fully autonomous, enabling it to handle large quantities of seed with minimal human intervention. To ensure optimal performance, ongoing experiments such as the angle of repose test are being conducted to determine how the angle of seed distribution impacts planting efficiency. We are also working to modify an existing planter designed by a previous capstone group, improving its maneuverability in field conditions to better suit our needs.

I'm currently a freshman studying electrical engineering. I'm currently working on our drones, keeping them in good repair. This experience will help me integrate vision-driven navigation with the drone hardware.

I'm currently a sophomore studying mechanical engineering. My team and I are currently working on building and optimizing a rover that will open the furrow, plant the corn seed, and then close the furrow back up. Since collecting data is so important for AI, we must make sure to reduce variability and increase the rate at which we plant seeds. I hope to use my new found-knowdeldge and experiences from this project to further my career as a mechanical engineer.