Progress 05/01/24 to 04/30/25
Outputs
Target Audience:The target audience reached by PD's efforts includes graduate and undergraduate students, as well as faculty. Changes/Problems:
Nothing Reported
What opportunities for training and professional development has the project provided?The project has offered vital professional development opportunities, greatly enhancing the team's expertise in agricultural AI. The Project Director and key team members attended several hands-on workshops and conferences, starting with the flagship "2025 AI in Agriculture and Natural Resources" conference, where they received comprehensive training on applying machine learning and computer vision techniques to crop and soil management. They further improved their skills through specialized events hosted by Texas A&M AgriLife, including the Automated Precision Phenotyping Greenhouse Data Pipeline Workshop and the "XGBoost Your Digital Ag Research," along with a Keras-based image classification workshop. Additional technical skills were developed through NVIDIA GTC workshops, which covered the fundamentals of Deep Learning, Accelerated Data Science, Modern CUDA C++, and Building Agentic AI Applications with large language models. To broaden collaborative networks, the Project Director also participated in the 1890/MSI Day, hosted by the AI Institute for Next Generation Food Systems (AIFS) at the University of California, Davis, and engaged with the AI Foundry for Ag Applications program at the University of Illinois' Center for Digital Agriculture. Collectively, these training activities have elevated the team's capacity to integrate state-of-the-art AI methods into research and teaching, positioning the project to deliver more innovative solutions and enriched learning experiences in the next reporting period. How have the results been disseminated to communities of interest?
Nothing Reported
What do you plan to do during the next reporting period to accomplish the goals?In the next reporting period, the team aims to complete the segmentation of the remaining flower images and expand the collection of non-image attributes from additional daylily plants. These data will be compiled into a unified, version-controlled repository to ensure consistency, accessibility, and reproducibility. At the same time, efforts will also concentrate on enriching the dataset by incorporating new trait variables and image-based features. Additionally, Pilot machine learning models will be developed and evaluated to validate their predictive potential, with ongoing refinements to model architectures designed to enhance accuracy and robustness across a broader range of phenotypic traits. At the same time, efforts will also be made to enhance the capacity of students and faculty in AI. Graduate students will receive hands-on training in machine learning methods, allowing them to apply AI techniques in their thesis research. Faculty and key personnel plan to attend advanced training workshops and national conferences related to deep learning and AI in agriculture to ensure they remain up-to-date with the latest practices. As part of curriculum development, AI learning modules will be designed for inclusion in select graduate-level courses. Instructional materials will be regularly updated to incorporate new tools, techniques, and real-world applications. Lastly, the team plans to publish peer-reviewed articles to share their findings.
Impacts
What was accomplished under these goals?
OBJECTIVE 1: Create daylily dataset with images and attributes, preprocess for AI analysis. The team has collected 2,651sets of kinship daylily flower images -- each consisting of the pod parent, pollen parent, and their hybrid offspring -- totaling approximately 3,037images. These images showcase a broad range of visual variation, including differences in color, shape, lighting, and background, which are crucial for training effective machine learning models. In support of this effort, the team has collaborated with local and regional growers, online communities, and national organizations such as the American Hemerocallis Society to ensure diversity in the cultivars represented. Alongside the image data, detailed non-image attributes have been collected from over 600 daylily plants. These include traits such as plant height, flower diameter, bloom season, foliage type, ploidy, rebloom frequency, branching patterns, bud count, and fragrance. Each data set corresponds to the same parent-progeny relationships represented in the image dataset. To prepare the images for AI analysis, preprocessing efforts have focused on segmentation-- isolating the flower from its background to enhance learning accuracy. So far, 2,875 images have been segmented. The segmentation process was performed using a custom-built interactive tool developed in-house that integrates automation with manual refinement features. Built in Python, the tool utilizes a TRACER convolutional neural network with an EfficientNet backbone and incorporates widely used libraries like TensorFlow, PyTorch, and OpenCV. A user-friendly GUI designed with Tkinter supports real-time image uploads, segmentation visualization, and customizable export options. Additionally, a built-in pencil tool allows users to manually adjust segmentation masks for improved accuracy. The tool ensured consistent preprocessing with automated resizing and normalization, helping to standardize inputs for downstream AI tasks. Initial tests show high segmentation accuracy even in complex conditions, highlighting the tool's effectiveness and flexibility. OBJECTIVE 2: Develop AI models to enhance the prediction of daylily hybrid traits. Several machine learning algorithms, including random forests, decision trees, XGBoost, and artificial neural networks (ANNs), have been tested to classify and predict phenotypic traits such as plant height, flower diameter, foliage type, blooming habit, ploidy level, and bloom sequence. To support model training, qualitative and categorical data fields, along with other input data, were preprocessed using normalization, categorical encoding, and formatting to ensure compatibility across various algorithms. Initial model development revealed challenges due to the complexity and variability of the data, including noise and overlapping patterns of traits. As a result, model performance has been mixed and remains under evaluation. While some algorithms demonstrated potential in capturing trait relationships, further refinement and validation are necessary to improve predictive accuracy. Objective 3: Develop New AI Courses and Course-Embedded Learning Modules Curriculum design for course-embedded modules is well underway for ten graduate-level instructional modules that will introduce students in agricultural science to essential Python programming and machine learning techniques. Each module covers a complete workflow--data preprocessing, model training, evaluation, and deployment--using real-world applications such as yield forecasting and food safety prediction. All ten modules are scheduled for rollout in BIOT 6513 "Computational Biology" (Fall term), where students will anchor hands-on, experiential learning. Students will address precision agriculture and agri-genomics challenges through Python exercises, annotated code notebooks, and case studies, working both individually and in teams to reflect industry practice. Comprehensive teaching materials--step-by-step guides, lecture slides, annotated code, and scoring rubrics--are being finalized to ensure consistent delivery and assessment. Additionally, the team has begun outlining a standalone graduate course in Artificial Intelligence for Agriculture, including learning objectives, a topical outline, and an assessment strategy that are currently in draft form. This course will expand the module content to cover advanced topics, including deep learning, model interpretability, and ethical AI. OBJECTIVE 4: Build cutting-edge AI infrastructure to support research and education. The project has successfully established a high-performance AI infrastructure to support advanced research and instruction in agricultural science. A Lambda Vector Pro AI workstation has been acquired and installed at the Agricultural Research Station (Room 211-B). This system is optimized for intensive deep learning tasks, featuring a 96-core AMD Threadripper PRO processor, 512 GB RAM, and three NVIDIA A800 GPUs with a total of 120 GB VRAM. The workstation comes with high-capacity SSD storage and is preloaded with the Lambda Stack, which includes essential AI frameworks such as TensorFlow, PyTorch, CUDA, and cuDNN. It provides a stable environment for developing and training complex models, particularly in fields such as precision agriculture and bioinformatics. Additionally, two Dell AI-enabled mobile workstations were acquired to support portable research and classroom instruction. These laptops allow students and researchers to utilize AI tools in both lab and field environments. This infrastructure significantly enhanced the institution's capacity for AI-driven innovation and experiential learning across the fields of agriculture and biology. OBJECTIVE 5: Develop FVSU faculty capacity in AI teaching and research. The Project Director and the research teamhave engaged in hands-on training focused on implementing deep learning technologies in the agricultural sector. Topics covered included computer vision, precision agriculture, and AI model development, equipping faculty with practical skills to incorporate AI into both research and teaching. As part of this capacity-building effort, team members also attended a national conference on AI in Agriculture, where they engaged with leading experts, explored emerging tools and methodologies, and formed collaborative networks with researchers from peer institutions.
Publications
- Type:
Conference Papers and Presentations
Status:
Published
Year Published:
2025
Citation:
Chunhua Dong, Priyanka Kumar, Ramana Gosukonda. 2025.Automated Daylily Flower Segmentation with Deep Learning GUI, in Agriculture and Natural Resource conference, Mississippi State University, March 31-April 2, 2025
|